Javier Flores
commited on
Commit
•
de28dc1
1
Parent(s):
58f6e24
- Crops_data.zip +3 -0
- Examples/Image1.jpg +0 -0
- Examples/Image2.jpg +0 -0
- Examples/Image7.jpg +0 -0
- Inference/a.txt +1 -0
- Temp_file/readme.txt +1 -0
- app.py +156 -5
- best.pt +3 -0
- custom.yaml +18 -0
- models/__init__.py +1 -0
- models/common.py +2019 -0
- models/experimental.py +262 -0
- models/yolo.py +843 -0
- requirements.txt +19 -0
- utils/__init__.py +1 -0
- utils/a.txt +1 -0
- utils/activations.py +72 -0
- utils/add_nms.py +155 -0
- utils/autoanchor.py +160 -0
- utils/datasets.py +1320 -0
- utils/general.py +891 -0
- utils/google_utils.py +123 -0
- utils/loss.py +1697 -0
- utils/metrics.py +223 -0
- utils/plots.py +489 -0
- utils/torch_utils.py +374 -0
Crops_data.zip
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8508a9b26a7e97cff5b61c55650b99a8a463a613d7c3f299f63324aaccb1ccef
|
3 |
+
size 263538132
|
Examples/Image1.jpg
ADDED
Examples/Image2.jpg
ADDED
Examples/Image7.jpg
ADDED
Inference/a.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
a
|
Temp_file/readme.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
This folder is to save the image temporarily.
|
app.py
CHANGED
@@ -1,10 +1,161 @@
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
-
def predict(image):
|
4 |
-
return image
|
5 |
|
|
|
6 |
|
7 |
-
inp = gr.inputs.Image(source="webcam", shape=(224,224))
|
8 |
|
9 |
-
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import argparse
|
3 |
import gradio as gr
|
4 |
+
from PIL import Image
|
5 |
+
from numpy import random
|
6 |
+
from pathlib import Path
|
7 |
+
import os
|
8 |
+
import time
|
9 |
+
import torch.backends.cudnn as cudnn
|
10 |
+
from models.experimental import attempt_load
|
11 |
+
import cv2
|
12 |
+
from utils.datasets import LoadStreams, LoadImages
|
13 |
+
from utils.general import check_img_size, check_requirements, check_imshow, non_max_suppression, apply_classifier,scale_coords, xyxy2xywh, strip_optimizer, set_logging, increment_path
|
14 |
+
from utils.plots import plot_one_box
|
15 |
+
from utils.torch_utils import select_device, load_classifier, time_synchronized, TracedModel
|
16 |
|
|
|
|
|
17 |
|
18 |
+
os.system('git clone https://github.com/WongKinYiu/yolov7')
|
19 |
|
|
|
20 |
|
21 |
+
|
22 |
+
def Custom_detect(img):
|
23 |
+
model='best'
|
24 |
+
parser = argparse.ArgumentParser()
|
25 |
+
parser.add_argument('--weights', nargs='+', type=str, default=model+".pt", help='model.pt path(s)')
|
26 |
+
parser.add_argument('--source', type=str, default='Temp_file/', help='source')
|
27 |
+
parser.add_argument('--img-size', type=int, default=100, help='inference size (pixels)')
|
28 |
+
parser.add_argument('--conf-thres', type=float, default=0.25, help='object confidence threshold')
|
29 |
+
parser.add_argument('--iou-thres', type=float, default=0.45, help='IOU threshold for NMS')
|
30 |
+
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
|
31 |
+
parser.add_argument('--view-img', action='store_true', help='display results')
|
32 |
+
parser.add_argument('--save-txt', action='store_true', help='save results to *.txt')
|
33 |
+
parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')
|
34 |
+
parser.add_argument('--nosave', action='store_true', help='do not save images/videos')
|
35 |
+
parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3')
|
36 |
+
parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')
|
37 |
+
parser.add_argument('--augment', action='store_true', help='augmented inference')
|
38 |
+
parser.add_argument('--update', action='store_true', help='update all models')
|
39 |
+
parser.add_argument('--project', default='runs/detect', help='save results to project/name')
|
40 |
+
parser.add_argument('--name', default='exp', help='save results to project/name')
|
41 |
+
parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')
|
42 |
+
parser.add_argument('--trace', action='store_true', help='trace model')
|
43 |
+
opt = parser.parse_args()
|
44 |
+
img.save("Temp_file/test.jpg")
|
45 |
+
source, weights, view_img, save_txt, imgsz, trace = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_size, opt.trace
|
46 |
+
save_img = True
|
47 |
+
webcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))
|
48 |
+
save_dir = Path(increment_path(Path(opt.project)/opt.name,exist_ok=opt.exist_ok))
|
49 |
+
|
50 |
+
(save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)
|
51 |
+
set_logging()
|
52 |
+
device = select_device(opt.device)
|
53 |
+
half = device.type != 'cpu'
|
54 |
+
model = attempt_load(weights, map_location=device)
|
55 |
+
stride = int(model.stride.max())
|
56 |
+
imgsz = check_img_size(imgsz, s=stride)
|
57 |
+
if trace:
|
58 |
+
model = TracedModel(model, device, opt.img_size)
|
59 |
+
if half:
|
60 |
+
model.half()
|
61 |
+
|
62 |
+
classify = False
|
63 |
+
if classify:
|
64 |
+
modelc = load_classifier(name='resnet101', n=2) # initialize
|
65 |
+
modelc.load_state_dict(torch.load('weights/resnet101.pt', map_location=device)['model']).to(device).eval()
|
66 |
+
vid_path, vid_writer = None, None
|
67 |
+
if webcam:
|
68 |
+
view_img = check_imshow()
|
69 |
+
cudnn.benchmark = True
|
70 |
+
dataset = LoadStreams(source, img_size=imgsz, stride=stride)
|
71 |
+
else:
|
72 |
+
dataset = LoadImages(source, img_size=imgsz, stride=stride)
|
73 |
+
names = model.module.names if hasattr(model, 'module') else model.names
|
74 |
+
colors = [[random.randint(0, 255) for _ in range(3)] for _ in names]
|
75 |
+
if device.type != 'cpu':
|
76 |
+
model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))
|
77 |
+
t0 = time.time()
|
78 |
+
for path, img, im0s, vid_cap in dataset:
|
79 |
+
img = torch.from_numpy(img).to(device)
|
80 |
+
img = img.half() if half else img.float()
|
81 |
+
img /= 255.0
|
82 |
+
if img.ndimension() == 3:
|
83 |
+
img = img.unsqueeze(0)
|
84 |
+
|
85 |
+
# Inference
|
86 |
+
t1 = time_synchronized()
|
87 |
+
pred = model(img, augment=opt.augment)[0]
|
88 |
+
|
89 |
+
pred = non_max_suppression(pred,opt.conf_thres,opt.iou_thres,classes=opt.classes, agnostic=opt.agnostic_nms)
|
90 |
+
t2 = time_synchronized()
|
91 |
+
|
92 |
+
|
93 |
+
# Apply Classifier
|
94 |
+
if classify:
|
95 |
+
pred = apply_classifier(pred, modelc, img, im0s)
|
96 |
+
|
97 |
+
for i, det in enumerate(pred):
|
98 |
+
if webcam:
|
99 |
+
p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.count
|
100 |
+
else:
|
101 |
+
p, s, im0, frame = path, '', im0s, getattr(dataset, 'frame', 0)
|
102 |
+
|
103 |
+
p = Path(p)
|
104 |
+
save_path = str(save_dir / p.name)
|
105 |
+
txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # img.txt
|
106 |
+
s += '%gx%g ' % img.shape[2:]
|
107 |
+
gn = torch.tensor(im0.shape)[[1, 0, 1, 0]]
|
108 |
+
if len(det):
|
109 |
+
det[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()
|
110 |
+
|
111 |
+
|
112 |
+
for c in det[:, -1].unique():
|
113 |
+
n = (det[:, -1] == c).sum()
|
114 |
+
s += f"{n} {names[int(c)]}{'s' * (n > 1)}, "
|
115 |
+
|
116 |
+
|
117 |
+
for *xyxy, conf, cls in reversed(det):
|
118 |
+
if save_txt:
|
119 |
+
xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()
|
120 |
+
line = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh)
|
121 |
+
with open(txt_path + '.txt', 'a') as f:
|
122 |
+
f.write(('%g ' * len(line)).rstrip() % line + '\n')
|
123 |
+
|
124 |
+
if save_img or view_img:
|
125 |
+
label = f'{names[int(cls)]} {conf:.2f}'
|
126 |
+
plot_one_box(xyxy, im0, label=label, color=colors[int(cls)], line_thickness=3)
|
127 |
+
if view_img:
|
128 |
+
cv2.imshow(str(p), im0)
|
129 |
+
cv2.waitKey(1)
|
130 |
+
|
131 |
+
if save_img:
|
132 |
+
if dataset.mode == 'image':
|
133 |
+
cv2.imwrite(save_path, im0)
|
134 |
+
else:
|
135 |
+
if vid_path != save_path:
|
136 |
+
vid_path = save_path
|
137 |
+
if isinstance(vid_writer, cv2.VideoWriter):
|
138 |
+
vid_writer.release()
|
139 |
+
if vid_cap:
|
140 |
+
fps = vid_cap.get(cv2.CAP_PROP_FPS)
|
141 |
+
w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
142 |
+
h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
143 |
+
else:
|
144 |
+
fps, w, h = 30, im0.shape[1], im0.shape[0]
|
145 |
+
save_path += '.mp4'
|
146 |
+
vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))
|
147 |
+
vid_writer.write(im0)
|
148 |
+
|
149 |
+
if save_txt or save_img:
|
150 |
+
s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
|
151 |
+
|
152 |
+
print(f'Done. ({time.time() - t0:.3f}s)')
|
153 |
+
|
154 |
+
return Image.fromarray(im0[:,:,::-1])
|
155 |
+
inp = gr.Image(type="pil")
|
156 |
+
output = gr.Image(type="pil")
|
157 |
+
|
158 |
+
examples=[["Examples/Image1.jpg","Image1"],["Examples/Image2.jpg","Image2"]]
|
159 |
+
|
160 |
+
io=gr.Interface(fn=Custom_detect, inputs=inp, outputs=output, title='Pot Hole Detection With Custom YOLOv7',examples=examples,cache_examples=False)
|
161 |
+
io.launch()
|
best.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f8c1af3a3efda26dd0f558580fabc257102163312cf1b7ed7729987f36e317aa
|
3 |
+
size 74810441
|
custom.yaml
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# COCO 2017 dataset http://cocodataset.org
|
2 |
+
|
3 |
+
# download command/URL (optional)
|
4 |
+
download: bash ./scripts/get_coco.sh
|
5 |
+
|
6 |
+
# train and val data as 1) directory: path/images/, 2) file: path/images.txt, or 3) list: [path1/images/, path2/images/]
|
7 |
+
train: ../train/images
|
8 |
+
val: ../valid/images
|
9 |
+
test: ../test/images
|
10 |
+
|
11 |
+
|
12 |
+
# number of classes
|
13 |
+
nc: 1
|
14 |
+
|
15 |
+
# class names
|
16 |
+
names: ['pothole']
|
17 |
+
|
18 |
+
|
models/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
# init
|
models/common.py
ADDED
@@ -0,0 +1,2019 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from copy import copy
|
3 |
+
from pathlib import Path
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import pandas as pd
|
7 |
+
import requests
|
8 |
+
import torch
|
9 |
+
import torch.nn as nn
|
10 |
+
import torch.nn.functional as F
|
11 |
+
from torchvision.ops import DeformConv2d
|
12 |
+
from PIL import Image
|
13 |
+
from torch.cuda import amp
|
14 |
+
|
15 |
+
from utils.datasets import letterbox
|
16 |
+
from utils.general import non_max_suppression, make_divisible, scale_coords, increment_path, xyxy2xywh
|
17 |
+
from utils.plots import color_list, plot_one_box
|
18 |
+
from utils.torch_utils import time_synchronized
|
19 |
+
|
20 |
+
|
21 |
+
##### basic ####
|
22 |
+
|
23 |
+
def autopad(k, p=None): # kernel, padding
|
24 |
+
# Pad to 'same'
|
25 |
+
if p is None:
|
26 |
+
p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad
|
27 |
+
return p
|
28 |
+
|
29 |
+
|
30 |
+
class MP(nn.Module):
|
31 |
+
def __init__(self, k=2):
|
32 |
+
super(MP, self).__init__()
|
33 |
+
self.m = nn.MaxPool2d(kernel_size=k, stride=k)
|
34 |
+
|
35 |
+
def forward(self, x):
|
36 |
+
return self.m(x)
|
37 |
+
|
38 |
+
|
39 |
+
class SP(nn.Module):
|
40 |
+
def __init__(self, k=3, s=1):
|
41 |
+
super(SP, self).__init__()
|
42 |
+
self.m = nn.MaxPool2d(kernel_size=k, stride=s, padding=k // 2)
|
43 |
+
|
44 |
+
def forward(self, x):
|
45 |
+
return self.m(x)
|
46 |
+
|
47 |
+
|
48 |
+
class ReOrg(nn.Module):
|
49 |
+
def __init__(self):
|
50 |
+
super(ReOrg, self).__init__()
|
51 |
+
|
52 |
+
def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2)
|
53 |
+
return torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1)
|
54 |
+
|
55 |
+
|
56 |
+
class Concat(nn.Module):
|
57 |
+
def __init__(self, dimension=1):
|
58 |
+
super(Concat, self).__init__()
|
59 |
+
self.d = dimension
|
60 |
+
|
61 |
+
def forward(self, x):
|
62 |
+
return torch.cat(x, self.d)
|
63 |
+
|
64 |
+
|
65 |
+
class Chuncat(nn.Module):
|
66 |
+
def __init__(self, dimension=1):
|
67 |
+
super(Chuncat, self).__init__()
|
68 |
+
self.d = dimension
|
69 |
+
|
70 |
+
def forward(self, x):
|
71 |
+
x1 = []
|
72 |
+
x2 = []
|
73 |
+
for xi in x:
|
74 |
+
xi1, xi2 = xi.chunk(2, self.d)
|
75 |
+
x1.append(xi1)
|
76 |
+
x2.append(xi2)
|
77 |
+
return torch.cat(x1+x2, self.d)
|
78 |
+
|
79 |
+
|
80 |
+
class Shortcut(nn.Module):
|
81 |
+
def __init__(self, dimension=0):
|
82 |
+
super(Shortcut, self).__init__()
|
83 |
+
self.d = dimension
|
84 |
+
|
85 |
+
def forward(self, x):
|
86 |
+
return x[0]+x[1]
|
87 |
+
|
88 |
+
|
89 |
+
class Foldcut(nn.Module):
|
90 |
+
def __init__(self, dimension=0):
|
91 |
+
super(Foldcut, self).__init__()
|
92 |
+
self.d = dimension
|
93 |
+
|
94 |
+
def forward(self, x):
|
95 |
+
x1, x2 = x.chunk(2, self.d)
|
96 |
+
return x1+x2
|
97 |
+
|
98 |
+
|
99 |
+
class Conv(nn.Module):
|
100 |
+
# Standard convolution
|
101 |
+
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
|
102 |
+
super(Conv, self).__init__()
|
103 |
+
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
|
104 |
+
self.bn = nn.BatchNorm2d(c2)
|
105 |
+
self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
|
106 |
+
|
107 |
+
def forward(self, x):
|
108 |
+
return self.act(self.bn(self.conv(x)))
|
109 |
+
|
110 |
+
def fuseforward(self, x):
|
111 |
+
return self.act(self.conv(x))
|
112 |
+
|
113 |
+
|
114 |
+
class RobustConv(nn.Module):
|
115 |
+
# Robust convolution (use high kernel size 7-11 for: downsampling and other layers). Train for 300 - 450 epochs.
|
116 |
+
def __init__(self, c1, c2, k=7, s=1, p=None, g=1, act=True, layer_scale_init_value=1e-6): # ch_in, ch_out, kernel, stride, padding, groups
|
117 |
+
super(RobustConv, self).__init__()
|
118 |
+
self.conv_dw = Conv(c1, c1, k=k, s=s, p=p, g=c1, act=act)
|
119 |
+
self.conv1x1 = nn.Conv2d(c1, c2, 1, 1, 0, groups=1, bias=True)
|
120 |
+
self.gamma = nn.Parameter(layer_scale_init_value * torch.ones(c2)) if layer_scale_init_value > 0 else None
|
121 |
+
|
122 |
+
def forward(self, x):
|
123 |
+
x = x.to(memory_format=torch.channels_last)
|
124 |
+
x = self.conv1x1(self.conv_dw(x))
|
125 |
+
if self.gamma is not None:
|
126 |
+
x = x.mul(self.gamma.reshape(1, -1, 1, 1))
|
127 |
+
return x
|
128 |
+
|
129 |
+
|
130 |
+
class RobustConv2(nn.Module):
|
131 |
+
# Robust convolution 2 (use [32, 5, 2] or [32, 7, 4] or [32, 11, 8] for one of the paths in CSP).
|
132 |
+
def __init__(self, c1, c2, k=7, s=4, p=None, g=1, act=True, layer_scale_init_value=1e-6): # ch_in, ch_out, kernel, stride, padding, groups
|
133 |
+
super(RobustConv2, self).__init__()
|
134 |
+
self.conv_strided = Conv(c1, c1, k=k, s=s, p=p, g=c1, act=act)
|
135 |
+
self.conv_deconv = nn.ConvTranspose2d(in_channels=c1, out_channels=c2, kernel_size=s, stride=s,
|
136 |
+
padding=0, bias=True, dilation=1, groups=1
|
137 |
+
)
|
138 |
+
self.gamma = nn.Parameter(layer_scale_init_value * torch.ones(c2)) if layer_scale_init_value > 0 else None
|
139 |
+
|
140 |
+
def forward(self, x):
|
141 |
+
x = self.conv_deconv(self.conv_strided(x))
|
142 |
+
if self.gamma is not None:
|
143 |
+
x = x.mul(self.gamma.reshape(1, -1, 1, 1))
|
144 |
+
return x
|
145 |
+
|
146 |
+
|
147 |
+
def DWConv(c1, c2, k=1, s=1, act=True):
|
148 |
+
# Depthwise convolution
|
149 |
+
return Conv(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
|
150 |
+
|
151 |
+
|
152 |
+
class GhostConv(nn.Module):
|
153 |
+
# Ghost Convolution https://github.com/huawei-noah/ghostnet
|
154 |
+
def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups
|
155 |
+
super(GhostConv, self).__init__()
|
156 |
+
c_ = c2 // 2 # hidden channels
|
157 |
+
self.cv1 = Conv(c1, c_, k, s, None, g, act)
|
158 |
+
self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
|
159 |
+
|
160 |
+
def forward(self, x):
|
161 |
+
y = self.cv1(x)
|
162 |
+
return torch.cat([y, self.cv2(y)], 1)
|
163 |
+
|
164 |
+
|
165 |
+
class Stem(nn.Module):
|
166 |
+
# Stem
|
167 |
+
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
|
168 |
+
super(Stem, self).__init__()
|
169 |
+
c_ = int(c2/2) # hidden channels
|
170 |
+
self.cv1 = Conv(c1, c_, 3, 2)
|
171 |
+
self.cv2 = Conv(c_, c_, 1, 1)
|
172 |
+
self.cv3 = Conv(c_, c_, 3, 2)
|
173 |
+
self.pool = torch.nn.MaxPool2d(2, stride=2)
|
174 |
+
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
175 |
+
|
176 |
+
def forward(self, x):
|
177 |
+
x = self.cv1(x)
|
178 |
+
return self.cv4(torch.cat((self.cv3(self.cv2(x)), self.pool(x)), dim=1))
|
179 |
+
|
180 |
+
|
181 |
+
class DownC(nn.Module):
|
182 |
+
# Spatial pyramid pooling layer used in YOLOv3-SPP
|
183 |
+
def __init__(self, c1, c2, n=1, k=2):
|
184 |
+
super(DownC, self).__init__()
|
185 |
+
c_ = int(c1) # hidden channels
|
186 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
187 |
+
self.cv2 = Conv(c_, c2//2, 3, k)
|
188 |
+
self.cv3 = Conv(c1, c2//2, 1, 1)
|
189 |
+
self.mp = nn.MaxPool2d(kernel_size=k, stride=k)
|
190 |
+
|
191 |
+
def forward(self, x):
|
192 |
+
return torch.cat((self.cv2(self.cv1(x)), self.cv3(self.mp(x))), dim=1)
|
193 |
+
|
194 |
+
|
195 |
+
class SPP(nn.Module):
|
196 |
+
# Spatial pyramid pooling layer used in YOLOv3-SPP
|
197 |
+
def __init__(self, c1, c2, k=(5, 9, 13)):
|
198 |
+
super(SPP, self).__init__()
|
199 |
+
c_ = c1 // 2 # hidden channels
|
200 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
201 |
+
self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
|
202 |
+
self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
|
203 |
+
|
204 |
+
def forward(self, x):
|
205 |
+
x = self.cv1(x)
|
206 |
+
return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
|
207 |
+
|
208 |
+
|
209 |
+
class Bottleneck(nn.Module):
|
210 |
+
# Darknet bottleneck
|
211 |
+
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
212 |
+
super(Bottleneck, self).__init__()
|
213 |
+
c_ = int(c2 * e) # hidden channels
|
214 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
215 |
+
self.cv2 = Conv(c_, c2, 3, 1, g=g)
|
216 |
+
self.add = shortcut and c1 == c2
|
217 |
+
|
218 |
+
def forward(self, x):
|
219 |
+
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
|
220 |
+
|
221 |
+
|
222 |
+
class Res(nn.Module):
|
223 |
+
# ResNet bottleneck
|
224 |
+
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
225 |
+
super(Res, self).__init__()
|
226 |
+
c_ = int(c2 * e) # hidden channels
|
227 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
228 |
+
self.cv2 = Conv(c_, c_, 3, 1, g=g)
|
229 |
+
self.cv3 = Conv(c_, c2, 1, 1)
|
230 |
+
self.add = shortcut and c1 == c2
|
231 |
+
|
232 |
+
def forward(self, x):
|
233 |
+
return x + self.cv3(self.cv2(self.cv1(x))) if self.add else self.cv3(self.cv2(self.cv1(x)))
|
234 |
+
|
235 |
+
|
236 |
+
class ResX(Res):
|
237 |
+
# ResNet bottleneck
|
238 |
+
def __init__(self, c1, c2, shortcut=True, g=32, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
239 |
+
super().__init__(c1, c2, shortcut, g, e)
|
240 |
+
c_ = int(c2 * e) # hidden channels
|
241 |
+
|
242 |
+
|
243 |
+
class Ghost(nn.Module):
|
244 |
+
# Ghost Bottleneck https://github.com/huawei-noah/ghostnet
|
245 |
+
def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride
|
246 |
+
super(Ghost, self).__init__()
|
247 |
+
c_ = c2 // 2
|
248 |
+
self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1), # pw
|
249 |
+
DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw
|
250 |
+
GhostConv(c_, c2, 1, 1, act=False)) # pw-linear
|
251 |
+
self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
|
252 |
+
Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
|
253 |
+
|
254 |
+
def forward(self, x):
|
255 |
+
return self.conv(x) + self.shortcut(x)
|
256 |
+
|
257 |
+
##### end of basic #####
|
258 |
+
|
259 |
+
|
260 |
+
##### cspnet #####
|
261 |
+
|
262 |
+
class SPPCSPC(nn.Module):
|
263 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
264 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
|
265 |
+
super(SPPCSPC, self).__init__()
|
266 |
+
c_ = int(2 * c2 * e) # hidden channels
|
267 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
268 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
269 |
+
self.cv3 = Conv(c_, c_, 3, 1)
|
270 |
+
self.cv4 = Conv(c_, c_, 1, 1)
|
271 |
+
self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
|
272 |
+
self.cv5 = Conv(4 * c_, c_, 1, 1)
|
273 |
+
self.cv6 = Conv(c_, c_, 3, 1)
|
274 |
+
self.cv7 = Conv(2 * c_, c2, 1, 1)
|
275 |
+
|
276 |
+
def forward(self, x):
|
277 |
+
x1 = self.cv4(self.cv3(self.cv1(x)))
|
278 |
+
y1 = self.cv6(self.cv5(torch.cat([x1] + [m(x1) for m in self.m], 1)))
|
279 |
+
y2 = self.cv2(x)
|
280 |
+
return self.cv7(torch.cat((y1, y2), dim=1))
|
281 |
+
|
282 |
+
class GhostSPPCSPC(SPPCSPC):
|
283 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
284 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5, k=(5, 9, 13)):
|
285 |
+
super().__init__(c1, c2, n, shortcut, g, e, k)
|
286 |
+
c_ = int(2 * c2 * e) # hidden channels
|
287 |
+
self.cv1 = GhostConv(c1, c_, 1, 1)
|
288 |
+
self.cv2 = GhostConv(c1, c_, 1, 1)
|
289 |
+
self.cv3 = GhostConv(c_, c_, 3, 1)
|
290 |
+
self.cv4 = GhostConv(c_, c_, 1, 1)
|
291 |
+
self.cv5 = GhostConv(4 * c_, c_, 1, 1)
|
292 |
+
self.cv6 = GhostConv(c_, c_, 3, 1)
|
293 |
+
self.cv7 = GhostConv(2 * c_, c2, 1, 1)
|
294 |
+
|
295 |
+
|
296 |
+
class GhostStem(Stem):
|
297 |
+
# Stem
|
298 |
+
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
|
299 |
+
super().__init__(c1, c2, k, s, p, g, act)
|
300 |
+
c_ = int(c2/2) # hidden channels
|
301 |
+
self.cv1 = GhostConv(c1, c_, 3, 2)
|
302 |
+
self.cv2 = GhostConv(c_, c_, 1, 1)
|
303 |
+
self.cv3 = GhostConv(c_, c_, 3, 2)
|
304 |
+
self.cv4 = GhostConv(2 * c_, c2, 1, 1)
|
305 |
+
|
306 |
+
|
307 |
+
class BottleneckCSPA(nn.Module):
|
308 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
309 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
310 |
+
super(BottleneckCSPA, self).__init__()
|
311 |
+
c_ = int(c2 * e) # hidden channels
|
312 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
313 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
314 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
315 |
+
self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
316 |
+
|
317 |
+
def forward(self, x):
|
318 |
+
y1 = self.m(self.cv1(x))
|
319 |
+
y2 = self.cv2(x)
|
320 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
321 |
+
|
322 |
+
|
323 |
+
class BottleneckCSPB(nn.Module):
|
324 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
325 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
326 |
+
super(BottleneckCSPB, self).__init__()
|
327 |
+
c_ = int(c2) # hidden channels
|
328 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
329 |
+
self.cv2 = Conv(c_, c_, 1, 1)
|
330 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
331 |
+
self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
332 |
+
|
333 |
+
def forward(self, x):
|
334 |
+
x1 = self.cv1(x)
|
335 |
+
y1 = self.m(x1)
|
336 |
+
y2 = self.cv2(x1)
|
337 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
338 |
+
|
339 |
+
|
340 |
+
class BottleneckCSPC(nn.Module):
|
341 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
342 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
343 |
+
super(BottleneckCSPC, self).__init__()
|
344 |
+
c_ = int(c2 * e) # hidden channels
|
345 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
346 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
347 |
+
self.cv3 = Conv(c_, c_, 1, 1)
|
348 |
+
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
349 |
+
self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
350 |
+
|
351 |
+
def forward(self, x):
|
352 |
+
y1 = self.cv3(self.m(self.cv1(x)))
|
353 |
+
y2 = self.cv2(x)
|
354 |
+
return self.cv4(torch.cat((y1, y2), dim=1))
|
355 |
+
|
356 |
+
|
357 |
+
class ResCSPA(BottleneckCSPA):
|
358 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
359 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
360 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
361 |
+
c_ = int(c2 * e) # hidden channels
|
362 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
363 |
+
|
364 |
+
|
365 |
+
class ResCSPB(BottleneckCSPB):
|
366 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
367 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
368 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
369 |
+
c_ = int(c2) # hidden channels
|
370 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
371 |
+
|
372 |
+
|
373 |
+
class ResCSPC(BottleneckCSPC):
|
374 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
375 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
376 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
377 |
+
c_ = int(c2 * e) # hidden channels
|
378 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
379 |
+
|
380 |
+
|
381 |
+
class ResXCSPA(ResCSPA):
|
382 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
383 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
384 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
385 |
+
c_ = int(c2 * e) # hidden channels
|
386 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
387 |
+
|
388 |
+
|
389 |
+
class ResXCSPB(ResCSPB):
|
390 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
391 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
392 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
393 |
+
c_ = int(c2) # hidden channels
|
394 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
395 |
+
|
396 |
+
|
397 |
+
class ResXCSPC(ResCSPC):
|
398 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
399 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
400 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
401 |
+
c_ = int(c2 * e) # hidden channels
|
402 |
+
self.m = nn.Sequential(*[Res(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
403 |
+
|
404 |
+
|
405 |
+
class GhostCSPA(BottleneckCSPA):
|
406 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
407 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
408 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
409 |
+
c_ = int(c2 * e) # hidden channels
|
410 |
+
self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
|
411 |
+
|
412 |
+
|
413 |
+
class GhostCSPB(BottleneckCSPB):
|
414 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
415 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
416 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
417 |
+
c_ = int(c2) # hidden channels
|
418 |
+
self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
|
419 |
+
|
420 |
+
|
421 |
+
class GhostCSPC(BottleneckCSPC):
|
422 |
+
# CSP https://github.com/WongKinYiu/CrossStagePartialNetworks
|
423 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
424 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
425 |
+
c_ = int(c2 * e) # hidden channels
|
426 |
+
self.m = nn.Sequential(*[Ghost(c_, c_) for _ in range(n)])
|
427 |
+
|
428 |
+
##### end of cspnet #####
|
429 |
+
|
430 |
+
|
431 |
+
##### yolor #####
|
432 |
+
|
433 |
+
class ImplicitA(nn.Module):
|
434 |
+
def __init__(self, channel, mean=0., std=.02):
|
435 |
+
super(ImplicitA, self).__init__()
|
436 |
+
self.channel = channel
|
437 |
+
self.mean = mean
|
438 |
+
self.std = std
|
439 |
+
self.implicit = nn.Parameter(torch.zeros(1, channel, 1, 1))
|
440 |
+
nn.init.normal_(self.implicit, mean=self.mean, std=self.std)
|
441 |
+
|
442 |
+
def forward(self, x):
|
443 |
+
return self.implicit + x
|
444 |
+
|
445 |
+
|
446 |
+
class ImplicitM(nn.Module):
|
447 |
+
def __init__(self, channel, mean=0., std=.02):
|
448 |
+
super(ImplicitM, self).__init__()
|
449 |
+
self.channel = channel
|
450 |
+
self.mean = mean
|
451 |
+
self.std = std
|
452 |
+
self.implicit = nn.Parameter(torch.ones(1, channel, 1, 1))
|
453 |
+
nn.init.normal_(self.implicit, mean=self.mean, std=self.std)
|
454 |
+
|
455 |
+
def forward(self, x):
|
456 |
+
return self.implicit * x
|
457 |
+
|
458 |
+
##### end of yolor #####
|
459 |
+
|
460 |
+
|
461 |
+
##### repvgg #####
|
462 |
+
|
463 |
+
class RepConv(nn.Module):
|
464 |
+
# Represented convolution
|
465 |
+
# https://arxiv.org/abs/2101.03697
|
466 |
+
|
467 |
+
def __init__(self, c1, c2, k=3, s=1, p=None, g=1, act=True, deploy=False):
|
468 |
+
super(RepConv, self).__init__()
|
469 |
+
|
470 |
+
self.deploy = deploy
|
471 |
+
self.groups = g
|
472 |
+
self.in_channels = c1
|
473 |
+
self.out_channels = c2
|
474 |
+
|
475 |
+
assert k == 3
|
476 |
+
assert autopad(k, p) == 1
|
477 |
+
|
478 |
+
padding_11 = autopad(k, p) - k // 2
|
479 |
+
|
480 |
+
self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
|
481 |
+
|
482 |
+
if deploy:
|
483 |
+
self.rbr_reparam = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=True)
|
484 |
+
|
485 |
+
else:
|
486 |
+
self.rbr_identity = (nn.BatchNorm2d(num_features=c1) if c2 == c1 and s == 1 else None)
|
487 |
+
|
488 |
+
self.rbr_dense = nn.Sequential(
|
489 |
+
nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False),
|
490 |
+
nn.BatchNorm2d(num_features=c2),
|
491 |
+
)
|
492 |
+
|
493 |
+
self.rbr_1x1 = nn.Sequential(
|
494 |
+
nn.Conv2d( c1, c2, 1, s, padding_11, groups=g, bias=False),
|
495 |
+
nn.BatchNorm2d(num_features=c2),
|
496 |
+
)
|
497 |
+
|
498 |
+
def forward(self, inputs):
|
499 |
+
if hasattr(self, "rbr_reparam"):
|
500 |
+
return self.act(self.rbr_reparam(inputs))
|
501 |
+
|
502 |
+
if self.rbr_identity is None:
|
503 |
+
id_out = 0
|
504 |
+
else:
|
505 |
+
id_out = self.rbr_identity(inputs)
|
506 |
+
|
507 |
+
return self.act(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)
|
508 |
+
|
509 |
+
def get_equivalent_kernel_bias(self):
|
510 |
+
kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
|
511 |
+
kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
|
512 |
+
kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
|
513 |
+
return (
|
514 |
+
kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid,
|
515 |
+
bias3x3 + bias1x1 + biasid,
|
516 |
+
)
|
517 |
+
|
518 |
+
def _pad_1x1_to_3x3_tensor(self, kernel1x1):
|
519 |
+
if kernel1x1 is None:
|
520 |
+
return 0
|
521 |
+
else:
|
522 |
+
return nn.functional.pad(kernel1x1, [1, 1, 1, 1])
|
523 |
+
|
524 |
+
def _fuse_bn_tensor(self, branch):
|
525 |
+
if branch is None:
|
526 |
+
return 0, 0
|
527 |
+
if isinstance(branch, nn.Sequential):
|
528 |
+
kernel = branch[0].weight
|
529 |
+
running_mean = branch[1].running_mean
|
530 |
+
running_var = branch[1].running_var
|
531 |
+
gamma = branch[1].weight
|
532 |
+
beta = branch[1].bias
|
533 |
+
eps = branch[1].eps
|
534 |
+
else:
|
535 |
+
assert isinstance(branch, nn.BatchNorm2d)
|
536 |
+
if not hasattr(self, "id_tensor"):
|
537 |
+
input_dim = self.in_channels // self.groups
|
538 |
+
kernel_value = np.zeros(
|
539 |
+
(self.in_channels, input_dim, 3, 3), dtype=np.float32
|
540 |
+
)
|
541 |
+
for i in range(self.in_channels):
|
542 |
+
kernel_value[i, i % input_dim, 1, 1] = 1
|
543 |
+
self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
|
544 |
+
kernel = self.id_tensor
|
545 |
+
running_mean = branch.running_mean
|
546 |
+
running_var = branch.running_var
|
547 |
+
gamma = branch.weight
|
548 |
+
beta = branch.bias
|
549 |
+
eps = branch.eps
|
550 |
+
std = (running_var + eps).sqrt()
|
551 |
+
t = (gamma / std).reshape(-1, 1, 1, 1)
|
552 |
+
return kernel * t, beta - running_mean * gamma / std
|
553 |
+
|
554 |
+
def repvgg_convert(self):
|
555 |
+
kernel, bias = self.get_equivalent_kernel_bias()
|
556 |
+
return (
|
557 |
+
kernel.detach().cpu().numpy(),
|
558 |
+
bias.detach().cpu().numpy(),
|
559 |
+
)
|
560 |
+
|
561 |
+
def fuse_conv_bn(self, conv, bn):
|
562 |
+
|
563 |
+
std = (bn.running_var + bn.eps).sqrt()
|
564 |
+
bias = bn.bias - bn.running_mean * bn.weight / std
|
565 |
+
|
566 |
+
t = (bn.weight / std).reshape(-1, 1, 1, 1)
|
567 |
+
weights = conv.weight * t
|
568 |
+
|
569 |
+
bn = nn.Identity()
|
570 |
+
conv = nn.Conv2d(in_channels = conv.in_channels,
|
571 |
+
out_channels = conv.out_channels,
|
572 |
+
kernel_size = conv.kernel_size,
|
573 |
+
stride=conv.stride,
|
574 |
+
padding = conv.padding,
|
575 |
+
dilation = conv.dilation,
|
576 |
+
groups = conv.groups,
|
577 |
+
bias = True,
|
578 |
+
padding_mode = conv.padding_mode)
|
579 |
+
|
580 |
+
conv.weight = torch.nn.Parameter(weights)
|
581 |
+
conv.bias = torch.nn.Parameter(bias)
|
582 |
+
return conv
|
583 |
+
|
584 |
+
def fuse_repvgg_block(self):
|
585 |
+
if self.deploy:
|
586 |
+
return
|
587 |
+
print(f"RepConv.fuse_repvgg_block")
|
588 |
+
|
589 |
+
self.rbr_dense = self.fuse_conv_bn(self.rbr_dense[0], self.rbr_dense[1])
|
590 |
+
|
591 |
+
self.rbr_1x1 = self.fuse_conv_bn(self.rbr_1x1[0], self.rbr_1x1[1])
|
592 |
+
rbr_1x1_bias = self.rbr_1x1.bias
|
593 |
+
weight_1x1_expanded = torch.nn.functional.pad(self.rbr_1x1.weight, [1, 1, 1, 1])
|
594 |
+
|
595 |
+
# Fuse self.rbr_identity
|
596 |
+
if (isinstance(self.rbr_identity, nn.BatchNorm2d) or isinstance(self.rbr_identity, nn.modules.batchnorm.SyncBatchNorm)):
|
597 |
+
# print(f"fuse: rbr_identity == BatchNorm2d or SyncBatchNorm")
|
598 |
+
identity_conv_1x1 = nn.Conv2d(
|
599 |
+
in_channels=self.in_channels,
|
600 |
+
out_channels=self.out_channels,
|
601 |
+
kernel_size=1,
|
602 |
+
stride=1,
|
603 |
+
padding=0,
|
604 |
+
groups=self.groups,
|
605 |
+
bias=False)
|
606 |
+
identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.to(self.rbr_1x1.weight.data.device)
|
607 |
+
identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.squeeze().squeeze()
|
608 |
+
# print(f" identity_conv_1x1.weight = {identity_conv_1x1.weight.shape}")
|
609 |
+
identity_conv_1x1.weight.data.fill_(0.0)
|
610 |
+
identity_conv_1x1.weight.data.fill_diagonal_(1.0)
|
611 |
+
identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.unsqueeze(2).unsqueeze(3)
|
612 |
+
# print(f" identity_conv_1x1.weight = {identity_conv_1x1.weight.shape}")
|
613 |
+
|
614 |
+
identity_conv_1x1 = self.fuse_conv_bn(identity_conv_1x1, self.rbr_identity)
|
615 |
+
bias_identity_expanded = identity_conv_1x1.bias
|
616 |
+
weight_identity_expanded = torch.nn.functional.pad(identity_conv_1x1.weight, [1, 1, 1, 1])
|
617 |
+
else:
|
618 |
+
# print(f"fuse: rbr_identity != BatchNorm2d, rbr_identity = {self.rbr_identity}")
|
619 |
+
bias_identity_expanded = torch.nn.Parameter( torch.zeros_like(rbr_1x1_bias) )
|
620 |
+
weight_identity_expanded = torch.nn.Parameter( torch.zeros_like(weight_1x1_expanded) )
|
621 |
+
|
622 |
+
|
623 |
+
#print(f"self.rbr_1x1.weight = {self.rbr_1x1.weight.shape}, ")
|
624 |
+
#print(f"weight_1x1_expanded = {weight_1x1_expanded.shape}, ")
|
625 |
+
#print(f"self.rbr_dense.weight = {self.rbr_dense.weight.shape}, ")
|
626 |
+
|
627 |
+
self.rbr_dense.weight = torch.nn.Parameter(self.rbr_dense.weight + weight_1x1_expanded + weight_identity_expanded)
|
628 |
+
self.rbr_dense.bias = torch.nn.Parameter(self.rbr_dense.bias + rbr_1x1_bias + bias_identity_expanded)
|
629 |
+
|
630 |
+
self.rbr_reparam = self.rbr_dense
|
631 |
+
self.deploy = True
|
632 |
+
|
633 |
+
if self.rbr_identity is not None:
|
634 |
+
del self.rbr_identity
|
635 |
+
self.rbr_identity = None
|
636 |
+
|
637 |
+
if self.rbr_1x1 is not None:
|
638 |
+
del self.rbr_1x1
|
639 |
+
self.rbr_1x1 = None
|
640 |
+
|
641 |
+
if self.rbr_dense is not None:
|
642 |
+
del self.rbr_dense
|
643 |
+
self.rbr_dense = None
|
644 |
+
|
645 |
+
|
646 |
+
class RepBottleneck(Bottleneck):
|
647 |
+
# Standard bottleneck
|
648 |
+
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
649 |
+
super().__init__(c1, c2, shortcut=True, g=1, e=0.5)
|
650 |
+
c_ = int(c2 * e) # hidden channels
|
651 |
+
self.cv2 = RepConv(c_, c2, 3, 1, g=g)
|
652 |
+
|
653 |
+
|
654 |
+
class RepBottleneckCSPA(BottleneckCSPA):
|
655 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
656 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
657 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
658 |
+
c_ = int(c2 * e) # hidden channels
|
659 |
+
self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
660 |
+
|
661 |
+
|
662 |
+
class RepBottleneckCSPB(BottleneckCSPB):
|
663 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
664 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
665 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
666 |
+
c_ = int(c2) # hidden channels
|
667 |
+
self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
668 |
+
|
669 |
+
|
670 |
+
class RepBottleneckCSPC(BottleneckCSPC):
|
671 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
672 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
673 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
674 |
+
c_ = int(c2 * e) # hidden channels
|
675 |
+
self.m = nn.Sequential(*[RepBottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
676 |
+
|
677 |
+
|
678 |
+
class RepRes(Res):
|
679 |
+
# Standard bottleneck
|
680 |
+
def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
681 |
+
super().__init__(c1, c2, shortcut, g, e)
|
682 |
+
c_ = int(c2 * e) # hidden channels
|
683 |
+
self.cv2 = RepConv(c_, c_, 3, 1, g=g)
|
684 |
+
|
685 |
+
|
686 |
+
class RepResCSPA(ResCSPA):
|
687 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
688 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
689 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
690 |
+
c_ = int(c2 * e) # hidden channels
|
691 |
+
self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
692 |
+
|
693 |
+
|
694 |
+
class RepResCSPB(ResCSPB):
|
695 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
696 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
697 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
698 |
+
c_ = int(c2) # hidden channels
|
699 |
+
self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
700 |
+
|
701 |
+
|
702 |
+
class RepResCSPC(ResCSPC):
|
703 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
704 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
705 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
706 |
+
c_ = int(c2 * e) # hidden channels
|
707 |
+
self.m = nn.Sequential(*[RepRes(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
708 |
+
|
709 |
+
|
710 |
+
class RepResX(ResX):
|
711 |
+
# Standard bottleneck
|
712 |
+
def __init__(self, c1, c2, shortcut=True, g=32, e=0.5): # ch_in, ch_out, shortcut, groups, expansion
|
713 |
+
super().__init__(c1, c2, shortcut, g, e)
|
714 |
+
c_ = int(c2 * e) # hidden channels
|
715 |
+
self.cv2 = RepConv(c_, c_, 3, 1, g=g)
|
716 |
+
|
717 |
+
|
718 |
+
class RepResXCSPA(ResXCSPA):
|
719 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
720 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
721 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
722 |
+
c_ = int(c2 * e) # hidden channels
|
723 |
+
self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
724 |
+
|
725 |
+
|
726 |
+
class RepResXCSPB(ResXCSPB):
|
727 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
728 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
729 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
730 |
+
c_ = int(c2) # hidden channels
|
731 |
+
self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
732 |
+
|
733 |
+
|
734 |
+
class RepResXCSPC(ResXCSPC):
|
735 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
736 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=32, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
737 |
+
super().__init__(c1, c2, n, shortcut, g, e)
|
738 |
+
c_ = int(c2 * e) # hidden channels
|
739 |
+
self.m = nn.Sequential(*[RepResX(c_, c_, shortcut, g, e=0.5) for _ in range(n)])
|
740 |
+
|
741 |
+
##### end of repvgg #####
|
742 |
+
|
743 |
+
|
744 |
+
##### transformer #####
|
745 |
+
|
746 |
+
class TransformerLayer(nn.Module):
|
747 |
+
# Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
|
748 |
+
def __init__(self, c, num_heads):
|
749 |
+
super().__init__()
|
750 |
+
self.q = nn.Linear(c, c, bias=False)
|
751 |
+
self.k = nn.Linear(c, c, bias=False)
|
752 |
+
self.v = nn.Linear(c, c, bias=False)
|
753 |
+
self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
|
754 |
+
self.fc1 = nn.Linear(c, c, bias=False)
|
755 |
+
self.fc2 = nn.Linear(c, c, bias=False)
|
756 |
+
|
757 |
+
def forward(self, x):
|
758 |
+
x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x
|
759 |
+
x = self.fc2(self.fc1(x)) + x
|
760 |
+
return x
|
761 |
+
|
762 |
+
|
763 |
+
class TransformerBlock(nn.Module):
|
764 |
+
# Vision Transformer https://arxiv.org/abs/2010.11929
|
765 |
+
def __init__(self, c1, c2, num_heads, num_layers):
|
766 |
+
super().__init__()
|
767 |
+
self.conv = None
|
768 |
+
if c1 != c2:
|
769 |
+
self.conv = Conv(c1, c2)
|
770 |
+
self.linear = nn.Linear(c2, c2) # learnable position embedding
|
771 |
+
self.tr = nn.Sequential(*[TransformerLayer(c2, num_heads) for _ in range(num_layers)])
|
772 |
+
self.c2 = c2
|
773 |
+
|
774 |
+
def forward(self, x):
|
775 |
+
if self.conv is not None:
|
776 |
+
x = self.conv(x)
|
777 |
+
b, _, w, h = x.shape
|
778 |
+
p = x.flatten(2)
|
779 |
+
p = p.unsqueeze(0)
|
780 |
+
p = p.transpose(0, 3)
|
781 |
+
p = p.squeeze(3)
|
782 |
+
e = self.linear(p)
|
783 |
+
x = p + e
|
784 |
+
|
785 |
+
x = self.tr(x)
|
786 |
+
x = x.unsqueeze(3)
|
787 |
+
x = x.transpose(0, 3)
|
788 |
+
x = x.reshape(b, self.c2, w, h)
|
789 |
+
return x
|
790 |
+
|
791 |
+
##### end of transformer #####
|
792 |
+
|
793 |
+
|
794 |
+
##### yolov5 #####
|
795 |
+
|
796 |
+
class Focus(nn.Module):
|
797 |
+
# Focus wh information into c-space
|
798 |
+
def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups
|
799 |
+
super(Focus, self).__init__()
|
800 |
+
self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
|
801 |
+
# self.contract = Contract(gain=2)
|
802 |
+
|
803 |
+
def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2)
|
804 |
+
return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
|
805 |
+
# return self.conv(self.contract(x))
|
806 |
+
|
807 |
+
|
808 |
+
class SPPF(nn.Module):
|
809 |
+
# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
|
810 |
+
def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13))
|
811 |
+
super().__init__()
|
812 |
+
c_ = c1 // 2 # hidden channels
|
813 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
814 |
+
self.cv2 = Conv(c_ * 4, c2, 1, 1)
|
815 |
+
self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
|
816 |
+
|
817 |
+
def forward(self, x):
|
818 |
+
x = self.cv1(x)
|
819 |
+
y1 = self.m(x)
|
820 |
+
y2 = self.m(y1)
|
821 |
+
return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
|
822 |
+
|
823 |
+
|
824 |
+
class Contract(nn.Module):
|
825 |
+
# Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
|
826 |
+
def __init__(self, gain=2):
|
827 |
+
super().__init__()
|
828 |
+
self.gain = gain
|
829 |
+
|
830 |
+
def forward(self, x):
|
831 |
+
N, C, H, W = x.size() # assert (H / s == 0) and (W / s == 0), 'Indivisible gain'
|
832 |
+
s = self.gain
|
833 |
+
x = x.view(N, C, H // s, s, W // s, s) # x(1,64,40,2,40,2)
|
834 |
+
x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40)
|
835 |
+
return x.view(N, C * s * s, H // s, W // s) # x(1,256,40,40)
|
836 |
+
|
837 |
+
|
838 |
+
class Expand(nn.Module):
|
839 |
+
# Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
|
840 |
+
def __init__(self, gain=2):
|
841 |
+
super().__init__()
|
842 |
+
self.gain = gain
|
843 |
+
|
844 |
+
def forward(self, x):
|
845 |
+
N, C, H, W = x.size() # assert C / s ** 2 == 0, 'Indivisible gain'
|
846 |
+
s = self.gain
|
847 |
+
x = x.view(N, s, s, C // s ** 2, H, W) # x(1,2,2,16,80,80)
|
848 |
+
x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2)
|
849 |
+
return x.view(N, C // s ** 2, H * s, W * s) # x(1,16,160,160)
|
850 |
+
|
851 |
+
|
852 |
+
class NMS(nn.Module):
|
853 |
+
# Non-Maximum Suppression (NMS) module
|
854 |
+
conf = 0.25 # confidence threshold
|
855 |
+
iou = 0.45 # IoU threshold
|
856 |
+
classes = None # (optional list) filter by class
|
857 |
+
|
858 |
+
def __init__(self):
|
859 |
+
super(NMS, self).__init__()
|
860 |
+
|
861 |
+
def forward(self, x):
|
862 |
+
return non_max_suppression(x[0], conf_thres=self.conf, iou_thres=self.iou, classes=self.classes)
|
863 |
+
|
864 |
+
|
865 |
+
class autoShape(nn.Module):
|
866 |
+
# input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS
|
867 |
+
conf = 0.25 # NMS confidence threshold
|
868 |
+
iou = 0.45 # NMS IoU threshold
|
869 |
+
classes = None # (optional list) filter by class
|
870 |
+
|
871 |
+
def __init__(self, model):
|
872 |
+
super(autoShape, self).__init__()
|
873 |
+
self.model = model.eval()
|
874 |
+
|
875 |
+
def autoshape(self):
|
876 |
+
print('autoShape already enabled, skipping... ') # model already converted to model.autoshape()
|
877 |
+
return self
|
878 |
+
|
879 |
+
@torch.no_grad()
|
880 |
+
def forward(self, imgs, size=640, augment=False, profile=False):
|
881 |
+
# Inference from various sources. For height=640, width=1280, RGB images example inputs are:
|
882 |
+
# filename: imgs = 'data/samples/zidane.jpg'
|
883 |
+
# URI: = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/zidane.jpg'
|
884 |
+
# OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3)
|
885 |
+
# PIL: = Image.open('image.jpg') # HWC x(640,1280,3)
|
886 |
+
# numpy: = np.zeros((640,1280,3)) # HWC
|
887 |
+
# torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values)
|
888 |
+
# multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images
|
889 |
+
|
890 |
+
t = [time_synchronized()]
|
891 |
+
p = next(self.model.parameters()) # for device and type
|
892 |
+
if isinstance(imgs, torch.Tensor): # torch
|
893 |
+
with amp.autocast(enabled=p.device.type != 'cpu'):
|
894 |
+
return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference
|
895 |
+
|
896 |
+
# Pre-process
|
897 |
+
n, imgs = (len(imgs), imgs) if isinstance(imgs, list) else (1, [imgs]) # number of images, list of images
|
898 |
+
shape0, shape1, files = [], [], [] # image and inference shapes, filenames
|
899 |
+
for i, im in enumerate(imgs):
|
900 |
+
f = f'image{i}' # filename
|
901 |
+
if isinstance(im, str): # filename or uri
|
902 |
+
im, f = np.asarray(Image.open(requests.get(im, stream=True).raw if im.startswith('http') else im)), im
|
903 |
+
elif isinstance(im, Image.Image): # PIL Image
|
904 |
+
im, f = np.asarray(im), getattr(im, 'filename', f) or f
|
905 |
+
files.append(Path(f).with_suffix('.jpg').name)
|
906 |
+
if im.shape[0] < 5: # image in CHW
|
907 |
+
im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1)
|
908 |
+
im = im[:, :, :3] if im.ndim == 3 else np.tile(im[:, :, None], 3) # enforce 3ch input
|
909 |
+
s = im.shape[:2] # HWC
|
910 |
+
shape0.append(s) # image shape
|
911 |
+
g = (size / max(s)) # gain
|
912 |
+
shape1.append([y * g for y in s])
|
913 |
+
imgs[i] = im # update
|
914 |
+
shape1 = [make_divisible(x, int(self.stride.max())) for x in np.stack(shape1, 0).max(0)] # inference shape
|
915 |
+
x = [letterbox(im, new_shape=shape1, auto=False)[0] for im in imgs] # pad
|
916 |
+
x = np.stack(x, 0) if n > 1 else x[0][None] # stack
|
917 |
+
x = np.ascontiguousarray(x.transpose((0, 3, 1, 2))) # BHWC to BCHW
|
918 |
+
x = torch.from_numpy(x).to(p.device).type_as(p) / 255. # uint8 to fp16/32
|
919 |
+
t.append(time_synchronized())
|
920 |
+
|
921 |
+
with amp.autocast(enabled=p.device.type != 'cpu'):
|
922 |
+
# Inference
|
923 |
+
y = self.model(x, augment, profile)[0] # forward
|
924 |
+
t.append(time_synchronized())
|
925 |
+
|
926 |
+
# Post-process
|
927 |
+
y = non_max_suppression(y, conf_thres=self.conf, iou_thres=self.iou, classes=self.classes) # NMS
|
928 |
+
for i in range(n):
|
929 |
+
scale_coords(shape1, y[i][:, :4], shape0[i])
|
930 |
+
|
931 |
+
t.append(time_synchronized())
|
932 |
+
return Detections(imgs, y, files, t, self.names, x.shape)
|
933 |
+
|
934 |
+
|
935 |
+
class Detections:
|
936 |
+
# detections class for YOLOv5 inference results
|
937 |
+
def __init__(self, imgs, pred, files, times=None, names=None, shape=None):
|
938 |
+
super(Detections, self).__init__()
|
939 |
+
d = pred[0].device # device
|
940 |
+
gn = [torch.tensor([*[im.shape[i] for i in [1, 0, 1, 0]], 1., 1.], device=d) for im in imgs] # normalizations
|
941 |
+
self.imgs = imgs # list of images as numpy arrays
|
942 |
+
self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls)
|
943 |
+
self.names = names # class names
|
944 |
+
self.files = files # image filenames
|
945 |
+
self.xyxy = pred # xyxy pixels
|
946 |
+
self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels
|
947 |
+
self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized
|
948 |
+
self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized
|
949 |
+
self.n = len(self.pred) # number of images (batch size)
|
950 |
+
self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms)
|
951 |
+
self.s = shape # inference BCHW shape
|
952 |
+
|
953 |
+
def display(self, pprint=False, show=False, save=False, render=False, save_dir=''):
|
954 |
+
colors = color_list()
|
955 |
+
for i, (img, pred) in enumerate(zip(self.imgs, self.pred)):
|
956 |
+
str = f'image {i + 1}/{len(self.pred)}: {img.shape[0]}x{img.shape[1]} '
|
957 |
+
if pred is not None:
|
958 |
+
for c in pred[:, -1].unique():
|
959 |
+
n = (pred[:, -1] == c).sum() # detections per class
|
960 |
+
str += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string
|
961 |
+
if show or save or render:
|
962 |
+
for *box, conf, cls in pred: # xyxy, confidence, class
|
963 |
+
label = f'{self.names[int(cls)]} {conf:.2f}'
|
964 |
+
plot_one_box(box, img, label=label, color=colors[int(cls) % 10])
|
965 |
+
img = Image.fromarray(img.astype(np.uint8)) if isinstance(img, np.ndarray) else img # from np
|
966 |
+
if pprint:
|
967 |
+
print(str.rstrip(', '))
|
968 |
+
if show:
|
969 |
+
img.show(self.files[i]) # show
|
970 |
+
if save:
|
971 |
+
f = self.files[i]
|
972 |
+
img.save(Path(save_dir) / f) # save
|
973 |
+
print(f"{'Saved' * (i == 0)} {f}", end=',' if i < self.n - 1 else f' to {save_dir}\n')
|
974 |
+
if render:
|
975 |
+
self.imgs[i] = np.asarray(img)
|
976 |
+
|
977 |
+
def print(self):
|
978 |
+
self.display(pprint=True) # print results
|
979 |
+
print(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' % self.t)
|
980 |
+
|
981 |
+
def show(self):
|
982 |
+
self.display(show=True) # show results
|
983 |
+
|
984 |
+
def save(self, save_dir='runs/hub/exp'):
|
985 |
+
save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/hub/exp') # increment save_dir
|
986 |
+
Path(save_dir).mkdir(parents=True, exist_ok=True)
|
987 |
+
self.display(save=True, save_dir=save_dir) # save results
|
988 |
+
|
989 |
+
def render(self):
|
990 |
+
self.display(render=True) # render results
|
991 |
+
return self.imgs
|
992 |
+
|
993 |
+
def pandas(self):
|
994 |
+
# return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0])
|
995 |
+
new = copy(self) # return copy
|
996 |
+
ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns
|
997 |
+
cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns
|
998 |
+
for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]):
|
999 |
+
a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update
|
1000 |
+
setattr(new, k, [pd.DataFrame(x, columns=c) for x in a])
|
1001 |
+
return new
|
1002 |
+
|
1003 |
+
def tolist(self):
|
1004 |
+
# return a list of Detections objects, i.e. 'for result in results.tolist():'
|
1005 |
+
x = [Detections([self.imgs[i]], [self.pred[i]], self.names, self.s) for i in range(self.n)]
|
1006 |
+
for d in x:
|
1007 |
+
for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']:
|
1008 |
+
setattr(d, k, getattr(d, k)[0]) # pop out of list
|
1009 |
+
return x
|
1010 |
+
|
1011 |
+
def __len__(self):
|
1012 |
+
return self.n
|
1013 |
+
|
1014 |
+
|
1015 |
+
class Classify(nn.Module):
|
1016 |
+
# Classification head, i.e. x(b,c1,20,20) to x(b,c2)
|
1017 |
+
def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups
|
1018 |
+
super(Classify, self).__init__()
|
1019 |
+
self.aap = nn.AdaptiveAvgPool2d(1) # to x(b,c1,1,1)
|
1020 |
+
self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g) # to x(b,c2,1,1)
|
1021 |
+
self.flat = nn.Flatten()
|
1022 |
+
|
1023 |
+
def forward(self, x):
|
1024 |
+
z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1) # cat if list
|
1025 |
+
return self.flat(self.conv(z)) # flatten to x(b,c2)
|
1026 |
+
|
1027 |
+
##### end of yolov5 ######
|
1028 |
+
|
1029 |
+
|
1030 |
+
##### orepa #####
|
1031 |
+
|
1032 |
+
def transI_fusebn(kernel, bn):
|
1033 |
+
gamma = bn.weight
|
1034 |
+
std = (bn.running_var + bn.eps).sqrt()
|
1035 |
+
return kernel * ((gamma / std).reshape(-1, 1, 1, 1)), bn.bias - bn.running_mean * gamma / std
|
1036 |
+
|
1037 |
+
|
1038 |
+
class ConvBN(nn.Module):
|
1039 |
+
def __init__(self, in_channels, out_channels, kernel_size,
|
1040 |
+
stride=1, padding=0, dilation=1, groups=1, deploy=False, nonlinear=None):
|
1041 |
+
super().__init__()
|
1042 |
+
if nonlinear is None:
|
1043 |
+
self.nonlinear = nn.Identity()
|
1044 |
+
else:
|
1045 |
+
self.nonlinear = nonlinear
|
1046 |
+
if deploy:
|
1047 |
+
self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
|
1048 |
+
stride=stride, padding=padding, dilation=dilation, groups=groups, bias=True)
|
1049 |
+
else:
|
1050 |
+
self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
|
1051 |
+
stride=stride, padding=padding, dilation=dilation, groups=groups, bias=False)
|
1052 |
+
self.bn = nn.BatchNorm2d(num_features=out_channels)
|
1053 |
+
|
1054 |
+
def forward(self, x):
|
1055 |
+
if hasattr(self, 'bn'):
|
1056 |
+
return self.nonlinear(self.bn(self.conv(x)))
|
1057 |
+
else:
|
1058 |
+
return self.nonlinear(self.conv(x))
|
1059 |
+
|
1060 |
+
def switch_to_deploy(self):
|
1061 |
+
kernel, bias = transI_fusebn(self.conv.weight, self.bn)
|
1062 |
+
conv = nn.Conv2d(in_channels=self.conv.in_channels, out_channels=self.conv.out_channels, kernel_size=self.conv.kernel_size,
|
1063 |
+
stride=self.conv.stride, padding=self.conv.padding, dilation=self.conv.dilation, groups=self.conv.groups, bias=True)
|
1064 |
+
conv.weight.data = kernel
|
1065 |
+
conv.bias.data = bias
|
1066 |
+
for para in self.parameters():
|
1067 |
+
para.detach_()
|
1068 |
+
self.__delattr__('conv')
|
1069 |
+
self.__delattr__('bn')
|
1070 |
+
self.conv = conv
|
1071 |
+
|
1072 |
+
class OREPA_3x3_RepConv(nn.Module):
|
1073 |
+
|
1074 |
+
def __init__(self, in_channels, out_channels, kernel_size,
|
1075 |
+
stride=1, padding=0, dilation=1, groups=1,
|
1076 |
+
internal_channels_1x1_3x3=None,
|
1077 |
+
deploy=False, nonlinear=None, single_init=False):
|
1078 |
+
super(OREPA_3x3_RepConv, self).__init__()
|
1079 |
+
self.deploy = deploy
|
1080 |
+
|
1081 |
+
if nonlinear is None:
|
1082 |
+
self.nonlinear = nn.Identity()
|
1083 |
+
else:
|
1084 |
+
self.nonlinear = nonlinear
|
1085 |
+
|
1086 |
+
self.kernel_size = kernel_size
|
1087 |
+
self.in_channels = in_channels
|
1088 |
+
self.out_channels = out_channels
|
1089 |
+
self.groups = groups
|
1090 |
+
assert padding == kernel_size // 2
|
1091 |
+
|
1092 |
+
self.stride = stride
|
1093 |
+
self.padding = padding
|
1094 |
+
self.dilation = dilation
|
1095 |
+
|
1096 |
+
self.branch_counter = 0
|
1097 |
+
|
1098 |
+
self.weight_rbr_origin = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), kernel_size, kernel_size))
|
1099 |
+
nn.init.kaiming_uniform_(self.weight_rbr_origin, a=math.sqrt(1.0))
|
1100 |
+
self.branch_counter += 1
|
1101 |
+
|
1102 |
+
|
1103 |
+
if groups < out_channels:
|
1104 |
+
self.weight_rbr_avg_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
|
1105 |
+
self.weight_rbr_pfir_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
|
1106 |
+
nn.init.kaiming_uniform_(self.weight_rbr_avg_conv, a=1.0)
|
1107 |
+
nn.init.kaiming_uniform_(self.weight_rbr_pfir_conv, a=1.0)
|
1108 |
+
self.weight_rbr_avg_conv.data
|
1109 |
+
self.weight_rbr_pfir_conv.data
|
1110 |
+
self.register_buffer('weight_rbr_avg_avg', torch.ones(kernel_size, kernel_size).mul(1.0/kernel_size/kernel_size))
|
1111 |
+
self.branch_counter += 1
|
1112 |
+
|
1113 |
+
else:
|
1114 |
+
raise NotImplementedError
|
1115 |
+
self.branch_counter += 1
|
1116 |
+
|
1117 |
+
if internal_channels_1x1_3x3 is None:
|
1118 |
+
internal_channels_1x1_3x3 = in_channels if groups < out_channels else 2 * in_channels # For mobilenet, it is better to have 2X internal channels
|
1119 |
+
|
1120 |
+
if internal_channels_1x1_3x3 == in_channels:
|
1121 |
+
self.weight_rbr_1x1_kxk_idconv1 = nn.Parameter(torch.zeros(in_channels, int(in_channels/self.groups), 1, 1))
|
1122 |
+
id_value = np.zeros((in_channels, int(in_channels/self.groups), 1, 1))
|
1123 |
+
for i in range(in_channels):
|
1124 |
+
id_value[i, i % int(in_channels/self.groups), 0, 0] = 1
|
1125 |
+
id_tensor = torch.from_numpy(id_value).type_as(self.weight_rbr_1x1_kxk_idconv1)
|
1126 |
+
self.register_buffer('id_tensor', id_tensor)
|
1127 |
+
|
1128 |
+
else:
|
1129 |
+
self.weight_rbr_1x1_kxk_conv1 = nn.Parameter(torch.Tensor(internal_channels_1x1_3x3, int(in_channels/self.groups), 1, 1))
|
1130 |
+
nn.init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv1, a=math.sqrt(1.0))
|
1131 |
+
self.weight_rbr_1x1_kxk_conv2 = nn.Parameter(torch.Tensor(out_channels, int(internal_channels_1x1_3x3/self.groups), kernel_size, kernel_size))
|
1132 |
+
nn.init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv2, a=math.sqrt(1.0))
|
1133 |
+
self.branch_counter += 1
|
1134 |
+
|
1135 |
+
expand_ratio = 8
|
1136 |
+
self.weight_rbr_gconv_dw = nn.Parameter(torch.Tensor(in_channels*expand_ratio, 1, kernel_size, kernel_size))
|
1137 |
+
self.weight_rbr_gconv_pw = nn.Parameter(torch.Tensor(out_channels, in_channels*expand_ratio, 1, 1))
|
1138 |
+
nn.init.kaiming_uniform_(self.weight_rbr_gconv_dw, a=math.sqrt(1.0))
|
1139 |
+
nn.init.kaiming_uniform_(self.weight_rbr_gconv_pw, a=math.sqrt(1.0))
|
1140 |
+
self.branch_counter += 1
|
1141 |
+
|
1142 |
+
if out_channels == in_channels and stride == 1:
|
1143 |
+
self.branch_counter += 1
|
1144 |
+
|
1145 |
+
self.vector = nn.Parameter(torch.Tensor(self.branch_counter, self.out_channels))
|
1146 |
+
self.bn = nn.BatchNorm2d(out_channels)
|
1147 |
+
|
1148 |
+
self.fre_init()
|
1149 |
+
|
1150 |
+
nn.init.constant_(self.vector[0, :], 0.25) #origin
|
1151 |
+
nn.init.constant_(self.vector[1, :], 0.25) #avg
|
1152 |
+
nn.init.constant_(self.vector[2, :], 0.0) #prior
|
1153 |
+
nn.init.constant_(self.vector[3, :], 0.5) #1x1_kxk
|
1154 |
+
nn.init.constant_(self.vector[4, :], 0.5) #dws_conv
|
1155 |
+
|
1156 |
+
|
1157 |
+
def fre_init(self):
|
1158 |
+
prior_tensor = torch.Tensor(self.out_channels, self.kernel_size, self.kernel_size)
|
1159 |
+
half_fg = self.out_channels/2
|
1160 |
+
for i in range(self.out_channels):
|
1161 |
+
for h in range(3):
|
1162 |
+
for w in range(3):
|
1163 |
+
if i < half_fg:
|
1164 |
+
prior_tensor[i, h, w] = math.cos(math.pi*(h+0.5)*(i+1)/3)
|
1165 |
+
else:
|
1166 |
+
prior_tensor[i, h, w] = math.cos(math.pi*(w+0.5)*(i+1-half_fg)/3)
|
1167 |
+
|
1168 |
+
self.register_buffer('weight_rbr_prior', prior_tensor)
|
1169 |
+
|
1170 |
+
def weight_gen(self):
|
1171 |
+
|
1172 |
+
weight_rbr_origin = torch.einsum('oihw,o->oihw', self.weight_rbr_origin, self.vector[0, :])
|
1173 |
+
|
1174 |
+
weight_rbr_avg = torch.einsum('oihw,o->oihw', torch.einsum('oihw,hw->oihw', self.weight_rbr_avg_conv, self.weight_rbr_avg_avg), self.vector[1, :])
|
1175 |
+
|
1176 |
+
weight_rbr_pfir = torch.einsum('oihw,o->oihw', torch.einsum('oihw,ohw->oihw', self.weight_rbr_pfir_conv, self.weight_rbr_prior), self.vector[2, :])
|
1177 |
+
|
1178 |
+
weight_rbr_1x1_kxk_conv1 = None
|
1179 |
+
if hasattr(self, 'weight_rbr_1x1_kxk_idconv1'):
|
1180 |
+
weight_rbr_1x1_kxk_conv1 = (self.weight_rbr_1x1_kxk_idconv1 + self.id_tensor).squeeze()
|
1181 |
+
elif hasattr(self, 'weight_rbr_1x1_kxk_conv1'):
|
1182 |
+
weight_rbr_1x1_kxk_conv1 = self.weight_rbr_1x1_kxk_conv1.squeeze()
|
1183 |
+
else:
|
1184 |
+
raise NotImplementedError
|
1185 |
+
weight_rbr_1x1_kxk_conv2 = self.weight_rbr_1x1_kxk_conv2
|
1186 |
+
|
1187 |
+
if self.groups > 1:
|
1188 |
+
g = self.groups
|
1189 |
+
t, ig = weight_rbr_1x1_kxk_conv1.size()
|
1190 |
+
o, tg, h, w = weight_rbr_1x1_kxk_conv2.size()
|
1191 |
+
weight_rbr_1x1_kxk_conv1 = weight_rbr_1x1_kxk_conv1.view(g, int(t/g), ig)
|
1192 |
+
weight_rbr_1x1_kxk_conv2 = weight_rbr_1x1_kxk_conv2.view(g, int(o/g), tg, h, w)
|
1193 |
+
weight_rbr_1x1_kxk = torch.einsum('gti,gothw->goihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2).view(o, ig, h, w)
|
1194 |
+
else:
|
1195 |
+
weight_rbr_1x1_kxk = torch.einsum('ti,othw->oihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2)
|
1196 |
+
|
1197 |
+
weight_rbr_1x1_kxk = torch.einsum('oihw,o->oihw', weight_rbr_1x1_kxk, self.vector[3, :])
|
1198 |
+
|
1199 |
+
weight_rbr_gconv = self.dwsc2full(self.weight_rbr_gconv_dw, self.weight_rbr_gconv_pw, self.in_channels)
|
1200 |
+
weight_rbr_gconv = torch.einsum('oihw,o->oihw', weight_rbr_gconv, self.vector[4, :])
|
1201 |
+
|
1202 |
+
weight = weight_rbr_origin + weight_rbr_avg + weight_rbr_1x1_kxk + weight_rbr_pfir + weight_rbr_gconv
|
1203 |
+
|
1204 |
+
return weight
|
1205 |
+
|
1206 |
+
def dwsc2full(self, weight_dw, weight_pw, groups):
|
1207 |
+
|
1208 |
+
t, ig, h, w = weight_dw.size()
|
1209 |
+
o, _, _, _ = weight_pw.size()
|
1210 |
+
tg = int(t/groups)
|
1211 |
+
i = int(ig*groups)
|
1212 |
+
weight_dw = weight_dw.view(groups, tg, ig, h, w)
|
1213 |
+
weight_pw = weight_pw.squeeze().view(o, groups, tg)
|
1214 |
+
|
1215 |
+
weight_dsc = torch.einsum('gtihw,ogt->ogihw', weight_dw, weight_pw)
|
1216 |
+
return weight_dsc.view(o, i, h, w)
|
1217 |
+
|
1218 |
+
def forward(self, inputs):
|
1219 |
+
weight = self.weight_gen()
|
1220 |
+
out = F.conv2d(inputs, weight, bias=None, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)
|
1221 |
+
|
1222 |
+
return self.nonlinear(self.bn(out))
|
1223 |
+
|
1224 |
+
class RepConv_OREPA(nn.Module):
|
1225 |
+
|
1226 |
+
def __init__(self, c1, c2, k=3, s=1, padding=1, dilation=1, groups=1, padding_mode='zeros', deploy=False, use_se=False, nonlinear=nn.SiLU()):
|
1227 |
+
super(RepConv_OREPA, self).__init__()
|
1228 |
+
self.deploy = deploy
|
1229 |
+
self.groups = groups
|
1230 |
+
self.in_channels = c1
|
1231 |
+
self.out_channels = c2
|
1232 |
+
|
1233 |
+
self.padding = padding
|
1234 |
+
self.dilation = dilation
|
1235 |
+
self.groups = groups
|
1236 |
+
|
1237 |
+
assert k == 3
|
1238 |
+
assert padding == 1
|
1239 |
+
|
1240 |
+
padding_11 = padding - k // 2
|
1241 |
+
|
1242 |
+
if nonlinear is None:
|
1243 |
+
self.nonlinearity = nn.Identity()
|
1244 |
+
else:
|
1245 |
+
self.nonlinearity = nonlinear
|
1246 |
+
|
1247 |
+
if use_se:
|
1248 |
+
self.se = SEBlock(self.out_channels, internal_neurons=self.out_channels // 16)
|
1249 |
+
else:
|
1250 |
+
self.se = nn.Identity()
|
1251 |
+
|
1252 |
+
if deploy:
|
1253 |
+
self.rbr_reparam = nn.Conv2d(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=k, stride=s,
|
1254 |
+
padding=padding, dilation=dilation, groups=groups, bias=True, padding_mode=padding_mode)
|
1255 |
+
|
1256 |
+
else:
|
1257 |
+
self.rbr_identity = nn.BatchNorm2d(num_features=self.in_channels) if self.out_channels == self.in_channels and s == 1 else None
|
1258 |
+
self.rbr_dense = OREPA_3x3_RepConv(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=k, stride=s, padding=padding, groups=groups, dilation=1)
|
1259 |
+
self.rbr_1x1 = ConvBN(in_channels=self.in_channels, out_channels=self.out_channels, kernel_size=1, stride=s, padding=padding_11, groups=groups, dilation=1)
|
1260 |
+
print('RepVGG Block, identity = ', self.rbr_identity)
|
1261 |
+
|
1262 |
+
|
1263 |
+
def forward(self, inputs):
|
1264 |
+
if hasattr(self, 'rbr_reparam'):
|
1265 |
+
return self.nonlinearity(self.se(self.rbr_reparam(inputs)))
|
1266 |
+
|
1267 |
+
if self.rbr_identity is None:
|
1268 |
+
id_out = 0
|
1269 |
+
else:
|
1270 |
+
id_out = self.rbr_identity(inputs)
|
1271 |
+
|
1272 |
+
out1 = self.rbr_dense(inputs)
|
1273 |
+
out2 = self.rbr_1x1(inputs)
|
1274 |
+
out3 = id_out
|
1275 |
+
out = out1 + out2 + out3
|
1276 |
+
|
1277 |
+
return self.nonlinearity(self.se(out))
|
1278 |
+
|
1279 |
+
|
1280 |
+
# Optional. This improves the accuracy and facilitates quantization.
|
1281 |
+
# 1. Cancel the original weight decay on rbr_dense.conv.weight and rbr_1x1.conv.weight.
|
1282 |
+
# 2. Use like this.
|
1283 |
+
# loss = criterion(....)
|
1284 |
+
# for every RepVGGBlock blk:
|
1285 |
+
# loss += weight_decay_coefficient * 0.5 * blk.get_cust_L2()
|
1286 |
+
# optimizer.zero_grad()
|
1287 |
+
# loss.backward()
|
1288 |
+
|
1289 |
+
# Not used for OREPA
|
1290 |
+
def get_custom_L2(self):
|
1291 |
+
K3 = self.rbr_dense.weight_gen()
|
1292 |
+
K1 = self.rbr_1x1.conv.weight
|
1293 |
+
t3 = (self.rbr_dense.bn.weight / ((self.rbr_dense.bn.running_var + self.rbr_dense.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
|
1294 |
+
t1 = (self.rbr_1x1.bn.weight / ((self.rbr_1x1.bn.running_var + self.rbr_1x1.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
|
1295 |
+
|
1296 |
+
l2_loss_circle = (K3 ** 2).sum() - (K3[:, :, 1:2, 1:2] ** 2).sum() # The L2 loss of the "circle" of weights in 3x3 kernel. Use regular L2 on them.
|
1297 |
+
eq_kernel = K3[:, :, 1:2, 1:2] * t3 + K1 * t1 # The equivalent resultant central point of 3x3 kernel.
|
1298 |
+
l2_loss_eq_kernel = (eq_kernel ** 2 / (t3 ** 2 + t1 ** 2)).sum() # Normalize for an L2 coefficient comparable to regular L2.
|
1299 |
+
return l2_loss_eq_kernel + l2_loss_circle
|
1300 |
+
|
1301 |
+
def get_equivalent_kernel_bias(self):
|
1302 |
+
kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
|
1303 |
+
kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
|
1304 |
+
kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
|
1305 |
+
return kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid, bias3x3 + bias1x1 + biasid
|
1306 |
+
|
1307 |
+
def _pad_1x1_to_3x3_tensor(self, kernel1x1):
|
1308 |
+
if kernel1x1 is None:
|
1309 |
+
return 0
|
1310 |
+
else:
|
1311 |
+
return torch.nn.functional.pad(kernel1x1, [1,1,1,1])
|
1312 |
+
|
1313 |
+
def _fuse_bn_tensor(self, branch):
|
1314 |
+
if branch is None:
|
1315 |
+
return 0, 0
|
1316 |
+
if not isinstance(branch, nn.BatchNorm2d):
|
1317 |
+
if isinstance(branch, OREPA_3x3_RepConv):
|
1318 |
+
kernel = branch.weight_gen()
|
1319 |
+
elif isinstance(branch, ConvBN):
|
1320 |
+
kernel = branch.conv.weight
|
1321 |
+
else:
|
1322 |
+
raise NotImplementedError
|
1323 |
+
running_mean = branch.bn.running_mean
|
1324 |
+
running_var = branch.bn.running_var
|
1325 |
+
gamma = branch.bn.weight
|
1326 |
+
beta = branch.bn.bias
|
1327 |
+
eps = branch.bn.eps
|
1328 |
+
else:
|
1329 |
+
if not hasattr(self, 'id_tensor'):
|
1330 |
+
input_dim = self.in_channels // self.groups
|
1331 |
+
kernel_value = np.zeros((self.in_channels, input_dim, 3, 3), dtype=np.float32)
|
1332 |
+
for i in range(self.in_channels):
|
1333 |
+
kernel_value[i, i % input_dim, 1, 1] = 1
|
1334 |
+
self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
|
1335 |
+
kernel = self.id_tensor
|
1336 |
+
running_mean = branch.running_mean
|
1337 |
+
running_var = branch.running_var
|
1338 |
+
gamma = branch.weight
|
1339 |
+
beta = branch.bias
|
1340 |
+
eps = branch.eps
|
1341 |
+
std = (running_var + eps).sqrt()
|
1342 |
+
t = (gamma / std).reshape(-1, 1, 1, 1)
|
1343 |
+
return kernel * t, beta - running_mean * gamma / std
|
1344 |
+
|
1345 |
+
def switch_to_deploy(self):
|
1346 |
+
if hasattr(self, 'rbr_reparam'):
|
1347 |
+
return
|
1348 |
+
print(f"RepConv_OREPA.switch_to_deploy")
|
1349 |
+
kernel, bias = self.get_equivalent_kernel_bias()
|
1350 |
+
self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.in_channels, out_channels=self.rbr_dense.out_channels,
|
1351 |
+
kernel_size=self.rbr_dense.kernel_size, stride=self.rbr_dense.stride,
|
1352 |
+
padding=self.rbr_dense.padding, dilation=self.rbr_dense.dilation, groups=self.rbr_dense.groups, bias=True)
|
1353 |
+
self.rbr_reparam.weight.data = kernel
|
1354 |
+
self.rbr_reparam.bias.data = bias
|
1355 |
+
for para in self.parameters():
|
1356 |
+
para.detach_()
|
1357 |
+
self.__delattr__('rbr_dense')
|
1358 |
+
self.__delattr__('rbr_1x1')
|
1359 |
+
if hasattr(self, 'rbr_identity'):
|
1360 |
+
self.__delattr__('rbr_identity')
|
1361 |
+
|
1362 |
+
##### end of orepa #####
|
1363 |
+
|
1364 |
+
|
1365 |
+
##### swin transformer #####
|
1366 |
+
|
1367 |
+
class WindowAttention(nn.Module):
|
1368 |
+
|
1369 |
+
def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
|
1370 |
+
|
1371 |
+
super().__init__()
|
1372 |
+
self.dim = dim
|
1373 |
+
self.window_size = window_size # Wh, Ww
|
1374 |
+
self.num_heads = num_heads
|
1375 |
+
head_dim = dim // num_heads
|
1376 |
+
self.scale = qk_scale or head_dim ** -0.5
|
1377 |
+
|
1378 |
+
# define a parameter table of relative position bias
|
1379 |
+
self.relative_position_bias_table = nn.Parameter(
|
1380 |
+
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
|
1381 |
+
|
1382 |
+
# get pair-wise relative position index for each token inside the window
|
1383 |
+
coords_h = torch.arange(self.window_size[0])
|
1384 |
+
coords_w = torch.arange(self.window_size[1])
|
1385 |
+
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
|
1386 |
+
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
|
1387 |
+
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
|
1388 |
+
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
|
1389 |
+
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
|
1390 |
+
relative_coords[:, :, 1] += self.window_size[1] - 1
|
1391 |
+
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
|
1392 |
+
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
|
1393 |
+
self.register_buffer("relative_position_index", relative_position_index)
|
1394 |
+
|
1395 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
|
1396 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
1397 |
+
self.proj = nn.Linear(dim, dim)
|
1398 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
1399 |
+
|
1400 |
+
nn.init.normal_(self.relative_position_bias_table, std=.02)
|
1401 |
+
self.softmax = nn.Softmax(dim=-1)
|
1402 |
+
|
1403 |
+
def forward(self, x, mask=None):
|
1404 |
+
|
1405 |
+
B_, N, C = x.shape
|
1406 |
+
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
1407 |
+
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
|
1408 |
+
|
1409 |
+
q = q * self.scale
|
1410 |
+
attn = (q @ k.transpose(-2, -1))
|
1411 |
+
|
1412 |
+
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
|
1413 |
+
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
|
1414 |
+
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
|
1415 |
+
attn = attn + relative_position_bias.unsqueeze(0)
|
1416 |
+
|
1417 |
+
if mask is not None:
|
1418 |
+
nW = mask.shape[0]
|
1419 |
+
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
|
1420 |
+
attn = attn.view(-1, self.num_heads, N, N)
|
1421 |
+
attn = self.softmax(attn)
|
1422 |
+
else:
|
1423 |
+
attn = self.softmax(attn)
|
1424 |
+
|
1425 |
+
attn = self.attn_drop(attn)
|
1426 |
+
|
1427 |
+
# print(attn.dtype, v.dtype)
|
1428 |
+
try:
|
1429 |
+
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
|
1430 |
+
except:
|
1431 |
+
#print(attn.dtype, v.dtype)
|
1432 |
+
x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)
|
1433 |
+
x = self.proj(x)
|
1434 |
+
x = self.proj_drop(x)
|
1435 |
+
return x
|
1436 |
+
|
1437 |
+
class Mlp(nn.Module):
|
1438 |
+
|
1439 |
+
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):
|
1440 |
+
super().__init__()
|
1441 |
+
out_features = out_features or in_features
|
1442 |
+
hidden_features = hidden_features or in_features
|
1443 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
1444 |
+
self.act = act_layer()
|
1445 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
1446 |
+
self.drop = nn.Dropout(drop)
|
1447 |
+
|
1448 |
+
def forward(self, x):
|
1449 |
+
x = self.fc1(x)
|
1450 |
+
x = self.act(x)
|
1451 |
+
x = self.drop(x)
|
1452 |
+
x = self.fc2(x)
|
1453 |
+
x = self.drop(x)
|
1454 |
+
return x
|
1455 |
+
|
1456 |
+
def window_partition(x, window_size):
|
1457 |
+
|
1458 |
+
B, H, W, C = x.shape
|
1459 |
+
assert H % window_size == 0, 'feature map h and w can not divide by window size'
|
1460 |
+
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
|
1461 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
1462 |
+
return windows
|
1463 |
+
|
1464 |
+
def window_reverse(windows, window_size, H, W):
|
1465 |
+
|
1466 |
+
B = int(windows.shape[0] / (H * W / window_size / window_size))
|
1467 |
+
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
|
1468 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
|
1469 |
+
return x
|
1470 |
+
|
1471 |
+
|
1472 |
+
class SwinTransformerLayer(nn.Module):
|
1473 |
+
|
1474 |
+
def __init__(self, dim, num_heads, window_size=8, shift_size=0,
|
1475 |
+
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
|
1476 |
+
act_layer=nn.SiLU, norm_layer=nn.LayerNorm):
|
1477 |
+
super().__init__()
|
1478 |
+
self.dim = dim
|
1479 |
+
self.num_heads = num_heads
|
1480 |
+
self.window_size = window_size
|
1481 |
+
self.shift_size = shift_size
|
1482 |
+
self.mlp_ratio = mlp_ratio
|
1483 |
+
# if min(self.input_resolution) <= self.window_size:
|
1484 |
+
# # if window size is larger than input resolution, we don't partition windows
|
1485 |
+
# self.shift_size = 0
|
1486 |
+
# self.window_size = min(self.input_resolution)
|
1487 |
+
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
|
1488 |
+
|
1489 |
+
self.norm1 = norm_layer(dim)
|
1490 |
+
self.attn = WindowAttention(
|
1491 |
+
dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
|
1492 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
|
1493 |
+
|
1494 |
+
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
1495 |
+
self.norm2 = norm_layer(dim)
|
1496 |
+
mlp_hidden_dim = int(dim * mlp_ratio)
|
1497 |
+
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
|
1498 |
+
|
1499 |
+
def create_mask(self, H, W):
|
1500 |
+
# calculate attention mask for SW-MSA
|
1501 |
+
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
|
1502 |
+
h_slices = (slice(0, -self.window_size),
|
1503 |
+
slice(-self.window_size, -self.shift_size),
|
1504 |
+
slice(-self.shift_size, None))
|
1505 |
+
w_slices = (slice(0, -self.window_size),
|
1506 |
+
slice(-self.window_size, -self.shift_size),
|
1507 |
+
slice(-self.shift_size, None))
|
1508 |
+
cnt = 0
|
1509 |
+
for h in h_slices:
|
1510 |
+
for w in w_slices:
|
1511 |
+
img_mask[:, h, w, :] = cnt
|
1512 |
+
cnt += 1
|
1513 |
+
|
1514 |
+
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
|
1515 |
+
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
|
1516 |
+
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
|
1517 |
+
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
|
1518 |
+
|
1519 |
+
return attn_mask
|
1520 |
+
|
1521 |
+
def forward(self, x):
|
1522 |
+
# reshape x[b c h w] to x[b l c]
|
1523 |
+
_, _, H_, W_ = x.shape
|
1524 |
+
|
1525 |
+
Padding = False
|
1526 |
+
if min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:
|
1527 |
+
Padding = True
|
1528 |
+
# print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')
|
1529 |
+
pad_r = (self.window_size - W_ % self.window_size) % self.window_size
|
1530 |
+
pad_b = (self.window_size - H_ % self.window_size) % self.window_size
|
1531 |
+
x = F.pad(x, (0, pad_r, 0, pad_b))
|
1532 |
+
|
1533 |
+
# print('2', x.shape)
|
1534 |
+
B, C, H, W = x.shape
|
1535 |
+
L = H * W
|
1536 |
+
x = x.permute(0, 2, 3, 1).contiguous().view(B, L, C) # b, L, c
|
1537 |
+
|
1538 |
+
# create mask from init to forward
|
1539 |
+
if self.shift_size > 0:
|
1540 |
+
attn_mask = self.create_mask(H, W).to(x.device)
|
1541 |
+
else:
|
1542 |
+
attn_mask = None
|
1543 |
+
|
1544 |
+
shortcut = x
|
1545 |
+
x = self.norm1(x)
|
1546 |
+
x = x.view(B, H, W, C)
|
1547 |
+
|
1548 |
+
# cyclic shift
|
1549 |
+
if self.shift_size > 0:
|
1550 |
+
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
|
1551 |
+
else:
|
1552 |
+
shifted_x = x
|
1553 |
+
|
1554 |
+
# partition windows
|
1555 |
+
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
|
1556 |
+
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
|
1557 |
+
|
1558 |
+
# W-MSA/SW-MSA
|
1559 |
+
attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
|
1560 |
+
|
1561 |
+
# merge windows
|
1562 |
+
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
|
1563 |
+
shifted_x = window_reverse(attn_windows, self.window_size, H, W) # B H' W' C
|
1564 |
+
|
1565 |
+
# reverse cyclic shift
|
1566 |
+
if self.shift_size > 0:
|
1567 |
+
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
|
1568 |
+
else:
|
1569 |
+
x = shifted_x
|
1570 |
+
x = x.view(B, H * W, C)
|
1571 |
+
|
1572 |
+
# FFN
|
1573 |
+
x = shortcut + self.drop_path(x)
|
1574 |
+
x = x + self.drop_path(self.mlp(self.norm2(x)))
|
1575 |
+
|
1576 |
+
x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W) # b c h w
|
1577 |
+
|
1578 |
+
if Padding:
|
1579 |
+
x = x[:, :, :H_, :W_] # reverse padding
|
1580 |
+
|
1581 |
+
return x
|
1582 |
+
|
1583 |
+
|
1584 |
+
class SwinTransformerBlock(nn.Module):
|
1585 |
+
def __init__(self, c1, c2, num_heads, num_layers, window_size=8):
|
1586 |
+
super().__init__()
|
1587 |
+
self.conv = None
|
1588 |
+
if c1 != c2:
|
1589 |
+
self.conv = Conv(c1, c2)
|
1590 |
+
|
1591 |
+
# remove input_resolution
|
1592 |
+
self.blocks = nn.Sequential(*[SwinTransformerLayer(dim=c2, num_heads=num_heads, window_size=window_size,
|
1593 |
+
shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])
|
1594 |
+
|
1595 |
+
def forward(self, x):
|
1596 |
+
if self.conv is not None:
|
1597 |
+
x = self.conv(x)
|
1598 |
+
x = self.blocks(x)
|
1599 |
+
return x
|
1600 |
+
|
1601 |
+
|
1602 |
+
class STCSPA(nn.Module):
|
1603 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
1604 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
1605 |
+
super(STCSPA, self).__init__()
|
1606 |
+
c_ = int(c2 * e) # hidden channels
|
1607 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
1608 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
1609 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
1610 |
+
num_heads = c_ // 32
|
1611 |
+
self.m = SwinTransformerBlock(c_, c_, num_heads, n)
|
1612 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
1613 |
+
|
1614 |
+
def forward(self, x):
|
1615 |
+
y1 = self.m(self.cv1(x))
|
1616 |
+
y2 = self.cv2(x)
|
1617 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
1618 |
+
|
1619 |
+
|
1620 |
+
class STCSPB(nn.Module):
|
1621 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
1622 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
1623 |
+
super(STCSPB, self).__init__()
|
1624 |
+
c_ = int(c2) # hidden channels
|
1625 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
1626 |
+
self.cv2 = Conv(c_, c_, 1, 1)
|
1627 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
1628 |
+
num_heads = c_ // 32
|
1629 |
+
self.m = SwinTransformerBlock(c_, c_, num_heads, n)
|
1630 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
1631 |
+
|
1632 |
+
def forward(self, x):
|
1633 |
+
x1 = self.cv1(x)
|
1634 |
+
y1 = self.m(x1)
|
1635 |
+
y2 = self.cv2(x1)
|
1636 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
1637 |
+
|
1638 |
+
|
1639 |
+
class STCSPC(nn.Module):
|
1640 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
1641 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
1642 |
+
super(STCSPC, self).__init__()
|
1643 |
+
c_ = int(c2 * e) # hidden channels
|
1644 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
1645 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
1646 |
+
self.cv3 = Conv(c_, c_, 1, 1)
|
1647 |
+
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
1648 |
+
num_heads = c_ // 32
|
1649 |
+
self.m = SwinTransformerBlock(c_, c_, num_heads, n)
|
1650 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
1651 |
+
|
1652 |
+
def forward(self, x):
|
1653 |
+
y1 = self.cv3(self.m(self.cv1(x)))
|
1654 |
+
y2 = self.cv2(x)
|
1655 |
+
return self.cv4(torch.cat((y1, y2), dim=1))
|
1656 |
+
|
1657 |
+
##### end of swin transformer #####
|
1658 |
+
|
1659 |
+
|
1660 |
+
##### swin transformer v2 #####
|
1661 |
+
|
1662 |
+
class WindowAttention_v2(nn.Module):
|
1663 |
+
|
1664 |
+
def __init__(self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
|
1665 |
+
pretrained_window_size=[0, 0]):
|
1666 |
+
|
1667 |
+
super().__init__()
|
1668 |
+
self.dim = dim
|
1669 |
+
self.window_size = window_size # Wh, Ww
|
1670 |
+
self.pretrained_window_size = pretrained_window_size
|
1671 |
+
self.num_heads = num_heads
|
1672 |
+
|
1673 |
+
self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True)
|
1674 |
+
|
1675 |
+
# mlp to generate continuous relative position bias
|
1676 |
+
self.cpb_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),
|
1677 |
+
nn.ReLU(inplace=True),
|
1678 |
+
nn.Linear(512, num_heads, bias=False))
|
1679 |
+
|
1680 |
+
# get relative_coords_table
|
1681 |
+
relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
|
1682 |
+
relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
|
1683 |
+
relative_coords_table = torch.stack(
|
1684 |
+
torch.meshgrid([relative_coords_h,
|
1685 |
+
relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0) # 1, 2*Wh-1, 2*Ww-1, 2
|
1686 |
+
if pretrained_window_size[0] > 0:
|
1687 |
+
relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
|
1688 |
+
relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
|
1689 |
+
else:
|
1690 |
+
relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
|
1691 |
+
relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
|
1692 |
+
relative_coords_table *= 8 # normalize to -8, 8
|
1693 |
+
relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
|
1694 |
+
torch.abs(relative_coords_table) + 1.0) / np.log2(8)
|
1695 |
+
|
1696 |
+
self.register_buffer("relative_coords_table", relative_coords_table)
|
1697 |
+
|
1698 |
+
# get pair-wise relative position index for each token inside the window
|
1699 |
+
coords_h = torch.arange(self.window_size[0])
|
1700 |
+
coords_w = torch.arange(self.window_size[1])
|
1701 |
+
coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww
|
1702 |
+
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
|
1703 |
+
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
|
1704 |
+
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
|
1705 |
+
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
|
1706 |
+
relative_coords[:, :, 1] += self.window_size[1] - 1
|
1707 |
+
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
|
1708 |
+
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
|
1709 |
+
self.register_buffer("relative_position_index", relative_position_index)
|
1710 |
+
|
1711 |
+
self.qkv = nn.Linear(dim, dim * 3, bias=False)
|
1712 |
+
if qkv_bias:
|
1713 |
+
self.q_bias = nn.Parameter(torch.zeros(dim))
|
1714 |
+
self.v_bias = nn.Parameter(torch.zeros(dim))
|
1715 |
+
else:
|
1716 |
+
self.q_bias = None
|
1717 |
+
self.v_bias = None
|
1718 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
1719 |
+
self.proj = nn.Linear(dim, dim)
|
1720 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
1721 |
+
self.softmax = nn.Softmax(dim=-1)
|
1722 |
+
|
1723 |
+
def forward(self, x, mask=None):
|
1724 |
+
|
1725 |
+
B_, N, C = x.shape
|
1726 |
+
qkv_bias = None
|
1727 |
+
if self.q_bias is not None:
|
1728 |
+
qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
|
1729 |
+
qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
|
1730 |
+
qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
|
1731 |
+
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
|
1732 |
+
|
1733 |
+
# cosine attention
|
1734 |
+
attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
|
1735 |
+
logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01))).exp()
|
1736 |
+
attn = attn * logit_scale
|
1737 |
+
|
1738 |
+
relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
|
1739 |
+
relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
|
1740 |
+
self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH
|
1741 |
+
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
|
1742 |
+
relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
|
1743 |
+
attn = attn + relative_position_bias.unsqueeze(0)
|
1744 |
+
|
1745 |
+
if mask is not None:
|
1746 |
+
nW = mask.shape[0]
|
1747 |
+
attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
|
1748 |
+
attn = attn.view(-1, self.num_heads, N, N)
|
1749 |
+
attn = self.softmax(attn)
|
1750 |
+
else:
|
1751 |
+
attn = self.softmax(attn)
|
1752 |
+
|
1753 |
+
attn = self.attn_drop(attn)
|
1754 |
+
|
1755 |
+
try:
|
1756 |
+
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
|
1757 |
+
except:
|
1758 |
+
x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)
|
1759 |
+
|
1760 |
+
x = self.proj(x)
|
1761 |
+
x = self.proj_drop(x)
|
1762 |
+
return x
|
1763 |
+
|
1764 |
+
def extra_repr(self) -> str:
|
1765 |
+
return f'dim={self.dim}, window_size={self.window_size}, ' \
|
1766 |
+
f'pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}'
|
1767 |
+
|
1768 |
+
def flops(self, N):
|
1769 |
+
# calculate flops for 1 window with token length of N
|
1770 |
+
flops = 0
|
1771 |
+
# qkv = self.qkv(x)
|
1772 |
+
flops += N * self.dim * 3 * self.dim
|
1773 |
+
# attn = (q @ k.transpose(-2, -1))
|
1774 |
+
flops += self.num_heads * N * (self.dim // self.num_heads) * N
|
1775 |
+
# x = (attn @ v)
|
1776 |
+
flops += self.num_heads * N * N * (self.dim // self.num_heads)
|
1777 |
+
# x = self.proj(x)
|
1778 |
+
flops += N * self.dim * self.dim
|
1779 |
+
return flops
|
1780 |
+
|
1781 |
+
class Mlp_v2(nn.Module):
|
1782 |
+
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):
|
1783 |
+
super().__init__()
|
1784 |
+
out_features = out_features or in_features
|
1785 |
+
hidden_features = hidden_features or in_features
|
1786 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
1787 |
+
self.act = act_layer()
|
1788 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
1789 |
+
self.drop = nn.Dropout(drop)
|
1790 |
+
|
1791 |
+
def forward(self, x):
|
1792 |
+
x = self.fc1(x)
|
1793 |
+
x = self.act(x)
|
1794 |
+
x = self.drop(x)
|
1795 |
+
x = self.fc2(x)
|
1796 |
+
x = self.drop(x)
|
1797 |
+
return x
|
1798 |
+
|
1799 |
+
|
1800 |
+
def window_partition_v2(x, window_size):
|
1801 |
+
|
1802 |
+
B, H, W, C = x.shape
|
1803 |
+
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
|
1804 |
+
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
|
1805 |
+
return windows
|
1806 |
+
|
1807 |
+
|
1808 |
+
def window_reverse_v2(windows, window_size, H, W):
|
1809 |
+
|
1810 |
+
B = int(windows.shape[0] / (H * W / window_size / window_size))
|
1811 |
+
x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
|
1812 |
+
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
|
1813 |
+
return x
|
1814 |
+
|
1815 |
+
|
1816 |
+
class SwinTransformerLayer_v2(nn.Module):
|
1817 |
+
|
1818 |
+
def __init__(self, dim, num_heads, window_size=7, shift_size=0,
|
1819 |
+
mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
|
1820 |
+
act_layer=nn.SiLU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
|
1821 |
+
super().__init__()
|
1822 |
+
self.dim = dim
|
1823 |
+
#self.input_resolution = input_resolution
|
1824 |
+
self.num_heads = num_heads
|
1825 |
+
self.window_size = window_size
|
1826 |
+
self.shift_size = shift_size
|
1827 |
+
self.mlp_ratio = mlp_ratio
|
1828 |
+
#if min(self.input_resolution) <= self.window_size:
|
1829 |
+
# # if window size is larger than input resolution, we don't partition windows
|
1830 |
+
# self.shift_size = 0
|
1831 |
+
# self.window_size = min(self.input_resolution)
|
1832 |
+
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
|
1833 |
+
|
1834 |
+
self.norm1 = norm_layer(dim)
|
1835 |
+
self.attn = WindowAttention_v2(
|
1836 |
+
dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
|
1837 |
+
qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
|
1838 |
+
pretrained_window_size=(pretrained_window_size, pretrained_window_size))
|
1839 |
+
|
1840 |
+
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
1841 |
+
self.norm2 = norm_layer(dim)
|
1842 |
+
mlp_hidden_dim = int(dim * mlp_ratio)
|
1843 |
+
self.mlp = Mlp_v2(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
|
1844 |
+
|
1845 |
+
def create_mask(self, H, W):
|
1846 |
+
# calculate attention mask for SW-MSA
|
1847 |
+
img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1
|
1848 |
+
h_slices = (slice(0, -self.window_size),
|
1849 |
+
slice(-self.window_size, -self.shift_size),
|
1850 |
+
slice(-self.shift_size, None))
|
1851 |
+
w_slices = (slice(0, -self.window_size),
|
1852 |
+
slice(-self.window_size, -self.shift_size),
|
1853 |
+
slice(-self.shift_size, None))
|
1854 |
+
cnt = 0
|
1855 |
+
for h in h_slices:
|
1856 |
+
for w in w_slices:
|
1857 |
+
img_mask[:, h, w, :] = cnt
|
1858 |
+
cnt += 1
|
1859 |
+
|
1860 |
+
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
|
1861 |
+
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
|
1862 |
+
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
|
1863 |
+
attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
|
1864 |
+
|
1865 |
+
return attn_mask
|
1866 |
+
|
1867 |
+
def forward(self, x):
|
1868 |
+
# reshape x[b c h w] to x[b l c]
|
1869 |
+
_, _, H_, W_ = x.shape
|
1870 |
+
|
1871 |
+
Padding = False
|
1872 |
+
if min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:
|
1873 |
+
Padding = True
|
1874 |
+
# print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')
|
1875 |
+
pad_r = (self.window_size - W_ % self.window_size) % self.window_size
|
1876 |
+
pad_b = (self.window_size - H_ % self.window_size) % self.window_size
|
1877 |
+
x = F.pad(x, (0, pad_r, 0, pad_b))
|
1878 |
+
|
1879 |
+
# print('2', x.shape)
|
1880 |
+
B, C, H, W = x.shape
|
1881 |
+
L = H * W
|
1882 |
+
x = x.permute(0, 2, 3, 1).contiguous().view(B, L, C) # b, L, c
|
1883 |
+
|
1884 |
+
# create mask from init to forward
|
1885 |
+
if self.shift_size > 0:
|
1886 |
+
attn_mask = self.create_mask(H, W).to(x.device)
|
1887 |
+
else:
|
1888 |
+
attn_mask = None
|
1889 |
+
|
1890 |
+
shortcut = x
|
1891 |
+
x = x.view(B, H, W, C)
|
1892 |
+
|
1893 |
+
# cyclic shift
|
1894 |
+
if self.shift_size > 0:
|
1895 |
+
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
|
1896 |
+
else:
|
1897 |
+
shifted_x = x
|
1898 |
+
|
1899 |
+
# partition windows
|
1900 |
+
x_windows = window_partition_v2(shifted_x, self.window_size) # nW*B, window_size, window_size, C
|
1901 |
+
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
|
1902 |
+
|
1903 |
+
# W-MSA/SW-MSA
|
1904 |
+
attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
|
1905 |
+
|
1906 |
+
# merge windows
|
1907 |
+
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
|
1908 |
+
shifted_x = window_reverse_v2(attn_windows, self.window_size, H, W) # B H' W' C
|
1909 |
+
|
1910 |
+
# reverse cyclic shift
|
1911 |
+
if self.shift_size > 0:
|
1912 |
+
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
|
1913 |
+
else:
|
1914 |
+
x = shifted_x
|
1915 |
+
x = x.view(B, H * W, C)
|
1916 |
+
x = shortcut + self.drop_path(self.norm1(x))
|
1917 |
+
|
1918 |
+
# FFN
|
1919 |
+
x = x + self.drop_path(self.norm2(self.mlp(x)))
|
1920 |
+
x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W) # b c h w
|
1921 |
+
|
1922 |
+
if Padding:
|
1923 |
+
x = x[:, :, :H_, :W_] # reverse padding
|
1924 |
+
|
1925 |
+
return x
|
1926 |
+
|
1927 |
+
def extra_repr(self) -> str:
|
1928 |
+
return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
|
1929 |
+
f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
|
1930 |
+
|
1931 |
+
def flops(self):
|
1932 |
+
flops = 0
|
1933 |
+
H, W = self.input_resolution
|
1934 |
+
# norm1
|
1935 |
+
flops += self.dim * H * W
|
1936 |
+
# W-MSA/SW-MSA
|
1937 |
+
nW = H * W / self.window_size / self.window_size
|
1938 |
+
flops += nW * self.attn.flops(self.window_size * self.window_size)
|
1939 |
+
# mlp
|
1940 |
+
flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
|
1941 |
+
# norm2
|
1942 |
+
flops += self.dim * H * W
|
1943 |
+
return flops
|
1944 |
+
|
1945 |
+
|
1946 |
+
class SwinTransformer2Block(nn.Module):
|
1947 |
+
def __init__(self, c1, c2, num_heads, num_layers, window_size=7):
|
1948 |
+
super().__init__()
|
1949 |
+
self.conv = None
|
1950 |
+
if c1 != c2:
|
1951 |
+
self.conv = Conv(c1, c2)
|
1952 |
+
|
1953 |
+
# remove input_resolution
|
1954 |
+
self.blocks = nn.Sequential(*[SwinTransformerLayer_v2(dim=c2, num_heads=num_heads, window_size=window_size,
|
1955 |
+
shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])
|
1956 |
+
|
1957 |
+
def forward(self, x):
|
1958 |
+
if self.conv is not None:
|
1959 |
+
x = self.conv(x)
|
1960 |
+
x = self.blocks(x)
|
1961 |
+
return x
|
1962 |
+
|
1963 |
+
|
1964 |
+
class ST2CSPA(nn.Module):
|
1965 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
1966 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
1967 |
+
super(ST2CSPA, self).__init__()
|
1968 |
+
c_ = int(c2 * e) # hidden channels
|
1969 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
1970 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
1971 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
1972 |
+
num_heads = c_ // 32
|
1973 |
+
self.m = SwinTransformer2Block(c_, c_, num_heads, n)
|
1974 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
1975 |
+
|
1976 |
+
def forward(self, x):
|
1977 |
+
y1 = self.m(self.cv1(x))
|
1978 |
+
y2 = self.cv2(x)
|
1979 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
1980 |
+
|
1981 |
+
|
1982 |
+
class ST2CSPB(nn.Module):
|
1983 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
1984 |
+
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
1985 |
+
super(ST2CSPB, self).__init__()
|
1986 |
+
c_ = int(c2) # hidden channels
|
1987 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
1988 |
+
self.cv2 = Conv(c_, c_, 1, 1)
|
1989 |
+
self.cv3 = Conv(2 * c_, c2, 1, 1)
|
1990 |
+
num_heads = c_ // 32
|
1991 |
+
self.m = SwinTransformer2Block(c_, c_, num_heads, n)
|
1992 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
1993 |
+
|
1994 |
+
def forward(self, x):
|
1995 |
+
x1 = self.cv1(x)
|
1996 |
+
y1 = self.m(x1)
|
1997 |
+
y2 = self.cv2(x1)
|
1998 |
+
return self.cv3(torch.cat((y1, y2), dim=1))
|
1999 |
+
|
2000 |
+
|
2001 |
+
class ST2CSPC(nn.Module):
|
2002 |
+
# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
|
2003 |
+
def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion
|
2004 |
+
super(ST2CSPC, self).__init__()
|
2005 |
+
c_ = int(c2 * e) # hidden channels
|
2006 |
+
self.cv1 = Conv(c1, c_, 1, 1)
|
2007 |
+
self.cv2 = Conv(c1, c_, 1, 1)
|
2008 |
+
self.cv3 = Conv(c_, c_, 1, 1)
|
2009 |
+
self.cv4 = Conv(2 * c_, c2, 1, 1)
|
2010 |
+
num_heads = c_ // 32
|
2011 |
+
self.m = SwinTransformer2Block(c_, c_, num_heads, n)
|
2012 |
+
#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])
|
2013 |
+
|
2014 |
+
def forward(self, x):
|
2015 |
+
y1 = self.cv3(self.m(self.cv1(x)))
|
2016 |
+
y2 = self.cv2(x)
|
2017 |
+
return self.cv4(torch.cat((y1, y2), dim=1))
|
2018 |
+
|
2019 |
+
##### end of swin transformer v2 #####
|
models/experimental.py
ADDED
@@ -0,0 +1,262 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import random
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
|
6 |
+
from models.common import Conv, DWConv
|
7 |
+
from utils.google_utils import attempt_download
|
8 |
+
|
9 |
+
|
10 |
+
class CrossConv(nn.Module):
|
11 |
+
# Cross Convolution Downsample
|
12 |
+
def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False):
|
13 |
+
# ch_in, ch_out, kernel, stride, groups, expansion, shortcut
|
14 |
+
super(CrossConv, self).__init__()
|
15 |
+
c_ = int(c2 * e) # hidden channels
|
16 |
+
self.cv1 = Conv(c1, c_, (1, k), (1, s))
|
17 |
+
self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g)
|
18 |
+
self.add = shortcut and c1 == c2
|
19 |
+
|
20 |
+
def forward(self, x):
|
21 |
+
return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
|
22 |
+
|
23 |
+
|
24 |
+
class Sum(nn.Module):
|
25 |
+
# Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070
|
26 |
+
def __init__(self, n, weight=False): # n: number of inputs
|
27 |
+
super(Sum, self).__init__()
|
28 |
+
self.weight = weight # apply weights boolean
|
29 |
+
self.iter = range(n - 1) # iter object
|
30 |
+
if weight:
|
31 |
+
self.w = nn.Parameter(-torch.arange(1., n) / 2, requires_grad=True) # layer weights
|
32 |
+
|
33 |
+
def forward(self, x):
|
34 |
+
y = x[0] # no weight
|
35 |
+
if self.weight:
|
36 |
+
w = torch.sigmoid(self.w) * 2
|
37 |
+
for i in self.iter:
|
38 |
+
y = y + x[i + 1] * w[i]
|
39 |
+
else:
|
40 |
+
for i in self.iter:
|
41 |
+
y = y + x[i + 1]
|
42 |
+
return y
|
43 |
+
|
44 |
+
|
45 |
+
class MixConv2d(nn.Module):
|
46 |
+
# Mixed Depthwise Conv https://arxiv.org/abs/1907.09595
|
47 |
+
def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True):
|
48 |
+
super(MixConv2d, self).__init__()
|
49 |
+
groups = len(k)
|
50 |
+
if equal_ch: # equal c_ per group
|
51 |
+
i = torch.linspace(0, groups - 1E-6, c2).floor() # c2 indices
|
52 |
+
c_ = [(i == g).sum() for g in range(groups)] # intermediate channels
|
53 |
+
else: # equal weight.numel() per group
|
54 |
+
b = [c2] + [0] * groups
|
55 |
+
a = np.eye(groups + 1, groups, k=-1)
|
56 |
+
a -= np.roll(a, 1, axis=1)
|
57 |
+
a *= np.array(k) ** 2
|
58 |
+
a[0] = 1
|
59 |
+
c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b
|
60 |
+
|
61 |
+
self.m = nn.ModuleList([nn.Conv2d(c1, int(c_[g]), k[g], s, k[g] // 2, bias=False) for g in range(groups)])
|
62 |
+
self.bn = nn.BatchNorm2d(c2)
|
63 |
+
self.act = nn.LeakyReLU(0.1, inplace=True)
|
64 |
+
|
65 |
+
def forward(self, x):
|
66 |
+
return x + self.act(self.bn(torch.cat([m(x) for m in self.m], 1)))
|
67 |
+
|
68 |
+
|
69 |
+
class Ensemble(nn.ModuleList):
|
70 |
+
# Ensemble of models
|
71 |
+
def __init__(self):
|
72 |
+
super(Ensemble, self).__init__()
|
73 |
+
|
74 |
+
def forward(self, x, augment=False):
|
75 |
+
y = []
|
76 |
+
for module in self:
|
77 |
+
y.append(module(x, augment)[0])
|
78 |
+
# y = torch.stack(y).max(0)[0] # max ensemble
|
79 |
+
# y = torch.stack(y).mean(0) # mean ensemble
|
80 |
+
y = torch.cat(y, 1) # nms ensemble
|
81 |
+
return y, None # inference, train output
|
82 |
+
|
83 |
+
|
84 |
+
|
85 |
+
|
86 |
+
|
87 |
+
class ORT_NMS(torch.autograd.Function):
|
88 |
+
'''ONNX-Runtime NMS operation'''
|
89 |
+
@staticmethod
|
90 |
+
def forward(ctx,
|
91 |
+
boxes,
|
92 |
+
scores,
|
93 |
+
max_output_boxes_per_class=torch.tensor([100]),
|
94 |
+
iou_threshold=torch.tensor([0.45]),
|
95 |
+
score_threshold=torch.tensor([0.25])):
|
96 |
+
device = boxes.device
|
97 |
+
batch = scores.shape[0]
|
98 |
+
num_det = random.randint(0, 100)
|
99 |
+
batches = torch.randint(0, batch, (num_det,)).sort()[0].to(device)
|
100 |
+
idxs = torch.arange(100, 100 + num_det).to(device)
|
101 |
+
zeros = torch.zeros((num_det,), dtype=torch.int64).to(device)
|
102 |
+
selected_indices = torch.cat([batches[None], zeros[None], idxs[None]], 0).T.contiguous()
|
103 |
+
selected_indices = selected_indices.to(torch.int64)
|
104 |
+
return selected_indices
|
105 |
+
|
106 |
+
@staticmethod
|
107 |
+
def symbolic(g, boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold):
|
108 |
+
return g.op("NonMaxSuppression", boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold)
|
109 |
+
|
110 |
+
|
111 |
+
class TRT_NMS(torch.autograd.Function):
|
112 |
+
'''TensorRT NMS operation'''
|
113 |
+
@staticmethod
|
114 |
+
def forward(
|
115 |
+
ctx,
|
116 |
+
boxes,
|
117 |
+
scores,
|
118 |
+
background_class=-1,
|
119 |
+
box_coding=1,
|
120 |
+
iou_threshold=0.45,
|
121 |
+
max_output_boxes=100,
|
122 |
+
plugin_version="1",
|
123 |
+
score_activation=0,
|
124 |
+
score_threshold=0.25,
|
125 |
+
):
|
126 |
+
batch_size, num_boxes, num_classes = scores.shape
|
127 |
+
num_det = torch.randint(0, max_output_boxes, (batch_size, 1), dtype=torch.int32)
|
128 |
+
det_boxes = torch.randn(batch_size, max_output_boxes, 4)
|
129 |
+
det_scores = torch.randn(batch_size, max_output_boxes)
|
130 |
+
det_classes = torch.randint(0, num_classes, (batch_size, max_output_boxes), dtype=torch.int32)
|
131 |
+
return num_det, det_boxes, det_scores, det_classes
|
132 |
+
|
133 |
+
@staticmethod
|
134 |
+
def symbolic(g,
|
135 |
+
boxes,
|
136 |
+
scores,
|
137 |
+
background_class=-1,
|
138 |
+
box_coding=1,
|
139 |
+
iou_threshold=0.45,
|
140 |
+
max_output_boxes=100,
|
141 |
+
plugin_version="1",
|
142 |
+
score_activation=0,
|
143 |
+
score_threshold=0.25):
|
144 |
+
out = g.op("TRT::EfficientNMS_TRT",
|
145 |
+
boxes,
|
146 |
+
scores,
|
147 |
+
background_class_i=background_class,
|
148 |
+
box_coding_i=box_coding,
|
149 |
+
iou_threshold_f=iou_threshold,
|
150 |
+
max_output_boxes_i=max_output_boxes,
|
151 |
+
plugin_version_s=plugin_version,
|
152 |
+
score_activation_i=score_activation,
|
153 |
+
score_threshold_f=score_threshold,
|
154 |
+
outputs=4)
|
155 |
+
nums, boxes, scores, classes = out
|
156 |
+
return nums, boxes, scores, classes
|
157 |
+
|
158 |
+
|
159 |
+
class ONNX_ORT(nn.Module):
|
160 |
+
'''onnx module with ONNX-Runtime NMS operation.'''
|
161 |
+
def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=640, device=None):
|
162 |
+
super().__init__()
|
163 |
+
self.device = device if device else torch.device("cpu")
|
164 |
+
self.max_obj = torch.tensor([max_obj]).to(device)
|
165 |
+
self.iou_threshold = torch.tensor([iou_thres]).to(device)
|
166 |
+
self.score_threshold = torch.tensor([score_thres]).to(device)
|
167 |
+
self.max_wh = max_wh # if max_wh != 0 : non-agnostic else : agnostic
|
168 |
+
self.convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
|
169 |
+
dtype=torch.float32,
|
170 |
+
device=self.device)
|
171 |
+
|
172 |
+
def forward(self, x):
|
173 |
+
boxes = x[:, :, :4]
|
174 |
+
conf = x[:, :, 4:5]
|
175 |
+
scores = x[:, :, 5:]
|
176 |
+
scores *= conf
|
177 |
+
boxes @= self.convert_matrix
|
178 |
+
max_score, category_id = scores.max(2, keepdim=True)
|
179 |
+
dis = category_id.float() * self.max_wh
|
180 |
+
nmsbox = boxes + dis
|
181 |
+
max_score_tp = max_score.transpose(1, 2).contiguous()
|
182 |
+
selected_indices = ORT_NMS.apply(nmsbox, max_score_tp, self.max_obj, self.iou_threshold, self.score_threshold)
|
183 |
+
X, Y = selected_indices[:, 0], selected_indices[:, 2]
|
184 |
+
selected_boxes = boxes[X, Y, :]
|
185 |
+
selected_categories = category_id[X, Y, :].float()
|
186 |
+
selected_scores = max_score[X, Y, :]
|
187 |
+
X = X.unsqueeze(1).float()
|
188 |
+
return torch.cat([X, selected_boxes, selected_categories, selected_scores], 1)
|
189 |
+
|
190 |
+
class ONNX_TRT(nn.Module):
|
191 |
+
'''onnx module with TensorRT NMS operation.'''
|
192 |
+
def __init__(self, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None ,device=None):
|
193 |
+
super().__init__()
|
194 |
+
assert max_wh is None
|
195 |
+
self.device = device if device else torch.device('cpu')
|
196 |
+
self.background_class = -1,
|
197 |
+
self.box_coding = 1,
|
198 |
+
self.iou_threshold = iou_thres
|
199 |
+
self.max_obj = max_obj
|
200 |
+
self.plugin_version = '1'
|
201 |
+
self.score_activation = 0
|
202 |
+
self.score_threshold = score_thres
|
203 |
+
|
204 |
+
def forward(self, x):
|
205 |
+
boxes = x[:, :, :4]
|
206 |
+
conf = x[:, :, 4:5]
|
207 |
+
scores = x[:, :, 5:]
|
208 |
+
scores *= conf
|
209 |
+
num_det, det_boxes, det_scores, det_classes = TRT_NMS.apply(boxes, scores, self.background_class, self.box_coding,
|
210 |
+
self.iou_threshold, self.max_obj,
|
211 |
+
self.plugin_version, self.score_activation,
|
212 |
+
self.score_threshold)
|
213 |
+
return num_det, det_boxes, det_scores, det_classes
|
214 |
+
|
215 |
+
|
216 |
+
class End2End(nn.Module):
|
217 |
+
'''export onnx or tensorrt model with NMS operation.'''
|
218 |
+
def __init__(self, model, max_obj=100, iou_thres=0.45, score_thres=0.25, max_wh=None, device=None):
|
219 |
+
super().__init__()
|
220 |
+
device = device if device else torch.device('cpu')
|
221 |
+
assert isinstance(max_wh,(int)) or max_wh is None
|
222 |
+
self.model = model.to(device)
|
223 |
+
self.model.model[-1].end2end = True
|
224 |
+
self.patch_model = ONNX_TRT if max_wh is None else ONNX_ORT
|
225 |
+
self.end2end = self.patch_model(max_obj, iou_thres, score_thres, max_wh, device)
|
226 |
+
self.end2end.eval()
|
227 |
+
|
228 |
+
def forward(self, x):
|
229 |
+
x = self.model(x)
|
230 |
+
x = self.end2end(x)
|
231 |
+
return x
|
232 |
+
|
233 |
+
|
234 |
+
|
235 |
+
|
236 |
+
|
237 |
+
def attempt_load(weights, map_location=None):
|
238 |
+
# Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a
|
239 |
+
model = Ensemble()
|
240 |
+
for w in weights if isinstance(weights, list) else [weights]:
|
241 |
+
attempt_download(w)
|
242 |
+
ckpt = torch.load(w, map_location=map_location) # load
|
243 |
+
model.append(ckpt['ema' if ckpt.get('ema') else 'model'].float().fuse().eval()) # FP32 model
|
244 |
+
|
245 |
+
# Compatibility updates
|
246 |
+
for m in model.modules():
|
247 |
+
if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
|
248 |
+
m.inplace = True # pytorch 1.7.0 compatibility
|
249 |
+
elif type(m) is nn.Upsample:
|
250 |
+
m.recompute_scale_factor = None # torch 1.11.0 compatibility
|
251 |
+
elif type(m) is Conv:
|
252 |
+
m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility
|
253 |
+
|
254 |
+
if len(model) == 1:
|
255 |
+
return model[-1] # return model
|
256 |
+
else:
|
257 |
+
print('Ensemble created with %s\n' % weights)
|
258 |
+
for k in ['names', 'stride']:
|
259 |
+
setattr(model, k, getattr(model[-1], k))
|
260 |
+
return model # return ensemble
|
261 |
+
|
262 |
+
|
models/yolo.py
ADDED
@@ -0,0 +1,843 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
import logging
|
3 |
+
import sys
|
4 |
+
from copy import deepcopy
|
5 |
+
|
6 |
+
sys.path.append('./') # to run '$ python *.py' files in subdirectories
|
7 |
+
logger = logging.getLogger(__name__)
|
8 |
+
import torch
|
9 |
+
from models.common import *
|
10 |
+
from models.experimental import *
|
11 |
+
from utils.autoanchor import check_anchor_order
|
12 |
+
from utils.general import make_divisible, check_file, set_logging
|
13 |
+
from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \
|
14 |
+
select_device, copy_attr
|
15 |
+
from utils.loss import SigmoidBin
|
16 |
+
|
17 |
+
try:
|
18 |
+
import thop # for FLOPS computation
|
19 |
+
except ImportError:
|
20 |
+
thop = None
|
21 |
+
|
22 |
+
|
23 |
+
class Detect(nn.Module):
|
24 |
+
stride = None # strides computed during build
|
25 |
+
export = False # onnx export
|
26 |
+
end2end = False
|
27 |
+
include_nms = False
|
28 |
+
concat = False
|
29 |
+
|
30 |
+
def __init__(self, nc=80, anchors=(), ch=()): # detection layer
|
31 |
+
super(Detect, self).__init__()
|
32 |
+
self.nc = nc # number of classes
|
33 |
+
self.no = nc + 5 # number of outputs per anchor
|
34 |
+
self.nl = len(anchors) # number of detection layers
|
35 |
+
self.na = len(anchors[0]) // 2 # number of anchors
|
36 |
+
self.grid = [torch.zeros(1)] * self.nl # init grid
|
37 |
+
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
|
38 |
+
self.register_buffer('anchors', a) # shape(nl,na,2)
|
39 |
+
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
|
40 |
+
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
|
41 |
+
|
42 |
+
def forward(self, x):
|
43 |
+
# x = x.copy() # for profiling
|
44 |
+
z = [] # inference output
|
45 |
+
self.training |= self.export
|
46 |
+
for i in range(self.nl):
|
47 |
+
x[i] = self.m[i](x[i]) # conv
|
48 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
49 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
50 |
+
|
51 |
+
if not self.training: # inference
|
52 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
53 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
54 |
+
y = x[i].sigmoid()
|
55 |
+
if not torch.onnx.is_in_onnx_export():
|
56 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
57 |
+
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
58 |
+
else:
|
59 |
+
xy, wh, conf = y.split((2, 2, self.nc + 1), 4) # y.tensor_split((2, 4, 5), 4) # torch 1.8.0
|
60 |
+
xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5)) # new xy
|
61 |
+
wh = wh ** 2 * (4 * self.anchor_grid[i].data) # new wh
|
62 |
+
y = torch.cat((xy, wh, conf), 4)
|
63 |
+
z.append(y.view(bs, -1, self.no))
|
64 |
+
|
65 |
+
if self.training:
|
66 |
+
out = x
|
67 |
+
elif self.end2end:
|
68 |
+
out = torch.cat(z, 1)
|
69 |
+
elif self.include_nms:
|
70 |
+
z = self.convert(z)
|
71 |
+
out = (z, )
|
72 |
+
elif self.concat:
|
73 |
+
out = torch.cat(z, 1)
|
74 |
+
else:
|
75 |
+
out = (torch.cat(z, 1), x)
|
76 |
+
|
77 |
+
return out
|
78 |
+
|
79 |
+
@staticmethod
|
80 |
+
def _make_grid(nx=20, ny=20):
|
81 |
+
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
|
82 |
+
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
|
83 |
+
|
84 |
+
def convert(self, z):
|
85 |
+
z = torch.cat(z, 1)
|
86 |
+
box = z[:, :, :4]
|
87 |
+
conf = z[:, :, 4:5]
|
88 |
+
score = z[:, :, 5:]
|
89 |
+
score *= conf
|
90 |
+
convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
|
91 |
+
dtype=torch.float32,
|
92 |
+
device=z.device)
|
93 |
+
box @= convert_matrix
|
94 |
+
return (box, score)
|
95 |
+
|
96 |
+
|
97 |
+
class IDetect(nn.Module):
|
98 |
+
stride = None # strides computed during build
|
99 |
+
export = False # onnx export
|
100 |
+
end2end = False
|
101 |
+
include_nms = False
|
102 |
+
concat = False
|
103 |
+
|
104 |
+
def __init__(self, nc=80, anchors=(), ch=()): # detection layer
|
105 |
+
super(IDetect, self).__init__()
|
106 |
+
self.nc = nc # number of classes
|
107 |
+
self.no = nc + 5 # number of outputs per anchor
|
108 |
+
self.nl = len(anchors) # number of detection layers
|
109 |
+
self.na = len(anchors[0]) // 2 # number of anchors
|
110 |
+
self.grid = [torch.zeros(1)] * self.nl # init grid
|
111 |
+
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
|
112 |
+
self.register_buffer('anchors', a) # shape(nl,na,2)
|
113 |
+
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
|
114 |
+
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
|
115 |
+
|
116 |
+
self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
|
117 |
+
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch)
|
118 |
+
|
119 |
+
def forward(self, x):
|
120 |
+
# x = x.copy() # for profiling
|
121 |
+
z = [] # inference output
|
122 |
+
self.training |= self.export
|
123 |
+
for i in range(self.nl):
|
124 |
+
x[i] = self.m[i](self.ia[i](x[i])) # conv
|
125 |
+
x[i] = self.im[i](x[i])
|
126 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
127 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
128 |
+
|
129 |
+
if not self.training: # inference
|
130 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
131 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
132 |
+
|
133 |
+
y = x[i].sigmoid()
|
134 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
135 |
+
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
136 |
+
z.append(y.view(bs, -1, self.no))
|
137 |
+
|
138 |
+
return x if self.training else (torch.cat(z, 1), x)
|
139 |
+
|
140 |
+
def fuseforward(self, x):
|
141 |
+
# x = x.copy() # for profiling
|
142 |
+
z = [] # inference output
|
143 |
+
self.training |= self.export
|
144 |
+
for i in range(self.nl):
|
145 |
+
x[i] = self.m[i](x[i]) # conv
|
146 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
147 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
148 |
+
|
149 |
+
if not self.training: # inference
|
150 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
151 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
152 |
+
|
153 |
+
y = x[i].sigmoid()
|
154 |
+
if not torch.onnx.is_in_onnx_export():
|
155 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
156 |
+
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
157 |
+
else:
|
158 |
+
xy, wh, conf = y.split((2, 2, self.nc + 1), 4) # y.tensor_split((2, 4, 5), 4) # torch 1.8.0
|
159 |
+
xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5)) # new xy
|
160 |
+
wh = wh ** 2 * (4 * self.anchor_grid[i].data) # new wh
|
161 |
+
y = torch.cat((xy, wh, conf), 4)
|
162 |
+
z.append(y.view(bs, -1, self.no))
|
163 |
+
|
164 |
+
if self.training:
|
165 |
+
out = x
|
166 |
+
elif self.end2end:
|
167 |
+
out = torch.cat(z, 1)
|
168 |
+
elif self.include_nms:
|
169 |
+
z = self.convert(z)
|
170 |
+
out = (z, )
|
171 |
+
elif self.concat:
|
172 |
+
out = torch.cat(z, 1)
|
173 |
+
else:
|
174 |
+
out = (torch.cat(z, 1), x)
|
175 |
+
|
176 |
+
return out
|
177 |
+
|
178 |
+
def fuse(self):
|
179 |
+
print("IDetect.fuse")
|
180 |
+
# fuse ImplicitA and Convolution
|
181 |
+
for i in range(len(self.m)):
|
182 |
+
c1,c2,_,_ = self.m[i].weight.shape
|
183 |
+
c1_,c2_, _,_ = self.ia[i].implicit.shape
|
184 |
+
self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1)
|
185 |
+
|
186 |
+
# fuse ImplicitM and Convolution
|
187 |
+
for i in range(len(self.m)):
|
188 |
+
c1,c2, _,_ = self.im[i].implicit.shape
|
189 |
+
self.m[i].bias *= self.im[i].implicit.reshape(c2)
|
190 |
+
self.m[i].weight *= self.im[i].implicit.transpose(0,1)
|
191 |
+
|
192 |
+
@staticmethod
|
193 |
+
def _make_grid(nx=20, ny=20):
|
194 |
+
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
|
195 |
+
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
|
196 |
+
|
197 |
+
def convert(self, z):
|
198 |
+
z = torch.cat(z, 1)
|
199 |
+
box = z[:, :, :4]
|
200 |
+
conf = z[:, :, 4:5]
|
201 |
+
score = z[:, :, 5:]
|
202 |
+
score *= conf
|
203 |
+
convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
|
204 |
+
dtype=torch.float32,
|
205 |
+
device=z.device)
|
206 |
+
box @= convert_matrix
|
207 |
+
return (box, score)
|
208 |
+
|
209 |
+
|
210 |
+
class IKeypoint(nn.Module):
|
211 |
+
stride = None # strides computed during build
|
212 |
+
export = False # onnx export
|
213 |
+
|
214 |
+
def __init__(self, nc=80, anchors=(), nkpt=17, ch=(), inplace=True, dw_conv_kpt=False): # detection layer
|
215 |
+
super(IKeypoint, self).__init__()
|
216 |
+
self.nc = nc # number of classes
|
217 |
+
self.nkpt = nkpt
|
218 |
+
self.dw_conv_kpt = dw_conv_kpt
|
219 |
+
self.no_det=(nc + 5) # number of outputs per anchor for box and class
|
220 |
+
self.no_kpt = 3*self.nkpt ## number of outputs per anchor for keypoints
|
221 |
+
self.no = self.no_det+self.no_kpt
|
222 |
+
self.nl = len(anchors) # number of detection layers
|
223 |
+
self.na = len(anchors[0]) // 2 # number of anchors
|
224 |
+
self.grid = [torch.zeros(1)] * self.nl # init grid
|
225 |
+
self.flip_test = False
|
226 |
+
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
|
227 |
+
self.register_buffer('anchors', a) # shape(nl,na,2)
|
228 |
+
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
|
229 |
+
self.m = nn.ModuleList(nn.Conv2d(x, self.no_det * self.na, 1) for x in ch) # output conv
|
230 |
+
|
231 |
+
self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
|
232 |
+
self.im = nn.ModuleList(ImplicitM(self.no_det * self.na) for _ in ch)
|
233 |
+
|
234 |
+
if self.nkpt is not None:
|
235 |
+
if self.dw_conv_kpt: #keypoint head is slightly more complex
|
236 |
+
self.m_kpt = nn.ModuleList(
|
237 |
+
nn.Sequential(DWConv(x, x, k=3), Conv(x,x),
|
238 |
+
DWConv(x, x, k=3), Conv(x, x),
|
239 |
+
DWConv(x, x, k=3), Conv(x,x),
|
240 |
+
DWConv(x, x, k=3), Conv(x, x),
|
241 |
+
DWConv(x, x, k=3), Conv(x, x),
|
242 |
+
DWConv(x, x, k=3), nn.Conv2d(x, self.no_kpt * self.na, 1)) for x in ch)
|
243 |
+
else: #keypoint head is a single convolution
|
244 |
+
self.m_kpt = nn.ModuleList(nn.Conv2d(x, self.no_kpt * self.na, 1) for x in ch)
|
245 |
+
|
246 |
+
self.inplace = inplace # use in-place ops (e.g. slice assignment)
|
247 |
+
|
248 |
+
def forward(self, x):
|
249 |
+
# x = x.copy() # for profiling
|
250 |
+
z = [] # inference output
|
251 |
+
self.training |= self.export
|
252 |
+
for i in range(self.nl):
|
253 |
+
if self.nkpt is None or self.nkpt==0:
|
254 |
+
x[i] = self.im[i](self.m[i](self.ia[i](x[i]))) # conv
|
255 |
+
else :
|
256 |
+
x[i] = torch.cat((self.im[i](self.m[i](self.ia[i](x[i]))), self.m_kpt[i](x[i])), axis=1)
|
257 |
+
|
258 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
259 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
260 |
+
x_det = x[i][..., :6]
|
261 |
+
x_kpt = x[i][..., 6:]
|
262 |
+
|
263 |
+
if not self.training: # inference
|
264 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
265 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
266 |
+
kpt_grid_x = self.grid[i][..., 0:1]
|
267 |
+
kpt_grid_y = self.grid[i][..., 1:2]
|
268 |
+
|
269 |
+
if self.nkpt == 0:
|
270 |
+
y = x[i].sigmoid()
|
271 |
+
else:
|
272 |
+
y = x_det.sigmoid()
|
273 |
+
|
274 |
+
if self.inplace:
|
275 |
+
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
276 |
+
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2) # wh
|
277 |
+
if self.nkpt != 0:
|
278 |
+
x_kpt[..., 0::3] = (x_kpt[..., ::3] * 2. - 0.5 + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy
|
279 |
+
x_kpt[..., 1::3] = (x_kpt[..., 1::3] * 2. - 0.5 + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy
|
280 |
+
#x_kpt[..., 0::3] = (x_kpt[..., ::3] + kpt_grid_x.repeat(1,1,1,1,17)) * self.stride[i] # xy
|
281 |
+
#x_kpt[..., 1::3] = (x_kpt[..., 1::3] + kpt_grid_y.repeat(1,1,1,1,17)) * self.stride[i] # xy
|
282 |
+
#print('=============')
|
283 |
+
#print(self.anchor_grid[i].shape)
|
284 |
+
#print(self.anchor_grid[i][...,0].unsqueeze(4).shape)
|
285 |
+
#print(x_kpt[..., 0::3].shape)
|
286 |
+
#x_kpt[..., 0::3] = ((x_kpt[..., 0::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy
|
287 |
+
#x_kpt[..., 1::3] = ((x_kpt[..., 1::3].tanh() * 2.) ** 3 * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy
|
288 |
+
#x_kpt[..., 0::3] = (((x_kpt[..., 0::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,0].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_x.repeat(1,1,1,1,17) * self.stride[i] # xy
|
289 |
+
#x_kpt[..., 1::3] = (((x_kpt[..., 1::3].sigmoid() * 4.) ** 2 - 8.) * self.anchor_grid[i][...,1].unsqueeze(4).repeat(1,1,1,1,self.nkpt)) + kpt_grid_y.repeat(1,1,1,1,17) * self.stride[i] # xy
|
290 |
+
x_kpt[..., 2::3] = x_kpt[..., 2::3].sigmoid()
|
291 |
+
|
292 |
+
y = torch.cat((xy, wh, y[..., 4:], x_kpt), dim = -1)
|
293 |
+
|
294 |
+
else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
|
295 |
+
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
296 |
+
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
297 |
+
if self.nkpt != 0:
|
298 |
+
y[..., 6:] = (y[..., 6:] * 2. - 0.5 + self.grid[i].repeat((1,1,1,1,self.nkpt))) * self.stride[i] # xy
|
299 |
+
y = torch.cat((xy, wh, y[..., 4:]), -1)
|
300 |
+
|
301 |
+
z.append(y.view(bs, -1, self.no))
|
302 |
+
|
303 |
+
return x if self.training else (torch.cat(z, 1), x)
|
304 |
+
|
305 |
+
@staticmethod
|
306 |
+
def _make_grid(nx=20, ny=20):
|
307 |
+
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
|
308 |
+
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
|
309 |
+
|
310 |
+
|
311 |
+
class IAuxDetect(nn.Module):
|
312 |
+
stride = None # strides computed during build
|
313 |
+
export = False # onnx export
|
314 |
+
end2end = False
|
315 |
+
include_nms = False
|
316 |
+
concat = False
|
317 |
+
|
318 |
+
def __init__(self, nc=80, anchors=(), ch=()): # detection layer
|
319 |
+
super(IAuxDetect, self).__init__()
|
320 |
+
self.nc = nc # number of classes
|
321 |
+
self.no = nc + 5 # number of outputs per anchor
|
322 |
+
self.nl = len(anchors) # number of detection layers
|
323 |
+
self.na = len(anchors[0]) // 2 # number of anchors
|
324 |
+
self.grid = [torch.zeros(1)] * self.nl # init grid
|
325 |
+
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
|
326 |
+
self.register_buffer('anchors', a) # shape(nl,na,2)
|
327 |
+
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
|
328 |
+
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[:self.nl]) # output conv
|
329 |
+
self.m2 = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch[self.nl:]) # output conv
|
330 |
+
|
331 |
+
self.ia = nn.ModuleList(ImplicitA(x) for x in ch[:self.nl])
|
332 |
+
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch[:self.nl])
|
333 |
+
|
334 |
+
def forward(self, x):
|
335 |
+
# x = x.copy() # for profiling
|
336 |
+
z = [] # inference output
|
337 |
+
self.training |= self.export
|
338 |
+
for i in range(self.nl):
|
339 |
+
x[i] = self.m[i](self.ia[i](x[i])) # conv
|
340 |
+
x[i] = self.im[i](x[i])
|
341 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
342 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
343 |
+
|
344 |
+
x[i+self.nl] = self.m2[i](x[i+self.nl])
|
345 |
+
x[i+self.nl] = x[i+self.nl].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
346 |
+
|
347 |
+
if not self.training: # inference
|
348 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
349 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
350 |
+
|
351 |
+
y = x[i].sigmoid()
|
352 |
+
if not torch.onnx.is_in_onnx_export():
|
353 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
354 |
+
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
355 |
+
else:
|
356 |
+
xy, wh, conf = y.split((2, 2, self.nc + 1), 4) # y.tensor_split((2, 4, 5), 4) # torch 1.8.0
|
357 |
+
xy = xy * (2. * self.stride[i]) + (self.stride[i] * (self.grid[i] - 0.5)) # new xy
|
358 |
+
wh = wh ** 2 * (4 * self.anchor_grid[i].data) # new wh
|
359 |
+
y = torch.cat((xy, wh, conf), 4)
|
360 |
+
z.append(y.view(bs, -1, self.no))
|
361 |
+
|
362 |
+
return x if self.training else (torch.cat(z, 1), x[:self.nl])
|
363 |
+
|
364 |
+
def fuseforward(self, x):
|
365 |
+
# x = x.copy() # for profiling
|
366 |
+
z = [] # inference output
|
367 |
+
self.training |= self.export
|
368 |
+
for i in range(self.nl):
|
369 |
+
x[i] = self.m[i](x[i]) # conv
|
370 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
371 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
372 |
+
|
373 |
+
if not self.training: # inference
|
374 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
375 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
376 |
+
|
377 |
+
y = x[i].sigmoid()
|
378 |
+
if not torch.onnx.is_in_onnx_export():
|
379 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
380 |
+
y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
381 |
+
else:
|
382 |
+
xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
383 |
+
wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].data # wh
|
384 |
+
y = torch.cat((xy, wh, y[..., 4:]), -1)
|
385 |
+
z.append(y.view(bs, -1, self.no))
|
386 |
+
|
387 |
+
if self.training:
|
388 |
+
out = x
|
389 |
+
elif self.end2end:
|
390 |
+
out = torch.cat(z, 1)
|
391 |
+
elif self.include_nms:
|
392 |
+
z = self.convert(z)
|
393 |
+
out = (z, )
|
394 |
+
elif self.concat:
|
395 |
+
out = torch.cat(z, 1)
|
396 |
+
else:
|
397 |
+
out = (torch.cat(z, 1), x)
|
398 |
+
|
399 |
+
return out
|
400 |
+
|
401 |
+
def fuse(self):
|
402 |
+
print("IAuxDetect.fuse")
|
403 |
+
# fuse ImplicitA and Convolution
|
404 |
+
for i in range(len(self.m)):
|
405 |
+
c1,c2,_,_ = self.m[i].weight.shape
|
406 |
+
c1_,c2_, _,_ = self.ia[i].implicit.shape
|
407 |
+
self.m[i].bias += torch.matmul(self.m[i].weight.reshape(c1,c2),self.ia[i].implicit.reshape(c2_,c1_)).squeeze(1)
|
408 |
+
|
409 |
+
# fuse ImplicitM and Convolution
|
410 |
+
for i in range(len(self.m)):
|
411 |
+
c1,c2, _,_ = self.im[i].implicit.shape
|
412 |
+
self.m[i].bias *= self.im[i].implicit.reshape(c2)
|
413 |
+
self.m[i].weight *= self.im[i].implicit.transpose(0,1)
|
414 |
+
|
415 |
+
@staticmethod
|
416 |
+
def _make_grid(nx=20, ny=20):
|
417 |
+
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
|
418 |
+
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
|
419 |
+
|
420 |
+
def convert(self, z):
|
421 |
+
z = torch.cat(z, 1)
|
422 |
+
box = z[:, :, :4]
|
423 |
+
conf = z[:, :, 4:5]
|
424 |
+
score = z[:, :, 5:]
|
425 |
+
score *= conf
|
426 |
+
convert_matrix = torch.tensor([[1, 0, 1, 0], [0, 1, 0, 1], [-0.5, 0, 0.5, 0], [0, -0.5, 0, 0.5]],
|
427 |
+
dtype=torch.float32,
|
428 |
+
device=z.device)
|
429 |
+
box @= convert_matrix
|
430 |
+
return (box, score)
|
431 |
+
|
432 |
+
|
433 |
+
class IBin(nn.Module):
|
434 |
+
stride = None # strides computed during build
|
435 |
+
export = False # onnx export
|
436 |
+
|
437 |
+
def __init__(self, nc=80, anchors=(), ch=(), bin_count=21): # detection layer
|
438 |
+
super(IBin, self).__init__()
|
439 |
+
self.nc = nc # number of classes
|
440 |
+
self.bin_count = bin_count
|
441 |
+
|
442 |
+
self.w_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0)
|
443 |
+
self.h_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0)
|
444 |
+
# classes, x,y,obj
|
445 |
+
self.no = nc + 3 + \
|
446 |
+
self.w_bin_sigmoid.get_length() + self.h_bin_sigmoid.get_length() # w-bce, h-bce
|
447 |
+
# + self.x_bin_sigmoid.get_length() + self.y_bin_sigmoid.get_length()
|
448 |
+
|
449 |
+
self.nl = len(anchors) # number of detection layers
|
450 |
+
self.na = len(anchors[0]) // 2 # number of anchors
|
451 |
+
self.grid = [torch.zeros(1)] * self.nl # init grid
|
452 |
+
a = torch.tensor(anchors).float().view(self.nl, -1, 2)
|
453 |
+
self.register_buffer('anchors', a) # shape(nl,na,2)
|
454 |
+
self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2)) # shape(nl,1,na,1,1,2)
|
455 |
+
self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv
|
456 |
+
|
457 |
+
self.ia = nn.ModuleList(ImplicitA(x) for x in ch)
|
458 |
+
self.im = nn.ModuleList(ImplicitM(self.no * self.na) for _ in ch)
|
459 |
+
|
460 |
+
def forward(self, x):
|
461 |
+
|
462 |
+
#self.x_bin_sigmoid.use_fw_regression = True
|
463 |
+
#self.y_bin_sigmoid.use_fw_regression = True
|
464 |
+
self.w_bin_sigmoid.use_fw_regression = True
|
465 |
+
self.h_bin_sigmoid.use_fw_regression = True
|
466 |
+
|
467 |
+
# x = x.copy() # for profiling
|
468 |
+
z = [] # inference output
|
469 |
+
self.training |= self.export
|
470 |
+
for i in range(self.nl):
|
471 |
+
x[i] = self.m[i](self.ia[i](x[i])) # conv
|
472 |
+
x[i] = self.im[i](x[i])
|
473 |
+
bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85)
|
474 |
+
x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
|
475 |
+
|
476 |
+
if not self.training: # inference
|
477 |
+
if self.grid[i].shape[2:4] != x[i].shape[2:4]:
|
478 |
+
self.grid[i] = self._make_grid(nx, ny).to(x[i].device)
|
479 |
+
|
480 |
+
y = x[i].sigmoid()
|
481 |
+
y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i] # xy
|
482 |
+
#y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh
|
483 |
+
|
484 |
+
|
485 |
+
#px = (self.x_bin_sigmoid.forward(y[..., 0:12]) + self.grid[i][..., 0]) * self.stride[i]
|
486 |
+
#py = (self.y_bin_sigmoid.forward(y[..., 12:24]) + self.grid[i][..., 1]) * self.stride[i]
|
487 |
+
|
488 |
+
pw = self.w_bin_sigmoid.forward(y[..., 2:24]) * self.anchor_grid[i][..., 0]
|
489 |
+
ph = self.h_bin_sigmoid.forward(y[..., 24:46]) * self.anchor_grid[i][..., 1]
|
490 |
+
|
491 |
+
#y[..., 0] = px
|
492 |
+
#y[..., 1] = py
|
493 |
+
y[..., 2] = pw
|
494 |
+
y[..., 3] = ph
|
495 |
+
|
496 |
+
y = torch.cat((y[..., 0:4], y[..., 46:]), dim=-1)
|
497 |
+
|
498 |
+
z.append(y.view(bs, -1, y.shape[-1]))
|
499 |
+
|
500 |
+
return x if self.training else (torch.cat(z, 1), x)
|
501 |
+
|
502 |
+
@staticmethod
|
503 |
+
def _make_grid(nx=20, ny=20):
|
504 |
+
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
|
505 |
+
return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()
|
506 |
+
|
507 |
+
|
508 |
+
class Model(nn.Module):
|
509 |
+
def __init__(self, cfg='yolor-csp-c.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes
|
510 |
+
super(Model, self).__init__()
|
511 |
+
self.traced = False
|
512 |
+
if isinstance(cfg, dict):
|
513 |
+
self.yaml = cfg # model dict
|
514 |
+
else: # is *.yaml
|
515 |
+
import yaml # for torch hub
|
516 |
+
self.yaml_file = Path(cfg).name
|
517 |
+
with open(cfg) as f:
|
518 |
+
self.yaml = yaml.load(f, Loader=yaml.SafeLoader) # model dict
|
519 |
+
|
520 |
+
# Define model
|
521 |
+
ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels
|
522 |
+
if nc and nc != self.yaml['nc']:
|
523 |
+
logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
|
524 |
+
self.yaml['nc'] = nc # override yaml value
|
525 |
+
if anchors:
|
526 |
+
logger.info(f'Overriding model.yaml anchors with anchors={anchors}')
|
527 |
+
self.yaml['anchors'] = round(anchors) # override yaml value
|
528 |
+
self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist
|
529 |
+
self.names = [str(i) for i in range(self.yaml['nc'])] # default names
|
530 |
+
# print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])
|
531 |
+
|
532 |
+
# Build strides, anchors
|
533 |
+
m = self.model[-1] # Detect()
|
534 |
+
if isinstance(m, Detect):
|
535 |
+
s = 256 # 2x min stride
|
536 |
+
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward
|
537 |
+
check_anchor_order(m)
|
538 |
+
m.anchors /= m.stride.view(-1, 1, 1)
|
539 |
+
self.stride = m.stride
|
540 |
+
self._initialize_biases() # only run once
|
541 |
+
# print('Strides: %s' % m.stride.tolist())
|
542 |
+
if isinstance(m, IDetect):
|
543 |
+
s = 256 # 2x min stride
|
544 |
+
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward
|
545 |
+
check_anchor_order(m)
|
546 |
+
m.anchors /= m.stride.view(-1, 1, 1)
|
547 |
+
self.stride = m.stride
|
548 |
+
self._initialize_biases() # only run once
|
549 |
+
# print('Strides: %s' % m.stride.tolist())
|
550 |
+
if isinstance(m, IAuxDetect):
|
551 |
+
s = 256 # 2x min stride
|
552 |
+
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))[:4]]) # forward
|
553 |
+
#print(m.stride)
|
554 |
+
check_anchor_order(m)
|
555 |
+
m.anchors /= m.stride.view(-1, 1, 1)
|
556 |
+
self.stride = m.stride
|
557 |
+
self._initialize_aux_biases() # only run once
|
558 |
+
# print('Strides: %s' % m.stride.tolist())
|
559 |
+
if isinstance(m, IBin):
|
560 |
+
s = 256 # 2x min stride
|
561 |
+
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward
|
562 |
+
check_anchor_order(m)
|
563 |
+
m.anchors /= m.stride.view(-1, 1, 1)
|
564 |
+
self.stride = m.stride
|
565 |
+
self._initialize_biases_bin() # only run once
|
566 |
+
# print('Strides: %s' % m.stride.tolist())
|
567 |
+
if isinstance(m, IKeypoint):
|
568 |
+
s = 256 # 2x min stride
|
569 |
+
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward
|
570 |
+
check_anchor_order(m)
|
571 |
+
m.anchors /= m.stride.view(-1, 1, 1)
|
572 |
+
self.stride = m.stride
|
573 |
+
self._initialize_biases_kpt() # only run once
|
574 |
+
# print('Strides: %s' % m.stride.tolist())
|
575 |
+
|
576 |
+
# Init weights, biases
|
577 |
+
initialize_weights(self)
|
578 |
+
self.info()
|
579 |
+
logger.info('')
|
580 |
+
|
581 |
+
def forward(self, x, augment=False, profile=False):
|
582 |
+
if augment:
|
583 |
+
img_size = x.shape[-2:] # height, width
|
584 |
+
s = [1, 0.83, 0.67] # scales
|
585 |
+
f = [None, 3, None] # flips (2-ud, 3-lr)
|
586 |
+
y = [] # outputs
|
587 |
+
for si, fi in zip(s, f):
|
588 |
+
xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
|
589 |
+
yi = self.forward_once(xi)[0] # forward
|
590 |
+
# cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save
|
591 |
+
yi[..., :4] /= si # de-scale
|
592 |
+
if fi == 2:
|
593 |
+
yi[..., 1] = img_size[0] - yi[..., 1] # de-flip ud
|
594 |
+
elif fi == 3:
|
595 |
+
yi[..., 0] = img_size[1] - yi[..., 0] # de-flip lr
|
596 |
+
y.append(yi)
|
597 |
+
return torch.cat(y, 1), None # augmented inference, train
|
598 |
+
else:
|
599 |
+
return self.forward_once(x, profile) # single-scale inference, train
|
600 |
+
|
601 |
+
def forward_once(self, x, profile=False):
|
602 |
+
y, dt = [], [] # outputs
|
603 |
+
for m in self.model:
|
604 |
+
if m.f != -1: # if not from previous layer
|
605 |
+
x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers
|
606 |
+
|
607 |
+
if not hasattr(self, 'traced'):
|
608 |
+
self.traced=False
|
609 |
+
|
610 |
+
if self.traced:
|
611 |
+
if isinstance(m, Detect) or isinstance(m, IDetect) or isinstance(m, IAuxDetect) or isinstance(m, IKeypoint):
|
612 |
+
break
|
613 |
+
|
614 |
+
if profile:
|
615 |
+
c = isinstance(m, (Detect, IDetect, IAuxDetect, IBin))
|
616 |
+
o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPS
|
617 |
+
for _ in range(10):
|
618 |
+
m(x.copy() if c else x)
|
619 |
+
t = time_synchronized()
|
620 |
+
for _ in range(10):
|
621 |
+
m(x.copy() if c else x)
|
622 |
+
dt.append((time_synchronized() - t) * 100)
|
623 |
+
print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type))
|
624 |
+
|
625 |
+
x = m(x) # run
|
626 |
+
|
627 |
+
y.append(x if m.i in self.save else None) # save output
|
628 |
+
|
629 |
+
if profile:
|
630 |
+
print('%.1fms total' % sum(dt))
|
631 |
+
return x
|
632 |
+
|
633 |
+
def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency
|
634 |
+
# https://arxiv.org/abs/1708.02002 section 3.3
|
635 |
+
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
|
636 |
+
m = self.model[-1] # Detect() module
|
637 |
+
for mi, s in zip(m.m, m.stride): # from
|
638 |
+
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85)
|
639 |
+
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image)
|
640 |
+
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
|
641 |
+
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
|
642 |
+
|
643 |
+
def _initialize_aux_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency
|
644 |
+
# https://arxiv.org/abs/1708.02002 section 3.3
|
645 |
+
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
|
646 |
+
m = self.model[-1] # Detect() module
|
647 |
+
for mi, mi2, s in zip(m.m, m.m2, m.stride): # from
|
648 |
+
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85)
|
649 |
+
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image)
|
650 |
+
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
|
651 |
+
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
|
652 |
+
b2 = mi2.bias.view(m.na, -1) # conv.bias(255) to (3,85)
|
653 |
+
b2.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image)
|
654 |
+
b2.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
|
655 |
+
mi2.bias = torch.nn.Parameter(b2.view(-1), requires_grad=True)
|
656 |
+
|
657 |
+
def _initialize_biases_bin(self, cf=None): # initialize biases into Detect(), cf is class frequency
|
658 |
+
# https://arxiv.org/abs/1708.02002 section 3.3
|
659 |
+
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
|
660 |
+
m = self.model[-1] # Bin() module
|
661 |
+
bc = m.bin_count
|
662 |
+
for mi, s in zip(m.m, m.stride): # from
|
663 |
+
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85)
|
664 |
+
old = b[:, (0,1,2,bc+3)].data
|
665 |
+
obj_idx = 2*bc+4
|
666 |
+
b[:, :obj_idx].data += math.log(0.6 / (bc + 1 - 0.99))
|
667 |
+
b[:, obj_idx].data += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image)
|
668 |
+
b[:, (obj_idx+1):].data += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
|
669 |
+
b[:, (0,1,2,bc+3)].data = old
|
670 |
+
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
|
671 |
+
|
672 |
+
def _initialize_biases_kpt(self, cf=None): # initialize biases into Detect(), cf is class frequency
|
673 |
+
# https://arxiv.org/abs/1708.02002 section 3.3
|
674 |
+
# cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
|
675 |
+
m = self.model[-1] # Detect() module
|
676 |
+
for mi, s in zip(m.m, m.stride): # from
|
677 |
+
b = mi.bias.view(m.na, -1) # conv.bias(255) to (3,85)
|
678 |
+
b.data[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image)
|
679 |
+
b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()) # cls
|
680 |
+
mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
|
681 |
+
|
682 |
+
def _print_biases(self):
|
683 |
+
m = self.model[-1] # Detect() module
|
684 |
+
for mi in m.m: # from
|
685 |
+
b = mi.bias.detach().view(m.na, -1).T # conv.bias(255) to (3,85)
|
686 |
+
print(('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean()))
|
687 |
+
|
688 |
+
# def _print_weights(self):
|
689 |
+
# for m in self.model.modules():
|
690 |
+
# if type(m) is Bottleneck:
|
691 |
+
# print('%10.3g' % (m.w.detach().sigmoid() * 2)) # shortcut weights
|
692 |
+
|
693 |
+
def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers
|
694 |
+
print('Fusing layers... ')
|
695 |
+
for m in self.model.modules():
|
696 |
+
if isinstance(m, RepConv):
|
697 |
+
#print(f" fuse_repvgg_block")
|
698 |
+
m.fuse_repvgg_block()
|
699 |
+
elif isinstance(m, RepConv_OREPA):
|
700 |
+
#print(f" switch_to_deploy")
|
701 |
+
m.switch_to_deploy()
|
702 |
+
elif type(m) is Conv and hasattr(m, 'bn'):
|
703 |
+
m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv
|
704 |
+
delattr(m, 'bn') # remove batchnorm
|
705 |
+
m.forward = m.fuseforward # update forward
|
706 |
+
elif isinstance(m, (IDetect, IAuxDetect)):
|
707 |
+
m.fuse()
|
708 |
+
m.forward = m.fuseforward
|
709 |
+
self.info()
|
710 |
+
return self
|
711 |
+
|
712 |
+
def nms(self, mode=True): # add or remove NMS module
|
713 |
+
present = type(self.model[-1]) is NMS # last layer is NMS
|
714 |
+
if mode and not present:
|
715 |
+
print('Adding NMS... ')
|
716 |
+
m = NMS() # module
|
717 |
+
m.f = -1 # from
|
718 |
+
m.i = self.model[-1].i + 1 # index
|
719 |
+
self.model.add_module(name='%s' % m.i, module=m) # add
|
720 |
+
self.eval()
|
721 |
+
elif not mode and present:
|
722 |
+
print('Removing NMS... ')
|
723 |
+
self.model = self.model[:-1] # remove
|
724 |
+
return self
|
725 |
+
|
726 |
+
def autoshape(self): # add autoShape module
|
727 |
+
print('Adding autoShape... ')
|
728 |
+
m = autoShape(self) # wrap model
|
729 |
+
copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=()) # copy attributes
|
730 |
+
return m
|
731 |
+
|
732 |
+
def info(self, verbose=False, img_size=640): # print model information
|
733 |
+
model_info(self, verbose, img_size)
|
734 |
+
|
735 |
+
|
736 |
+
def parse_model(d, ch): # model_dict, input_channels(3)
|
737 |
+
logger.info('\n%3s%18s%3s%10s %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
|
738 |
+
anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
|
739 |
+
na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors
|
740 |
+
no = na * (nc + 5) # number of outputs = anchors * (classes + 5)
|
741 |
+
|
742 |
+
layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out
|
743 |
+
for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args
|
744 |
+
m = eval(m) if isinstance(m, str) else m # eval strings
|
745 |
+
for j, a in enumerate(args):
|
746 |
+
try:
|
747 |
+
args[j] = eval(a) if isinstance(a, str) else a # eval strings
|
748 |
+
except:
|
749 |
+
pass
|
750 |
+
|
751 |
+
n = max(round(n * gd), 1) if n > 1 else n # depth gain
|
752 |
+
if m in [nn.Conv2d, Conv, RobustConv, RobustConv2, DWConv, GhostConv, RepConv, RepConv_OREPA, DownC,
|
753 |
+
SPP, SPPF, SPPCSPC, GhostSPPCSPC, MixConv2d, Focus, Stem, GhostStem, CrossConv,
|
754 |
+
Bottleneck, BottleneckCSPA, BottleneckCSPB, BottleneckCSPC,
|
755 |
+
RepBottleneck, RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC,
|
756 |
+
Res, ResCSPA, ResCSPB, ResCSPC,
|
757 |
+
RepRes, RepResCSPA, RepResCSPB, RepResCSPC,
|
758 |
+
ResX, ResXCSPA, ResXCSPB, ResXCSPC,
|
759 |
+
RepResX, RepResXCSPA, RepResXCSPB, RepResXCSPC,
|
760 |
+
Ghost, GhostCSPA, GhostCSPB, GhostCSPC,
|
761 |
+
SwinTransformerBlock, STCSPA, STCSPB, STCSPC,
|
762 |
+
SwinTransformer2Block, ST2CSPA, ST2CSPB, ST2CSPC]:
|
763 |
+
c1, c2 = ch[f], args[0]
|
764 |
+
if c2 != no: # if not output
|
765 |
+
c2 = make_divisible(c2 * gw, 8)
|
766 |
+
|
767 |
+
args = [c1, c2, *args[1:]]
|
768 |
+
if m in [DownC, SPPCSPC, GhostSPPCSPC,
|
769 |
+
BottleneckCSPA, BottleneckCSPB, BottleneckCSPC,
|
770 |
+
RepBottleneckCSPA, RepBottleneckCSPB, RepBottleneckCSPC,
|
771 |
+
ResCSPA, ResCSPB, ResCSPC,
|
772 |
+
RepResCSPA, RepResCSPB, RepResCSPC,
|
773 |
+
ResXCSPA, ResXCSPB, ResXCSPC,
|
774 |
+
RepResXCSPA, RepResXCSPB, RepResXCSPC,
|
775 |
+
GhostCSPA, GhostCSPB, GhostCSPC,
|
776 |
+
STCSPA, STCSPB, STCSPC,
|
777 |
+
ST2CSPA, ST2CSPB, ST2CSPC]:
|
778 |
+
args.insert(2, n) # number of repeats
|
779 |
+
n = 1
|
780 |
+
elif m is nn.BatchNorm2d:
|
781 |
+
args = [ch[f]]
|
782 |
+
elif m is Concat:
|
783 |
+
c2 = sum([ch[x] for x in f])
|
784 |
+
elif m is Chuncat:
|
785 |
+
c2 = sum([ch[x] for x in f])
|
786 |
+
elif m is Shortcut:
|
787 |
+
c2 = ch[f[0]]
|
788 |
+
elif m is Foldcut:
|
789 |
+
c2 = ch[f] // 2
|
790 |
+
elif m in [Detect, IDetect, IAuxDetect, IBin, IKeypoint]:
|
791 |
+
args.append([ch[x] for x in f])
|
792 |
+
if isinstance(args[1], int): # number of anchors
|
793 |
+
args[1] = [list(range(args[1] * 2))] * len(f)
|
794 |
+
elif m is ReOrg:
|
795 |
+
c2 = ch[f] * 4
|
796 |
+
elif m is Contract:
|
797 |
+
c2 = ch[f] * args[0] ** 2
|
798 |
+
elif m is Expand:
|
799 |
+
c2 = ch[f] // args[0] ** 2
|
800 |
+
else:
|
801 |
+
c2 = ch[f]
|
802 |
+
|
803 |
+
m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args) # module
|
804 |
+
t = str(m)[8:-2].replace('__main__.', '') # module type
|
805 |
+
np = sum([x.numel() for x in m_.parameters()]) # number params
|
806 |
+
m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params
|
807 |
+
logger.info('%3s%18s%3s%10.0f %-40s%-30s' % (i, f, n, np, t, args)) # print
|
808 |
+
save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist
|
809 |
+
layers.append(m_)
|
810 |
+
if i == 0:
|
811 |
+
ch = []
|
812 |
+
ch.append(c2)
|
813 |
+
return nn.Sequential(*layers), sorted(save)
|
814 |
+
|
815 |
+
|
816 |
+
if __name__ == '__main__':
|
817 |
+
parser = argparse.ArgumentParser()
|
818 |
+
parser.add_argument('--cfg', type=str, default='yolor-csp-c.yaml', help='model.yaml')
|
819 |
+
parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
|
820 |
+
parser.add_argument('--profile', action='store_true', help='profile model speed')
|
821 |
+
opt = parser.parse_args()
|
822 |
+
opt.cfg = check_file(opt.cfg) # check file
|
823 |
+
set_logging()
|
824 |
+
device = select_device(opt.device)
|
825 |
+
|
826 |
+
# Create model
|
827 |
+
model = Model(opt.cfg).to(device)
|
828 |
+
model.train()
|
829 |
+
|
830 |
+
if opt.profile:
|
831 |
+
img = torch.rand(1, 3, 640, 640).to(device)
|
832 |
+
y = model(img, profile=True)
|
833 |
+
|
834 |
+
# Profile
|
835 |
+
# img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 640, 640).to(device)
|
836 |
+
# y = model(img, profile=True)
|
837 |
+
|
838 |
+
# Tensorboard
|
839 |
+
# from torch.utils.tensorboard import SummaryWriter
|
840 |
+
# tb_writer = SummaryWriter()
|
841 |
+
# print("Run 'tensorboard --logdir=models/runs' to view tensorboard at http://localhost:6006/")
|
842 |
+
# tb_writer.add_graph(model.model, img) # add model to tensorboard
|
843 |
+
# tb_writer.add_image('test', img[0], dataformats='CWH') # add model to tensorboard
|
requirements.txt
CHANGED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
matplotlib>=3.2.2
|
4 |
+
numpy>=1.18.5
|
5 |
+
opencv-python>=4.1.1
|
6 |
+
Pillow>=7.1.2
|
7 |
+
PyYAML>=5.3.1
|
8 |
+
requests>=2.23.0
|
9 |
+
scipy>=1.4.1
|
10 |
+
torch>=1.7.0,!=1.12.0
|
11 |
+
torchvision>=0.8.1,!=0.13.0
|
12 |
+
tqdm>=4.41.0
|
13 |
+
protobuf<4.21.3
|
14 |
+
|
15 |
+
|
16 |
+
tensorboard>=2.4.1
|
17 |
+
|
18 |
+
pandas>=1.1.4
|
19 |
+
seaborn>=0.11.0
|
utils/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
# init
|
utils/a.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
k
|
utils/activations.py
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Activation functions
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
|
8 |
+
# SiLU https://arxiv.org/pdf/1606.08415.pdf ----------------------------------------------------------------------------
|
9 |
+
class SiLU(nn.Module): # export-friendly version of nn.SiLU()
|
10 |
+
@staticmethod
|
11 |
+
def forward(x):
|
12 |
+
return x * torch.sigmoid(x)
|
13 |
+
|
14 |
+
|
15 |
+
class Hardswish(nn.Module): # export-friendly version of nn.Hardswish()
|
16 |
+
@staticmethod
|
17 |
+
def forward(x):
|
18 |
+
# return x * F.hardsigmoid(x) # for torchscript and CoreML
|
19 |
+
return x * F.hardtanh(x + 3, 0., 6.) / 6. # for torchscript, CoreML and ONNX
|
20 |
+
|
21 |
+
|
22 |
+
class MemoryEfficientSwish(nn.Module):
|
23 |
+
class F(torch.autograd.Function):
|
24 |
+
@staticmethod
|
25 |
+
def forward(ctx, x):
|
26 |
+
ctx.save_for_backward(x)
|
27 |
+
return x * torch.sigmoid(x)
|
28 |
+
|
29 |
+
@staticmethod
|
30 |
+
def backward(ctx, grad_output):
|
31 |
+
x = ctx.saved_tensors[0]
|
32 |
+
sx = torch.sigmoid(x)
|
33 |
+
return grad_output * (sx * (1 + x * (1 - sx)))
|
34 |
+
|
35 |
+
def forward(self, x):
|
36 |
+
return self.F.apply(x)
|
37 |
+
|
38 |
+
|
39 |
+
# Mish https://github.com/digantamisra98/Mish --------------------------------------------------------------------------
|
40 |
+
class Mish(nn.Module):
|
41 |
+
@staticmethod
|
42 |
+
def forward(x):
|
43 |
+
return x * F.softplus(x).tanh()
|
44 |
+
|
45 |
+
|
46 |
+
class MemoryEfficientMish(nn.Module):
|
47 |
+
class F(torch.autograd.Function):
|
48 |
+
@staticmethod
|
49 |
+
def forward(ctx, x):
|
50 |
+
ctx.save_for_backward(x)
|
51 |
+
return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x)))
|
52 |
+
|
53 |
+
@staticmethod
|
54 |
+
def backward(ctx, grad_output):
|
55 |
+
x = ctx.saved_tensors[0]
|
56 |
+
sx = torch.sigmoid(x)
|
57 |
+
fx = F.softplus(x).tanh()
|
58 |
+
return grad_output * (fx + x * sx * (1 - fx * fx))
|
59 |
+
|
60 |
+
def forward(self, x):
|
61 |
+
return self.F.apply(x)
|
62 |
+
|
63 |
+
|
64 |
+
# FReLU https://arxiv.org/abs/2007.11824 -------------------------------------------------------------------------------
|
65 |
+
class FReLU(nn.Module):
|
66 |
+
def __init__(self, c1, k=3): # ch_in, kernel
|
67 |
+
super().__init__()
|
68 |
+
self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False)
|
69 |
+
self.bn = nn.BatchNorm2d(c1)
|
70 |
+
|
71 |
+
def forward(self, x):
|
72 |
+
return torch.max(x, self.bn(self.conv(x)))
|
utils/add_nms.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import onnx
|
3 |
+
from onnx import shape_inference
|
4 |
+
try:
|
5 |
+
import onnx_graphsurgeon as gs
|
6 |
+
except Exception as e:
|
7 |
+
print('Import onnx_graphsurgeon failure: %s' % e)
|
8 |
+
|
9 |
+
import logging
|
10 |
+
|
11 |
+
LOGGER = logging.getLogger(__name__)
|
12 |
+
|
13 |
+
class RegisterNMS(object):
|
14 |
+
def __init__(
|
15 |
+
self,
|
16 |
+
onnx_model_path: str,
|
17 |
+
precision: str = "fp32",
|
18 |
+
):
|
19 |
+
|
20 |
+
self.graph = gs.import_onnx(onnx.load(onnx_model_path))
|
21 |
+
assert self.graph
|
22 |
+
LOGGER.info("ONNX graph created successfully")
|
23 |
+
# Fold constants via ONNX-GS that PyTorch2ONNX may have missed
|
24 |
+
self.graph.fold_constants()
|
25 |
+
self.precision = precision
|
26 |
+
self.batch_size = 1
|
27 |
+
def infer(self):
|
28 |
+
"""
|
29 |
+
Sanitize the graph by cleaning any unconnected nodes, do a topological resort,
|
30 |
+
and fold constant inputs values. When possible, run shape inference on the
|
31 |
+
ONNX graph to determine tensor shapes.
|
32 |
+
"""
|
33 |
+
for _ in range(3):
|
34 |
+
count_before = len(self.graph.nodes)
|
35 |
+
|
36 |
+
self.graph.cleanup().toposort()
|
37 |
+
try:
|
38 |
+
for node in self.graph.nodes:
|
39 |
+
for o in node.outputs:
|
40 |
+
o.shape = None
|
41 |
+
model = gs.export_onnx(self.graph)
|
42 |
+
model = shape_inference.infer_shapes(model)
|
43 |
+
self.graph = gs.import_onnx(model)
|
44 |
+
except Exception as e:
|
45 |
+
LOGGER.info(f"Shape inference could not be performed at this time:\n{e}")
|
46 |
+
try:
|
47 |
+
self.graph.fold_constants(fold_shapes=True)
|
48 |
+
except TypeError as e:
|
49 |
+
LOGGER.error(
|
50 |
+
"This version of ONNX GraphSurgeon does not support folding shapes, "
|
51 |
+
f"please upgrade your onnx_graphsurgeon module. Error:\n{e}"
|
52 |
+
)
|
53 |
+
raise
|
54 |
+
|
55 |
+
count_after = len(self.graph.nodes)
|
56 |
+
if count_before == count_after:
|
57 |
+
# No new folding occurred in this iteration, so we can stop for now.
|
58 |
+
break
|
59 |
+
|
60 |
+
def save(self, output_path):
|
61 |
+
"""
|
62 |
+
Save the ONNX model to the given location.
|
63 |
+
Args:
|
64 |
+
output_path: Path pointing to the location where to write
|
65 |
+
out the updated ONNX model.
|
66 |
+
"""
|
67 |
+
self.graph.cleanup().toposort()
|
68 |
+
model = gs.export_onnx(self.graph)
|
69 |
+
onnx.save(model, output_path)
|
70 |
+
LOGGER.info(f"Saved ONNX model to {output_path}")
|
71 |
+
|
72 |
+
def register_nms(
|
73 |
+
self,
|
74 |
+
*,
|
75 |
+
score_thresh: float = 0.25,
|
76 |
+
nms_thresh: float = 0.45,
|
77 |
+
detections_per_img: int = 100,
|
78 |
+
):
|
79 |
+
"""
|
80 |
+
Register the ``EfficientNMS_TRT`` plugin node.
|
81 |
+
NMS expects these shapes for its input tensors:
|
82 |
+
- box_net: [batch_size, number_boxes, 4]
|
83 |
+
- class_net: [batch_size, number_boxes, number_labels]
|
84 |
+
Args:
|
85 |
+
score_thresh (float): The scalar threshold for score (low scoring boxes are removed).
|
86 |
+
nms_thresh (float): The scalar threshold for IOU (new boxes that have high IOU
|
87 |
+
overlap with previously selected boxes are removed).
|
88 |
+
detections_per_img (int): Number of best detections to keep after NMS.
|
89 |
+
"""
|
90 |
+
|
91 |
+
self.infer()
|
92 |
+
# Find the concat node at the end of the network
|
93 |
+
op_inputs = self.graph.outputs
|
94 |
+
op = "EfficientNMS_TRT"
|
95 |
+
attrs = {
|
96 |
+
"plugin_version": "1",
|
97 |
+
"background_class": -1, # no background class
|
98 |
+
"max_output_boxes": detections_per_img,
|
99 |
+
"score_threshold": score_thresh,
|
100 |
+
"iou_threshold": nms_thresh,
|
101 |
+
"score_activation": False,
|
102 |
+
"box_coding": 0,
|
103 |
+
}
|
104 |
+
|
105 |
+
if self.precision == "fp32":
|
106 |
+
dtype_output = np.float32
|
107 |
+
elif self.precision == "fp16":
|
108 |
+
dtype_output = np.float16
|
109 |
+
else:
|
110 |
+
raise NotImplementedError(f"Currently not supports precision: {self.precision}")
|
111 |
+
|
112 |
+
# NMS Outputs
|
113 |
+
output_num_detections = gs.Variable(
|
114 |
+
name="num_dets",
|
115 |
+
dtype=np.int32,
|
116 |
+
shape=[self.batch_size, 1],
|
117 |
+
) # A scalar indicating the number of valid detections per batch image.
|
118 |
+
output_boxes = gs.Variable(
|
119 |
+
name="det_boxes",
|
120 |
+
dtype=dtype_output,
|
121 |
+
shape=[self.batch_size, detections_per_img, 4],
|
122 |
+
)
|
123 |
+
output_scores = gs.Variable(
|
124 |
+
name="det_scores",
|
125 |
+
dtype=dtype_output,
|
126 |
+
shape=[self.batch_size, detections_per_img],
|
127 |
+
)
|
128 |
+
output_labels = gs.Variable(
|
129 |
+
name="det_classes",
|
130 |
+
dtype=np.int32,
|
131 |
+
shape=[self.batch_size, detections_per_img],
|
132 |
+
)
|
133 |
+
|
134 |
+
op_outputs = [output_num_detections, output_boxes, output_scores, output_labels]
|
135 |
+
|
136 |
+
# Create the NMS Plugin node with the selected inputs. The outputs of the node will also
|
137 |
+
# become the final outputs of the graph.
|
138 |
+
self.graph.layer(op=op, name="batched_nms", inputs=op_inputs, outputs=op_outputs, attrs=attrs)
|
139 |
+
LOGGER.info(f"Created NMS plugin '{op}' with attributes: {attrs}")
|
140 |
+
|
141 |
+
self.graph.outputs = op_outputs
|
142 |
+
|
143 |
+
self.infer()
|
144 |
+
|
145 |
+
def save(self, output_path):
|
146 |
+
"""
|
147 |
+
Save the ONNX model to the given location.
|
148 |
+
Args:
|
149 |
+
output_path: Path pointing to the location where to write
|
150 |
+
out the updated ONNX model.
|
151 |
+
"""
|
152 |
+
self.graph.cleanup().toposort()
|
153 |
+
model = gs.export_onnx(self.graph)
|
154 |
+
onnx.save(model, output_path)
|
155 |
+
LOGGER.info(f"Saved ONNX model to {output_path}")
|
utils/autoanchor.py
ADDED
@@ -0,0 +1,160 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Auto-anchor utils
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import yaml
|
6 |
+
from scipy.cluster.vq import kmeans
|
7 |
+
from tqdm import tqdm
|
8 |
+
|
9 |
+
from utils.general import colorstr
|
10 |
+
|
11 |
+
|
12 |
+
def check_anchor_order(m):
|
13 |
+
# Check anchor order against stride order for YOLO Detect() module m, and correct if necessary
|
14 |
+
a = m.anchor_grid.prod(-1).view(-1) # anchor area
|
15 |
+
da = a[-1] - a[0] # delta a
|
16 |
+
ds = m.stride[-1] - m.stride[0] # delta s
|
17 |
+
if da.sign() != ds.sign(): # same order
|
18 |
+
print('Reversing anchor order')
|
19 |
+
m.anchors[:] = m.anchors.flip(0)
|
20 |
+
m.anchor_grid[:] = m.anchor_grid.flip(0)
|
21 |
+
|
22 |
+
|
23 |
+
def check_anchors(dataset, model, thr=4.0, imgsz=640):
|
24 |
+
# Check anchor fit to data, recompute if necessary
|
25 |
+
prefix = colorstr('autoanchor: ')
|
26 |
+
print(f'\n{prefix}Analyzing anchors... ', end='')
|
27 |
+
m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect()
|
28 |
+
shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
|
29 |
+
scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale
|
30 |
+
wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh
|
31 |
+
|
32 |
+
def metric(k): # compute metric
|
33 |
+
r = wh[:, None] / k[None]
|
34 |
+
x = torch.min(r, 1. / r).min(2)[0] # ratio metric
|
35 |
+
best = x.max(1)[0] # best_x
|
36 |
+
aat = (x > 1. / thr).float().sum(1).mean() # anchors above threshold
|
37 |
+
bpr = (best > 1. / thr).float().mean() # best possible recall
|
38 |
+
return bpr, aat
|
39 |
+
|
40 |
+
anchors = m.anchor_grid.clone().cpu().view(-1, 2) # current anchors
|
41 |
+
bpr, aat = metric(anchors)
|
42 |
+
print(f'anchors/target = {aat:.2f}, Best Possible Recall (BPR) = {bpr:.4f}', end='')
|
43 |
+
if bpr < 0.98: # threshold to recompute
|
44 |
+
print('. Attempting to improve anchors, please wait...')
|
45 |
+
na = m.anchor_grid.numel() // 2 # number of anchors
|
46 |
+
try:
|
47 |
+
anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
|
48 |
+
except Exception as e:
|
49 |
+
print(f'{prefix}ERROR: {e}')
|
50 |
+
new_bpr = metric(anchors)[0]
|
51 |
+
if new_bpr > bpr: # replace anchors
|
52 |
+
anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
|
53 |
+
m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid) # for inference
|
54 |
+
check_anchor_order(m)
|
55 |
+
m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1) # loss
|
56 |
+
print(f'{prefix}New anchors saved to model. Update model *.yaml to use these anchors in the future.')
|
57 |
+
else:
|
58 |
+
print(f'{prefix}Original anchors better than new anchors. Proceeding with original anchors.')
|
59 |
+
print('') # newline
|
60 |
+
|
61 |
+
|
62 |
+
def kmean_anchors(path='./data/coco.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
|
63 |
+
""" Creates kmeans-evolved anchors from training dataset
|
64 |
+
|
65 |
+
Arguments:
|
66 |
+
path: path to dataset *.yaml, or a loaded dataset
|
67 |
+
n: number of anchors
|
68 |
+
img_size: image size used for training
|
69 |
+
thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
|
70 |
+
gen: generations to evolve anchors using genetic algorithm
|
71 |
+
verbose: print all results
|
72 |
+
|
73 |
+
Return:
|
74 |
+
k: kmeans evolved anchors
|
75 |
+
|
76 |
+
Usage:
|
77 |
+
from utils.autoanchor import *; _ = kmean_anchors()
|
78 |
+
"""
|
79 |
+
thr = 1. / thr
|
80 |
+
prefix = colorstr('autoanchor: ')
|
81 |
+
|
82 |
+
def metric(k, wh): # compute metrics
|
83 |
+
r = wh[:, None] / k[None]
|
84 |
+
x = torch.min(r, 1. / r).min(2)[0] # ratio metric
|
85 |
+
# x = wh_iou(wh, torch.tensor(k)) # iou metric
|
86 |
+
return x, x.max(1)[0] # x, best_x
|
87 |
+
|
88 |
+
def anchor_fitness(k): # mutation fitness
|
89 |
+
_, best = metric(torch.tensor(k, dtype=torch.float32), wh)
|
90 |
+
return (best * (best > thr).float()).mean() # fitness
|
91 |
+
|
92 |
+
def print_results(k):
|
93 |
+
k = k[np.argsort(k.prod(1))] # sort small to large
|
94 |
+
x, best = metric(k, wh0)
|
95 |
+
bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr
|
96 |
+
print(f'{prefix}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr')
|
97 |
+
print(f'{prefix}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, '
|
98 |
+
f'past_thr={x[x > thr].mean():.3f}-mean: ', end='')
|
99 |
+
for i, x in enumerate(k):
|
100 |
+
print('%i,%i' % (round(x[0]), round(x[1])), end=', ' if i < len(k) - 1 else '\n') # use in *.cfg
|
101 |
+
return k
|
102 |
+
|
103 |
+
if isinstance(path, str): # *.yaml file
|
104 |
+
with open(path) as f:
|
105 |
+
data_dict = yaml.load(f, Loader=yaml.SafeLoader) # model dict
|
106 |
+
from utils.datasets import LoadImagesAndLabels
|
107 |
+
dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
|
108 |
+
else:
|
109 |
+
dataset = path # dataset
|
110 |
+
|
111 |
+
# Get label wh
|
112 |
+
shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
|
113 |
+
wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh
|
114 |
+
|
115 |
+
# Filter
|
116 |
+
i = (wh0 < 3.0).any(1).sum()
|
117 |
+
if i:
|
118 |
+
print(f'{prefix}WARNING: Extremely small objects found. {i} of {len(wh0)} labels are < 3 pixels in size.')
|
119 |
+
wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels
|
120 |
+
# wh = wh * (np.random.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1
|
121 |
+
|
122 |
+
# Kmeans calculation
|
123 |
+
print(f'{prefix}Running kmeans for {n} anchors on {len(wh)} points...')
|
124 |
+
s = wh.std(0) # sigmas for whitening
|
125 |
+
k, dist = kmeans(wh / s, n, iter=30) # points, mean distance
|
126 |
+
assert len(k) == n, print(f'{prefix}ERROR: scipy.cluster.vq.kmeans requested {n} points but returned only {len(k)}')
|
127 |
+
k *= s
|
128 |
+
wh = torch.tensor(wh, dtype=torch.float32) # filtered
|
129 |
+
wh0 = torch.tensor(wh0, dtype=torch.float32) # unfiltered
|
130 |
+
k = print_results(k)
|
131 |
+
|
132 |
+
# Plot
|
133 |
+
# k, d = [None] * 20, [None] * 20
|
134 |
+
# for i in tqdm(range(1, 21)):
|
135 |
+
# k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance
|
136 |
+
# fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
|
137 |
+
# ax = ax.ravel()
|
138 |
+
# ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
|
139 |
+
# fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh
|
140 |
+
# ax[0].hist(wh[wh[:, 0]<100, 0],400)
|
141 |
+
# ax[1].hist(wh[wh[:, 1]<100, 1],400)
|
142 |
+
# fig.savefig('wh.png', dpi=200)
|
143 |
+
|
144 |
+
# Evolve
|
145 |
+
npr = np.random
|
146 |
+
f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma
|
147 |
+
pbar = tqdm(range(gen), desc=f'{prefix}Evolving anchors with Genetic Algorithm:') # progress bar
|
148 |
+
for _ in pbar:
|
149 |
+
v = np.ones(sh)
|
150 |
+
while (v == 1).all(): # mutate until a change occurs (prevent duplicates)
|
151 |
+
v = ((npr.random(sh) < mp) * npr.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
|
152 |
+
kg = (k.copy() * v).clip(min=2.0)
|
153 |
+
fg = anchor_fitness(kg)
|
154 |
+
if fg > f:
|
155 |
+
f, k = fg, kg.copy()
|
156 |
+
pbar.desc = f'{prefix}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
|
157 |
+
if verbose:
|
158 |
+
print_results(k)
|
159 |
+
|
160 |
+
return print_results(k)
|
utils/datasets.py
ADDED
@@ -0,0 +1,1320 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Dataset utils and dataloaders
|
2 |
+
|
3 |
+
import glob
|
4 |
+
import logging
|
5 |
+
import math
|
6 |
+
import os
|
7 |
+
import random
|
8 |
+
import shutil
|
9 |
+
import time
|
10 |
+
from itertools import repeat
|
11 |
+
from multiprocessing.pool import ThreadPool
|
12 |
+
from pathlib import Path
|
13 |
+
from threading import Thread
|
14 |
+
|
15 |
+
import cv2
|
16 |
+
import numpy as np
|
17 |
+
import torch
|
18 |
+
import torch.nn.functional as F
|
19 |
+
from PIL import Image, ExifTags
|
20 |
+
from torch.utils.data import Dataset
|
21 |
+
from tqdm import tqdm
|
22 |
+
|
23 |
+
import pickle
|
24 |
+
from copy import deepcopy
|
25 |
+
#from pycocotools import mask as maskUtils
|
26 |
+
from torchvision.utils import save_image
|
27 |
+
from torchvision.ops import roi_pool, roi_align, ps_roi_pool, ps_roi_align
|
28 |
+
|
29 |
+
from utils.general import check_requirements, xyxy2xywh, xywh2xyxy, xywhn2xyxy, xyn2xy, segment2box, segments2boxes, \
|
30 |
+
resample_segments, clean_str
|
31 |
+
from utils.torch_utils import torch_distributed_zero_first
|
32 |
+
|
33 |
+
# Parameters
|
34 |
+
help_url = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data'
|
35 |
+
img_formats = ['bmp', 'jpg', 'jpeg', 'png', 'tif', 'tiff', 'dng', 'webp', 'mpo'] # acceptable image suffixes
|
36 |
+
vid_formats = ['mov', 'avi', 'mp4', 'mpg', 'mpeg', 'm4v', 'wmv', 'mkv'] # acceptable video suffixes
|
37 |
+
logger = logging.getLogger(__name__)
|
38 |
+
|
39 |
+
# Get orientation exif tag
|
40 |
+
for orientation in ExifTags.TAGS.keys():
|
41 |
+
if ExifTags.TAGS[orientation] == 'Orientation':
|
42 |
+
break
|
43 |
+
|
44 |
+
|
45 |
+
def get_hash(files):
|
46 |
+
# Returns a single hash value of a list of files
|
47 |
+
return sum(os.path.getsize(f) for f in files if os.path.isfile(f))
|
48 |
+
|
49 |
+
|
50 |
+
def exif_size(img):
|
51 |
+
# Returns exif-corrected PIL size
|
52 |
+
s = img.size # (width, height)
|
53 |
+
try:
|
54 |
+
rotation = dict(img._getexif().items())[orientation]
|
55 |
+
if rotation == 6: # rotation 270
|
56 |
+
s = (s[1], s[0])
|
57 |
+
elif rotation == 8: # rotation 90
|
58 |
+
s = (s[1], s[0])
|
59 |
+
except:
|
60 |
+
pass
|
61 |
+
|
62 |
+
return s
|
63 |
+
|
64 |
+
|
65 |
+
def create_dataloader(path, imgsz, batch_size, stride, opt, hyp=None, augment=False, cache=False, pad=0.0, rect=False,
|
66 |
+
rank=-1, world_size=1, workers=8, image_weights=False, quad=False, prefix=''):
|
67 |
+
# Make sure only the first process in DDP process the dataset first, and the following others can use the cache
|
68 |
+
with torch_distributed_zero_first(rank):
|
69 |
+
dataset = LoadImagesAndLabels(path, imgsz, batch_size,
|
70 |
+
augment=augment, # augment images
|
71 |
+
hyp=hyp, # augmentation hyperparameters
|
72 |
+
rect=rect, # rectangular training
|
73 |
+
cache_images=cache,
|
74 |
+
single_cls=opt.single_cls,
|
75 |
+
stride=int(stride),
|
76 |
+
pad=pad,
|
77 |
+
image_weights=image_weights,
|
78 |
+
prefix=prefix)
|
79 |
+
|
80 |
+
batch_size = min(batch_size, len(dataset))
|
81 |
+
nw = min([os.cpu_count() // world_size, batch_size if batch_size > 1 else 0, workers]) # number of workers
|
82 |
+
sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None
|
83 |
+
loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader
|
84 |
+
# Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader()
|
85 |
+
dataloader = loader(dataset,
|
86 |
+
batch_size=batch_size,
|
87 |
+
num_workers=nw,
|
88 |
+
sampler=sampler,
|
89 |
+
pin_memory=True,
|
90 |
+
collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn)
|
91 |
+
return dataloader, dataset
|
92 |
+
|
93 |
+
|
94 |
+
class InfiniteDataLoader(torch.utils.data.dataloader.DataLoader):
|
95 |
+
""" Dataloader that reuses workers
|
96 |
+
|
97 |
+
Uses same syntax as vanilla DataLoader
|
98 |
+
"""
|
99 |
+
|
100 |
+
def __init__(self, *args, **kwargs):
|
101 |
+
super().__init__(*args, **kwargs)
|
102 |
+
object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler))
|
103 |
+
self.iterator = super().__iter__()
|
104 |
+
|
105 |
+
def __len__(self):
|
106 |
+
return len(self.batch_sampler.sampler)
|
107 |
+
|
108 |
+
def __iter__(self):
|
109 |
+
for i in range(len(self)):
|
110 |
+
yield next(self.iterator)
|
111 |
+
|
112 |
+
|
113 |
+
class _RepeatSampler(object):
|
114 |
+
""" Sampler that repeats forever
|
115 |
+
|
116 |
+
Args:
|
117 |
+
sampler (Sampler)
|
118 |
+
"""
|
119 |
+
|
120 |
+
def __init__(self, sampler):
|
121 |
+
self.sampler = sampler
|
122 |
+
|
123 |
+
def __iter__(self):
|
124 |
+
while True:
|
125 |
+
yield from iter(self.sampler)
|
126 |
+
|
127 |
+
|
128 |
+
class LoadImages: # for inference
|
129 |
+
def __init__(self, path, img_size=640, stride=32):
|
130 |
+
p = str(Path(path).absolute()) # os-agnostic absolute path
|
131 |
+
if '*' in p:
|
132 |
+
files = sorted(glob.glob(p, recursive=True)) # glob
|
133 |
+
elif os.path.isdir(p):
|
134 |
+
files = sorted(glob.glob(os.path.join(p, '*.*'))) # dir
|
135 |
+
elif os.path.isfile(p):
|
136 |
+
files = [p] # files
|
137 |
+
else:
|
138 |
+
raise Exception(f'ERROR: {p} does not exist')
|
139 |
+
|
140 |
+
images = [x for x in files if x.split('.')[-1].lower() in img_formats]
|
141 |
+
videos = [x for x in files if x.split('.')[-1].lower() in vid_formats]
|
142 |
+
ni, nv = len(images), len(videos)
|
143 |
+
|
144 |
+
self.img_size = img_size
|
145 |
+
self.stride = stride
|
146 |
+
self.files = images + videos
|
147 |
+
self.nf = ni + nv # number of files
|
148 |
+
self.video_flag = [False] * ni + [True] * nv
|
149 |
+
self.mode = 'image'
|
150 |
+
if any(videos):
|
151 |
+
self.new_video(videos[0]) # new video
|
152 |
+
else:
|
153 |
+
self.cap = None
|
154 |
+
assert self.nf > 0, f'No images or videos found in {p}. ' \
|
155 |
+
f'Supported formats are:\nimages: {img_formats}\nvideos: {vid_formats}'
|
156 |
+
|
157 |
+
def __iter__(self):
|
158 |
+
self.count = 0
|
159 |
+
return self
|
160 |
+
|
161 |
+
def __next__(self):
|
162 |
+
if self.count == self.nf:
|
163 |
+
raise StopIteration
|
164 |
+
path = self.files[self.count]
|
165 |
+
|
166 |
+
if self.video_flag[self.count]:
|
167 |
+
# Read video
|
168 |
+
self.mode = 'video'
|
169 |
+
ret_val, img0 = self.cap.read()
|
170 |
+
if not ret_val:
|
171 |
+
self.count += 1
|
172 |
+
self.cap.release()
|
173 |
+
if self.count == self.nf: # last video
|
174 |
+
raise StopIteration
|
175 |
+
else:
|
176 |
+
path = self.files[self.count]
|
177 |
+
self.new_video(path)
|
178 |
+
ret_val, img0 = self.cap.read()
|
179 |
+
|
180 |
+
self.frame += 1
|
181 |
+
print(f'video {self.count + 1}/{self.nf} ({self.frame}/{self.nframes}) {path}: ', end='')
|
182 |
+
|
183 |
+
else:
|
184 |
+
# Read image
|
185 |
+
self.count += 1
|
186 |
+
img0 = cv2.imread(path) # BGR
|
187 |
+
assert img0 is not None, 'Image Not Found ' + path
|
188 |
+
#print(f'image {self.count}/{self.nf} {path}: ', end='')
|
189 |
+
|
190 |
+
# Padded resize
|
191 |
+
img = letterbox(img0, self.img_size, stride=self.stride)[0]
|
192 |
+
|
193 |
+
# Convert
|
194 |
+
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
|
195 |
+
img = np.ascontiguousarray(img)
|
196 |
+
|
197 |
+
return path, img, img0, self.cap
|
198 |
+
|
199 |
+
def new_video(self, path):
|
200 |
+
self.frame = 0
|
201 |
+
self.cap = cv2.VideoCapture(path)
|
202 |
+
self.nframes = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
203 |
+
|
204 |
+
def __len__(self):
|
205 |
+
return self.nf # number of files
|
206 |
+
|
207 |
+
|
208 |
+
class LoadWebcam: # for inference
|
209 |
+
def __init__(self, pipe='0', img_size=640, stride=32):
|
210 |
+
self.img_size = img_size
|
211 |
+
self.stride = stride
|
212 |
+
|
213 |
+
if pipe.isnumeric():
|
214 |
+
pipe = eval(pipe) # local camera
|
215 |
+
# pipe = 'rtsp://192.168.1.64/1' # IP camera
|
216 |
+
# pipe = 'rtsp://username:[email protected]/1' # IP camera with login
|
217 |
+
# pipe = 'http://wmccpinetop.axiscam.net/mjpg/video.mjpg' # IP golf camera
|
218 |
+
|
219 |
+
self.pipe = pipe
|
220 |
+
self.cap = cv2.VideoCapture(pipe) # video capture object
|
221 |
+
self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3) # set buffer size
|
222 |
+
|
223 |
+
def __iter__(self):
|
224 |
+
self.count = -1
|
225 |
+
return self
|
226 |
+
|
227 |
+
def __next__(self):
|
228 |
+
self.count += 1
|
229 |
+
if cv2.waitKey(1) == ord('q'): # q to quit
|
230 |
+
self.cap.release()
|
231 |
+
cv2.destroyAllWindows()
|
232 |
+
raise StopIteration
|
233 |
+
|
234 |
+
# Read frame
|
235 |
+
if self.pipe == 0: # local camera
|
236 |
+
ret_val, img0 = self.cap.read()
|
237 |
+
img0 = cv2.flip(img0, 1) # flip left-right
|
238 |
+
else: # IP camera
|
239 |
+
n = 0
|
240 |
+
while True:
|
241 |
+
n += 1
|
242 |
+
self.cap.grab()
|
243 |
+
if n % 30 == 0: # skip frames
|
244 |
+
ret_val, img0 = self.cap.retrieve()
|
245 |
+
if ret_val:
|
246 |
+
break
|
247 |
+
|
248 |
+
# Print
|
249 |
+
assert ret_val, f'Camera Error {self.pipe}'
|
250 |
+
img_path = 'webcam.jpg'
|
251 |
+
print(f'webcam {self.count}: ', end='')
|
252 |
+
|
253 |
+
# Padded resize
|
254 |
+
img = letterbox(img0, self.img_size, stride=self.stride)[0]
|
255 |
+
|
256 |
+
# Convert
|
257 |
+
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
|
258 |
+
img = np.ascontiguousarray(img)
|
259 |
+
|
260 |
+
return img_path, img, img0, None
|
261 |
+
|
262 |
+
def __len__(self):
|
263 |
+
return 0
|
264 |
+
|
265 |
+
|
266 |
+
class LoadStreams: # multiple IP or RTSP cameras
|
267 |
+
def __init__(self, sources='streams.txt', img_size=640, stride=32):
|
268 |
+
self.mode = 'stream'
|
269 |
+
self.img_size = img_size
|
270 |
+
self.stride = stride
|
271 |
+
|
272 |
+
if os.path.isfile(sources):
|
273 |
+
with open(sources, 'r') as f:
|
274 |
+
sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())]
|
275 |
+
else:
|
276 |
+
sources = [sources]
|
277 |
+
|
278 |
+
n = len(sources)
|
279 |
+
self.imgs = [None] * n
|
280 |
+
self.sources = [clean_str(x) for x in sources] # clean source names for later
|
281 |
+
for i, s in enumerate(sources):
|
282 |
+
# Start the thread to read frames from the video stream
|
283 |
+
print(f'{i + 1}/{n}: {s}... ', end='')
|
284 |
+
url = eval(s) if s.isnumeric() else s
|
285 |
+
if 'youtube.com/' in str(url) or 'youtu.be/' in str(url): # if source is YouTube video
|
286 |
+
check_requirements(('pafy', 'youtube_dl'))
|
287 |
+
import pafy
|
288 |
+
url = pafy.new(url).getbest(preftype="mp4").url
|
289 |
+
cap = cv2.VideoCapture(url)
|
290 |
+
assert cap.isOpened(), f'Failed to open {s}'
|
291 |
+
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
292 |
+
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
293 |
+
self.fps = cap.get(cv2.CAP_PROP_FPS) % 100
|
294 |
+
|
295 |
+
_, self.imgs[i] = cap.read() # guarantee first frame
|
296 |
+
thread = Thread(target=self.update, args=([i, cap]), daemon=True)
|
297 |
+
print(f' success ({w}x{h} at {self.fps:.2f} FPS).')
|
298 |
+
thread.start()
|
299 |
+
print('') # newline
|
300 |
+
|
301 |
+
# check for common shapes
|
302 |
+
s = np.stack([letterbox(x, self.img_size, stride=self.stride)[0].shape for x in self.imgs], 0) # shapes
|
303 |
+
self.rect = np.unique(s, axis=0).shape[0] == 1 # rect inference if all shapes equal
|
304 |
+
if not self.rect:
|
305 |
+
print('WARNING: Different stream shapes detected. For optimal performance supply similarly-shaped streams.')
|
306 |
+
|
307 |
+
def update(self, index, cap):
|
308 |
+
# Read next stream frame in a daemon thread
|
309 |
+
n = 0
|
310 |
+
while cap.isOpened():
|
311 |
+
n += 1
|
312 |
+
# _, self.imgs[index] = cap.read()
|
313 |
+
cap.grab()
|
314 |
+
if n == 4: # read every 4th frame
|
315 |
+
success, im = cap.retrieve()
|
316 |
+
self.imgs[index] = im if success else self.imgs[index] * 0
|
317 |
+
n = 0
|
318 |
+
time.sleep(1 / self.fps) # wait time
|
319 |
+
|
320 |
+
def __iter__(self):
|
321 |
+
self.count = -1
|
322 |
+
return self
|
323 |
+
|
324 |
+
def __next__(self):
|
325 |
+
self.count += 1
|
326 |
+
img0 = self.imgs.copy()
|
327 |
+
if cv2.waitKey(1) == ord('q'): # q to quit
|
328 |
+
cv2.destroyAllWindows()
|
329 |
+
raise StopIteration
|
330 |
+
|
331 |
+
# Letterbox
|
332 |
+
img = [letterbox(x, self.img_size, auto=self.rect, stride=self.stride)[0] for x in img0]
|
333 |
+
|
334 |
+
# Stack
|
335 |
+
img = np.stack(img, 0)
|
336 |
+
|
337 |
+
# Convert
|
338 |
+
img = img[:, :, :, ::-1].transpose(0, 3, 1, 2) # BGR to RGB, to bsx3x416x416
|
339 |
+
img = np.ascontiguousarray(img)
|
340 |
+
|
341 |
+
return self.sources, img, img0, None
|
342 |
+
|
343 |
+
def __len__(self):
|
344 |
+
return 0 # 1E12 frames = 32 streams at 30 FPS for 30 years
|
345 |
+
|
346 |
+
|
347 |
+
def img2label_paths(img_paths):
|
348 |
+
# Define label paths as a function of image paths
|
349 |
+
sa, sb = os.sep + 'images' + os.sep, os.sep + 'labels' + os.sep # /images/, /labels/ substrings
|
350 |
+
return ['txt'.join(x.replace(sa, sb, 1).rsplit(x.split('.')[-1], 1)) for x in img_paths]
|
351 |
+
|
352 |
+
|
353 |
+
class LoadImagesAndLabels(Dataset): # for training/testing
|
354 |
+
def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
|
355 |
+
cache_images=False, single_cls=False, stride=32, pad=0.0, prefix=''):
|
356 |
+
self.img_size = img_size
|
357 |
+
self.augment = augment
|
358 |
+
self.hyp = hyp
|
359 |
+
self.image_weights = image_weights
|
360 |
+
self.rect = False if image_weights else rect
|
361 |
+
self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training)
|
362 |
+
self.mosaic_border = [-img_size // 2, -img_size // 2]
|
363 |
+
self.stride = stride
|
364 |
+
self.path = path
|
365 |
+
#self.albumentations = Albumentations() if augment else None
|
366 |
+
|
367 |
+
try:
|
368 |
+
f = [] # image files
|
369 |
+
for p in path if isinstance(path, list) else [path]:
|
370 |
+
p = Path(p) # os-agnostic
|
371 |
+
if p.is_dir(): # dir
|
372 |
+
f += glob.glob(str(p / '**' / '*.*'), recursive=True)
|
373 |
+
# f = list(p.rglob('**/*.*')) # pathlib
|
374 |
+
elif p.is_file(): # file
|
375 |
+
with open(p, 'r') as t:
|
376 |
+
t = t.read().strip().splitlines()
|
377 |
+
parent = str(p.parent) + os.sep
|
378 |
+
f += [x.replace('./', parent) if x.startswith('./') else x for x in t] # local to global path
|
379 |
+
# f += [p.parent / x.lstrip(os.sep) for x in t] # local to global path (pathlib)
|
380 |
+
else:
|
381 |
+
raise Exception(f'{prefix}{p} does not exist')
|
382 |
+
self.img_files = sorted([x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in img_formats])
|
383 |
+
# self.img_files = sorted([x for x in f if x.suffix[1:].lower() in img_formats]) # pathlib
|
384 |
+
assert self.img_files, f'{prefix}No images found'
|
385 |
+
except Exception as e:
|
386 |
+
raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {help_url}')
|
387 |
+
|
388 |
+
# Check cache
|
389 |
+
self.label_files = img2label_paths(self.img_files) # labels
|
390 |
+
cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache') # cached labels
|
391 |
+
if cache_path.is_file():
|
392 |
+
cache, exists = torch.load(cache_path), True # load
|
393 |
+
#if cache['hash'] != get_hash(self.label_files + self.img_files) or 'version' not in cache: # changed
|
394 |
+
# cache, exists = self.cache_labels(cache_path, prefix), False # re-cache
|
395 |
+
else:
|
396 |
+
cache, exists = self.cache_labels(cache_path, prefix), False # cache
|
397 |
+
|
398 |
+
# Display cache
|
399 |
+
nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupted, total
|
400 |
+
if exists:
|
401 |
+
d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupted"
|
402 |
+
tqdm(None, desc=prefix + d, total=n, initial=n) # display cache results
|
403 |
+
assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {help_url}'
|
404 |
+
|
405 |
+
# Read cache
|
406 |
+
cache.pop('hash') # remove hash
|
407 |
+
cache.pop('version') # remove version
|
408 |
+
labels, shapes, self.segments = zip(*cache.values())
|
409 |
+
self.labels = list(labels)
|
410 |
+
self.shapes = np.array(shapes, dtype=np.float64)
|
411 |
+
self.img_files = list(cache.keys()) # update
|
412 |
+
self.label_files = img2label_paths(cache.keys()) # update
|
413 |
+
if single_cls:
|
414 |
+
for x in self.labels:
|
415 |
+
x[:, 0] = 0
|
416 |
+
|
417 |
+
n = len(shapes) # number of images
|
418 |
+
bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index
|
419 |
+
nb = bi[-1] + 1 # number of batches
|
420 |
+
self.batch = bi # batch index of image
|
421 |
+
self.n = n
|
422 |
+
self.indices = range(n)
|
423 |
+
|
424 |
+
# Rectangular Training
|
425 |
+
if self.rect:
|
426 |
+
# Sort by aspect ratio
|
427 |
+
s = self.shapes # wh
|
428 |
+
ar = s[:, 1] / s[:, 0] # aspect ratio
|
429 |
+
irect = ar.argsort()
|
430 |
+
self.img_files = [self.img_files[i] for i in irect]
|
431 |
+
self.label_files = [self.label_files[i] for i in irect]
|
432 |
+
self.labels = [self.labels[i] for i in irect]
|
433 |
+
self.shapes = s[irect] # wh
|
434 |
+
ar = ar[irect]
|
435 |
+
|
436 |
+
# Set training image shapes
|
437 |
+
shapes = [[1, 1]] * nb
|
438 |
+
for i in range(nb):
|
439 |
+
ari = ar[bi == i]
|
440 |
+
mini, maxi = ari.min(), ari.max()
|
441 |
+
if maxi < 1:
|
442 |
+
shapes[i] = [maxi, 1]
|
443 |
+
elif mini > 1:
|
444 |
+
shapes[i] = [1, 1 / mini]
|
445 |
+
|
446 |
+
self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride
|
447 |
+
|
448 |
+
# Cache images into memory for faster training (WARNING: large datasets may exceed system RAM)
|
449 |
+
self.imgs = [None] * n
|
450 |
+
if cache_images:
|
451 |
+
if cache_images == 'disk':
|
452 |
+
self.im_cache_dir = Path(Path(self.img_files[0]).parent.as_posix() + '_npy')
|
453 |
+
self.img_npy = [self.im_cache_dir / Path(f).with_suffix('.npy').name for f in self.img_files]
|
454 |
+
self.im_cache_dir.mkdir(parents=True, exist_ok=True)
|
455 |
+
gb = 0 # Gigabytes of cached images
|
456 |
+
self.img_hw0, self.img_hw = [None] * n, [None] * n
|
457 |
+
results = ThreadPool(8).imap(lambda x: load_image(*x), zip(repeat(self), range(n)))
|
458 |
+
pbar = tqdm(enumerate(results), total=n)
|
459 |
+
for i, x in pbar:
|
460 |
+
if cache_images == 'disk':
|
461 |
+
if not self.img_npy[i].exists():
|
462 |
+
np.save(self.img_npy[i].as_posix(), x[0])
|
463 |
+
gb += self.img_npy[i].stat().st_size
|
464 |
+
else:
|
465 |
+
self.imgs[i], self.img_hw0[i], self.img_hw[i] = x
|
466 |
+
gb += self.imgs[i].nbytes
|
467 |
+
pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB)'
|
468 |
+
pbar.close()
|
469 |
+
|
470 |
+
def cache_labels(self, path=Path('./labels.cache'), prefix=''):
|
471 |
+
# Cache dataset labels, check images and read shapes
|
472 |
+
x = {} # dict
|
473 |
+
nm, nf, ne, nc = 0, 0, 0, 0 # number missing, found, empty, duplicate
|
474 |
+
pbar = tqdm(zip(self.img_files, self.label_files), desc='Scanning images', total=len(self.img_files))
|
475 |
+
for i, (im_file, lb_file) in enumerate(pbar):
|
476 |
+
try:
|
477 |
+
# verify images
|
478 |
+
im = Image.open(im_file)
|
479 |
+
im.verify() # PIL verify
|
480 |
+
shape = exif_size(im) # image size
|
481 |
+
segments = [] # instance segments
|
482 |
+
assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels'
|
483 |
+
assert im.format.lower() in img_formats, f'invalid image format {im.format}'
|
484 |
+
|
485 |
+
# verify labels
|
486 |
+
if os.path.isfile(lb_file):
|
487 |
+
nf += 1 # label found
|
488 |
+
with open(lb_file, 'r') as f:
|
489 |
+
l = [x.split() for x in f.read().strip().splitlines()]
|
490 |
+
if any([len(x) > 8 for x in l]): # is segment
|
491 |
+
classes = np.array([x[0] for x in l], dtype=np.float32)
|
492 |
+
segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in l] # (cls, xy1...)
|
493 |
+
l = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1) # (cls, xywh)
|
494 |
+
l = np.array(l, dtype=np.float32)
|
495 |
+
if len(l):
|
496 |
+
assert l.shape[1] == 5, 'labels require 5 columns each'
|
497 |
+
assert (l >= 0).all(), 'negative labels'
|
498 |
+
assert (l[:, 1:] <= 1).all(), 'non-normalized or out of bounds coordinate labels'
|
499 |
+
assert np.unique(l, axis=0).shape[0] == l.shape[0], 'duplicate labels'
|
500 |
+
else:
|
501 |
+
ne += 1 # label empty
|
502 |
+
l = np.zeros((0, 5), dtype=np.float32)
|
503 |
+
else:
|
504 |
+
nm += 1 # label missing
|
505 |
+
l = np.zeros((0, 5), dtype=np.float32)
|
506 |
+
x[im_file] = [l, shape, segments]
|
507 |
+
except Exception as e:
|
508 |
+
nc += 1
|
509 |
+
print(f'{prefix}WARNING: Ignoring corrupted image and/or label {im_file}: {e}')
|
510 |
+
|
511 |
+
pbar.desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels... " \
|
512 |
+
f"{nf} found, {nm} missing, {ne} empty, {nc} corrupted"
|
513 |
+
pbar.close()
|
514 |
+
|
515 |
+
if nf == 0:
|
516 |
+
print(f'{prefix}WARNING: No labels found in {path}. See {help_url}')
|
517 |
+
|
518 |
+
x['hash'] = get_hash(self.label_files + self.img_files)
|
519 |
+
x['results'] = nf, nm, ne, nc, i + 1
|
520 |
+
x['version'] = 0.1 # cache version
|
521 |
+
torch.save(x, path) # save for next time
|
522 |
+
logging.info(f'{prefix}New cache created: {path}')
|
523 |
+
return x
|
524 |
+
|
525 |
+
def __len__(self):
|
526 |
+
return len(self.img_files)
|
527 |
+
|
528 |
+
# def __iter__(self):
|
529 |
+
# self.count = -1
|
530 |
+
# print('ran dataset iter')
|
531 |
+
# #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
|
532 |
+
# return self
|
533 |
+
|
534 |
+
def __getitem__(self, index):
|
535 |
+
index = self.indices[index] # linear, shuffled, or image_weights
|
536 |
+
|
537 |
+
hyp = self.hyp
|
538 |
+
mosaic = self.mosaic and random.random() < hyp['mosaic']
|
539 |
+
if mosaic:
|
540 |
+
# Load mosaic
|
541 |
+
if random.random() < 0.8:
|
542 |
+
img, labels = load_mosaic(self, index)
|
543 |
+
else:
|
544 |
+
img, labels = load_mosaic9(self, index)
|
545 |
+
shapes = None
|
546 |
+
|
547 |
+
# MixUp https://arxiv.org/pdf/1710.09412.pdf
|
548 |
+
if random.random() < hyp['mixup']:
|
549 |
+
if random.random() < 0.8:
|
550 |
+
img2, labels2 = load_mosaic(self, random.randint(0, len(self.labels) - 1))
|
551 |
+
else:
|
552 |
+
img2, labels2 = load_mosaic9(self, random.randint(0, len(self.labels) - 1))
|
553 |
+
r = np.random.beta(8.0, 8.0) # mixup ratio, alpha=beta=8.0
|
554 |
+
img = (img * r + img2 * (1 - r)).astype(np.uint8)
|
555 |
+
labels = np.concatenate((labels, labels2), 0)
|
556 |
+
|
557 |
+
else:
|
558 |
+
# Load image
|
559 |
+
img, (h0, w0), (h, w) = load_image(self, index)
|
560 |
+
|
561 |
+
# Letterbox
|
562 |
+
shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape
|
563 |
+
img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
|
564 |
+
shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling
|
565 |
+
|
566 |
+
labels = self.labels[index].copy()
|
567 |
+
if labels.size: # normalized xywh to pixel xyxy format
|
568 |
+
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
|
569 |
+
|
570 |
+
if self.augment:
|
571 |
+
# Augment imagespace
|
572 |
+
if not mosaic:
|
573 |
+
img, labels = random_perspective(img, labels,
|
574 |
+
degrees=hyp['degrees'],
|
575 |
+
translate=hyp['translate'],
|
576 |
+
scale=hyp['scale'],
|
577 |
+
shear=hyp['shear'],
|
578 |
+
perspective=hyp['perspective'])
|
579 |
+
|
580 |
+
|
581 |
+
#img, labels = self.albumentations(img, labels)
|
582 |
+
|
583 |
+
# Augment colorspace
|
584 |
+
augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])
|
585 |
+
|
586 |
+
# Apply cutouts
|
587 |
+
# if random.random() < 0.9:
|
588 |
+
# labels = cutout(img, labels)
|
589 |
+
|
590 |
+
if random.random() < hyp['paste_in']:
|
591 |
+
sample_labels, sample_images, sample_masks = [], [], []
|
592 |
+
while len(sample_labels) < 30:
|
593 |
+
sample_labels_, sample_images_, sample_masks_ = load_samples(self, random.randint(0, len(self.labels) - 1))
|
594 |
+
sample_labels += sample_labels_
|
595 |
+
sample_images += sample_images_
|
596 |
+
sample_masks += sample_masks_
|
597 |
+
#print(len(sample_labels))
|
598 |
+
if len(sample_labels) == 0:
|
599 |
+
break
|
600 |
+
labels = pastein(img, labels, sample_labels, sample_images, sample_masks)
|
601 |
+
|
602 |
+
nL = len(labels) # number of labels
|
603 |
+
if nL:
|
604 |
+
labels[:, 1:5] = xyxy2xywh(labels[:, 1:5]) # convert xyxy to xywh
|
605 |
+
labels[:, [2, 4]] /= img.shape[0] # normalized height 0-1
|
606 |
+
labels[:, [1, 3]] /= img.shape[1] # normalized width 0-1
|
607 |
+
|
608 |
+
if self.augment:
|
609 |
+
# flip up-down
|
610 |
+
if random.random() < hyp['flipud']:
|
611 |
+
img = np.flipud(img)
|
612 |
+
if nL:
|
613 |
+
labels[:, 2] = 1 - labels[:, 2]
|
614 |
+
|
615 |
+
# flip left-right
|
616 |
+
if random.random() < hyp['fliplr']:
|
617 |
+
img = np.fliplr(img)
|
618 |
+
if nL:
|
619 |
+
labels[:, 1] = 1 - labels[:, 1]
|
620 |
+
|
621 |
+
labels_out = torch.zeros((nL, 6))
|
622 |
+
if nL:
|
623 |
+
labels_out[:, 1:] = torch.from_numpy(labels)
|
624 |
+
|
625 |
+
# Convert
|
626 |
+
img = img[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
|
627 |
+
img = np.ascontiguousarray(img)
|
628 |
+
|
629 |
+
return torch.from_numpy(img), labels_out, self.img_files[index], shapes
|
630 |
+
|
631 |
+
@staticmethod
|
632 |
+
def collate_fn(batch):
|
633 |
+
img, label, path, shapes = zip(*batch) # transposed
|
634 |
+
for i, l in enumerate(label):
|
635 |
+
l[:, 0] = i # add target image index for build_targets()
|
636 |
+
return torch.stack(img, 0), torch.cat(label, 0), path, shapes
|
637 |
+
|
638 |
+
@staticmethod
|
639 |
+
def collate_fn4(batch):
|
640 |
+
img, label, path, shapes = zip(*batch) # transposed
|
641 |
+
n = len(shapes) // 4
|
642 |
+
img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n]
|
643 |
+
|
644 |
+
ho = torch.tensor([[0., 0, 0, 1, 0, 0]])
|
645 |
+
wo = torch.tensor([[0., 0, 1, 0, 0, 0]])
|
646 |
+
s = torch.tensor([[1, 1, .5, .5, .5, .5]]) # scale
|
647 |
+
for i in range(n): # zidane torch.zeros(16,3,720,1280) # BCHW
|
648 |
+
i *= 4
|
649 |
+
if random.random() < 0.5:
|
650 |
+
im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2., mode='bilinear', align_corners=False)[
|
651 |
+
0].type(img[i].type())
|
652 |
+
l = label[i]
|
653 |
+
else:
|
654 |
+
im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2)
|
655 |
+
l = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s
|
656 |
+
img4.append(im)
|
657 |
+
label4.append(l)
|
658 |
+
|
659 |
+
for i, l in enumerate(label4):
|
660 |
+
l[:, 0] = i # add target image index for build_targets()
|
661 |
+
|
662 |
+
return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4
|
663 |
+
|
664 |
+
|
665 |
+
# Ancillary functions --------------------------------------------------------------------------------------------------
|
666 |
+
def load_image(self, index):
|
667 |
+
# loads 1 image from dataset, returns img, original hw, resized hw
|
668 |
+
img = self.imgs[index]
|
669 |
+
if img is None: # not cached
|
670 |
+
path = self.img_files[index]
|
671 |
+
img = cv2.imread(path) # BGR
|
672 |
+
assert img is not None, 'Image Not Found ' + path
|
673 |
+
h0, w0 = img.shape[:2] # orig hw
|
674 |
+
r = self.img_size / max(h0, w0) # resize image to img_size
|
675 |
+
if r != 1: # always resize down, only resize up if training with augmentation
|
676 |
+
interp = cv2.INTER_AREA if r < 1 and not self.augment else cv2.INTER_LINEAR
|
677 |
+
img = cv2.resize(img, (int(w0 * r), int(h0 * r)), interpolation=interp)
|
678 |
+
return img, (h0, w0), img.shape[:2] # img, hw_original, hw_resized
|
679 |
+
else:
|
680 |
+
return self.imgs[index], self.img_hw0[index], self.img_hw[index] # img, hw_original, hw_resized
|
681 |
+
|
682 |
+
|
683 |
+
def augment_hsv(img, hgain=0.5, sgain=0.5, vgain=0.5):
|
684 |
+
r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains
|
685 |
+
hue, sat, val = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV))
|
686 |
+
dtype = img.dtype # uint8
|
687 |
+
|
688 |
+
x = np.arange(0, 256, dtype=np.int16)
|
689 |
+
lut_hue = ((x * r[0]) % 180).astype(dtype)
|
690 |
+
lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
|
691 |
+
lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
|
692 |
+
|
693 |
+
img_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))).astype(dtype)
|
694 |
+
cv2.cvtColor(img_hsv, cv2.COLOR_HSV2BGR, dst=img) # no return needed
|
695 |
+
|
696 |
+
|
697 |
+
def hist_equalize(img, clahe=True, bgr=False):
|
698 |
+
# Equalize histogram on BGR image 'img' with img.shape(n,m,3) and range 0-255
|
699 |
+
yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)
|
700 |
+
if clahe:
|
701 |
+
c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
|
702 |
+
yuv[:, :, 0] = c.apply(yuv[:, :, 0])
|
703 |
+
else:
|
704 |
+
yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram
|
705 |
+
return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB
|
706 |
+
|
707 |
+
|
708 |
+
def load_mosaic(self, index):
|
709 |
+
# loads images in a 4-mosaic
|
710 |
+
|
711 |
+
labels4, segments4 = [], []
|
712 |
+
s = self.img_size
|
713 |
+
yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border] # mosaic center x, y
|
714 |
+
indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices
|
715 |
+
for i, index in enumerate(indices):
|
716 |
+
# Load image
|
717 |
+
img, _, (h, w) = load_image(self, index)
|
718 |
+
|
719 |
+
# place img in img4
|
720 |
+
if i == 0: # top left
|
721 |
+
img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
|
722 |
+
x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image)
|
723 |
+
x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image)
|
724 |
+
elif i == 1: # top right
|
725 |
+
x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
|
726 |
+
x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
|
727 |
+
elif i == 2: # bottom left
|
728 |
+
x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
|
729 |
+
x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
|
730 |
+
elif i == 3: # bottom right
|
731 |
+
x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
|
732 |
+
x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
|
733 |
+
|
734 |
+
img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
|
735 |
+
padw = x1a - x1b
|
736 |
+
padh = y1a - y1b
|
737 |
+
|
738 |
+
# Labels
|
739 |
+
labels, segments = self.labels[index].copy(), self.segments[index].copy()
|
740 |
+
if labels.size:
|
741 |
+
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format
|
742 |
+
segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
|
743 |
+
labels4.append(labels)
|
744 |
+
segments4.extend(segments)
|
745 |
+
|
746 |
+
# Concat/clip labels
|
747 |
+
labels4 = np.concatenate(labels4, 0)
|
748 |
+
for x in (labels4[:, 1:], *segments4):
|
749 |
+
np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
|
750 |
+
# img4, labels4 = replicate(img4, labels4) # replicate
|
751 |
+
|
752 |
+
# Augment
|
753 |
+
#img4, labels4, segments4 = remove_background(img4, labels4, segments4)
|
754 |
+
#sample_segments(img4, labels4, segments4, probability=self.hyp['copy_paste'])
|
755 |
+
img4, labels4, segments4 = copy_paste(img4, labels4, segments4, probability=self.hyp['copy_paste'])
|
756 |
+
img4, labels4 = random_perspective(img4, labels4, segments4,
|
757 |
+
degrees=self.hyp['degrees'],
|
758 |
+
translate=self.hyp['translate'],
|
759 |
+
scale=self.hyp['scale'],
|
760 |
+
shear=self.hyp['shear'],
|
761 |
+
perspective=self.hyp['perspective'],
|
762 |
+
border=self.mosaic_border) # border to remove
|
763 |
+
|
764 |
+
return img4, labels4
|
765 |
+
|
766 |
+
|
767 |
+
def load_mosaic9(self, index):
|
768 |
+
# loads images in a 9-mosaic
|
769 |
+
|
770 |
+
labels9, segments9 = [], []
|
771 |
+
s = self.img_size
|
772 |
+
indices = [index] + random.choices(self.indices, k=8) # 8 additional image indices
|
773 |
+
for i, index in enumerate(indices):
|
774 |
+
# Load image
|
775 |
+
img, _, (h, w) = load_image(self, index)
|
776 |
+
|
777 |
+
# place img in img9
|
778 |
+
if i == 0: # center
|
779 |
+
img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
|
780 |
+
h0, w0 = h, w
|
781 |
+
c = s, s, s + w, s + h # xmin, ymin, xmax, ymax (base) coordinates
|
782 |
+
elif i == 1: # top
|
783 |
+
c = s, s - h, s + w, s
|
784 |
+
elif i == 2: # top right
|
785 |
+
c = s + wp, s - h, s + wp + w, s
|
786 |
+
elif i == 3: # right
|
787 |
+
c = s + w0, s, s + w0 + w, s + h
|
788 |
+
elif i == 4: # bottom right
|
789 |
+
c = s + w0, s + hp, s + w0 + w, s + hp + h
|
790 |
+
elif i == 5: # bottom
|
791 |
+
c = s + w0 - w, s + h0, s + w0, s + h0 + h
|
792 |
+
elif i == 6: # bottom left
|
793 |
+
c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
|
794 |
+
elif i == 7: # left
|
795 |
+
c = s - w, s + h0 - h, s, s + h0
|
796 |
+
elif i == 8: # top left
|
797 |
+
c = s - w, s + h0 - hp - h, s, s + h0 - hp
|
798 |
+
|
799 |
+
padx, pady = c[:2]
|
800 |
+
x1, y1, x2, y2 = [max(x, 0) for x in c] # allocate coords
|
801 |
+
|
802 |
+
# Labels
|
803 |
+
labels, segments = self.labels[index].copy(), self.segments[index].copy()
|
804 |
+
if labels.size:
|
805 |
+
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady) # normalized xywh to pixel xyxy format
|
806 |
+
segments = [xyn2xy(x, w, h, padx, pady) for x in segments]
|
807 |
+
labels9.append(labels)
|
808 |
+
segments9.extend(segments)
|
809 |
+
|
810 |
+
# Image
|
811 |
+
img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:] # img9[ymin:ymax, xmin:xmax]
|
812 |
+
hp, wp = h, w # height, width previous
|
813 |
+
|
814 |
+
# Offset
|
815 |
+
yc, xc = [int(random.uniform(0, s)) for _ in self.mosaic_border] # mosaic center x, y
|
816 |
+
img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s]
|
817 |
+
|
818 |
+
# Concat/clip labels
|
819 |
+
labels9 = np.concatenate(labels9, 0)
|
820 |
+
labels9[:, [1, 3]] -= xc
|
821 |
+
labels9[:, [2, 4]] -= yc
|
822 |
+
c = np.array([xc, yc]) # centers
|
823 |
+
segments9 = [x - c for x in segments9]
|
824 |
+
|
825 |
+
for x in (labels9[:, 1:], *segments9):
|
826 |
+
np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
|
827 |
+
# img9, labels9 = replicate(img9, labels9) # replicate
|
828 |
+
|
829 |
+
# Augment
|
830 |
+
#img9, labels9, segments9 = remove_background(img9, labels9, segments9)
|
831 |
+
img9, labels9, segments9 = copy_paste(img9, labels9, segments9, probability=self.hyp['copy_paste'])
|
832 |
+
img9, labels9 = random_perspective(img9, labels9, segments9,
|
833 |
+
degrees=self.hyp['degrees'],
|
834 |
+
translate=self.hyp['translate'],
|
835 |
+
scale=self.hyp['scale'],
|
836 |
+
shear=self.hyp['shear'],
|
837 |
+
perspective=self.hyp['perspective'],
|
838 |
+
border=self.mosaic_border) # border to remove
|
839 |
+
|
840 |
+
return img9, labels9
|
841 |
+
|
842 |
+
|
843 |
+
def load_samples(self, index):
|
844 |
+
# loads images in a 4-mosaic
|
845 |
+
|
846 |
+
labels4, segments4 = [], []
|
847 |
+
s = self.img_size
|
848 |
+
yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border] # mosaic center x, y
|
849 |
+
indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices
|
850 |
+
for i, index in enumerate(indices):
|
851 |
+
# Load image
|
852 |
+
img, _, (h, w) = load_image(self, index)
|
853 |
+
|
854 |
+
# place img in img4
|
855 |
+
if i == 0: # top left
|
856 |
+
img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles
|
857 |
+
x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image)
|
858 |
+
x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image)
|
859 |
+
elif i == 1: # top right
|
860 |
+
x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
|
861 |
+
x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
|
862 |
+
elif i == 2: # bottom left
|
863 |
+
x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
|
864 |
+
x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
|
865 |
+
elif i == 3: # bottom right
|
866 |
+
x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
|
867 |
+
x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
|
868 |
+
|
869 |
+
img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
|
870 |
+
padw = x1a - x1b
|
871 |
+
padh = y1a - y1b
|
872 |
+
|
873 |
+
# Labels
|
874 |
+
labels, segments = self.labels[index].copy(), self.segments[index].copy()
|
875 |
+
if labels.size:
|
876 |
+
labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format
|
877 |
+
segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
|
878 |
+
labels4.append(labels)
|
879 |
+
segments4.extend(segments)
|
880 |
+
|
881 |
+
# Concat/clip labels
|
882 |
+
labels4 = np.concatenate(labels4, 0)
|
883 |
+
for x in (labels4[:, 1:], *segments4):
|
884 |
+
np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective()
|
885 |
+
# img4, labels4 = replicate(img4, labels4) # replicate
|
886 |
+
|
887 |
+
# Augment
|
888 |
+
#img4, labels4, segments4 = remove_background(img4, labels4, segments4)
|
889 |
+
sample_labels, sample_images, sample_masks = sample_segments(img4, labels4, segments4, probability=0.5)
|
890 |
+
|
891 |
+
return sample_labels, sample_images, sample_masks
|
892 |
+
|
893 |
+
|
894 |
+
def copy_paste(img, labels, segments, probability=0.5):
|
895 |
+
# Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
|
896 |
+
n = len(segments)
|
897 |
+
if probability and n:
|
898 |
+
h, w, c = img.shape # height, width, channels
|
899 |
+
im_new = np.zeros(img.shape, np.uint8)
|
900 |
+
for j in random.sample(range(n), k=round(probability * n)):
|
901 |
+
l, s = labels[j], segments[j]
|
902 |
+
box = w - l[3], l[2], w - l[1], l[4]
|
903 |
+
ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area
|
904 |
+
if (ioa < 0.30).all(): # allow 30% obscuration of existing labels
|
905 |
+
labels = np.concatenate((labels, [[l[0], *box]]), 0)
|
906 |
+
segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
|
907 |
+
cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
|
908 |
+
|
909 |
+
result = cv2.bitwise_and(src1=img, src2=im_new)
|
910 |
+
result = cv2.flip(result, 1) # augment segments (flip left-right)
|
911 |
+
i = result > 0 # pixels to replace
|
912 |
+
# i[:, :] = result.max(2).reshape(h, w, 1) # act over ch
|
913 |
+
img[i] = result[i] # cv2.imwrite('debug.jpg', img) # debug
|
914 |
+
|
915 |
+
return img, labels, segments
|
916 |
+
|
917 |
+
|
918 |
+
def remove_background(img, labels, segments):
|
919 |
+
# Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
|
920 |
+
n = len(segments)
|
921 |
+
h, w, c = img.shape # height, width, channels
|
922 |
+
im_new = np.zeros(img.shape, np.uint8)
|
923 |
+
img_new = np.ones(img.shape, np.uint8) * 114
|
924 |
+
for j in range(n):
|
925 |
+
cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
|
926 |
+
|
927 |
+
result = cv2.bitwise_and(src1=img, src2=im_new)
|
928 |
+
|
929 |
+
i = result > 0 # pixels to replace
|
930 |
+
img_new[i] = result[i] # cv2.imwrite('debug.jpg', img) # debug
|
931 |
+
|
932 |
+
return img_new, labels, segments
|
933 |
+
|
934 |
+
|
935 |
+
def sample_segments(img, labels, segments, probability=0.5):
|
936 |
+
# Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
|
937 |
+
n = len(segments)
|
938 |
+
sample_labels = []
|
939 |
+
sample_images = []
|
940 |
+
sample_masks = []
|
941 |
+
if probability and n:
|
942 |
+
h, w, c = img.shape # height, width, channels
|
943 |
+
for j in random.sample(range(n), k=round(probability * n)):
|
944 |
+
l, s = labels[j], segments[j]
|
945 |
+
box = l[1].astype(int).clip(0,w-1), l[2].astype(int).clip(0,h-1), l[3].astype(int).clip(0,w-1), l[4].astype(int).clip(0,h-1)
|
946 |
+
|
947 |
+
#print(box)
|
948 |
+
if (box[2] <= box[0]) or (box[3] <= box[1]):
|
949 |
+
continue
|
950 |
+
|
951 |
+
sample_labels.append(l[0])
|
952 |
+
|
953 |
+
mask = np.zeros(img.shape, np.uint8)
|
954 |
+
|
955 |
+
cv2.drawContours(mask, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)
|
956 |
+
sample_masks.append(mask[box[1]:box[3],box[0]:box[2],:])
|
957 |
+
|
958 |
+
result = cv2.bitwise_and(src1=img, src2=mask)
|
959 |
+
i = result > 0 # pixels to replace
|
960 |
+
mask[i] = result[i] # cv2.imwrite('debug.jpg', img) # debug
|
961 |
+
#print(box)
|
962 |
+
sample_images.append(mask[box[1]:box[3],box[0]:box[2],:])
|
963 |
+
|
964 |
+
return sample_labels, sample_images, sample_masks
|
965 |
+
|
966 |
+
|
967 |
+
def replicate(img, labels):
|
968 |
+
# Replicate labels
|
969 |
+
h, w = img.shape[:2]
|
970 |
+
boxes = labels[:, 1:].astype(int)
|
971 |
+
x1, y1, x2, y2 = boxes.T
|
972 |
+
s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels)
|
973 |
+
for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices
|
974 |
+
x1b, y1b, x2b, y2b = boxes[i]
|
975 |
+
bh, bw = y2b - y1b, x2b - x1b
|
976 |
+
yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y
|
977 |
+
x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]
|
978 |
+
img[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax]
|
979 |
+
labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)
|
980 |
+
|
981 |
+
return img, labels
|
982 |
+
|
983 |
+
|
984 |
+
def letterbox(img, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
|
985 |
+
# Resize and pad image while meeting stride-multiple constraints
|
986 |
+
shape = img.shape[:2] # current shape [height, width]
|
987 |
+
if isinstance(new_shape, int):
|
988 |
+
new_shape = (new_shape, new_shape)
|
989 |
+
|
990 |
+
# Scale ratio (new / old)
|
991 |
+
r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
|
992 |
+
if not scaleup: # only scale down, do not scale up (for better test mAP)
|
993 |
+
r = min(r, 1.0)
|
994 |
+
|
995 |
+
# Compute padding
|
996 |
+
ratio = r, r # width, height ratios
|
997 |
+
new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
|
998 |
+
dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding
|
999 |
+
if auto: # minimum rectangle
|
1000 |
+
dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding
|
1001 |
+
elif scaleFill: # stretch
|
1002 |
+
dw, dh = 0.0, 0.0
|
1003 |
+
new_unpad = (new_shape[1], new_shape[0])
|
1004 |
+
ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios
|
1005 |
+
|
1006 |
+
dw /= 2 # divide padding into 2 sides
|
1007 |
+
dh /= 2
|
1008 |
+
|
1009 |
+
if shape[::-1] != new_unpad: # resize
|
1010 |
+
img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
|
1011 |
+
top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
|
1012 |
+
left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
|
1013 |
+
img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border
|
1014 |
+
return img, ratio, (dw, dh)
|
1015 |
+
|
1016 |
+
|
1017 |
+
def random_perspective(img, targets=(), segments=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0,
|
1018 |
+
border=(0, 0)):
|
1019 |
+
# torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
|
1020 |
+
# targets = [cls, xyxy]
|
1021 |
+
|
1022 |
+
height = img.shape[0] + border[0] * 2 # shape(h,w,c)
|
1023 |
+
width = img.shape[1] + border[1] * 2
|
1024 |
+
|
1025 |
+
# Center
|
1026 |
+
C = np.eye(3)
|
1027 |
+
C[0, 2] = -img.shape[1] / 2 # x translation (pixels)
|
1028 |
+
C[1, 2] = -img.shape[0] / 2 # y translation (pixels)
|
1029 |
+
|
1030 |
+
# Perspective
|
1031 |
+
P = np.eye(3)
|
1032 |
+
P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y)
|
1033 |
+
P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x)
|
1034 |
+
|
1035 |
+
# Rotation and Scale
|
1036 |
+
R = np.eye(3)
|
1037 |
+
a = random.uniform(-degrees, degrees)
|
1038 |
+
# a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations
|
1039 |
+
s = random.uniform(1 - scale, 1.1 + scale)
|
1040 |
+
# s = 2 ** random.uniform(-scale, scale)
|
1041 |
+
R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
|
1042 |
+
|
1043 |
+
# Shear
|
1044 |
+
S = np.eye(3)
|
1045 |
+
S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)
|
1046 |
+
S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)
|
1047 |
+
|
1048 |
+
# Translation
|
1049 |
+
T = np.eye(3)
|
1050 |
+
T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels)
|
1051 |
+
T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels)
|
1052 |
+
|
1053 |
+
# Combined rotation matrix
|
1054 |
+
M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT
|
1055 |
+
if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed
|
1056 |
+
if perspective:
|
1057 |
+
img = cv2.warpPerspective(img, M, dsize=(width, height), borderValue=(114, 114, 114))
|
1058 |
+
else: # affine
|
1059 |
+
img = cv2.warpAffine(img, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
|
1060 |
+
|
1061 |
+
# Visualize
|
1062 |
+
# import matplotlib.pyplot as plt
|
1063 |
+
# ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
|
1064 |
+
# ax[0].imshow(img[:, :, ::-1]) # base
|
1065 |
+
# ax[1].imshow(img2[:, :, ::-1]) # warped
|
1066 |
+
|
1067 |
+
# Transform label coordinates
|
1068 |
+
n = len(targets)
|
1069 |
+
if n:
|
1070 |
+
use_segments = any(x.any() for x in segments)
|
1071 |
+
new = np.zeros((n, 4))
|
1072 |
+
if use_segments: # warp segments
|
1073 |
+
segments = resample_segments(segments) # upsample
|
1074 |
+
for i, segment in enumerate(segments):
|
1075 |
+
xy = np.ones((len(segment), 3))
|
1076 |
+
xy[:, :2] = segment
|
1077 |
+
xy = xy @ M.T # transform
|
1078 |
+
xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine
|
1079 |
+
|
1080 |
+
# clip
|
1081 |
+
new[i] = segment2box(xy, width, height)
|
1082 |
+
|
1083 |
+
else: # warp boxes
|
1084 |
+
xy = np.ones((n * 4, 3))
|
1085 |
+
xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1
|
1086 |
+
xy = xy @ M.T # transform
|
1087 |
+
xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine
|
1088 |
+
|
1089 |
+
# create new boxes
|
1090 |
+
x = xy[:, [0, 2, 4, 6]]
|
1091 |
+
y = xy[:, [1, 3, 5, 7]]
|
1092 |
+
new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
|
1093 |
+
|
1094 |
+
# clip
|
1095 |
+
new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
|
1096 |
+
new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)
|
1097 |
+
|
1098 |
+
# filter candidates
|
1099 |
+
i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
|
1100 |
+
targets = targets[i]
|
1101 |
+
targets[:, 1:5] = new[i]
|
1102 |
+
|
1103 |
+
return img, targets
|
1104 |
+
|
1105 |
+
|
1106 |
+
def box_candidates(box1, box2, wh_thr=2, ar_thr=20, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n)
|
1107 |
+
# Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
|
1108 |
+
w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
|
1109 |
+
w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
|
1110 |
+
ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio
|
1111 |
+
return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates
|
1112 |
+
|
1113 |
+
|
1114 |
+
def bbox_ioa(box1, box2):
|
1115 |
+
# Returns the intersection over box2 area given box1, box2. box1 is 4, box2 is nx4. boxes are x1y1x2y2
|
1116 |
+
box2 = box2.transpose()
|
1117 |
+
|
1118 |
+
# Get the coordinates of bounding boxes
|
1119 |
+
b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
|
1120 |
+
b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
|
1121 |
+
|
1122 |
+
# Intersection area
|
1123 |
+
inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
|
1124 |
+
(np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)
|
1125 |
+
|
1126 |
+
# box2 area
|
1127 |
+
box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + 1e-16
|
1128 |
+
|
1129 |
+
# Intersection over box2 area
|
1130 |
+
return inter_area / box2_area
|
1131 |
+
|
1132 |
+
|
1133 |
+
def cutout(image, labels):
|
1134 |
+
# Applies image cutout augmentation https://arxiv.org/abs/1708.04552
|
1135 |
+
h, w = image.shape[:2]
|
1136 |
+
|
1137 |
+
# create random masks
|
1138 |
+
scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction
|
1139 |
+
for s in scales:
|
1140 |
+
mask_h = random.randint(1, int(h * s))
|
1141 |
+
mask_w = random.randint(1, int(w * s))
|
1142 |
+
|
1143 |
+
# box
|
1144 |
+
xmin = max(0, random.randint(0, w) - mask_w // 2)
|
1145 |
+
ymin = max(0, random.randint(0, h) - mask_h // 2)
|
1146 |
+
xmax = min(w, xmin + mask_w)
|
1147 |
+
ymax = min(h, ymin + mask_h)
|
1148 |
+
|
1149 |
+
# apply random color mask
|
1150 |
+
image[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]
|
1151 |
+
|
1152 |
+
# return unobscured labels
|
1153 |
+
if len(labels) and s > 0.03:
|
1154 |
+
box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
|
1155 |
+
ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area
|
1156 |
+
labels = labels[ioa < 0.60] # remove >60% obscured labels
|
1157 |
+
|
1158 |
+
return labels
|
1159 |
+
|
1160 |
+
|
1161 |
+
def pastein(image, labels, sample_labels, sample_images, sample_masks):
|
1162 |
+
# Applies image cutout augmentation https://arxiv.org/abs/1708.04552
|
1163 |
+
h, w = image.shape[:2]
|
1164 |
+
|
1165 |
+
# create random masks
|
1166 |
+
scales = [0.75] * 2 + [0.5] * 4 + [0.25] * 4 + [0.125] * 4 + [0.0625] * 6 # image size fraction
|
1167 |
+
for s in scales:
|
1168 |
+
if random.random() < 0.2:
|
1169 |
+
continue
|
1170 |
+
mask_h = random.randint(1, int(h * s))
|
1171 |
+
mask_w = random.randint(1, int(w * s))
|
1172 |
+
|
1173 |
+
# box
|
1174 |
+
xmin = max(0, random.randint(0, w) - mask_w // 2)
|
1175 |
+
ymin = max(0, random.randint(0, h) - mask_h // 2)
|
1176 |
+
xmax = min(w, xmin + mask_w)
|
1177 |
+
ymax = min(h, ymin + mask_h)
|
1178 |
+
|
1179 |
+
box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
|
1180 |
+
if len(labels):
|
1181 |
+
ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area
|
1182 |
+
else:
|
1183 |
+
ioa = np.zeros(1)
|
1184 |
+
|
1185 |
+
if (ioa < 0.30).all() and len(sample_labels) and (xmax > xmin+20) and (ymax > ymin+20): # allow 30% obscuration of existing labels
|
1186 |
+
sel_ind = random.randint(0, len(sample_labels)-1)
|
1187 |
+
#print(len(sample_labels))
|
1188 |
+
#print(sel_ind)
|
1189 |
+
#print((xmax-xmin, ymax-ymin))
|
1190 |
+
#print(image[ymin:ymax, xmin:xmax].shape)
|
1191 |
+
#print([[sample_labels[sel_ind], *box]])
|
1192 |
+
#print(labels.shape)
|
1193 |
+
hs, ws, cs = sample_images[sel_ind].shape
|
1194 |
+
r_scale = min((ymax-ymin)/hs, (xmax-xmin)/ws)
|
1195 |
+
r_w = int(ws*r_scale)
|
1196 |
+
r_h = int(hs*r_scale)
|
1197 |
+
|
1198 |
+
if (r_w > 10) and (r_h > 10):
|
1199 |
+
r_mask = cv2.resize(sample_masks[sel_ind], (r_w, r_h))
|
1200 |
+
r_image = cv2.resize(sample_images[sel_ind], (r_w, r_h))
|
1201 |
+
temp_crop = image[ymin:ymin+r_h, xmin:xmin+r_w]
|
1202 |
+
m_ind = r_mask > 0
|
1203 |
+
if m_ind.astype(np.int).sum() > 60:
|
1204 |
+
temp_crop[m_ind] = r_image[m_ind]
|
1205 |
+
#print(sample_labels[sel_ind])
|
1206 |
+
#print(sample_images[sel_ind].shape)
|
1207 |
+
#print(temp_crop.shape)
|
1208 |
+
box = np.array([xmin, ymin, xmin+r_w, ymin+r_h], dtype=np.float32)
|
1209 |
+
if len(labels):
|
1210 |
+
labels = np.concatenate((labels, [[sample_labels[sel_ind], *box]]), 0)
|
1211 |
+
else:
|
1212 |
+
labels = np.array([[sample_labels[sel_ind], *box]])
|
1213 |
+
|
1214 |
+
image[ymin:ymin+r_h, xmin:xmin+r_w] = temp_crop
|
1215 |
+
|
1216 |
+
return labels
|
1217 |
+
|
1218 |
+
class Albumentations:
|
1219 |
+
# YOLOv5 Albumentations class (optional, only used if package is installed)
|
1220 |
+
def __init__(self):
|
1221 |
+
self.transform = None
|
1222 |
+
import albumentations as A
|
1223 |
+
|
1224 |
+
self.transform = A.Compose([
|
1225 |
+
A.CLAHE(p=0.01),
|
1226 |
+
A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.01),
|
1227 |
+
A.RandomGamma(gamma_limit=[80, 120], p=0.01),
|
1228 |
+
A.Blur(p=0.01),
|
1229 |
+
A.MedianBlur(p=0.01),
|
1230 |
+
A.ToGray(p=0.01),
|
1231 |
+
A.ImageCompression(quality_lower=75, p=0.01),],
|
1232 |
+
bbox_params=A.BboxParams(format='pascal_voc', label_fields=['class_labels']))
|
1233 |
+
|
1234 |
+
#logging.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p))
|
1235 |
+
|
1236 |
+
def __call__(self, im, labels, p=1.0):
|
1237 |
+
if self.transform and random.random() < p:
|
1238 |
+
new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0]) # transformed
|
1239 |
+
im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
|
1240 |
+
return im, labels
|
1241 |
+
|
1242 |
+
|
1243 |
+
def create_folder(path='./new'):
|
1244 |
+
# Create folder
|
1245 |
+
if os.path.exists(path):
|
1246 |
+
shutil.rmtree(path) # delete output folder
|
1247 |
+
os.makedirs(path) # make new output folder
|
1248 |
+
|
1249 |
+
|
1250 |
+
def flatten_recursive(path='../coco'):
|
1251 |
+
# Flatten a recursive directory by bringing all files to top level
|
1252 |
+
new_path = Path(path + '_flat')
|
1253 |
+
create_folder(new_path)
|
1254 |
+
for file in tqdm(glob.glob(str(Path(path)) + '/**/*.*', recursive=True)):
|
1255 |
+
shutil.copyfile(file, new_path / Path(file).name)
|
1256 |
+
|
1257 |
+
|
1258 |
+
def extract_boxes(path='../coco/'): # from utils.datasets import *; extract_boxes('../coco128')
|
1259 |
+
# Convert detection dataset into classification dataset, with one directory per class
|
1260 |
+
|
1261 |
+
path = Path(path) # images dir
|
1262 |
+
shutil.rmtree(path / 'classifier') if (path / 'classifier').is_dir() else None # remove existing
|
1263 |
+
files = list(path.rglob('*.*'))
|
1264 |
+
n = len(files) # number of files
|
1265 |
+
for im_file in tqdm(files, total=n):
|
1266 |
+
if im_file.suffix[1:] in img_formats:
|
1267 |
+
# image
|
1268 |
+
im = cv2.imread(str(im_file))[..., ::-1] # BGR to RGB
|
1269 |
+
h, w = im.shape[:2]
|
1270 |
+
|
1271 |
+
# labels
|
1272 |
+
lb_file = Path(img2label_paths([str(im_file)])[0])
|
1273 |
+
if Path(lb_file).exists():
|
1274 |
+
with open(lb_file, 'r') as f:
|
1275 |
+
lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32) # labels
|
1276 |
+
|
1277 |
+
for j, x in enumerate(lb):
|
1278 |
+
c = int(x[0]) # class
|
1279 |
+
f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg' # new filename
|
1280 |
+
if not f.parent.is_dir():
|
1281 |
+
f.parent.mkdir(parents=True)
|
1282 |
+
|
1283 |
+
b = x[1:] * [w, h, w, h] # box
|
1284 |
+
# b[2:] = b[2:].max() # rectangle to square
|
1285 |
+
b[2:] = b[2:] * 1.2 + 3 # pad
|
1286 |
+
b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int)
|
1287 |
+
|
1288 |
+
b[[0, 2]] = np.clip(b[[0, 2]], 0, w) # clip boxes outside of image
|
1289 |
+
b[[1, 3]] = np.clip(b[[1, 3]], 0, h)
|
1290 |
+
assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}'
|
1291 |
+
|
1292 |
+
|
1293 |
+
def autosplit(path='../coco', weights=(0.9, 0.1, 0.0), annotated_only=False):
|
1294 |
+
""" Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files
|
1295 |
+
Usage: from utils.datasets import *; autosplit('../coco')
|
1296 |
+
Arguments
|
1297 |
+
path: Path to images directory
|
1298 |
+
weights: Train, val, test weights (list)
|
1299 |
+
annotated_only: Only use images with an annotated txt file
|
1300 |
+
"""
|
1301 |
+
path = Path(path) # images dir
|
1302 |
+
files = sum([list(path.rglob(f"*.{img_ext}")) for img_ext in img_formats], []) # image files only
|
1303 |
+
n = len(files) # number of files
|
1304 |
+
indices = random.choices([0, 1, 2], weights=weights, k=n) # assign each image to a split
|
1305 |
+
|
1306 |
+
txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt'] # 3 txt files
|
1307 |
+
[(path / x).unlink() for x in txt if (path / x).exists()] # remove existing
|
1308 |
+
|
1309 |
+
print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only)
|
1310 |
+
for i, img in tqdm(zip(indices, files), total=n):
|
1311 |
+
if not annotated_only or Path(img2label_paths([str(img)])[0]).exists(): # check label
|
1312 |
+
with open(path / txt[i], 'a') as f:
|
1313 |
+
f.write(str(img) + '\n') # add image to txt file
|
1314 |
+
|
1315 |
+
|
1316 |
+
def load_segmentations(self, index):
|
1317 |
+
key = '/work/handsomejw66/coco17/' + self.img_files[index]
|
1318 |
+
#print(key)
|
1319 |
+
# /work/handsomejw66/coco17/
|
1320 |
+
return self.segs[key]
|
utils/general.py
ADDED
@@ -0,0 +1,891 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# YOLOR general utils
|
2 |
+
|
3 |
+
import glob
|
4 |
+
import logging
|
5 |
+
import math
|
6 |
+
import os
|
7 |
+
import platform
|
8 |
+
import random
|
9 |
+
import re
|
10 |
+
import subprocess
|
11 |
+
import time
|
12 |
+
from pathlib import Path
|
13 |
+
|
14 |
+
import cv2
|
15 |
+
import numpy as np
|
16 |
+
import pandas as pd
|
17 |
+
import torch
|
18 |
+
import torchvision
|
19 |
+
import yaml
|
20 |
+
|
21 |
+
from utils.google_utils import gsutil_getsize
|
22 |
+
from utils.metrics import fitness
|
23 |
+
from utils.torch_utils import init_torch_seeds
|
24 |
+
|
25 |
+
# Settings
|
26 |
+
torch.set_printoptions(linewidth=320, precision=5, profile='long')
|
27 |
+
np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5
|
28 |
+
pd.options.display.max_columns = 10
|
29 |
+
cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader)
|
30 |
+
os.environ['NUMEXPR_MAX_THREADS'] = str(min(os.cpu_count(), 8)) # NumExpr max threads
|
31 |
+
|
32 |
+
|
33 |
+
def set_logging(rank=-1):
|
34 |
+
logging.basicConfig(
|
35 |
+
format="%(message)s",
|
36 |
+
level=logging.INFO if rank in [-1, 0] else logging.WARN)
|
37 |
+
|
38 |
+
|
39 |
+
def init_seeds(seed=0):
|
40 |
+
# Initialize random number generator (RNG) seeds
|
41 |
+
random.seed(seed)
|
42 |
+
np.random.seed(seed)
|
43 |
+
init_torch_seeds(seed)
|
44 |
+
|
45 |
+
|
46 |
+
def get_latest_run(search_dir='.'):
|
47 |
+
# Return path to most recent 'last.pt' in /runs (i.e. to --resume from)
|
48 |
+
last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True)
|
49 |
+
return max(last_list, key=os.path.getctime) if last_list else ''
|
50 |
+
|
51 |
+
|
52 |
+
def isdocker():
|
53 |
+
# Is environment a Docker container
|
54 |
+
return Path('/workspace').exists() # or Path('/.dockerenv').exists()
|
55 |
+
|
56 |
+
|
57 |
+
def emojis(str=''):
|
58 |
+
# Return platform-dependent emoji-safe version of string
|
59 |
+
return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str
|
60 |
+
|
61 |
+
|
62 |
+
def check_online():
|
63 |
+
# Check internet connectivity
|
64 |
+
import socket
|
65 |
+
try:
|
66 |
+
socket.create_connection(("1.1.1.1", 443), 5) # check host accesability
|
67 |
+
return True
|
68 |
+
except OSError:
|
69 |
+
return False
|
70 |
+
|
71 |
+
|
72 |
+
def check_git_status():
|
73 |
+
# Recommend 'git pull' if code is out of date
|
74 |
+
print(colorstr('github: '), end='')
|
75 |
+
try:
|
76 |
+
assert Path('.git').exists(), 'skipping check (not a git repository)'
|
77 |
+
assert not isdocker(), 'skipping check (Docker image)'
|
78 |
+
assert check_online(), 'skipping check (offline)'
|
79 |
+
|
80 |
+
cmd = 'git fetch && git config --get remote.origin.url'
|
81 |
+
url = subprocess.check_output(cmd, shell=True).decode().strip().rstrip('.git') # github repo url
|
82 |
+
branch = subprocess.check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip() # checked out
|
83 |
+
n = int(subprocess.check_output(f'git rev-list {branch}..origin/master --count', shell=True)) # commits behind
|
84 |
+
if n > 0:
|
85 |
+
s = f"⚠️ WARNING: code is out of date by {n} commit{'s' * (n > 1)}. " \
|
86 |
+
f"Use 'git pull' to update or 'git clone {url}' to download latest."
|
87 |
+
else:
|
88 |
+
s = f'up to date with {url} ✅'
|
89 |
+
print(emojis(s)) # emoji-safe
|
90 |
+
except Exception as e:
|
91 |
+
print(e)
|
92 |
+
|
93 |
+
|
94 |
+
def check_requirements(requirements='requirements.txt', exclude=()):
|
95 |
+
# Check installed dependencies meet requirements (pass *.txt file or list of packages)
|
96 |
+
import pkg_resources as pkg
|
97 |
+
prefix = colorstr('red', 'bold', 'requirements:')
|
98 |
+
if isinstance(requirements, (str, Path)): # requirements.txt file
|
99 |
+
file = Path(requirements)
|
100 |
+
if not file.exists():
|
101 |
+
print(f"{prefix} {file.resolve()} not found, check failed.")
|
102 |
+
return
|
103 |
+
requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(file.open()) if x.name not in exclude]
|
104 |
+
else: # list or tuple of packages
|
105 |
+
requirements = [x for x in requirements if x not in exclude]
|
106 |
+
|
107 |
+
n = 0 # number of packages updates
|
108 |
+
for r in requirements:
|
109 |
+
try:
|
110 |
+
pkg.require(r)
|
111 |
+
except Exception as e: # DistributionNotFound or VersionConflict if requirements not met
|
112 |
+
n += 1
|
113 |
+
print(f"{prefix} {e.req} not found and is required by YOLOR, attempting auto-update...")
|
114 |
+
print(subprocess.check_output(f"pip install '{e.req}'", shell=True).decode())
|
115 |
+
|
116 |
+
if n: # if packages updated
|
117 |
+
source = file.resolve() if 'file' in locals() else requirements
|
118 |
+
s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \
|
119 |
+
f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n"
|
120 |
+
print(emojis(s)) # emoji-safe
|
121 |
+
|
122 |
+
|
123 |
+
def check_img_size(img_size, s=32):
|
124 |
+
# Verify img_size is a multiple of stride s
|
125 |
+
new_size = make_divisible(img_size, int(s)) # ceil gs-multiple
|
126 |
+
if new_size != img_size:
|
127 |
+
print('WARNING: --img-size %g must be multiple of max stride %g, updating to %g' % (img_size, s, new_size))
|
128 |
+
return new_size
|
129 |
+
|
130 |
+
|
131 |
+
def check_imshow():
|
132 |
+
# Check if environment supports image displays
|
133 |
+
try:
|
134 |
+
assert not isdocker(), 'cv2.imshow() is disabled in Docker environments'
|
135 |
+
cv2.imshow('test', np.zeros((1, 1, 3)))
|
136 |
+
cv2.waitKey(1)
|
137 |
+
cv2.destroyAllWindows()
|
138 |
+
cv2.waitKey(1)
|
139 |
+
return True
|
140 |
+
except Exception as e:
|
141 |
+
print(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}')
|
142 |
+
return False
|
143 |
+
|
144 |
+
|
145 |
+
def check_file(file):
|
146 |
+
# Search for file if not found
|
147 |
+
if Path(file).is_file() or file == '':
|
148 |
+
return file
|
149 |
+
else:
|
150 |
+
files = glob.glob('./**/' + file, recursive=True) # find file
|
151 |
+
assert len(files), f'File Not Found: {file}' # assert file was found
|
152 |
+
assert len(files) == 1, f"Multiple files match '{file}', specify exact path: {files}" # assert unique
|
153 |
+
return files[0] # return file
|
154 |
+
|
155 |
+
|
156 |
+
def check_dataset(dict):
|
157 |
+
# Download dataset if not found locally
|
158 |
+
val, s = dict.get('val'), dict.get('download')
|
159 |
+
if val and len(val):
|
160 |
+
val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path
|
161 |
+
if not all(x.exists() for x in val):
|
162 |
+
print('\nWARNING: Dataset not found, nonexistent paths: %s' % [str(x) for x in val if not x.exists()])
|
163 |
+
if s and len(s): # download script
|
164 |
+
print('Downloading %s ...' % s)
|
165 |
+
if s.startswith('http') and s.endswith('.zip'): # URL
|
166 |
+
f = Path(s).name # filename
|
167 |
+
torch.hub.download_url_to_file(s, f)
|
168 |
+
r = os.system('unzip -q %s -d ../ && rm %s' % (f, f)) # unzip
|
169 |
+
else: # bash script
|
170 |
+
r = os.system(s)
|
171 |
+
print('Dataset autodownload %s\n' % ('success' if r == 0 else 'failure')) # analyze return value
|
172 |
+
else:
|
173 |
+
raise Exception('Dataset not found.')
|
174 |
+
|
175 |
+
|
176 |
+
def make_divisible(x, divisor):
|
177 |
+
# Returns x evenly divisible by divisor
|
178 |
+
return math.ceil(x / divisor) * divisor
|
179 |
+
|
180 |
+
|
181 |
+
def clean_str(s):
|
182 |
+
# Cleans a string by replacing special characters with underscore _
|
183 |
+
return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s)
|
184 |
+
|
185 |
+
|
186 |
+
def one_cycle(y1=0.0, y2=1.0, steps=100):
|
187 |
+
# lambda function for sinusoidal ramp from y1 to y2
|
188 |
+
return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
|
189 |
+
|
190 |
+
|
191 |
+
def colorstr(*input):
|
192 |
+
# Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world')
|
193 |
+
*args, string = input if len(input) > 1 else ('blue', 'bold', input[0]) # color arguments, string
|
194 |
+
colors = {'black': '\033[30m', # basic colors
|
195 |
+
'red': '\033[31m',
|
196 |
+
'green': '\033[32m',
|
197 |
+
'yellow': '\033[33m',
|
198 |
+
'blue': '\033[34m',
|
199 |
+
'magenta': '\033[35m',
|
200 |
+
'cyan': '\033[36m',
|
201 |
+
'white': '\033[37m',
|
202 |
+
'bright_black': '\033[90m', # bright colors
|
203 |
+
'bright_red': '\033[91m',
|
204 |
+
'bright_green': '\033[92m',
|
205 |
+
'bright_yellow': '\033[93m',
|
206 |
+
'bright_blue': '\033[94m',
|
207 |
+
'bright_magenta': '\033[95m',
|
208 |
+
'bright_cyan': '\033[96m',
|
209 |
+
'bright_white': '\033[97m',
|
210 |
+
'end': '\033[0m', # misc
|
211 |
+
'bold': '\033[1m',
|
212 |
+
'underline': '\033[4m'}
|
213 |
+
return ''.join(colors[x] for x in args) + f'{string}' + colors['end']
|
214 |
+
|
215 |
+
|
216 |
+
def labels_to_class_weights(labels, nc=80):
|
217 |
+
# Get class weights (inverse frequency) from training labels
|
218 |
+
if labels[0] is None: # no labels loaded
|
219 |
+
return torch.Tensor()
|
220 |
+
|
221 |
+
labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO
|
222 |
+
classes = labels[:, 0].astype(np.int) # labels = [class xywh]
|
223 |
+
weights = np.bincount(classes, minlength=nc) # occurrences per class
|
224 |
+
|
225 |
+
# Prepend gridpoint count (for uCE training)
|
226 |
+
# gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image
|
227 |
+
# weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start
|
228 |
+
|
229 |
+
weights[weights == 0] = 1 # replace empty bins with 1
|
230 |
+
weights = 1 / weights # number of targets per class
|
231 |
+
weights /= weights.sum() # normalize
|
232 |
+
return torch.from_numpy(weights)
|
233 |
+
|
234 |
+
|
235 |
+
def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)):
|
236 |
+
# Produces image weights based on class_weights and image contents
|
237 |
+
class_counts = np.array([np.bincount(x[:, 0].astype(np.int), minlength=nc) for x in labels])
|
238 |
+
image_weights = (class_weights.reshape(1, nc) * class_counts).sum(1)
|
239 |
+
# index = random.choices(range(n), weights=image_weights, k=1) # weight image sample
|
240 |
+
return image_weights
|
241 |
+
|
242 |
+
|
243 |
+
def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper)
|
244 |
+
# https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/
|
245 |
+
# a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')
|
246 |
+
# b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')
|
247 |
+
# x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco
|
248 |
+
# x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet
|
249 |
+
x = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34,
|
250 |
+
35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
|
251 |
+
64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90]
|
252 |
+
return x
|
253 |
+
|
254 |
+
|
255 |
+
def xyxy2xywh(x):
|
256 |
+
# Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right
|
257 |
+
y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
|
258 |
+
y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center
|
259 |
+
y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center
|
260 |
+
y[:, 2] = x[:, 2] - x[:, 0] # width
|
261 |
+
y[:, 3] = x[:, 3] - x[:, 1] # height
|
262 |
+
return y
|
263 |
+
|
264 |
+
|
265 |
+
def xywh2xyxy(x):
|
266 |
+
# Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
|
267 |
+
y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
|
268 |
+
y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x
|
269 |
+
y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y
|
270 |
+
y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x
|
271 |
+
y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y
|
272 |
+
return y
|
273 |
+
|
274 |
+
|
275 |
+
def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0):
|
276 |
+
# Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
|
277 |
+
y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
|
278 |
+
y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw # top left x
|
279 |
+
y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh # top left y
|
280 |
+
y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw # bottom right x
|
281 |
+
y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh # bottom right y
|
282 |
+
return y
|
283 |
+
|
284 |
+
|
285 |
+
def xyn2xy(x, w=640, h=640, padw=0, padh=0):
|
286 |
+
# Convert normalized segments into pixel segments, shape (n,2)
|
287 |
+
y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
|
288 |
+
y[:, 0] = w * x[:, 0] + padw # top left x
|
289 |
+
y[:, 1] = h * x[:, 1] + padh # top left y
|
290 |
+
return y
|
291 |
+
|
292 |
+
|
293 |
+
def segment2box(segment, width=640, height=640):
|
294 |
+
# Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy)
|
295 |
+
x, y = segment.T # segment xy
|
296 |
+
inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height)
|
297 |
+
x, y, = x[inside], y[inside]
|
298 |
+
return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4)) # xyxy
|
299 |
+
|
300 |
+
|
301 |
+
def segments2boxes(segments):
|
302 |
+
# Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh)
|
303 |
+
boxes = []
|
304 |
+
for s in segments:
|
305 |
+
x, y = s.T # segment xy
|
306 |
+
boxes.append([x.min(), y.min(), x.max(), y.max()]) # cls, xyxy
|
307 |
+
return xyxy2xywh(np.array(boxes)) # cls, xywh
|
308 |
+
|
309 |
+
|
310 |
+
def resample_segments(segments, n=1000):
|
311 |
+
# Up-sample an (n,2) segment
|
312 |
+
for i, s in enumerate(segments):
|
313 |
+
x = np.linspace(0, len(s) - 1, n)
|
314 |
+
xp = np.arange(len(s))
|
315 |
+
segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T # segment xy
|
316 |
+
return segments
|
317 |
+
|
318 |
+
|
319 |
+
def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None):
|
320 |
+
# Rescale coords (xyxy) from img1_shape to img0_shape
|
321 |
+
if ratio_pad is None: # calculate from img0_shape
|
322 |
+
gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new
|
323 |
+
pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding
|
324 |
+
else:
|
325 |
+
gain = ratio_pad[0][0]
|
326 |
+
pad = ratio_pad[1]
|
327 |
+
|
328 |
+
coords[:, [0, 2]] -= pad[0] # x padding
|
329 |
+
coords[:, [1, 3]] -= pad[1] # y padding
|
330 |
+
coords[:, :4] /= gain
|
331 |
+
clip_coords(coords, img0_shape)
|
332 |
+
return coords
|
333 |
+
|
334 |
+
|
335 |
+
def clip_coords(boxes, img_shape):
|
336 |
+
# Clip bounding xyxy bounding boxes to image shape (height, width)
|
337 |
+
boxes[:, 0].clamp_(0, img_shape[1]) # x1
|
338 |
+
boxes[:, 1].clamp_(0, img_shape[0]) # y1
|
339 |
+
boxes[:, 2].clamp_(0, img_shape[1]) # x2
|
340 |
+
boxes[:, 3].clamp_(0, img_shape[0]) # y2
|
341 |
+
|
342 |
+
|
343 |
+
def bbox_iou(box1, box2, x1y1x2y2=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
|
344 |
+
# Returns the IoU of box1 to box2. box1 is 4, box2 is nx4
|
345 |
+
box2 = box2.T
|
346 |
+
|
347 |
+
# Get the coordinates of bounding boxes
|
348 |
+
if x1y1x2y2: # x1, y1, x2, y2 = box1
|
349 |
+
b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
|
350 |
+
b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
|
351 |
+
else: # transform from xywh to xyxy
|
352 |
+
b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
|
353 |
+
b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
|
354 |
+
b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
|
355 |
+
b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
|
356 |
+
|
357 |
+
# Intersection area
|
358 |
+
inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
|
359 |
+
(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
|
360 |
+
|
361 |
+
# Union Area
|
362 |
+
w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
|
363 |
+
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
|
364 |
+
union = w1 * h1 + w2 * h2 - inter + eps
|
365 |
+
|
366 |
+
iou = inter / union
|
367 |
+
|
368 |
+
if GIoU or DIoU or CIoU:
|
369 |
+
cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width
|
370 |
+
ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height
|
371 |
+
if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
|
372 |
+
c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared
|
373 |
+
rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 +
|
374 |
+
(b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center distance squared
|
375 |
+
if DIoU:
|
376 |
+
return iou - rho2 / c2 # DIoU
|
377 |
+
elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
|
378 |
+
v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / (h2 + eps)) - torch.atan(w1 / (h1 + eps)), 2)
|
379 |
+
with torch.no_grad():
|
380 |
+
alpha = v / (v - iou + (1 + eps))
|
381 |
+
return iou - (rho2 / c2 + v * alpha) # CIoU
|
382 |
+
else: # GIoU https://arxiv.org/pdf/1902.09630.pdf
|
383 |
+
c_area = cw * ch + eps # convex area
|
384 |
+
return iou - (c_area - union) / c_area # GIoU
|
385 |
+
else:
|
386 |
+
return iou # IoU
|
387 |
+
|
388 |
+
|
389 |
+
|
390 |
+
|
391 |
+
def bbox_alpha_iou(box1, box2, x1y1x2y2=False, GIoU=False, DIoU=False, CIoU=False, alpha=2, eps=1e-9):
|
392 |
+
# Returns tsqrt_he IoU of box1 to box2. box1 is 4, box2 is nx4
|
393 |
+
box2 = box2.T
|
394 |
+
|
395 |
+
# Get the coordinates of bounding boxes
|
396 |
+
if x1y1x2y2: # x1, y1, x2, y2 = box1
|
397 |
+
b1_x1, b1_y1, b1_x2, b1_y2 = box1[0], box1[1], box1[2], box1[3]
|
398 |
+
b2_x1, b2_y1, b2_x2, b2_y2 = box2[0], box2[1], box2[2], box2[3]
|
399 |
+
else: # transform from xywh to xyxy
|
400 |
+
b1_x1, b1_x2 = box1[0] - box1[2] / 2, box1[0] + box1[2] / 2
|
401 |
+
b1_y1, b1_y2 = box1[1] - box1[3] / 2, box1[1] + box1[3] / 2
|
402 |
+
b2_x1, b2_x2 = box2[0] - box2[2] / 2, box2[0] + box2[2] / 2
|
403 |
+
b2_y1, b2_y2 = box2[1] - box2[3] / 2, box2[1] + box2[3] / 2
|
404 |
+
|
405 |
+
# Intersection area
|
406 |
+
inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \
|
407 |
+
(torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0)
|
408 |
+
|
409 |
+
# Union Area
|
410 |
+
w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + eps
|
411 |
+
w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + eps
|
412 |
+
union = w1 * h1 + w2 * h2 - inter + eps
|
413 |
+
|
414 |
+
# change iou into pow(iou+eps)
|
415 |
+
# iou = inter / union
|
416 |
+
iou = torch.pow(inter/union + eps, alpha)
|
417 |
+
# beta = 2 * alpha
|
418 |
+
if GIoU or DIoU or CIoU:
|
419 |
+
cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width
|
420 |
+
ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height
|
421 |
+
if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
|
422 |
+
c2 = (cw ** 2 + ch ** 2) ** alpha + eps # convex diagonal
|
423 |
+
rho_x = torch.abs(b2_x1 + b2_x2 - b1_x1 - b1_x2)
|
424 |
+
rho_y = torch.abs(b2_y1 + b2_y2 - b1_y1 - b1_y2)
|
425 |
+
rho2 = ((rho_x ** 2 + rho_y ** 2) / 4) ** alpha # center distance
|
426 |
+
if DIoU:
|
427 |
+
return iou - rho2 / c2 # DIoU
|
428 |
+
elif CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
|
429 |
+
v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
|
430 |
+
with torch.no_grad():
|
431 |
+
alpha_ciou = v / ((1 + eps) - inter / union + v)
|
432 |
+
# return iou - (rho2 / c2 + v * alpha_ciou) # CIoU
|
433 |
+
return iou - (rho2 / c2 + torch.pow(v * alpha_ciou + eps, alpha)) # CIoU
|
434 |
+
else: # GIoU https://arxiv.org/pdf/1902.09630.pdf
|
435 |
+
# c_area = cw * ch + eps # convex area
|
436 |
+
# return iou - (c_area - union) / c_area # GIoU
|
437 |
+
c_area = torch.max(cw * ch + eps, union) # convex area
|
438 |
+
return iou - torch.pow((c_area - union) / c_area + eps, alpha) # GIoU
|
439 |
+
else:
|
440 |
+
return iou # torch.log(iou+eps) or iou
|
441 |
+
|
442 |
+
|
443 |
+
def box_iou(box1, box2):
|
444 |
+
# https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
|
445 |
+
"""
|
446 |
+
Return intersection-over-union (Jaccard index) of boxes.
|
447 |
+
Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
|
448 |
+
Arguments:
|
449 |
+
box1 (Tensor[N, 4])
|
450 |
+
box2 (Tensor[M, 4])
|
451 |
+
Returns:
|
452 |
+
iou (Tensor[N, M]): the NxM matrix containing the pairwise
|
453 |
+
IoU values for every element in boxes1 and boxes2
|
454 |
+
"""
|
455 |
+
|
456 |
+
def box_area(box):
|
457 |
+
# box = 4xn
|
458 |
+
return (box[2] - box[0]) * (box[3] - box[1])
|
459 |
+
|
460 |
+
area1 = box_area(box1.T)
|
461 |
+
area2 = box_area(box2.T)
|
462 |
+
|
463 |
+
# inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
|
464 |
+
inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
|
465 |
+
return inter / (area1[:, None] + area2 - inter) # iou = inter / (area1 + area2 - inter)
|
466 |
+
|
467 |
+
|
468 |
+
def wh_iou(wh1, wh2):
|
469 |
+
# Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2
|
470 |
+
wh1 = wh1[:, None] # [N,1,2]
|
471 |
+
wh2 = wh2[None] # [1,M,2]
|
472 |
+
inter = torch.min(wh1, wh2).prod(2) # [N,M]
|
473 |
+
return inter / (wh1.prod(2) + wh2.prod(2) - inter) # iou = inter / (area1 + area2 - inter)
|
474 |
+
|
475 |
+
|
476 |
+
def box_giou(box1, box2):
|
477 |
+
"""
|
478 |
+
Return generalized intersection-over-union (Jaccard index) between two sets of boxes.
|
479 |
+
Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
|
480 |
+
``0 <= x1 < x2`` and ``0 <= y1 < y2``.
|
481 |
+
Args:
|
482 |
+
boxes1 (Tensor[N, 4]): first set of boxes
|
483 |
+
boxes2 (Tensor[M, 4]): second set of boxes
|
484 |
+
Returns:
|
485 |
+
Tensor[N, M]: the NxM matrix containing the pairwise generalized IoU values
|
486 |
+
for every element in boxes1 and boxes2
|
487 |
+
"""
|
488 |
+
|
489 |
+
def box_area(box):
|
490 |
+
# box = 4xn
|
491 |
+
return (box[2] - box[0]) * (box[3] - box[1])
|
492 |
+
|
493 |
+
area1 = box_area(box1.T)
|
494 |
+
area2 = box_area(box2.T)
|
495 |
+
|
496 |
+
inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
|
497 |
+
union = (area1[:, None] + area2 - inter)
|
498 |
+
|
499 |
+
iou = inter / union
|
500 |
+
|
501 |
+
lti = torch.min(box1[:, None, :2], box2[:, :2])
|
502 |
+
rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
|
503 |
+
|
504 |
+
whi = (rbi - lti).clamp(min=0) # [N,M,2]
|
505 |
+
areai = whi[:, :, 0] * whi[:, :, 1]
|
506 |
+
|
507 |
+
return iou - (areai - union) / areai
|
508 |
+
|
509 |
+
|
510 |
+
def box_ciou(box1, box2, eps: float = 1e-7):
|
511 |
+
"""
|
512 |
+
Return complete intersection-over-union (Jaccard index) between two sets of boxes.
|
513 |
+
Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
|
514 |
+
``0 <= x1 < x2`` and ``0 <= y1 < y2``.
|
515 |
+
Args:
|
516 |
+
boxes1 (Tensor[N, 4]): first set of boxes
|
517 |
+
boxes2 (Tensor[M, 4]): second set of boxes
|
518 |
+
eps (float, optional): small number to prevent division by zero. Default: 1e-7
|
519 |
+
Returns:
|
520 |
+
Tensor[N, M]: the NxM matrix containing the pairwise complete IoU values
|
521 |
+
for every element in boxes1 and boxes2
|
522 |
+
"""
|
523 |
+
|
524 |
+
def box_area(box):
|
525 |
+
# box = 4xn
|
526 |
+
return (box[2] - box[0]) * (box[3] - box[1])
|
527 |
+
|
528 |
+
area1 = box_area(box1.T)
|
529 |
+
area2 = box_area(box2.T)
|
530 |
+
|
531 |
+
inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
|
532 |
+
union = (area1[:, None] + area2 - inter)
|
533 |
+
|
534 |
+
iou = inter / union
|
535 |
+
|
536 |
+
lti = torch.min(box1[:, None, :2], box2[:, :2])
|
537 |
+
rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
|
538 |
+
|
539 |
+
whi = (rbi - lti).clamp(min=0) # [N,M,2]
|
540 |
+
diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps
|
541 |
+
|
542 |
+
# centers of boxes
|
543 |
+
x_p = (box1[:, None, 0] + box1[:, None, 2]) / 2
|
544 |
+
y_p = (box1[:, None, 1] + box1[:, None, 3]) / 2
|
545 |
+
x_g = (box2[:, 0] + box2[:, 2]) / 2
|
546 |
+
y_g = (box2[:, 1] + box2[:, 3]) / 2
|
547 |
+
# The distance between boxes' centers squared.
|
548 |
+
centers_distance_squared = (x_p - x_g) ** 2 + (y_p - y_g) ** 2
|
549 |
+
|
550 |
+
w_pred = box1[:, None, 2] - box1[:, None, 0]
|
551 |
+
h_pred = box1[:, None, 3] - box1[:, None, 1]
|
552 |
+
|
553 |
+
w_gt = box2[:, 2] - box2[:, 0]
|
554 |
+
h_gt = box2[:, 3] - box2[:, 1]
|
555 |
+
|
556 |
+
v = (4 / (torch.pi ** 2)) * torch.pow((torch.atan(w_gt / h_gt) - torch.atan(w_pred / h_pred)), 2)
|
557 |
+
with torch.no_grad():
|
558 |
+
alpha = v / (1 - iou + v + eps)
|
559 |
+
return iou - (centers_distance_squared / diagonal_distance_squared) - alpha * v
|
560 |
+
|
561 |
+
|
562 |
+
def box_diou(box1, box2, eps: float = 1e-7):
|
563 |
+
"""
|
564 |
+
Return distance intersection-over-union (Jaccard index) between two sets of boxes.
|
565 |
+
Both sets of boxes are expected to be in ``(x1, y1, x2, y2)`` format with
|
566 |
+
``0 <= x1 < x2`` and ``0 <= y1 < y2``.
|
567 |
+
Args:
|
568 |
+
boxes1 (Tensor[N, 4]): first set of boxes
|
569 |
+
boxes2 (Tensor[M, 4]): second set of boxes
|
570 |
+
eps (float, optional): small number to prevent division by zero. Default: 1e-7
|
571 |
+
Returns:
|
572 |
+
Tensor[N, M]: the NxM matrix containing the pairwise distance IoU values
|
573 |
+
for every element in boxes1 and boxes2
|
574 |
+
"""
|
575 |
+
|
576 |
+
def box_area(box):
|
577 |
+
# box = 4xn
|
578 |
+
return (box[2] - box[0]) * (box[3] - box[1])
|
579 |
+
|
580 |
+
area1 = box_area(box1.T)
|
581 |
+
area2 = box_area(box2.T)
|
582 |
+
|
583 |
+
inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
|
584 |
+
union = (area1[:, None] + area2 - inter)
|
585 |
+
|
586 |
+
iou = inter / union
|
587 |
+
|
588 |
+
lti = torch.min(box1[:, None, :2], box2[:, :2])
|
589 |
+
rbi = torch.max(box1[:, None, 2:], box2[:, 2:])
|
590 |
+
|
591 |
+
whi = (rbi - lti).clamp(min=0) # [N,M,2]
|
592 |
+
diagonal_distance_squared = (whi[:, :, 0] ** 2) + (whi[:, :, 1] ** 2) + eps
|
593 |
+
|
594 |
+
# centers of boxes
|
595 |
+
x_p = (box1[:, None, 0] + box1[:, None, 2]) / 2
|
596 |
+
y_p = (box1[:, None, 1] + box1[:, None, 3]) / 2
|
597 |
+
x_g = (box2[:, 0] + box2[:, 2]) / 2
|
598 |
+
y_g = (box2[:, 1] + box2[:, 3]) / 2
|
599 |
+
# The distance between boxes' centers squared.
|
600 |
+
centers_distance_squared = (x_p - x_g) ** 2 + (y_p - y_g) ** 2
|
601 |
+
|
602 |
+
# The distance IoU is the IoU penalized by a normalized
|
603 |
+
# distance between boxes' centers squared.
|
604 |
+
return iou - (centers_distance_squared / diagonal_distance_squared)
|
605 |
+
|
606 |
+
|
607 |
+
def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
|
608 |
+
labels=()):
|
609 |
+
"""Runs Non-Maximum Suppression (NMS) on inference results
|
610 |
+
|
611 |
+
Returns:
|
612 |
+
list of detections, on (n,6) tensor per image [xyxy, conf, cls]
|
613 |
+
"""
|
614 |
+
|
615 |
+
nc = prediction.shape[2] - 5 # number of classes
|
616 |
+
xc = prediction[..., 4] > conf_thres # candidates
|
617 |
+
|
618 |
+
# Settings
|
619 |
+
min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height
|
620 |
+
max_det = 300 # maximum number of detections per image
|
621 |
+
max_nms = 30000 # maximum number of boxes into torchvision.ops.nms()
|
622 |
+
time_limit = 10.0 # seconds to quit after
|
623 |
+
redundant = True # require redundant detections
|
624 |
+
multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img)
|
625 |
+
merge = False # use merge-NMS
|
626 |
+
|
627 |
+
t = time.time()
|
628 |
+
output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
|
629 |
+
for xi, x in enumerate(prediction): # image index, image inference
|
630 |
+
# Apply constraints
|
631 |
+
# x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height
|
632 |
+
x = x[xc[xi]] # confidence
|
633 |
+
|
634 |
+
# Cat apriori labels if autolabelling
|
635 |
+
if labels and len(labels[xi]):
|
636 |
+
l = labels[xi]
|
637 |
+
v = torch.zeros((len(l), nc + 5), device=x.device)
|
638 |
+
v[:, :4] = l[:, 1:5] # box
|
639 |
+
v[:, 4] = 1.0 # conf
|
640 |
+
v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls
|
641 |
+
x = torch.cat((x, v), 0)
|
642 |
+
|
643 |
+
# If none remain process next image
|
644 |
+
if not x.shape[0]:
|
645 |
+
continue
|
646 |
+
|
647 |
+
# Compute conf
|
648 |
+
if nc == 1:
|
649 |
+
x[:, 5:] = x[:, 4:5] # for models with one class, cls_loss is 0 and cls_conf is always 0.5,
|
650 |
+
# so there is no need to multiplicate.
|
651 |
+
else:
|
652 |
+
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
|
653 |
+
|
654 |
+
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
|
655 |
+
box = xywh2xyxy(x[:, :4])
|
656 |
+
|
657 |
+
# Detections matrix nx6 (xyxy, conf, cls)
|
658 |
+
if multi_label:
|
659 |
+
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
|
660 |
+
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
|
661 |
+
else: # best class only
|
662 |
+
conf, j = x[:, 5:].max(1, keepdim=True)
|
663 |
+
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
|
664 |
+
|
665 |
+
# Filter by class
|
666 |
+
if classes is not None:
|
667 |
+
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
|
668 |
+
|
669 |
+
# Apply finite constraint
|
670 |
+
# if not torch.isfinite(x).all():
|
671 |
+
# x = x[torch.isfinite(x).all(1)]
|
672 |
+
|
673 |
+
# Check shape
|
674 |
+
n = x.shape[0] # number of boxes
|
675 |
+
if not n: # no boxes
|
676 |
+
continue
|
677 |
+
elif n > max_nms: # excess boxes
|
678 |
+
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence
|
679 |
+
|
680 |
+
# Batched NMS
|
681 |
+
c = x[:, 5:6] * (0 if agnostic else max_wh) # classes
|
682 |
+
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
|
683 |
+
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
|
684 |
+
if i.shape[0] > max_det: # limit detections
|
685 |
+
i = i[:max_det]
|
686 |
+
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
|
687 |
+
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
|
688 |
+
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
|
689 |
+
weights = iou * scores[None] # box weights
|
690 |
+
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
|
691 |
+
if redundant:
|
692 |
+
i = i[iou.sum(1) > 1] # require redundancy
|
693 |
+
|
694 |
+
output[xi] = x[i]
|
695 |
+
if (time.time() - t) > time_limit:
|
696 |
+
print(f'WARNING: NMS time limit {time_limit}s exceeded')
|
697 |
+
break # time limit exceeded
|
698 |
+
|
699 |
+
return output
|
700 |
+
|
701 |
+
|
702 |
+
def non_max_suppression_kpt(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
|
703 |
+
labels=(), kpt_label=False, nc=None, nkpt=None):
|
704 |
+
"""Runs Non-Maximum Suppression (NMS) on inference results
|
705 |
+
|
706 |
+
Returns:
|
707 |
+
list of detections, on (n,6) tensor per image [xyxy, conf, cls]
|
708 |
+
"""
|
709 |
+
if nc is None:
|
710 |
+
nc = prediction.shape[2] - 5 if not kpt_label else prediction.shape[2] - 56 # number of classes
|
711 |
+
xc = prediction[..., 4] > conf_thres # candidates
|
712 |
+
|
713 |
+
# Settings
|
714 |
+
min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height
|
715 |
+
max_det = 300 # maximum number of detections per image
|
716 |
+
max_nms = 30000 # maximum number of boxes into torchvision.ops.nms()
|
717 |
+
time_limit = 10.0 # seconds to quit after
|
718 |
+
redundant = True # require redundant detections
|
719 |
+
multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img)
|
720 |
+
merge = False # use merge-NMS
|
721 |
+
|
722 |
+
t = time.time()
|
723 |
+
output = [torch.zeros((0,6), device=prediction.device)] * prediction.shape[0]
|
724 |
+
for xi, x in enumerate(prediction): # image index, image inference
|
725 |
+
# Apply constraints
|
726 |
+
# x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height
|
727 |
+
x = x[xc[xi]] # confidence
|
728 |
+
|
729 |
+
# Cat apriori labels if autolabelling
|
730 |
+
if labels and len(labels[xi]):
|
731 |
+
l = labels[xi]
|
732 |
+
v = torch.zeros((len(l), nc + 5), device=x.device)
|
733 |
+
v[:, :4] = l[:, 1:5] # box
|
734 |
+
v[:, 4] = 1.0 # conf
|
735 |
+
v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls
|
736 |
+
x = torch.cat((x, v), 0)
|
737 |
+
|
738 |
+
# If none remain process next image
|
739 |
+
if not x.shape[0]:
|
740 |
+
continue
|
741 |
+
|
742 |
+
# Compute conf
|
743 |
+
x[:, 5:5+nc] *= x[:, 4:5] # conf = obj_conf * cls_conf
|
744 |
+
|
745 |
+
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
|
746 |
+
box = xywh2xyxy(x[:, :4])
|
747 |
+
|
748 |
+
# Detections matrix nx6 (xyxy, conf, cls)
|
749 |
+
if multi_label:
|
750 |
+
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
|
751 |
+
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
|
752 |
+
else: # best class only
|
753 |
+
if not kpt_label:
|
754 |
+
conf, j = x[:, 5:].max(1, keepdim=True)
|
755 |
+
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
|
756 |
+
else:
|
757 |
+
kpts = x[:, 6:]
|
758 |
+
conf, j = x[:, 5:6].max(1, keepdim=True)
|
759 |
+
x = torch.cat((box, conf, j.float(), kpts), 1)[conf.view(-1) > conf_thres]
|
760 |
+
|
761 |
+
|
762 |
+
# Filter by class
|
763 |
+
if classes is not None:
|
764 |
+
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
|
765 |
+
|
766 |
+
# Apply finite constraint
|
767 |
+
# if not torch.isfinite(x).all():
|
768 |
+
# x = x[torch.isfinite(x).all(1)]
|
769 |
+
|
770 |
+
# Check shape
|
771 |
+
n = x.shape[0] # number of boxes
|
772 |
+
if not n: # no boxes
|
773 |
+
continue
|
774 |
+
elif n > max_nms: # excess boxes
|
775 |
+
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence
|
776 |
+
|
777 |
+
# Batched NMS
|
778 |
+
c = x[:, 5:6] * (0 if agnostic else max_wh) # classes
|
779 |
+
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
|
780 |
+
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
|
781 |
+
if i.shape[0] > max_det: # limit detections
|
782 |
+
i = i[:max_det]
|
783 |
+
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
|
784 |
+
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
|
785 |
+
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
|
786 |
+
weights = iou * scores[None] # box weights
|
787 |
+
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
|
788 |
+
if redundant:
|
789 |
+
i = i[iou.sum(1) > 1] # require redundancy
|
790 |
+
|
791 |
+
output[xi] = x[i]
|
792 |
+
if (time.time() - t) > time_limit:
|
793 |
+
print(f'WARNING: NMS time limit {time_limit}s exceeded')
|
794 |
+
break # time limit exceeded
|
795 |
+
|
796 |
+
return output
|
797 |
+
|
798 |
+
|
799 |
+
def strip_optimizer(f='best.pt', s=''): # from utils.general import *; strip_optimizer()
|
800 |
+
# Strip optimizer from 'f' to finalize training, optionally save as 's'
|
801 |
+
x = torch.load(f, map_location=torch.device('cpu'))
|
802 |
+
if x.get('ema'):
|
803 |
+
x['model'] = x['ema'] # replace model with ema
|
804 |
+
for k in 'optimizer', 'training_results', 'wandb_id', 'ema', 'updates': # keys
|
805 |
+
x[k] = None
|
806 |
+
x['epoch'] = -1
|
807 |
+
x['model'].half() # to FP16
|
808 |
+
for p in x['model'].parameters():
|
809 |
+
p.requires_grad = False
|
810 |
+
torch.save(x, s or f)
|
811 |
+
mb = os.path.getsize(s or f) / 1E6 # filesize
|
812 |
+
print(f"Optimizer stripped from {f},{(' saved as %s,' % s) if s else ''} {mb:.1f}MB")
|
813 |
+
|
814 |
+
|
815 |
+
def print_mutation(hyp, results, yaml_file='hyp_evolved.yaml', bucket=''):
|
816 |
+
# Print mutation results to evolve.txt (for use with train.py --evolve)
|
817 |
+
a = '%10s' * len(hyp) % tuple(hyp.keys()) # hyperparam keys
|
818 |
+
b = '%10.3g' * len(hyp) % tuple(hyp.values()) # hyperparam values
|
819 |
+
c = '%10.4g' * len(results) % results # results (P, R, [email protected], [email protected]:0.95, val_losses x 3)
|
820 |
+
print('\n%s\n%s\nEvolved fitness: %s\n' % (a, b, c))
|
821 |
+
|
822 |
+
if bucket:
|
823 |
+
url = 'gs://%s/evolve.txt' % bucket
|
824 |
+
if gsutil_getsize(url) > (os.path.getsize('evolve.txt') if os.path.exists('evolve.txt') else 0):
|
825 |
+
os.system('gsutil cp %s .' % url) # download evolve.txt if larger than local
|
826 |
+
|
827 |
+
with open('evolve.txt', 'a') as f: # append result
|
828 |
+
f.write(c + b + '\n')
|
829 |
+
x = np.unique(np.loadtxt('evolve.txt', ndmin=2), axis=0) # load unique rows
|
830 |
+
x = x[np.argsort(-fitness(x))] # sort
|
831 |
+
np.savetxt('evolve.txt', x, '%10.3g') # save sort by fitness
|
832 |
+
|
833 |
+
# Save yaml
|
834 |
+
for i, k in enumerate(hyp.keys()):
|
835 |
+
hyp[k] = float(x[0, i + 7])
|
836 |
+
with open(yaml_file, 'w') as f:
|
837 |
+
results = tuple(x[0, :7])
|
838 |
+
c = '%10.4g' * len(results) % results # results (P, R, [email protected], [email protected]:0.95, val_losses x 3)
|
839 |
+
f.write('# Hyperparameter Evolution Results\n# Generations: %g\n# Metrics: ' % len(x) + c + '\n\n')
|
840 |
+
yaml.dump(hyp, f, sort_keys=False)
|
841 |
+
|
842 |
+
if bucket:
|
843 |
+
os.system('gsutil cp evolve.txt %s gs://%s' % (yaml_file, bucket)) # upload
|
844 |
+
|
845 |
+
|
846 |
+
def apply_classifier(x, model, img, im0):
|
847 |
+
# applies a second stage classifier to yolo outputs
|
848 |
+
im0 = [im0] if isinstance(im0, np.ndarray) else im0
|
849 |
+
for i, d in enumerate(x): # per image
|
850 |
+
if d is not None and len(d):
|
851 |
+
d = d.clone()
|
852 |
+
|
853 |
+
# Reshape and pad cutouts
|
854 |
+
b = xyxy2xywh(d[:, :4]) # boxes
|
855 |
+
b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square
|
856 |
+
b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad
|
857 |
+
d[:, :4] = xywh2xyxy(b).long()
|
858 |
+
|
859 |
+
# Rescale boxes from img_size to im0 size
|
860 |
+
scale_coords(img.shape[2:], d[:, :4], im0[i].shape)
|
861 |
+
|
862 |
+
# Classes
|
863 |
+
pred_cls1 = d[:, 5].long()
|
864 |
+
ims = []
|
865 |
+
for j, a in enumerate(d): # per item
|
866 |
+
cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])]
|
867 |
+
im = cv2.resize(cutout, (224, 224)) # BGR
|
868 |
+
# cv2.imwrite('test%i.jpg' % j, cutout)
|
869 |
+
|
870 |
+
im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416
|
871 |
+
im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32
|
872 |
+
im /= 255.0 # 0 - 255 to 0.0 - 1.0
|
873 |
+
ims.append(im)
|
874 |
+
|
875 |
+
pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction
|
876 |
+
x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections
|
877 |
+
|
878 |
+
return x
|
879 |
+
|
880 |
+
|
881 |
+
def increment_path(path, exist_ok=True, sep=''):
|
882 |
+
# Increment path, i.e. runs/exp --> runs/exp{sep}0, runs/exp{sep}1 etc.
|
883 |
+
path = Path(path) # os-agnostic
|
884 |
+
if (path.exists() and exist_ok) or (not path.exists()):
|
885 |
+
return str(path)
|
886 |
+
else:
|
887 |
+
dirs = glob.glob(f"{path}{sep}*") # similar paths
|
888 |
+
matches = [re.search(rf"%s{sep}(\d+)" % path.stem, d) for d in dirs]
|
889 |
+
i = [int(m.groups()[0]) for m in matches if m] # indices
|
890 |
+
n = max(i) + 1 if i else 2 # increment number
|
891 |
+
return f"{path}{sep}{n}" # update path
|
utils/google_utils.py
ADDED
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Google utils: https://cloud.google.com/storage/docs/reference/libraries
|
2 |
+
|
3 |
+
import os
|
4 |
+
import platform
|
5 |
+
import subprocess
|
6 |
+
import time
|
7 |
+
from pathlib import Path
|
8 |
+
|
9 |
+
import requests
|
10 |
+
import torch
|
11 |
+
|
12 |
+
|
13 |
+
def gsutil_getsize(url=''):
|
14 |
+
# gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du
|
15 |
+
s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8')
|
16 |
+
return eval(s.split(' ')[0]) if len(s) else 0 # bytes
|
17 |
+
|
18 |
+
|
19 |
+
def attempt_download(file, repo='WongKinYiu/yolov7'):
|
20 |
+
# Attempt file download if does not exist
|
21 |
+
file = Path(str(file).strip().replace("'", '').lower())
|
22 |
+
|
23 |
+
if not file.exists():
|
24 |
+
try:
|
25 |
+
response = requests.get(f'https://api.github.com/repos/{repo}/releases/latest').json() # github api
|
26 |
+
assets = [x['name'] for x in response['assets']] # release assets
|
27 |
+
tag = response['tag_name'] # i.e. 'v1.0'
|
28 |
+
except: # fallback plan
|
29 |
+
assets = ['yolov7.pt', 'yolov7-tiny.pt', 'yolov7x.pt', 'yolov7-d6.pt', 'yolov7-e6.pt',
|
30 |
+
'yolov7-e6e.pt', 'yolov7-w6.pt']
|
31 |
+
tag = subprocess.check_output('git tag', shell=True).decode().split()[-1]
|
32 |
+
|
33 |
+
name = file.name
|
34 |
+
if name in assets:
|
35 |
+
msg = f'{file} missing, try downloading from https://github.com/{repo}/releases/'
|
36 |
+
redundant = False # second download option
|
37 |
+
try: # GitHub
|
38 |
+
url = f'https://github.com/{repo}/releases/download/{tag}/{name}'
|
39 |
+
print(f'Downloading {url} to {file}...')
|
40 |
+
torch.hub.download_url_to_file(url, file)
|
41 |
+
assert file.exists() and file.stat().st_size > 1E6 # check
|
42 |
+
except Exception as e: # GCP
|
43 |
+
print(f'Download error: {e}')
|
44 |
+
assert redundant, 'No secondary mirror'
|
45 |
+
url = f'https://storage.googleapis.com/{repo}/ckpt/{name}'
|
46 |
+
print(f'Downloading {url} to {file}...')
|
47 |
+
os.system(f'curl -L {url} -o {file}') # torch.hub.download_url_to_file(url, weights)
|
48 |
+
finally:
|
49 |
+
if not file.exists() or file.stat().st_size < 1E6: # check
|
50 |
+
file.unlink(missing_ok=True) # remove partial downloads
|
51 |
+
print(f'ERROR: Download failure: {msg}')
|
52 |
+
print('')
|
53 |
+
return
|
54 |
+
|
55 |
+
|
56 |
+
def gdrive_download(id='', file='tmp.zip'):
|
57 |
+
# Downloads a file from Google Drive. from yolov7.utils.google_utils import *; gdrive_download()
|
58 |
+
t = time.time()
|
59 |
+
file = Path(file)
|
60 |
+
cookie = Path('cookie') # gdrive cookie
|
61 |
+
print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='')
|
62 |
+
file.unlink(missing_ok=True) # remove existing file
|
63 |
+
cookie.unlink(missing_ok=True) # remove existing cookie
|
64 |
+
|
65 |
+
# Attempt file download
|
66 |
+
out = "NUL" if platform.system() == "Windows" else "/dev/null"
|
67 |
+
os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}')
|
68 |
+
if os.path.exists('cookie'): # large file
|
69 |
+
s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}'
|
70 |
+
else: # small file
|
71 |
+
s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"'
|
72 |
+
r = os.system(s) # execute, capture return
|
73 |
+
cookie.unlink(missing_ok=True) # remove existing cookie
|
74 |
+
|
75 |
+
# Error check
|
76 |
+
if r != 0:
|
77 |
+
file.unlink(missing_ok=True) # remove partial
|
78 |
+
print('Download error ') # raise Exception('Download error')
|
79 |
+
return r
|
80 |
+
|
81 |
+
# Unzip if archive
|
82 |
+
if file.suffix == '.zip':
|
83 |
+
print('unzipping... ', end='')
|
84 |
+
os.system(f'unzip -q {file}') # unzip
|
85 |
+
file.unlink() # remove zip to free space
|
86 |
+
|
87 |
+
print(f'Done ({time.time() - t:.1f}s)')
|
88 |
+
return r
|
89 |
+
|
90 |
+
|
91 |
+
def get_token(cookie="./cookie"):
|
92 |
+
with open(cookie) as f:
|
93 |
+
for line in f:
|
94 |
+
if "download" in line:
|
95 |
+
return line.split()[-1]
|
96 |
+
return ""
|
97 |
+
|
98 |
+
# def upload_blob(bucket_name, source_file_name, destination_blob_name):
|
99 |
+
# # Uploads a file to a bucket
|
100 |
+
# # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python
|
101 |
+
#
|
102 |
+
# storage_client = storage.Client()
|
103 |
+
# bucket = storage_client.get_bucket(bucket_name)
|
104 |
+
# blob = bucket.blob(destination_blob_name)
|
105 |
+
#
|
106 |
+
# blob.upload_from_filename(source_file_name)
|
107 |
+
#
|
108 |
+
# print('File {} uploaded to {}.'.format(
|
109 |
+
# source_file_name,
|
110 |
+
# destination_blob_name))
|
111 |
+
#
|
112 |
+
#
|
113 |
+
# def download_blob(bucket_name, source_blob_name, destination_file_name):
|
114 |
+
# # Uploads a blob from a bucket
|
115 |
+
# storage_client = storage.Client()
|
116 |
+
# bucket = storage_client.get_bucket(bucket_name)
|
117 |
+
# blob = bucket.blob(source_blob_name)
|
118 |
+
#
|
119 |
+
# blob.download_to_filename(destination_file_name)
|
120 |
+
#
|
121 |
+
# print('Blob {} downloaded to {}.'.format(
|
122 |
+
# source_blob_name,
|
123 |
+
# destination_file_name))
|
utils/loss.py
ADDED
@@ -0,0 +1,1697 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Loss functions
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
from utils.general import bbox_iou, bbox_alpha_iou, box_iou, box_giou, box_diou, box_ciou, xywh2xyxy
|
8 |
+
from utils.torch_utils import is_parallel
|
9 |
+
|
10 |
+
|
11 |
+
def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
|
12 |
+
# return positive, negative label smoothing BCE targets
|
13 |
+
return 1.0 - 0.5 * eps, 0.5 * eps
|
14 |
+
|
15 |
+
|
16 |
+
class BCEBlurWithLogitsLoss(nn.Module):
|
17 |
+
# BCEwithLogitLoss() with reduced missing label effects.
|
18 |
+
def __init__(self, alpha=0.05):
|
19 |
+
super(BCEBlurWithLogitsLoss, self).__init__()
|
20 |
+
self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss()
|
21 |
+
self.alpha = alpha
|
22 |
+
|
23 |
+
def forward(self, pred, true):
|
24 |
+
loss = self.loss_fcn(pred, true)
|
25 |
+
pred = torch.sigmoid(pred) # prob from logits
|
26 |
+
dx = pred - true # reduce only missing label effects
|
27 |
+
# dx = (pred - true).abs() # reduce missing label and false label effects
|
28 |
+
alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
|
29 |
+
loss *= alpha_factor
|
30 |
+
return loss.mean()
|
31 |
+
|
32 |
+
|
33 |
+
class SigmoidBin(nn.Module):
|
34 |
+
stride = None # strides computed during build
|
35 |
+
export = False # onnx export
|
36 |
+
|
37 |
+
def __init__(self, bin_count=10, min=0.0, max=1.0, reg_scale = 2.0, use_loss_regression=True, use_fw_regression=True, BCE_weight=1.0, smooth_eps=0.0):
|
38 |
+
super(SigmoidBin, self).__init__()
|
39 |
+
|
40 |
+
self.bin_count = bin_count
|
41 |
+
self.length = bin_count + 1
|
42 |
+
self.min = min
|
43 |
+
self.max = max
|
44 |
+
self.scale = float(max - min)
|
45 |
+
self.shift = self.scale / 2.0
|
46 |
+
|
47 |
+
self.use_loss_regression = use_loss_regression
|
48 |
+
self.use_fw_regression = use_fw_regression
|
49 |
+
self.reg_scale = reg_scale
|
50 |
+
self.BCE_weight = BCE_weight
|
51 |
+
|
52 |
+
start = min + (self.scale/2.0) / self.bin_count
|
53 |
+
end = max - (self.scale/2.0) / self.bin_count
|
54 |
+
step = self.scale / self.bin_count
|
55 |
+
self.step = step
|
56 |
+
#print(f" start = {start}, end = {end}, step = {step} ")
|
57 |
+
|
58 |
+
bins = torch.range(start, end + 0.0001, step).float()
|
59 |
+
self.register_buffer('bins', bins)
|
60 |
+
|
61 |
+
|
62 |
+
self.cp = 1.0 - 0.5 * smooth_eps
|
63 |
+
self.cn = 0.5 * smooth_eps
|
64 |
+
|
65 |
+
self.BCEbins = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([BCE_weight]))
|
66 |
+
self.MSELoss = nn.MSELoss()
|
67 |
+
|
68 |
+
def get_length(self):
|
69 |
+
return self.length
|
70 |
+
|
71 |
+
def forward(self, pred):
|
72 |
+
assert pred.shape[-1] == self.length, 'pred.shape[-1]=%d is not equal to self.length=%d' % (pred.shape[-1], self.length)
|
73 |
+
|
74 |
+
pred_reg = (pred[..., 0] * self.reg_scale - self.reg_scale/2.0) * self.step
|
75 |
+
pred_bin = pred[..., 1:(1+self.bin_count)]
|
76 |
+
|
77 |
+
_, bin_idx = torch.max(pred_bin, dim=-1)
|
78 |
+
bin_bias = self.bins[bin_idx]
|
79 |
+
|
80 |
+
if self.use_fw_regression:
|
81 |
+
result = pred_reg + bin_bias
|
82 |
+
else:
|
83 |
+
result = bin_bias
|
84 |
+
result = result.clamp(min=self.min, max=self.max)
|
85 |
+
|
86 |
+
return result
|
87 |
+
|
88 |
+
|
89 |
+
def training_loss(self, pred, target):
|
90 |
+
assert pred.shape[-1] == self.length, 'pred.shape[-1]=%d is not equal to self.length=%d' % (pred.shape[-1], self.length)
|
91 |
+
assert pred.shape[0] == target.shape[0], 'pred.shape=%d is not equal to the target.shape=%d' % (pred.shape[0], target.shape[0])
|
92 |
+
device = pred.device
|
93 |
+
|
94 |
+
pred_reg = (pred[..., 0].sigmoid() * self.reg_scale - self.reg_scale/2.0) * self.step
|
95 |
+
pred_bin = pred[..., 1:(1+self.bin_count)]
|
96 |
+
|
97 |
+
diff_bin_target = torch.abs(target[..., None] - self.bins)
|
98 |
+
_, bin_idx = torch.min(diff_bin_target, dim=-1)
|
99 |
+
|
100 |
+
bin_bias = self.bins[bin_idx]
|
101 |
+
bin_bias.requires_grad = False
|
102 |
+
result = pred_reg + bin_bias
|
103 |
+
|
104 |
+
target_bins = torch.full_like(pred_bin, self.cn, device=device) # targets
|
105 |
+
n = pred.shape[0]
|
106 |
+
target_bins[range(n), bin_idx] = self.cp
|
107 |
+
|
108 |
+
loss_bin = self.BCEbins(pred_bin, target_bins) # BCE
|
109 |
+
|
110 |
+
if self.use_loss_regression:
|
111 |
+
loss_regression = self.MSELoss(result, target) # MSE
|
112 |
+
loss = loss_bin + loss_regression
|
113 |
+
else:
|
114 |
+
loss = loss_bin
|
115 |
+
|
116 |
+
out_result = result.clamp(min=self.min, max=self.max)
|
117 |
+
|
118 |
+
return loss, out_result
|
119 |
+
|
120 |
+
|
121 |
+
class FocalLoss(nn.Module):
|
122 |
+
# Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
|
123 |
+
def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
|
124 |
+
super(FocalLoss, self).__init__()
|
125 |
+
self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
|
126 |
+
self.gamma = gamma
|
127 |
+
self.alpha = alpha
|
128 |
+
self.reduction = loss_fcn.reduction
|
129 |
+
self.loss_fcn.reduction = 'none' # required to apply FL to each element
|
130 |
+
|
131 |
+
def forward(self, pred, true):
|
132 |
+
loss = self.loss_fcn(pred, true)
|
133 |
+
# p_t = torch.exp(-loss)
|
134 |
+
# loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability
|
135 |
+
|
136 |
+
# TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
|
137 |
+
pred_prob = torch.sigmoid(pred) # prob from logits
|
138 |
+
p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
|
139 |
+
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
|
140 |
+
modulating_factor = (1.0 - p_t) ** self.gamma
|
141 |
+
loss *= alpha_factor * modulating_factor
|
142 |
+
|
143 |
+
if self.reduction == 'mean':
|
144 |
+
return loss.mean()
|
145 |
+
elif self.reduction == 'sum':
|
146 |
+
return loss.sum()
|
147 |
+
else: # 'none'
|
148 |
+
return loss
|
149 |
+
|
150 |
+
|
151 |
+
class QFocalLoss(nn.Module):
|
152 |
+
# Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
|
153 |
+
def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
|
154 |
+
super(QFocalLoss, self).__init__()
|
155 |
+
self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss()
|
156 |
+
self.gamma = gamma
|
157 |
+
self.alpha = alpha
|
158 |
+
self.reduction = loss_fcn.reduction
|
159 |
+
self.loss_fcn.reduction = 'none' # required to apply FL to each element
|
160 |
+
|
161 |
+
def forward(self, pred, true):
|
162 |
+
loss = self.loss_fcn(pred, true)
|
163 |
+
|
164 |
+
pred_prob = torch.sigmoid(pred) # prob from logits
|
165 |
+
alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
|
166 |
+
modulating_factor = torch.abs(true - pred_prob) ** self.gamma
|
167 |
+
loss *= alpha_factor * modulating_factor
|
168 |
+
|
169 |
+
if self.reduction == 'mean':
|
170 |
+
return loss.mean()
|
171 |
+
elif self.reduction == 'sum':
|
172 |
+
return loss.sum()
|
173 |
+
else: # 'none'
|
174 |
+
return loss
|
175 |
+
|
176 |
+
class RankSort(torch.autograd.Function):
|
177 |
+
@staticmethod
|
178 |
+
def forward(ctx, logits, targets, delta_RS=0.50, eps=1e-10):
|
179 |
+
|
180 |
+
classification_grads=torch.zeros(logits.shape).cuda()
|
181 |
+
|
182 |
+
#Filter fg logits
|
183 |
+
fg_labels = (targets > 0.)
|
184 |
+
fg_logits = logits[fg_labels]
|
185 |
+
fg_targets = targets[fg_labels]
|
186 |
+
fg_num = len(fg_logits)
|
187 |
+
|
188 |
+
#Do not use bg with scores less than minimum fg logit
|
189 |
+
#since changing its score does not have an effect on precision
|
190 |
+
threshold_logit = torch.min(fg_logits)-delta_RS
|
191 |
+
relevant_bg_labels=((targets==0) & (logits>=threshold_logit))
|
192 |
+
|
193 |
+
relevant_bg_logits = logits[relevant_bg_labels]
|
194 |
+
relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
|
195 |
+
sorting_error=torch.zeros(fg_num).cuda()
|
196 |
+
ranking_error=torch.zeros(fg_num).cuda()
|
197 |
+
fg_grad=torch.zeros(fg_num).cuda()
|
198 |
+
|
199 |
+
#sort the fg logits
|
200 |
+
order=torch.argsort(fg_logits)
|
201 |
+
#Loops over each positive following the order
|
202 |
+
for ii in order:
|
203 |
+
# Difference Transforms (x_ij)
|
204 |
+
fg_relations=fg_logits-fg_logits[ii]
|
205 |
+
bg_relations=relevant_bg_logits-fg_logits[ii]
|
206 |
+
|
207 |
+
if delta_RS > 0:
|
208 |
+
fg_relations=torch.clamp(fg_relations/(2*delta_RS)+0.5,min=0,max=1)
|
209 |
+
bg_relations=torch.clamp(bg_relations/(2*delta_RS)+0.5,min=0,max=1)
|
210 |
+
else:
|
211 |
+
fg_relations = (fg_relations >= 0).float()
|
212 |
+
bg_relations = (bg_relations >= 0).float()
|
213 |
+
|
214 |
+
# Rank of ii among pos and false positive number (bg with larger scores)
|
215 |
+
rank_pos=torch.sum(fg_relations)
|
216 |
+
FP_num=torch.sum(bg_relations)
|
217 |
+
|
218 |
+
# Rank of ii among all examples
|
219 |
+
rank=rank_pos+FP_num
|
220 |
+
|
221 |
+
# Ranking error of example ii. target_ranking_error is always 0. (Eq. 7)
|
222 |
+
ranking_error[ii]=FP_num/rank
|
223 |
+
|
224 |
+
# Current sorting error of example ii. (Eq. 7)
|
225 |
+
current_sorting_error = torch.sum(fg_relations*(1-fg_targets))/rank_pos
|
226 |
+
|
227 |
+
#Find examples in the target sorted order for example ii
|
228 |
+
iou_relations = (fg_targets >= fg_targets[ii])
|
229 |
+
target_sorted_order = iou_relations * fg_relations
|
230 |
+
|
231 |
+
#The rank of ii among positives in sorted order
|
232 |
+
rank_pos_target = torch.sum(target_sorted_order)
|
233 |
+
|
234 |
+
#Compute target sorting error. (Eq. 8)
|
235 |
+
#Since target ranking error is 0, this is also total target error
|
236 |
+
target_sorting_error= torch.sum(target_sorted_order*(1-fg_targets))/rank_pos_target
|
237 |
+
|
238 |
+
#Compute sorting error on example ii
|
239 |
+
sorting_error[ii] = current_sorting_error - target_sorting_error
|
240 |
+
|
241 |
+
#Identity Update for Ranking Error
|
242 |
+
if FP_num > eps:
|
243 |
+
#For ii the update is the ranking error
|
244 |
+
fg_grad[ii] -= ranking_error[ii]
|
245 |
+
#For negatives, distribute error via ranking pmf (i.e. bg_relations/FP_num)
|
246 |
+
relevant_bg_grad += (bg_relations*(ranking_error[ii]/FP_num))
|
247 |
+
|
248 |
+
#Find the positives that are misranked (the cause of the error)
|
249 |
+
#These are the ones with smaller IoU but larger logits
|
250 |
+
missorted_examples = (~ iou_relations) * fg_relations
|
251 |
+
|
252 |
+
#Denominotor of sorting pmf
|
253 |
+
sorting_pmf_denom = torch.sum(missorted_examples)
|
254 |
+
|
255 |
+
#Identity Update for Sorting Error
|
256 |
+
if sorting_pmf_denom > eps:
|
257 |
+
#For ii the update is the sorting error
|
258 |
+
fg_grad[ii] -= sorting_error[ii]
|
259 |
+
#For positives, distribute error via sorting pmf (i.e. missorted_examples/sorting_pmf_denom)
|
260 |
+
fg_grad += (missorted_examples*(sorting_error[ii]/sorting_pmf_denom))
|
261 |
+
|
262 |
+
#Normalize gradients by number of positives
|
263 |
+
classification_grads[fg_labels]= (fg_grad/fg_num)
|
264 |
+
classification_grads[relevant_bg_labels]= (relevant_bg_grad/fg_num)
|
265 |
+
|
266 |
+
ctx.save_for_backward(classification_grads)
|
267 |
+
|
268 |
+
return ranking_error.mean(), sorting_error.mean()
|
269 |
+
|
270 |
+
@staticmethod
|
271 |
+
def backward(ctx, out_grad1, out_grad2):
|
272 |
+
g1, =ctx.saved_tensors
|
273 |
+
return g1*out_grad1, None, None, None
|
274 |
+
|
275 |
+
class aLRPLoss(torch.autograd.Function):
|
276 |
+
@staticmethod
|
277 |
+
def forward(ctx, logits, targets, regression_losses, delta=1., eps=1e-5):
|
278 |
+
classification_grads=torch.zeros(logits.shape).cuda()
|
279 |
+
|
280 |
+
#Filter fg logits
|
281 |
+
fg_labels = (targets == 1)
|
282 |
+
fg_logits = logits[fg_labels]
|
283 |
+
fg_num = len(fg_logits)
|
284 |
+
|
285 |
+
#Do not use bg with scores less than minimum fg logit
|
286 |
+
#since changing its score does not have an effect on precision
|
287 |
+
threshold_logit = torch.min(fg_logits)-delta
|
288 |
+
|
289 |
+
#Get valid bg logits
|
290 |
+
relevant_bg_labels=((targets==0)&(logits>=threshold_logit))
|
291 |
+
relevant_bg_logits=logits[relevant_bg_labels]
|
292 |
+
relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
|
293 |
+
rank=torch.zeros(fg_num).cuda()
|
294 |
+
prec=torch.zeros(fg_num).cuda()
|
295 |
+
fg_grad=torch.zeros(fg_num).cuda()
|
296 |
+
|
297 |
+
max_prec=0
|
298 |
+
#sort the fg logits
|
299 |
+
order=torch.argsort(fg_logits)
|
300 |
+
#Loops over each positive following the order
|
301 |
+
for ii in order:
|
302 |
+
#x_ij s as score differences with fgs
|
303 |
+
fg_relations=fg_logits-fg_logits[ii]
|
304 |
+
#Apply piecewise linear function and determine relations with fgs
|
305 |
+
fg_relations=torch.clamp(fg_relations/(2*delta)+0.5,min=0,max=1)
|
306 |
+
#Discard i=j in the summation in rank_pos
|
307 |
+
fg_relations[ii]=0
|
308 |
+
|
309 |
+
#x_ij s as score differences with bgs
|
310 |
+
bg_relations=relevant_bg_logits-fg_logits[ii]
|
311 |
+
#Apply piecewise linear function and determine relations with bgs
|
312 |
+
bg_relations=torch.clamp(bg_relations/(2*delta)+0.5,min=0,max=1)
|
313 |
+
|
314 |
+
#Compute the rank of the example within fgs and number of bgs with larger scores
|
315 |
+
rank_pos=1+torch.sum(fg_relations)
|
316 |
+
FP_num=torch.sum(bg_relations)
|
317 |
+
#Store the total since it is normalizer also for aLRP Regression error
|
318 |
+
rank[ii]=rank_pos+FP_num
|
319 |
+
|
320 |
+
#Compute precision for this example to compute classification loss
|
321 |
+
prec[ii]=rank_pos/rank[ii]
|
322 |
+
#For stability, set eps to a infinitesmall value (e.g. 1e-6), then compute grads
|
323 |
+
if FP_num > eps:
|
324 |
+
fg_grad[ii] = -(torch.sum(fg_relations*regression_losses)+FP_num)/rank[ii]
|
325 |
+
relevant_bg_grad += (bg_relations*(-fg_grad[ii]/FP_num))
|
326 |
+
|
327 |
+
#aLRP with grad formulation fg gradient
|
328 |
+
classification_grads[fg_labels]= fg_grad
|
329 |
+
#aLRP with grad formulation bg gradient
|
330 |
+
classification_grads[relevant_bg_labels]= relevant_bg_grad
|
331 |
+
|
332 |
+
classification_grads /= (fg_num)
|
333 |
+
|
334 |
+
cls_loss=1-prec.mean()
|
335 |
+
ctx.save_for_backward(classification_grads)
|
336 |
+
|
337 |
+
return cls_loss, rank, order
|
338 |
+
|
339 |
+
@staticmethod
|
340 |
+
def backward(ctx, out_grad1, out_grad2, out_grad3):
|
341 |
+
g1, =ctx.saved_tensors
|
342 |
+
return g1*out_grad1, None, None, None, None
|
343 |
+
|
344 |
+
|
345 |
+
class APLoss(torch.autograd.Function):
|
346 |
+
@staticmethod
|
347 |
+
def forward(ctx, logits, targets, delta=1.):
|
348 |
+
classification_grads=torch.zeros(logits.shape).cuda()
|
349 |
+
|
350 |
+
#Filter fg logits
|
351 |
+
fg_labels = (targets == 1)
|
352 |
+
fg_logits = logits[fg_labels]
|
353 |
+
fg_num = len(fg_logits)
|
354 |
+
|
355 |
+
#Do not use bg with scores less than minimum fg logit
|
356 |
+
#since changing its score does not have an effect on precision
|
357 |
+
threshold_logit = torch.min(fg_logits)-delta
|
358 |
+
|
359 |
+
#Get valid bg logits
|
360 |
+
relevant_bg_labels=((targets==0)&(logits>=threshold_logit))
|
361 |
+
relevant_bg_logits=logits[relevant_bg_labels]
|
362 |
+
relevant_bg_grad=torch.zeros(len(relevant_bg_logits)).cuda()
|
363 |
+
rank=torch.zeros(fg_num).cuda()
|
364 |
+
prec=torch.zeros(fg_num).cuda()
|
365 |
+
fg_grad=torch.zeros(fg_num).cuda()
|
366 |
+
|
367 |
+
max_prec=0
|
368 |
+
#sort the fg logits
|
369 |
+
order=torch.argsort(fg_logits)
|
370 |
+
#Loops over each positive following the order
|
371 |
+
for ii in order:
|
372 |
+
#x_ij s as score differences with fgs
|
373 |
+
fg_relations=fg_logits-fg_logits[ii]
|
374 |
+
#Apply piecewise linear function and determine relations with fgs
|
375 |
+
fg_relations=torch.clamp(fg_relations/(2*delta)+0.5,min=0,max=1)
|
376 |
+
#Discard i=j in the summation in rank_pos
|
377 |
+
fg_relations[ii]=0
|
378 |
+
|
379 |
+
#x_ij s as score differences with bgs
|
380 |
+
bg_relations=relevant_bg_logits-fg_logits[ii]
|
381 |
+
#Apply piecewise linear function and determine relations with bgs
|
382 |
+
bg_relations=torch.clamp(bg_relations/(2*delta)+0.5,min=0,max=1)
|
383 |
+
|
384 |
+
#Compute the rank of the example within fgs and number of bgs with larger scores
|
385 |
+
rank_pos=1+torch.sum(fg_relations)
|
386 |
+
FP_num=torch.sum(bg_relations)
|
387 |
+
#Store the total since it is normalizer also for aLRP Regression error
|
388 |
+
rank[ii]=rank_pos+FP_num
|
389 |
+
|
390 |
+
#Compute precision for this example
|
391 |
+
current_prec=rank_pos/rank[ii]
|
392 |
+
|
393 |
+
#Compute interpolated AP and store gradients for relevant bg examples
|
394 |
+
if (max_prec<=current_prec):
|
395 |
+
max_prec=current_prec
|
396 |
+
relevant_bg_grad += (bg_relations/rank[ii])
|
397 |
+
else:
|
398 |
+
relevant_bg_grad += (bg_relations/rank[ii])*(((1-max_prec)/(1-current_prec)))
|
399 |
+
|
400 |
+
#Store fg gradients
|
401 |
+
fg_grad[ii]=-(1-max_prec)
|
402 |
+
prec[ii]=max_prec
|
403 |
+
|
404 |
+
#aLRP with grad formulation fg gradient
|
405 |
+
classification_grads[fg_labels]= fg_grad
|
406 |
+
#aLRP with grad formulation bg gradient
|
407 |
+
classification_grads[relevant_bg_labels]= relevant_bg_grad
|
408 |
+
|
409 |
+
classification_grads /= fg_num
|
410 |
+
|
411 |
+
cls_loss=1-prec.mean()
|
412 |
+
ctx.save_for_backward(classification_grads)
|
413 |
+
|
414 |
+
return cls_loss
|
415 |
+
|
416 |
+
@staticmethod
|
417 |
+
def backward(ctx, out_grad1):
|
418 |
+
g1, =ctx.saved_tensors
|
419 |
+
return g1*out_grad1, None, None
|
420 |
+
|
421 |
+
|
422 |
+
class ComputeLoss:
|
423 |
+
# Compute losses
|
424 |
+
def __init__(self, model, autobalance=False):
|
425 |
+
super(ComputeLoss, self).__init__()
|
426 |
+
device = next(model.parameters()).device # get model device
|
427 |
+
h = model.hyp # hyperparameters
|
428 |
+
|
429 |
+
# Define criteria
|
430 |
+
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
|
431 |
+
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
|
432 |
+
|
433 |
+
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
|
434 |
+
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
|
435 |
+
|
436 |
+
# Focal loss
|
437 |
+
g = h['fl_gamma'] # focal loss gamma
|
438 |
+
if g > 0:
|
439 |
+
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
|
440 |
+
|
441 |
+
det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
|
442 |
+
self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
|
443 |
+
#self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.1, .05]) # P3-P7
|
444 |
+
#self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.5, 0.4, .1]) # P3-P7
|
445 |
+
self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
|
446 |
+
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
|
447 |
+
for k in 'na', 'nc', 'nl', 'anchors':
|
448 |
+
setattr(self, k, getattr(det, k))
|
449 |
+
|
450 |
+
def __call__(self, p, targets): # predictions, targets, model
|
451 |
+
device = targets.device
|
452 |
+
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
|
453 |
+
tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets
|
454 |
+
|
455 |
+
# Losses
|
456 |
+
for i, pi in enumerate(p): # layer index, layer predictions
|
457 |
+
b, a, gj, gi = indices[i] # image, anchor, gridy, gridx
|
458 |
+
tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
|
459 |
+
|
460 |
+
n = b.shape[0] # number of targets
|
461 |
+
if n:
|
462 |
+
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
|
463 |
+
|
464 |
+
# Regression
|
465 |
+
pxy = ps[:, :2].sigmoid() * 2. - 0.5
|
466 |
+
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
|
467 |
+
pbox = torch.cat((pxy, pwh), 1) # predicted box
|
468 |
+
iou = bbox_iou(pbox.T, tbox[i], x1y1x2y2=False, CIoU=True) # iou(prediction, target)
|
469 |
+
lbox += (1.0 - iou).mean() # iou loss
|
470 |
+
|
471 |
+
# Objectness
|
472 |
+
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
|
473 |
+
|
474 |
+
# Classification
|
475 |
+
if self.nc > 1: # cls loss (only if multiple classes)
|
476 |
+
t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
|
477 |
+
t[range(n), tcls[i]] = self.cp
|
478 |
+
#t[t==self.cp] = iou.detach().clamp(0).type(t.dtype)
|
479 |
+
lcls += self.BCEcls(ps[:, 5:], t) # BCE
|
480 |
+
|
481 |
+
# Append targets to text file
|
482 |
+
# with open('targets.txt', 'a') as file:
|
483 |
+
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
|
484 |
+
|
485 |
+
obji = self.BCEobj(pi[..., 4], tobj)
|
486 |
+
lobj += obji * self.balance[i] # obj loss
|
487 |
+
if self.autobalance:
|
488 |
+
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
|
489 |
+
|
490 |
+
if self.autobalance:
|
491 |
+
self.balance = [x / self.balance[self.ssi] for x in self.balance]
|
492 |
+
lbox *= self.hyp['box']
|
493 |
+
lobj *= self.hyp['obj']
|
494 |
+
lcls *= self.hyp['cls']
|
495 |
+
bs = tobj.shape[0] # batch size
|
496 |
+
|
497 |
+
loss = lbox + lobj + lcls
|
498 |
+
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
|
499 |
+
|
500 |
+
def build_targets(self, p, targets):
|
501 |
+
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
|
502 |
+
na, nt = self.na, targets.shape[0] # number of anchors, targets
|
503 |
+
tcls, tbox, indices, anch = [], [], [], []
|
504 |
+
gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
|
505 |
+
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
|
506 |
+
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
|
507 |
+
|
508 |
+
g = 0.5 # bias
|
509 |
+
off = torch.tensor([[0, 0],
|
510 |
+
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
|
511 |
+
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
|
512 |
+
], device=targets.device).float() * g # offsets
|
513 |
+
|
514 |
+
for i in range(self.nl):
|
515 |
+
anchors = self.anchors[i]
|
516 |
+
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
|
517 |
+
|
518 |
+
# Match targets to anchors
|
519 |
+
t = targets * gain
|
520 |
+
if nt:
|
521 |
+
# Matches
|
522 |
+
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
|
523 |
+
j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
|
524 |
+
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
|
525 |
+
t = t[j] # filter
|
526 |
+
|
527 |
+
# Offsets
|
528 |
+
gxy = t[:, 2:4] # grid xy
|
529 |
+
gxi = gain[[2, 3]] - gxy # inverse
|
530 |
+
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
|
531 |
+
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
|
532 |
+
j = torch.stack((torch.ones_like(j), j, k, l, m))
|
533 |
+
t = t.repeat((5, 1, 1))[j]
|
534 |
+
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
|
535 |
+
else:
|
536 |
+
t = targets[0]
|
537 |
+
offsets = 0
|
538 |
+
|
539 |
+
# Define
|
540 |
+
b, c = t[:, :2].long().T # image, class
|
541 |
+
gxy = t[:, 2:4] # grid xy
|
542 |
+
gwh = t[:, 4:6] # grid wh
|
543 |
+
gij = (gxy - offsets).long()
|
544 |
+
gi, gj = gij.T # grid xy indices
|
545 |
+
|
546 |
+
# Append
|
547 |
+
a = t[:, 6].long() # anchor indices
|
548 |
+
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
|
549 |
+
tbox.append(torch.cat((gxy - gij, gwh), 1)) # box
|
550 |
+
anch.append(anchors[a]) # anchors
|
551 |
+
tcls.append(c) # class
|
552 |
+
|
553 |
+
return tcls, tbox, indices, anch
|
554 |
+
|
555 |
+
|
556 |
+
class ComputeLossOTA:
|
557 |
+
# Compute losses
|
558 |
+
def __init__(self, model, autobalance=False):
|
559 |
+
super(ComputeLossOTA, self).__init__()
|
560 |
+
device = next(model.parameters()).device # get model device
|
561 |
+
h = model.hyp # hyperparameters
|
562 |
+
|
563 |
+
# Define criteria
|
564 |
+
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
|
565 |
+
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
|
566 |
+
|
567 |
+
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
|
568 |
+
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
|
569 |
+
|
570 |
+
# Focal loss
|
571 |
+
g = h['fl_gamma'] # focal loss gamma
|
572 |
+
if g > 0:
|
573 |
+
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
|
574 |
+
|
575 |
+
det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
|
576 |
+
self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
|
577 |
+
self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
|
578 |
+
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
|
579 |
+
for k in 'na', 'nc', 'nl', 'anchors', 'stride':
|
580 |
+
setattr(self, k, getattr(det, k))
|
581 |
+
|
582 |
+
def __call__(self, p, targets, imgs): # predictions, targets, model
|
583 |
+
device = targets.device
|
584 |
+
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
|
585 |
+
bs, as_, gjs, gis, targets, anchors = self.build_targets(p, targets, imgs)
|
586 |
+
pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p]
|
587 |
+
|
588 |
+
|
589 |
+
# Losses
|
590 |
+
for i, pi in enumerate(p): # layer index, layer predictions
|
591 |
+
b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i] # image, anchor, gridy, gridx
|
592 |
+
tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
|
593 |
+
|
594 |
+
n = b.shape[0] # number of targets
|
595 |
+
if n:
|
596 |
+
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
|
597 |
+
|
598 |
+
# Regression
|
599 |
+
grid = torch.stack([gi, gj], dim=1)
|
600 |
+
pxy = ps[:, :2].sigmoid() * 2. - 0.5
|
601 |
+
#pxy = ps[:, :2].sigmoid() * 3. - 1.
|
602 |
+
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
|
603 |
+
pbox = torch.cat((pxy, pwh), 1) # predicted box
|
604 |
+
selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
|
605 |
+
selected_tbox[:, :2] -= grid
|
606 |
+
iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True) # iou(prediction, target)
|
607 |
+
lbox += (1.0 - iou).mean() # iou loss
|
608 |
+
|
609 |
+
# Objectness
|
610 |
+
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
|
611 |
+
|
612 |
+
# Classification
|
613 |
+
selected_tcls = targets[i][:, 1].long()
|
614 |
+
if self.nc > 1: # cls loss (only if multiple classes)
|
615 |
+
t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
|
616 |
+
t[range(n), selected_tcls] = self.cp
|
617 |
+
lcls += self.BCEcls(ps[:, 5:], t) # BCE
|
618 |
+
|
619 |
+
# Append targets to text file
|
620 |
+
# with open('targets.txt', 'a') as file:
|
621 |
+
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
|
622 |
+
|
623 |
+
obji = self.BCEobj(pi[..., 4], tobj)
|
624 |
+
lobj += obji * self.balance[i] # obj loss
|
625 |
+
if self.autobalance:
|
626 |
+
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
|
627 |
+
|
628 |
+
if self.autobalance:
|
629 |
+
self.balance = [x / self.balance[self.ssi] for x in self.balance]
|
630 |
+
lbox *= self.hyp['box']
|
631 |
+
lobj *= self.hyp['obj']
|
632 |
+
lcls *= self.hyp['cls']
|
633 |
+
bs = tobj.shape[0] # batch size
|
634 |
+
|
635 |
+
loss = lbox + lobj + lcls
|
636 |
+
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
|
637 |
+
|
638 |
+
def build_targets(self, p, targets, imgs):
|
639 |
+
|
640 |
+
#indices, anch = self.find_positive(p, targets)
|
641 |
+
indices, anch = self.find_3_positive(p, targets)
|
642 |
+
#indices, anch = self.find_4_positive(p, targets)
|
643 |
+
#indices, anch = self.find_5_positive(p, targets)
|
644 |
+
#indices, anch = self.find_9_positive(p, targets)
|
645 |
+
|
646 |
+
matching_bs = [[] for pp in p]
|
647 |
+
matching_as = [[] for pp in p]
|
648 |
+
matching_gjs = [[] for pp in p]
|
649 |
+
matching_gis = [[] for pp in p]
|
650 |
+
matching_targets = [[] for pp in p]
|
651 |
+
matching_anchs = [[] for pp in p]
|
652 |
+
|
653 |
+
nl = len(p)
|
654 |
+
|
655 |
+
for batch_idx in range(p[0].shape[0]):
|
656 |
+
|
657 |
+
b_idx = targets[:, 0]==batch_idx
|
658 |
+
this_target = targets[b_idx]
|
659 |
+
if this_target.shape[0] == 0:
|
660 |
+
continue
|
661 |
+
|
662 |
+
txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
|
663 |
+
txyxy = xywh2xyxy(txywh)
|
664 |
+
|
665 |
+
pxyxys = []
|
666 |
+
p_cls = []
|
667 |
+
p_obj = []
|
668 |
+
from_which_layer = []
|
669 |
+
all_b = []
|
670 |
+
all_a = []
|
671 |
+
all_gj = []
|
672 |
+
all_gi = []
|
673 |
+
all_anch = []
|
674 |
+
|
675 |
+
for i, pi in enumerate(p):
|
676 |
+
|
677 |
+
b, a, gj, gi = indices[i]
|
678 |
+
idx = (b == batch_idx)
|
679 |
+
b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]
|
680 |
+
all_b.append(b)
|
681 |
+
all_a.append(a)
|
682 |
+
all_gj.append(gj)
|
683 |
+
all_gi.append(gi)
|
684 |
+
all_anch.append(anch[i][idx])
|
685 |
+
from_which_layer.append(torch.ones(size=(len(b),)) * i)
|
686 |
+
|
687 |
+
fg_pred = pi[b, a, gj, gi]
|
688 |
+
p_obj.append(fg_pred[:, 4:5])
|
689 |
+
p_cls.append(fg_pred[:, 5:])
|
690 |
+
|
691 |
+
grid = torch.stack([gi, gj], dim=1)
|
692 |
+
pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
|
693 |
+
#pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
|
694 |
+
pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
|
695 |
+
pxywh = torch.cat([pxy, pwh], dim=-1)
|
696 |
+
pxyxy = xywh2xyxy(pxywh)
|
697 |
+
pxyxys.append(pxyxy)
|
698 |
+
|
699 |
+
pxyxys = torch.cat(pxyxys, dim=0)
|
700 |
+
if pxyxys.shape[0] == 0:
|
701 |
+
continue
|
702 |
+
p_obj = torch.cat(p_obj, dim=0)
|
703 |
+
p_cls = torch.cat(p_cls, dim=0)
|
704 |
+
from_which_layer = torch.cat(from_which_layer, dim=0)
|
705 |
+
all_b = torch.cat(all_b, dim=0)
|
706 |
+
all_a = torch.cat(all_a, dim=0)
|
707 |
+
all_gj = torch.cat(all_gj, dim=0)
|
708 |
+
all_gi = torch.cat(all_gi, dim=0)
|
709 |
+
all_anch = torch.cat(all_anch, dim=0)
|
710 |
+
|
711 |
+
pair_wise_iou = box_iou(txyxy, pxyxys)
|
712 |
+
|
713 |
+
pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
|
714 |
+
|
715 |
+
top_k, _ = torch.topk(pair_wise_iou, min(10, pair_wise_iou.shape[1]), dim=1)
|
716 |
+
dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
|
717 |
+
|
718 |
+
gt_cls_per_image = (
|
719 |
+
F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
|
720 |
+
.float()
|
721 |
+
.unsqueeze(1)
|
722 |
+
.repeat(1, pxyxys.shape[0], 1)
|
723 |
+
)
|
724 |
+
|
725 |
+
num_gt = this_target.shape[0]
|
726 |
+
cls_preds_ = (
|
727 |
+
p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
728 |
+
* p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
729 |
+
)
|
730 |
+
|
731 |
+
y = cls_preds_.sqrt_()
|
732 |
+
pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
|
733 |
+
torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
|
734 |
+
).sum(-1)
|
735 |
+
del cls_preds_
|
736 |
+
|
737 |
+
cost = (
|
738 |
+
pair_wise_cls_loss
|
739 |
+
+ 3.0 * pair_wise_iou_loss
|
740 |
+
)
|
741 |
+
|
742 |
+
matching_matrix = torch.zeros_like(cost)
|
743 |
+
|
744 |
+
for gt_idx in range(num_gt):
|
745 |
+
_, pos_idx = torch.topk(
|
746 |
+
cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
|
747 |
+
)
|
748 |
+
matching_matrix[gt_idx][pos_idx] = 1.0
|
749 |
+
|
750 |
+
del top_k, dynamic_ks
|
751 |
+
anchor_matching_gt = matching_matrix.sum(0)
|
752 |
+
if (anchor_matching_gt > 1).sum() > 0:
|
753 |
+
_, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
|
754 |
+
matching_matrix[:, anchor_matching_gt > 1] *= 0.0
|
755 |
+
matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
|
756 |
+
fg_mask_inboxes = matching_matrix.sum(0) > 0.0
|
757 |
+
matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
|
758 |
+
|
759 |
+
from_which_layer = from_which_layer[fg_mask_inboxes]
|
760 |
+
all_b = all_b[fg_mask_inboxes]
|
761 |
+
all_a = all_a[fg_mask_inboxes]
|
762 |
+
all_gj = all_gj[fg_mask_inboxes]
|
763 |
+
all_gi = all_gi[fg_mask_inboxes]
|
764 |
+
all_anch = all_anch[fg_mask_inboxes]
|
765 |
+
|
766 |
+
this_target = this_target[matched_gt_inds]
|
767 |
+
|
768 |
+
for i in range(nl):
|
769 |
+
layer_idx = from_which_layer == i
|
770 |
+
matching_bs[i].append(all_b[layer_idx])
|
771 |
+
matching_as[i].append(all_a[layer_idx])
|
772 |
+
matching_gjs[i].append(all_gj[layer_idx])
|
773 |
+
matching_gis[i].append(all_gi[layer_idx])
|
774 |
+
matching_targets[i].append(this_target[layer_idx])
|
775 |
+
matching_anchs[i].append(all_anch[layer_idx])
|
776 |
+
|
777 |
+
for i in range(nl):
|
778 |
+
if matching_targets[i] != []:
|
779 |
+
matching_bs[i] = torch.cat(matching_bs[i], dim=0)
|
780 |
+
matching_as[i] = torch.cat(matching_as[i], dim=0)
|
781 |
+
matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
|
782 |
+
matching_gis[i] = torch.cat(matching_gis[i], dim=0)
|
783 |
+
matching_targets[i] = torch.cat(matching_targets[i], dim=0)
|
784 |
+
matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
|
785 |
+
else:
|
786 |
+
matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
787 |
+
matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
788 |
+
matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
789 |
+
matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
790 |
+
matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
791 |
+
matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
792 |
+
|
793 |
+
return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs
|
794 |
+
|
795 |
+
def find_3_positive(self, p, targets):
|
796 |
+
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
|
797 |
+
na, nt = self.na, targets.shape[0] # number of anchors, targets
|
798 |
+
indices, anch = [], []
|
799 |
+
gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
|
800 |
+
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
|
801 |
+
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
|
802 |
+
|
803 |
+
g = 0.5 # bias
|
804 |
+
off = torch.tensor([[0, 0],
|
805 |
+
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
|
806 |
+
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
|
807 |
+
], device=targets.device).float() * g # offsets
|
808 |
+
|
809 |
+
for i in range(self.nl):
|
810 |
+
anchors = self.anchors[i]
|
811 |
+
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
|
812 |
+
|
813 |
+
# Match targets to anchors
|
814 |
+
t = targets * gain
|
815 |
+
if nt:
|
816 |
+
# Matches
|
817 |
+
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
|
818 |
+
j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
|
819 |
+
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
|
820 |
+
t = t[j] # filter
|
821 |
+
|
822 |
+
# Offsets
|
823 |
+
gxy = t[:, 2:4] # grid xy
|
824 |
+
gxi = gain[[2, 3]] - gxy # inverse
|
825 |
+
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
|
826 |
+
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
|
827 |
+
j = torch.stack((torch.ones_like(j), j, k, l, m))
|
828 |
+
t = t.repeat((5, 1, 1))[j]
|
829 |
+
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
|
830 |
+
else:
|
831 |
+
t = targets[0]
|
832 |
+
offsets = 0
|
833 |
+
|
834 |
+
# Define
|
835 |
+
b, c = t[:, :2].long().T # image, class
|
836 |
+
gxy = t[:, 2:4] # grid xy
|
837 |
+
gwh = t[:, 4:6] # grid wh
|
838 |
+
gij = (gxy - offsets).long()
|
839 |
+
gi, gj = gij.T # grid xy indices
|
840 |
+
|
841 |
+
# Append
|
842 |
+
a = t[:, 6].long() # anchor indices
|
843 |
+
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
|
844 |
+
anch.append(anchors[a]) # anchors
|
845 |
+
|
846 |
+
return indices, anch
|
847 |
+
|
848 |
+
|
849 |
+
class ComputeLossBinOTA:
|
850 |
+
# Compute losses
|
851 |
+
def __init__(self, model, autobalance=False):
|
852 |
+
super(ComputeLossBinOTA, self).__init__()
|
853 |
+
device = next(model.parameters()).device # get model device
|
854 |
+
h = model.hyp # hyperparameters
|
855 |
+
|
856 |
+
# Define criteria
|
857 |
+
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
|
858 |
+
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
|
859 |
+
#MSEangle = nn.MSELoss().to(device)
|
860 |
+
|
861 |
+
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
|
862 |
+
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
|
863 |
+
|
864 |
+
# Focal loss
|
865 |
+
g = h['fl_gamma'] # focal loss gamma
|
866 |
+
if g > 0:
|
867 |
+
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
|
868 |
+
|
869 |
+
det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
|
870 |
+
self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
|
871 |
+
self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
|
872 |
+
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
|
873 |
+
for k in 'na', 'nc', 'nl', 'anchors', 'stride', 'bin_count':
|
874 |
+
setattr(self, k, getattr(det, k))
|
875 |
+
|
876 |
+
#xy_bin_sigmoid = SigmoidBin(bin_count=11, min=-0.5, max=1.5, use_loss_regression=False).to(device)
|
877 |
+
wh_bin_sigmoid = SigmoidBin(bin_count=self.bin_count, min=0.0, max=4.0, use_loss_regression=False).to(device)
|
878 |
+
#angle_bin_sigmoid = SigmoidBin(bin_count=31, min=-1.1, max=1.1, use_loss_regression=False).to(device)
|
879 |
+
self.wh_bin_sigmoid = wh_bin_sigmoid
|
880 |
+
|
881 |
+
def __call__(self, p, targets, imgs): # predictions, targets, model
|
882 |
+
device = targets.device
|
883 |
+
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
|
884 |
+
bs, as_, gjs, gis, targets, anchors = self.build_targets(p, targets, imgs)
|
885 |
+
pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p]
|
886 |
+
|
887 |
+
|
888 |
+
# Losses
|
889 |
+
for i, pi in enumerate(p): # layer index, layer predictions
|
890 |
+
b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i] # image, anchor, gridy, gridx
|
891 |
+
tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
|
892 |
+
|
893 |
+
obj_idx = self.wh_bin_sigmoid.get_length()*2 + 2 # x,y, w-bce, h-bce # xy_bin_sigmoid.get_length()*2
|
894 |
+
|
895 |
+
n = b.shape[0] # number of targets
|
896 |
+
if n:
|
897 |
+
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
|
898 |
+
|
899 |
+
# Regression
|
900 |
+
grid = torch.stack([gi, gj], dim=1)
|
901 |
+
selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
|
902 |
+
selected_tbox[:, :2] -= grid
|
903 |
+
|
904 |
+
#pxy = ps[:, :2].sigmoid() * 2. - 0.5
|
905 |
+
##pxy = ps[:, :2].sigmoid() * 3. - 1.
|
906 |
+
#pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
|
907 |
+
#pbox = torch.cat((pxy, pwh), 1) # predicted box
|
908 |
+
|
909 |
+
#x_loss, px = xy_bin_sigmoid.training_loss(ps[..., 0:12], tbox[i][..., 0])
|
910 |
+
#y_loss, py = xy_bin_sigmoid.training_loss(ps[..., 12:24], tbox[i][..., 1])
|
911 |
+
w_loss, pw = self.wh_bin_sigmoid.training_loss(ps[..., 2:(3+self.bin_count)], selected_tbox[..., 2] / anchors[i][..., 0])
|
912 |
+
h_loss, ph = self.wh_bin_sigmoid.training_loss(ps[..., (3+self.bin_count):obj_idx], selected_tbox[..., 3] / anchors[i][..., 1])
|
913 |
+
|
914 |
+
pw *= anchors[i][..., 0]
|
915 |
+
ph *= anchors[i][..., 1]
|
916 |
+
|
917 |
+
px = ps[:, 0].sigmoid() * 2. - 0.5
|
918 |
+
py = ps[:, 1].sigmoid() * 2. - 0.5
|
919 |
+
|
920 |
+
lbox += w_loss + h_loss # + x_loss + y_loss
|
921 |
+
|
922 |
+
#print(f"\n px = {px.shape}, py = {py.shape}, pw = {pw.shape}, ph = {ph.shape} \n")
|
923 |
+
|
924 |
+
pbox = torch.cat((px.unsqueeze(1), py.unsqueeze(1), pw.unsqueeze(1), ph.unsqueeze(1)), 1).to(device) # predicted box
|
925 |
+
|
926 |
+
|
927 |
+
|
928 |
+
|
929 |
+
iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True) # iou(prediction, target)
|
930 |
+
lbox += (1.0 - iou).mean() # iou loss
|
931 |
+
|
932 |
+
# Objectness
|
933 |
+
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
|
934 |
+
|
935 |
+
# Classification
|
936 |
+
selected_tcls = targets[i][:, 1].long()
|
937 |
+
if self.nc > 1: # cls loss (only if multiple classes)
|
938 |
+
t = torch.full_like(ps[:, (1+obj_idx):], self.cn, device=device) # targets
|
939 |
+
t[range(n), selected_tcls] = self.cp
|
940 |
+
lcls += self.BCEcls(ps[:, (1+obj_idx):], t) # BCE
|
941 |
+
|
942 |
+
# Append targets to text file
|
943 |
+
# with open('targets.txt', 'a') as file:
|
944 |
+
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
|
945 |
+
|
946 |
+
obji = self.BCEobj(pi[..., obj_idx], tobj)
|
947 |
+
lobj += obji * self.balance[i] # obj loss
|
948 |
+
if self.autobalance:
|
949 |
+
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
|
950 |
+
|
951 |
+
if self.autobalance:
|
952 |
+
self.balance = [x / self.balance[self.ssi] for x in self.balance]
|
953 |
+
lbox *= self.hyp['box']
|
954 |
+
lobj *= self.hyp['obj']
|
955 |
+
lcls *= self.hyp['cls']
|
956 |
+
bs = tobj.shape[0] # batch size
|
957 |
+
|
958 |
+
loss = lbox + lobj + lcls
|
959 |
+
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
|
960 |
+
|
961 |
+
def build_targets(self, p, targets, imgs):
|
962 |
+
|
963 |
+
#indices, anch = self.find_positive(p, targets)
|
964 |
+
indices, anch = self.find_3_positive(p, targets)
|
965 |
+
#indices, anch = self.find_4_positive(p, targets)
|
966 |
+
#indices, anch = self.find_5_positive(p, targets)
|
967 |
+
#indices, anch = self.find_9_positive(p, targets)
|
968 |
+
|
969 |
+
matching_bs = [[] for pp in p]
|
970 |
+
matching_as = [[] for pp in p]
|
971 |
+
matching_gjs = [[] for pp in p]
|
972 |
+
matching_gis = [[] for pp in p]
|
973 |
+
matching_targets = [[] for pp in p]
|
974 |
+
matching_anchs = [[] for pp in p]
|
975 |
+
|
976 |
+
nl = len(p)
|
977 |
+
|
978 |
+
for batch_idx in range(p[0].shape[0]):
|
979 |
+
|
980 |
+
b_idx = targets[:, 0]==batch_idx
|
981 |
+
this_target = targets[b_idx]
|
982 |
+
if this_target.shape[0] == 0:
|
983 |
+
continue
|
984 |
+
|
985 |
+
txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
|
986 |
+
txyxy = xywh2xyxy(txywh)
|
987 |
+
|
988 |
+
pxyxys = []
|
989 |
+
p_cls = []
|
990 |
+
p_obj = []
|
991 |
+
from_which_layer = []
|
992 |
+
all_b = []
|
993 |
+
all_a = []
|
994 |
+
all_gj = []
|
995 |
+
all_gi = []
|
996 |
+
all_anch = []
|
997 |
+
|
998 |
+
for i, pi in enumerate(p):
|
999 |
+
|
1000 |
+
obj_idx = self.wh_bin_sigmoid.get_length()*2 + 2
|
1001 |
+
|
1002 |
+
b, a, gj, gi = indices[i]
|
1003 |
+
idx = (b == batch_idx)
|
1004 |
+
b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]
|
1005 |
+
all_b.append(b)
|
1006 |
+
all_a.append(a)
|
1007 |
+
all_gj.append(gj)
|
1008 |
+
all_gi.append(gi)
|
1009 |
+
all_anch.append(anch[i][idx])
|
1010 |
+
from_which_layer.append(torch.ones(size=(len(b),)) * i)
|
1011 |
+
|
1012 |
+
fg_pred = pi[b, a, gj, gi]
|
1013 |
+
p_obj.append(fg_pred[:, obj_idx:(obj_idx+1)])
|
1014 |
+
p_cls.append(fg_pred[:, (obj_idx+1):])
|
1015 |
+
|
1016 |
+
grid = torch.stack([gi, gj], dim=1)
|
1017 |
+
pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
|
1018 |
+
#pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
|
1019 |
+
pw = self.wh_bin_sigmoid.forward(fg_pred[..., 2:(3+self.bin_count)].sigmoid()) * anch[i][idx][:, 0] * self.stride[i]
|
1020 |
+
ph = self.wh_bin_sigmoid.forward(fg_pred[..., (3+self.bin_count):obj_idx].sigmoid()) * anch[i][idx][:, 1] * self.stride[i]
|
1021 |
+
|
1022 |
+
pxywh = torch.cat([pxy, pw.unsqueeze(1), ph.unsqueeze(1)], dim=-1)
|
1023 |
+
pxyxy = xywh2xyxy(pxywh)
|
1024 |
+
pxyxys.append(pxyxy)
|
1025 |
+
|
1026 |
+
pxyxys = torch.cat(pxyxys, dim=0)
|
1027 |
+
if pxyxys.shape[0] == 0:
|
1028 |
+
continue
|
1029 |
+
p_obj = torch.cat(p_obj, dim=0)
|
1030 |
+
p_cls = torch.cat(p_cls, dim=0)
|
1031 |
+
from_which_layer = torch.cat(from_which_layer, dim=0)
|
1032 |
+
all_b = torch.cat(all_b, dim=0)
|
1033 |
+
all_a = torch.cat(all_a, dim=0)
|
1034 |
+
all_gj = torch.cat(all_gj, dim=0)
|
1035 |
+
all_gi = torch.cat(all_gi, dim=0)
|
1036 |
+
all_anch = torch.cat(all_anch, dim=0)
|
1037 |
+
|
1038 |
+
pair_wise_iou = box_iou(txyxy, pxyxys)
|
1039 |
+
|
1040 |
+
pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
|
1041 |
+
|
1042 |
+
top_k, _ = torch.topk(pair_wise_iou, min(10, pair_wise_iou.shape[1]), dim=1)
|
1043 |
+
dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
|
1044 |
+
|
1045 |
+
gt_cls_per_image = (
|
1046 |
+
F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
|
1047 |
+
.float()
|
1048 |
+
.unsqueeze(1)
|
1049 |
+
.repeat(1, pxyxys.shape[0], 1)
|
1050 |
+
)
|
1051 |
+
|
1052 |
+
num_gt = this_target.shape[0]
|
1053 |
+
cls_preds_ = (
|
1054 |
+
p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1055 |
+
* p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1056 |
+
)
|
1057 |
+
|
1058 |
+
y = cls_preds_.sqrt_()
|
1059 |
+
pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
|
1060 |
+
torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
|
1061 |
+
).sum(-1)
|
1062 |
+
del cls_preds_
|
1063 |
+
|
1064 |
+
cost = (
|
1065 |
+
pair_wise_cls_loss
|
1066 |
+
+ 3.0 * pair_wise_iou_loss
|
1067 |
+
)
|
1068 |
+
|
1069 |
+
matching_matrix = torch.zeros_like(cost)
|
1070 |
+
|
1071 |
+
for gt_idx in range(num_gt):
|
1072 |
+
_, pos_idx = torch.topk(
|
1073 |
+
cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
|
1074 |
+
)
|
1075 |
+
matching_matrix[gt_idx][pos_idx] = 1.0
|
1076 |
+
|
1077 |
+
del top_k, dynamic_ks
|
1078 |
+
anchor_matching_gt = matching_matrix.sum(0)
|
1079 |
+
if (anchor_matching_gt > 1).sum() > 0:
|
1080 |
+
_, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
|
1081 |
+
matching_matrix[:, anchor_matching_gt > 1] *= 0.0
|
1082 |
+
matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
|
1083 |
+
fg_mask_inboxes = matching_matrix.sum(0) > 0.0
|
1084 |
+
matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
|
1085 |
+
|
1086 |
+
from_which_layer = from_which_layer[fg_mask_inboxes]
|
1087 |
+
all_b = all_b[fg_mask_inboxes]
|
1088 |
+
all_a = all_a[fg_mask_inboxes]
|
1089 |
+
all_gj = all_gj[fg_mask_inboxes]
|
1090 |
+
all_gi = all_gi[fg_mask_inboxes]
|
1091 |
+
all_anch = all_anch[fg_mask_inboxes]
|
1092 |
+
|
1093 |
+
this_target = this_target[matched_gt_inds]
|
1094 |
+
|
1095 |
+
for i in range(nl):
|
1096 |
+
layer_idx = from_which_layer == i
|
1097 |
+
matching_bs[i].append(all_b[layer_idx])
|
1098 |
+
matching_as[i].append(all_a[layer_idx])
|
1099 |
+
matching_gjs[i].append(all_gj[layer_idx])
|
1100 |
+
matching_gis[i].append(all_gi[layer_idx])
|
1101 |
+
matching_targets[i].append(this_target[layer_idx])
|
1102 |
+
matching_anchs[i].append(all_anch[layer_idx])
|
1103 |
+
|
1104 |
+
for i in range(nl):
|
1105 |
+
if matching_targets[i] != []:
|
1106 |
+
matching_bs[i] = torch.cat(matching_bs[i], dim=0)
|
1107 |
+
matching_as[i] = torch.cat(matching_as[i], dim=0)
|
1108 |
+
matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
|
1109 |
+
matching_gis[i] = torch.cat(matching_gis[i], dim=0)
|
1110 |
+
matching_targets[i] = torch.cat(matching_targets[i], dim=0)
|
1111 |
+
matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
|
1112 |
+
else:
|
1113 |
+
matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1114 |
+
matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1115 |
+
matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1116 |
+
matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1117 |
+
matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1118 |
+
matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1119 |
+
|
1120 |
+
return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs
|
1121 |
+
|
1122 |
+
def find_3_positive(self, p, targets):
|
1123 |
+
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
|
1124 |
+
na, nt = self.na, targets.shape[0] # number of anchors, targets
|
1125 |
+
indices, anch = [], []
|
1126 |
+
gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
|
1127 |
+
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
|
1128 |
+
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
|
1129 |
+
|
1130 |
+
g = 0.5 # bias
|
1131 |
+
off = torch.tensor([[0, 0],
|
1132 |
+
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
|
1133 |
+
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
|
1134 |
+
], device=targets.device).float() * g # offsets
|
1135 |
+
|
1136 |
+
for i in range(self.nl):
|
1137 |
+
anchors = self.anchors[i]
|
1138 |
+
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
|
1139 |
+
|
1140 |
+
# Match targets to anchors
|
1141 |
+
t = targets * gain
|
1142 |
+
if nt:
|
1143 |
+
# Matches
|
1144 |
+
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
|
1145 |
+
j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
|
1146 |
+
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
|
1147 |
+
t = t[j] # filter
|
1148 |
+
|
1149 |
+
# Offsets
|
1150 |
+
gxy = t[:, 2:4] # grid xy
|
1151 |
+
gxi = gain[[2, 3]] - gxy # inverse
|
1152 |
+
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
|
1153 |
+
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
|
1154 |
+
j = torch.stack((torch.ones_like(j), j, k, l, m))
|
1155 |
+
t = t.repeat((5, 1, 1))[j]
|
1156 |
+
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
|
1157 |
+
else:
|
1158 |
+
t = targets[0]
|
1159 |
+
offsets = 0
|
1160 |
+
|
1161 |
+
# Define
|
1162 |
+
b, c = t[:, :2].long().T # image, class
|
1163 |
+
gxy = t[:, 2:4] # grid xy
|
1164 |
+
gwh = t[:, 4:6] # grid wh
|
1165 |
+
gij = (gxy - offsets).long()
|
1166 |
+
gi, gj = gij.T # grid xy indices
|
1167 |
+
|
1168 |
+
# Append
|
1169 |
+
a = t[:, 6].long() # anchor indices
|
1170 |
+
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
|
1171 |
+
anch.append(anchors[a]) # anchors
|
1172 |
+
|
1173 |
+
return indices, anch
|
1174 |
+
|
1175 |
+
|
1176 |
+
class ComputeLossAuxOTA:
|
1177 |
+
# Compute losses
|
1178 |
+
def __init__(self, model, autobalance=False):
|
1179 |
+
super(ComputeLossAuxOTA, self).__init__()
|
1180 |
+
device = next(model.parameters()).device # get model device
|
1181 |
+
h = model.hyp # hyperparameters
|
1182 |
+
|
1183 |
+
# Define criteria
|
1184 |
+
BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
|
1185 |
+
BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
|
1186 |
+
|
1187 |
+
# Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
|
1188 |
+
self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets
|
1189 |
+
|
1190 |
+
# Focal loss
|
1191 |
+
g = h['fl_gamma'] # focal loss gamma
|
1192 |
+
if g > 0:
|
1193 |
+
BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
|
1194 |
+
|
1195 |
+
det = model.module.model[-1] if is_parallel(model) else model.model[-1] # Detect() module
|
1196 |
+
self.balance = {3: [4.0, 1.0, 0.4]}.get(det.nl, [4.0, 1.0, 0.25, 0.06, .02]) # P3-P7
|
1197 |
+
self.ssi = list(det.stride).index(16) if autobalance else 0 # stride 16 index
|
1198 |
+
self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, model.gr, h, autobalance
|
1199 |
+
for k in 'na', 'nc', 'nl', 'anchors', 'stride':
|
1200 |
+
setattr(self, k, getattr(det, k))
|
1201 |
+
|
1202 |
+
def __call__(self, p, targets, imgs): # predictions, targets, model
|
1203 |
+
device = targets.device
|
1204 |
+
lcls, lbox, lobj = torch.zeros(1, device=device), torch.zeros(1, device=device), torch.zeros(1, device=device)
|
1205 |
+
bs_aux, as_aux_, gjs_aux, gis_aux, targets_aux, anchors_aux = self.build_targets2(p[:self.nl], targets, imgs)
|
1206 |
+
bs, as_, gjs, gis, targets, anchors = self.build_targets(p[:self.nl], targets, imgs)
|
1207 |
+
pre_gen_gains_aux = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p[:self.nl]]
|
1208 |
+
pre_gen_gains = [torch.tensor(pp.shape, device=device)[[3, 2, 3, 2]] for pp in p[:self.nl]]
|
1209 |
+
|
1210 |
+
|
1211 |
+
# Losses
|
1212 |
+
for i in range(self.nl): # layer index, layer predictions
|
1213 |
+
pi = p[i]
|
1214 |
+
pi_aux = p[i+self.nl]
|
1215 |
+
b, a, gj, gi = bs[i], as_[i], gjs[i], gis[i] # image, anchor, gridy, gridx
|
1216 |
+
b_aux, a_aux, gj_aux, gi_aux = bs_aux[i], as_aux_[i], gjs_aux[i], gis_aux[i] # image, anchor, gridy, gridx
|
1217 |
+
tobj = torch.zeros_like(pi[..., 0], device=device) # target obj
|
1218 |
+
tobj_aux = torch.zeros_like(pi_aux[..., 0], device=device) # target obj
|
1219 |
+
|
1220 |
+
n = b.shape[0] # number of targets
|
1221 |
+
if n:
|
1222 |
+
ps = pi[b, a, gj, gi] # prediction subset corresponding to targets
|
1223 |
+
|
1224 |
+
# Regression
|
1225 |
+
grid = torch.stack([gi, gj], dim=1)
|
1226 |
+
pxy = ps[:, :2].sigmoid() * 2. - 0.5
|
1227 |
+
pwh = (ps[:, 2:4].sigmoid() * 2) ** 2 * anchors[i]
|
1228 |
+
pbox = torch.cat((pxy, pwh), 1) # predicted box
|
1229 |
+
selected_tbox = targets[i][:, 2:6] * pre_gen_gains[i]
|
1230 |
+
selected_tbox[:, :2] -= grid
|
1231 |
+
iou = bbox_iou(pbox.T, selected_tbox, x1y1x2y2=False, CIoU=True) # iou(prediction, target)
|
1232 |
+
lbox += (1.0 - iou).mean() # iou loss
|
1233 |
+
|
1234 |
+
# Objectness
|
1235 |
+
tobj[b, a, gj, gi] = (1.0 - self.gr) + self.gr * iou.detach().clamp(0).type(tobj.dtype) # iou ratio
|
1236 |
+
|
1237 |
+
# Classification
|
1238 |
+
selected_tcls = targets[i][:, 1].long()
|
1239 |
+
if self.nc > 1: # cls loss (only if multiple classes)
|
1240 |
+
t = torch.full_like(ps[:, 5:], self.cn, device=device) # targets
|
1241 |
+
t[range(n), selected_tcls] = self.cp
|
1242 |
+
lcls += self.BCEcls(ps[:, 5:], t) # BCE
|
1243 |
+
|
1244 |
+
# Append targets to text file
|
1245 |
+
# with open('targets.txt', 'a') as file:
|
1246 |
+
# [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
|
1247 |
+
|
1248 |
+
n_aux = b_aux.shape[0] # number of targets
|
1249 |
+
if n_aux:
|
1250 |
+
ps_aux = pi_aux[b_aux, a_aux, gj_aux, gi_aux] # prediction subset corresponding to targets
|
1251 |
+
grid_aux = torch.stack([gi_aux, gj_aux], dim=1)
|
1252 |
+
pxy_aux = ps_aux[:, :2].sigmoid() * 2. - 0.5
|
1253 |
+
#pxy_aux = ps_aux[:, :2].sigmoid() * 3. - 1.
|
1254 |
+
pwh_aux = (ps_aux[:, 2:4].sigmoid() * 2) ** 2 * anchors_aux[i]
|
1255 |
+
pbox_aux = torch.cat((pxy_aux, pwh_aux), 1) # predicted box
|
1256 |
+
selected_tbox_aux = targets_aux[i][:, 2:6] * pre_gen_gains_aux[i]
|
1257 |
+
selected_tbox_aux[:, :2] -= grid_aux
|
1258 |
+
iou_aux = bbox_iou(pbox_aux.T, selected_tbox_aux, x1y1x2y2=False, CIoU=True) # iou(prediction, target)
|
1259 |
+
lbox += 0.25 * (1.0 - iou_aux).mean() # iou loss
|
1260 |
+
|
1261 |
+
# Objectness
|
1262 |
+
tobj_aux[b_aux, a_aux, gj_aux, gi_aux] = (1.0 - self.gr) + self.gr * iou_aux.detach().clamp(0).type(tobj_aux.dtype) # iou ratio
|
1263 |
+
|
1264 |
+
# Classification
|
1265 |
+
selected_tcls_aux = targets_aux[i][:, 1].long()
|
1266 |
+
if self.nc > 1: # cls loss (only if multiple classes)
|
1267 |
+
t_aux = torch.full_like(ps_aux[:, 5:], self.cn, device=device) # targets
|
1268 |
+
t_aux[range(n_aux), selected_tcls_aux] = self.cp
|
1269 |
+
lcls += 0.25 * self.BCEcls(ps_aux[:, 5:], t_aux) # BCE
|
1270 |
+
|
1271 |
+
obji = self.BCEobj(pi[..., 4], tobj)
|
1272 |
+
obji_aux = self.BCEobj(pi_aux[..., 4], tobj_aux)
|
1273 |
+
lobj += obji * self.balance[i] + 0.25 * obji_aux * self.balance[i] # obj loss
|
1274 |
+
if self.autobalance:
|
1275 |
+
self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
|
1276 |
+
|
1277 |
+
if self.autobalance:
|
1278 |
+
self.balance = [x / self.balance[self.ssi] for x in self.balance]
|
1279 |
+
lbox *= self.hyp['box']
|
1280 |
+
lobj *= self.hyp['obj']
|
1281 |
+
lcls *= self.hyp['cls']
|
1282 |
+
bs = tobj.shape[0] # batch size
|
1283 |
+
|
1284 |
+
loss = lbox + lobj + lcls
|
1285 |
+
return loss * bs, torch.cat((lbox, lobj, lcls, loss)).detach()
|
1286 |
+
|
1287 |
+
def build_targets(self, p, targets, imgs):
|
1288 |
+
|
1289 |
+
indices, anch = self.find_3_positive(p, targets)
|
1290 |
+
|
1291 |
+
matching_bs = [[] for pp in p]
|
1292 |
+
matching_as = [[] for pp in p]
|
1293 |
+
matching_gjs = [[] for pp in p]
|
1294 |
+
matching_gis = [[] for pp in p]
|
1295 |
+
matching_targets = [[] for pp in p]
|
1296 |
+
matching_anchs = [[] for pp in p]
|
1297 |
+
|
1298 |
+
nl = len(p)
|
1299 |
+
|
1300 |
+
for batch_idx in range(p[0].shape[0]):
|
1301 |
+
|
1302 |
+
b_idx = targets[:, 0]==batch_idx
|
1303 |
+
this_target = targets[b_idx]
|
1304 |
+
if this_target.shape[0] == 0:
|
1305 |
+
continue
|
1306 |
+
|
1307 |
+
txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
|
1308 |
+
txyxy = xywh2xyxy(txywh)
|
1309 |
+
|
1310 |
+
pxyxys = []
|
1311 |
+
p_cls = []
|
1312 |
+
p_obj = []
|
1313 |
+
from_which_layer = []
|
1314 |
+
all_b = []
|
1315 |
+
all_a = []
|
1316 |
+
all_gj = []
|
1317 |
+
all_gi = []
|
1318 |
+
all_anch = []
|
1319 |
+
|
1320 |
+
for i, pi in enumerate(p):
|
1321 |
+
|
1322 |
+
b, a, gj, gi = indices[i]
|
1323 |
+
idx = (b == batch_idx)
|
1324 |
+
b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]
|
1325 |
+
all_b.append(b)
|
1326 |
+
all_a.append(a)
|
1327 |
+
all_gj.append(gj)
|
1328 |
+
all_gi.append(gi)
|
1329 |
+
all_anch.append(anch[i][idx])
|
1330 |
+
from_which_layer.append(torch.ones(size=(len(b),)) * i)
|
1331 |
+
|
1332 |
+
fg_pred = pi[b, a, gj, gi]
|
1333 |
+
p_obj.append(fg_pred[:, 4:5])
|
1334 |
+
p_cls.append(fg_pred[:, 5:])
|
1335 |
+
|
1336 |
+
grid = torch.stack([gi, gj], dim=1)
|
1337 |
+
pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
|
1338 |
+
#pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
|
1339 |
+
pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
|
1340 |
+
pxywh = torch.cat([pxy, pwh], dim=-1)
|
1341 |
+
pxyxy = xywh2xyxy(pxywh)
|
1342 |
+
pxyxys.append(pxyxy)
|
1343 |
+
|
1344 |
+
pxyxys = torch.cat(pxyxys, dim=0)
|
1345 |
+
if pxyxys.shape[0] == 0:
|
1346 |
+
continue
|
1347 |
+
p_obj = torch.cat(p_obj, dim=0)
|
1348 |
+
p_cls = torch.cat(p_cls, dim=0)
|
1349 |
+
from_which_layer = torch.cat(from_which_layer, dim=0)
|
1350 |
+
all_b = torch.cat(all_b, dim=0)
|
1351 |
+
all_a = torch.cat(all_a, dim=0)
|
1352 |
+
all_gj = torch.cat(all_gj, dim=0)
|
1353 |
+
all_gi = torch.cat(all_gi, dim=0)
|
1354 |
+
all_anch = torch.cat(all_anch, dim=0)
|
1355 |
+
|
1356 |
+
pair_wise_iou = box_iou(txyxy, pxyxys)
|
1357 |
+
|
1358 |
+
pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
|
1359 |
+
|
1360 |
+
top_k, _ = torch.topk(pair_wise_iou, min(20, pair_wise_iou.shape[1]), dim=1)
|
1361 |
+
dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
|
1362 |
+
|
1363 |
+
gt_cls_per_image = (
|
1364 |
+
F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
|
1365 |
+
.float()
|
1366 |
+
.unsqueeze(1)
|
1367 |
+
.repeat(1, pxyxys.shape[0], 1)
|
1368 |
+
)
|
1369 |
+
|
1370 |
+
num_gt = this_target.shape[0]
|
1371 |
+
cls_preds_ = (
|
1372 |
+
p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1373 |
+
* p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1374 |
+
)
|
1375 |
+
|
1376 |
+
y = cls_preds_.sqrt_()
|
1377 |
+
pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
|
1378 |
+
torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
|
1379 |
+
).sum(-1)
|
1380 |
+
del cls_preds_
|
1381 |
+
|
1382 |
+
cost = (
|
1383 |
+
pair_wise_cls_loss
|
1384 |
+
+ 3.0 * pair_wise_iou_loss
|
1385 |
+
)
|
1386 |
+
|
1387 |
+
matching_matrix = torch.zeros_like(cost)
|
1388 |
+
|
1389 |
+
for gt_idx in range(num_gt):
|
1390 |
+
_, pos_idx = torch.topk(
|
1391 |
+
cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
|
1392 |
+
)
|
1393 |
+
matching_matrix[gt_idx][pos_idx] = 1.0
|
1394 |
+
|
1395 |
+
del top_k, dynamic_ks
|
1396 |
+
anchor_matching_gt = matching_matrix.sum(0)
|
1397 |
+
if (anchor_matching_gt > 1).sum() > 0:
|
1398 |
+
_, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
|
1399 |
+
matching_matrix[:, anchor_matching_gt > 1] *= 0.0
|
1400 |
+
matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
|
1401 |
+
fg_mask_inboxes = matching_matrix.sum(0) > 0.0
|
1402 |
+
matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
|
1403 |
+
|
1404 |
+
from_which_layer = from_which_layer[fg_mask_inboxes]
|
1405 |
+
all_b = all_b[fg_mask_inboxes]
|
1406 |
+
all_a = all_a[fg_mask_inboxes]
|
1407 |
+
all_gj = all_gj[fg_mask_inboxes]
|
1408 |
+
all_gi = all_gi[fg_mask_inboxes]
|
1409 |
+
all_anch = all_anch[fg_mask_inboxes]
|
1410 |
+
|
1411 |
+
this_target = this_target[matched_gt_inds]
|
1412 |
+
|
1413 |
+
for i in range(nl):
|
1414 |
+
layer_idx = from_which_layer == i
|
1415 |
+
matching_bs[i].append(all_b[layer_idx])
|
1416 |
+
matching_as[i].append(all_a[layer_idx])
|
1417 |
+
matching_gjs[i].append(all_gj[layer_idx])
|
1418 |
+
matching_gis[i].append(all_gi[layer_idx])
|
1419 |
+
matching_targets[i].append(this_target[layer_idx])
|
1420 |
+
matching_anchs[i].append(all_anch[layer_idx])
|
1421 |
+
|
1422 |
+
for i in range(nl):
|
1423 |
+
if matching_targets[i] != []:
|
1424 |
+
matching_bs[i] = torch.cat(matching_bs[i], dim=0)
|
1425 |
+
matching_as[i] = torch.cat(matching_as[i], dim=0)
|
1426 |
+
matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
|
1427 |
+
matching_gis[i] = torch.cat(matching_gis[i], dim=0)
|
1428 |
+
matching_targets[i] = torch.cat(matching_targets[i], dim=0)
|
1429 |
+
matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
|
1430 |
+
else:
|
1431 |
+
matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1432 |
+
matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1433 |
+
matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1434 |
+
matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1435 |
+
matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1436 |
+
matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1437 |
+
|
1438 |
+
return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs
|
1439 |
+
|
1440 |
+
def build_targets2(self, p, targets, imgs):
|
1441 |
+
|
1442 |
+
indices, anch = self.find_5_positive(p, targets)
|
1443 |
+
|
1444 |
+
matching_bs = [[] for pp in p]
|
1445 |
+
matching_as = [[] for pp in p]
|
1446 |
+
matching_gjs = [[] for pp in p]
|
1447 |
+
matching_gis = [[] for pp in p]
|
1448 |
+
matching_targets = [[] for pp in p]
|
1449 |
+
matching_anchs = [[] for pp in p]
|
1450 |
+
|
1451 |
+
nl = len(p)
|
1452 |
+
|
1453 |
+
for batch_idx in range(p[0].shape[0]):
|
1454 |
+
|
1455 |
+
b_idx = targets[:, 0]==batch_idx
|
1456 |
+
this_target = targets[b_idx]
|
1457 |
+
if this_target.shape[0] == 0:
|
1458 |
+
continue
|
1459 |
+
|
1460 |
+
txywh = this_target[:, 2:6] * imgs[batch_idx].shape[1]
|
1461 |
+
txyxy = xywh2xyxy(txywh)
|
1462 |
+
|
1463 |
+
pxyxys = []
|
1464 |
+
p_cls = []
|
1465 |
+
p_obj = []
|
1466 |
+
from_which_layer = []
|
1467 |
+
all_b = []
|
1468 |
+
all_a = []
|
1469 |
+
all_gj = []
|
1470 |
+
all_gi = []
|
1471 |
+
all_anch = []
|
1472 |
+
|
1473 |
+
for i, pi in enumerate(p):
|
1474 |
+
|
1475 |
+
b, a, gj, gi = indices[i]
|
1476 |
+
idx = (b == batch_idx)
|
1477 |
+
b, a, gj, gi = b[idx], a[idx], gj[idx], gi[idx]
|
1478 |
+
all_b.append(b)
|
1479 |
+
all_a.append(a)
|
1480 |
+
all_gj.append(gj)
|
1481 |
+
all_gi.append(gi)
|
1482 |
+
all_anch.append(anch[i][idx])
|
1483 |
+
from_which_layer.append(torch.ones(size=(len(b),)) * i)
|
1484 |
+
|
1485 |
+
fg_pred = pi[b, a, gj, gi]
|
1486 |
+
p_obj.append(fg_pred[:, 4:5])
|
1487 |
+
p_cls.append(fg_pred[:, 5:])
|
1488 |
+
|
1489 |
+
grid = torch.stack([gi, gj], dim=1)
|
1490 |
+
pxy = (fg_pred[:, :2].sigmoid() * 2. - 0.5 + grid) * self.stride[i] #/ 8.
|
1491 |
+
#pxy = (fg_pred[:, :2].sigmoid() * 3. - 1. + grid) * self.stride[i]
|
1492 |
+
pwh = (fg_pred[:, 2:4].sigmoid() * 2) ** 2 * anch[i][idx] * self.stride[i] #/ 8.
|
1493 |
+
pxywh = torch.cat([pxy, pwh], dim=-1)
|
1494 |
+
pxyxy = xywh2xyxy(pxywh)
|
1495 |
+
pxyxys.append(pxyxy)
|
1496 |
+
|
1497 |
+
pxyxys = torch.cat(pxyxys, dim=0)
|
1498 |
+
if pxyxys.shape[0] == 0:
|
1499 |
+
continue
|
1500 |
+
p_obj = torch.cat(p_obj, dim=0)
|
1501 |
+
p_cls = torch.cat(p_cls, dim=0)
|
1502 |
+
from_which_layer = torch.cat(from_which_layer, dim=0)
|
1503 |
+
all_b = torch.cat(all_b, dim=0)
|
1504 |
+
all_a = torch.cat(all_a, dim=0)
|
1505 |
+
all_gj = torch.cat(all_gj, dim=0)
|
1506 |
+
all_gi = torch.cat(all_gi, dim=0)
|
1507 |
+
all_anch = torch.cat(all_anch, dim=0)
|
1508 |
+
|
1509 |
+
pair_wise_iou = box_iou(txyxy, pxyxys)
|
1510 |
+
|
1511 |
+
pair_wise_iou_loss = -torch.log(pair_wise_iou + 1e-8)
|
1512 |
+
|
1513 |
+
top_k, _ = torch.topk(pair_wise_iou, min(20, pair_wise_iou.shape[1]), dim=1)
|
1514 |
+
dynamic_ks = torch.clamp(top_k.sum(1).int(), min=1)
|
1515 |
+
|
1516 |
+
gt_cls_per_image = (
|
1517 |
+
F.one_hot(this_target[:, 1].to(torch.int64), self.nc)
|
1518 |
+
.float()
|
1519 |
+
.unsqueeze(1)
|
1520 |
+
.repeat(1, pxyxys.shape[0], 1)
|
1521 |
+
)
|
1522 |
+
|
1523 |
+
num_gt = this_target.shape[0]
|
1524 |
+
cls_preds_ = (
|
1525 |
+
p_cls.float().unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1526 |
+
* p_obj.unsqueeze(0).repeat(num_gt, 1, 1).sigmoid_()
|
1527 |
+
)
|
1528 |
+
|
1529 |
+
y = cls_preds_.sqrt_()
|
1530 |
+
pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
|
1531 |
+
torch.log(y/(1-y)) , gt_cls_per_image, reduction="none"
|
1532 |
+
).sum(-1)
|
1533 |
+
del cls_preds_
|
1534 |
+
|
1535 |
+
cost = (
|
1536 |
+
pair_wise_cls_loss
|
1537 |
+
+ 3.0 * pair_wise_iou_loss
|
1538 |
+
)
|
1539 |
+
|
1540 |
+
matching_matrix = torch.zeros_like(cost)
|
1541 |
+
|
1542 |
+
for gt_idx in range(num_gt):
|
1543 |
+
_, pos_idx = torch.topk(
|
1544 |
+
cost[gt_idx], k=dynamic_ks[gt_idx].item(), largest=False
|
1545 |
+
)
|
1546 |
+
matching_matrix[gt_idx][pos_idx] = 1.0
|
1547 |
+
|
1548 |
+
del top_k, dynamic_ks
|
1549 |
+
anchor_matching_gt = matching_matrix.sum(0)
|
1550 |
+
if (anchor_matching_gt > 1).sum() > 0:
|
1551 |
+
_, cost_argmin = torch.min(cost[:, anchor_matching_gt > 1], dim=0)
|
1552 |
+
matching_matrix[:, anchor_matching_gt > 1] *= 0.0
|
1553 |
+
matching_matrix[cost_argmin, anchor_matching_gt > 1] = 1.0
|
1554 |
+
fg_mask_inboxes = matching_matrix.sum(0) > 0.0
|
1555 |
+
matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
|
1556 |
+
|
1557 |
+
from_which_layer = from_which_layer[fg_mask_inboxes]
|
1558 |
+
all_b = all_b[fg_mask_inboxes]
|
1559 |
+
all_a = all_a[fg_mask_inboxes]
|
1560 |
+
all_gj = all_gj[fg_mask_inboxes]
|
1561 |
+
all_gi = all_gi[fg_mask_inboxes]
|
1562 |
+
all_anch = all_anch[fg_mask_inboxes]
|
1563 |
+
|
1564 |
+
this_target = this_target[matched_gt_inds]
|
1565 |
+
|
1566 |
+
for i in range(nl):
|
1567 |
+
layer_idx = from_which_layer == i
|
1568 |
+
matching_bs[i].append(all_b[layer_idx])
|
1569 |
+
matching_as[i].append(all_a[layer_idx])
|
1570 |
+
matching_gjs[i].append(all_gj[layer_idx])
|
1571 |
+
matching_gis[i].append(all_gi[layer_idx])
|
1572 |
+
matching_targets[i].append(this_target[layer_idx])
|
1573 |
+
matching_anchs[i].append(all_anch[layer_idx])
|
1574 |
+
|
1575 |
+
for i in range(nl):
|
1576 |
+
if matching_targets[i] != []:
|
1577 |
+
matching_bs[i] = torch.cat(matching_bs[i], dim=0)
|
1578 |
+
matching_as[i] = torch.cat(matching_as[i], dim=0)
|
1579 |
+
matching_gjs[i] = torch.cat(matching_gjs[i], dim=0)
|
1580 |
+
matching_gis[i] = torch.cat(matching_gis[i], dim=0)
|
1581 |
+
matching_targets[i] = torch.cat(matching_targets[i], dim=0)
|
1582 |
+
matching_anchs[i] = torch.cat(matching_anchs[i], dim=0)
|
1583 |
+
else:
|
1584 |
+
matching_bs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1585 |
+
matching_as[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1586 |
+
matching_gjs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1587 |
+
matching_gis[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1588 |
+
matching_targets[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1589 |
+
matching_anchs[i] = torch.tensor([], device='cuda:0', dtype=torch.int64)
|
1590 |
+
|
1591 |
+
return matching_bs, matching_as, matching_gjs, matching_gis, matching_targets, matching_anchs
|
1592 |
+
|
1593 |
+
def find_5_positive(self, p, targets):
|
1594 |
+
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
|
1595 |
+
na, nt = self.na, targets.shape[0] # number of anchors, targets
|
1596 |
+
indices, anch = [], []
|
1597 |
+
gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
|
1598 |
+
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
|
1599 |
+
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
|
1600 |
+
|
1601 |
+
g = 1.0 # bias
|
1602 |
+
off = torch.tensor([[0, 0],
|
1603 |
+
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
|
1604 |
+
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
|
1605 |
+
], device=targets.device).float() * g # offsets
|
1606 |
+
|
1607 |
+
for i in range(self.nl):
|
1608 |
+
anchors = self.anchors[i]
|
1609 |
+
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
|
1610 |
+
|
1611 |
+
# Match targets to anchors
|
1612 |
+
t = targets * gain
|
1613 |
+
if nt:
|
1614 |
+
# Matches
|
1615 |
+
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
|
1616 |
+
j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
|
1617 |
+
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
|
1618 |
+
t = t[j] # filter
|
1619 |
+
|
1620 |
+
# Offsets
|
1621 |
+
gxy = t[:, 2:4] # grid xy
|
1622 |
+
gxi = gain[[2, 3]] - gxy # inverse
|
1623 |
+
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
|
1624 |
+
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
|
1625 |
+
j = torch.stack((torch.ones_like(j), j, k, l, m))
|
1626 |
+
t = t.repeat((5, 1, 1))[j]
|
1627 |
+
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
|
1628 |
+
else:
|
1629 |
+
t = targets[0]
|
1630 |
+
offsets = 0
|
1631 |
+
|
1632 |
+
# Define
|
1633 |
+
b, c = t[:, :2].long().T # image, class
|
1634 |
+
gxy = t[:, 2:4] # grid xy
|
1635 |
+
gwh = t[:, 4:6] # grid wh
|
1636 |
+
gij = (gxy - offsets).long()
|
1637 |
+
gi, gj = gij.T # grid xy indices
|
1638 |
+
|
1639 |
+
# Append
|
1640 |
+
a = t[:, 6].long() # anchor indices
|
1641 |
+
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
|
1642 |
+
anch.append(anchors[a]) # anchors
|
1643 |
+
|
1644 |
+
return indices, anch
|
1645 |
+
|
1646 |
+
def find_3_positive(self, p, targets):
|
1647 |
+
# Build targets for compute_loss(), input targets(image,class,x,y,w,h)
|
1648 |
+
na, nt = self.na, targets.shape[0] # number of anchors, targets
|
1649 |
+
indices, anch = [], []
|
1650 |
+
gain = torch.ones(7, device=targets.device).long() # normalized to gridspace gain
|
1651 |
+
ai = torch.arange(na, device=targets.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt)
|
1652 |
+
targets = torch.cat((targets.repeat(na, 1, 1), ai[:, :, None]), 2) # append anchor indices
|
1653 |
+
|
1654 |
+
g = 0.5 # bias
|
1655 |
+
off = torch.tensor([[0, 0],
|
1656 |
+
[1, 0], [0, 1], [-1, 0], [0, -1], # j,k,l,m
|
1657 |
+
# [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm
|
1658 |
+
], device=targets.device).float() * g # offsets
|
1659 |
+
|
1660 |
+
for i in range(self.nl):
|
1661 |
+
anchors = self.anchors[i]
|
1662 |
+
gain[2:6] = torch.tensor(p[i].shape)[[3, 2, 3, 2]] # xyxy gain
|
1663 |
+
|
1664 |
+
# Match targets to anchors
|
1665 |
+
t = targets * gain
|
1666 |
+
if nt:
|
1667 |
+
# Matches
|
1668 |
+
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
|
1669 |
+
j = torch.max(r, 1. / r).max(2)[0] < self.hyp['anchor_t'] # compare
|
1670 |
+
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
|
1671 |
+
t = t[j] # filter
|
1672 |
+
|
1673 |
+
# Offsets
|
1674 |
+
gxy = t[:, 2:4] # grid xy
|
1675 |
+
gxi = gain[[2, 3]] - gxy # inverse
|
1676 |
+
j, k = ((gxy % 1. < g) & (gxy > 1.)).T
|
1677 |
+
l, m = ((gxi % 1. < g) & (gxi > 1.)).T
|
1678 |
+
j = torch.stack((torch.ones_like(j), j, k, l, m))
|
1679 |
+
t = t.repeat((5, 1, 1))[j]
|
1680 |
+
offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
|
1681 |
+
else:
|
1682 |
+
t = targets[0]
|
1683 |
+
offsets = 0
|
1684 |
+
|
1685 |
+
# Define
|
1686 |
+
b, c = t[:, :2].long().T # image, class
|
1687 |
+
gxy = t[:, 2:4] # grid xy
|
1688 |
+
gwh = t[:, 4:6] # grid wh
|
1689 |
+
gij = (gxy - offsets).long()
|
1690 |
+
gi, gj = gij.T # grid xy indices
|
1691 |
+
|
1692 |
+
# Append
|
1693 |
+
a = t[:, 6].long() # anchor indices
|
1694 |
+
indices.append((b, a, gj.clamp_(0, gain[3] - 1), gi.clamp_(0, gain[2] - 1))) # image, anchor, grid indices
|
1695 |
+
anch.append(anchors[a]) # anchors
|
1696 |
+
|
1697 |
+
return indices, anch
|
utils/metrics.py
ADDED
@@ -0,0 +1,223 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Model validation metrics
|
2 |
+
|
3 |
+
from pathlib import Path
|
4 |
+
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
|
9 |
+
from . import general
|
10 |
+
|
11 |
+
|
12 |
+
def fitness(x):
|
13 |
+
# Model fitness as a weighted combination of metrics
|
14 |
+
w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, [email protected], [email protected]:0.95]
|
15 |
+
return (x[:, :4] * w).sum(1)
|
16 |
+
|
17 |
+
|
18 |
+
def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=()):
|
19 |
+
""" Compute the average precision, given the recall and precision curves.
|
20 |
+
Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
|
21 |
+
# Arguments
|
22 |
+
tp: True positives (nparray, nx1 or nx10).
|
23 |
+
conf: Objectness value from 0-1 (nparray).
|
24 |
+
pred_cls: Predicted object classes (nparray).
|
25 |
+
target_cls: True object classes (nparray).
|
26 |
+
plot: Plot precision-recall curve at [email protected]
|
27 |
+
save_dir: Plot save directory
|
28 |
+
# Returns
|
29 |
+
The average precision as computed in py-faster-rcnn.
|
30 |
+
"""
|
31 |
+
|
32 |
+
# Sort by objectness
|
33 |
+
i = np.argsort(-conf)
|
34 |
+
tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
|
35 |
+
|
36 |
+
# Find unique classes
|
37 |
+
unique_classes = np.unique(target_cls)
|
38 |
+
nc = unique_classes.shape[0] # number of classes, number of detections
|
39 |
+
|
40 |
+
# Create Precision-Recall curve and compute AP for each class
|
41 |
+
px, py = np.linspace(0, 1, 1000), [] # for plotting
|
42 |
+
ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
|
43 |
+
for ci, c in enumerate(unique_classes):
|
44 |
+
i = pred_cls == c
|
45 |
+
n_l = (target_cls == c).sum() # number of labels
|
46 |
+
n_p = i.sum() # number of predictions
|
47 |
+
|
48 |
+
if n_p == 0 or n_l == 0:
|
49 |
+
continue
|
50 |
+
else:
|
51 |
+
# Accumulate FPs and TPs
|
52 |
+
fpc = (1 - tp[i]).cumsum(0)
|
53 |
+
tpc = tp[i].cumsum(0)
|
54 |
+
|
55 |
+
# Recall
|
56 |
+
recall = tpc / (n_l + 1e-16) # recall curve
|
57 |
+
r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases
|
58 |
+
|
59 |
+
# Precision
|
60 |
+
precision = tpc / (tpc + fpc) # precision curve
|
61 |
+
p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score
|
62 |
+
|
63 |
+
# AP from recall-precision curve
|
64 |
+
for j in range(tp.shape[1]):
|
65 |
+
ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
|
66 |
+
if plot and j == 0:
|
67 |
+
py.append(np.interp(px, mrec, mpre)) # precision at [email protected]
|
68 |
+
|
69 |
+
# Compute F1 (harmonic mean of precision and recall)
|
70 |
+
f1 = 2 * p * r / (p + r + 1e-16)
|
71 |
+
if plot:
|
72 |
+
plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names)
|
73 |
+
plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1')
|
74 |
+
plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision')
|
75 |
+
plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall')
|
76 |
+
|
77 |
+
i = f1.mean(0).argmax() # max F1 index
|
78 |
+
return p[:, i], r[:, i], ap, f1[:, i], unique_classes.astype('int32')
|
79 |
+
|
80 |
+
|
81 |
+
def compute_ap(recall, precision):
|
82 |
+
""" Compute the average precision, given the recall and precision curves
|
83 |
+
# Arguments
|
84 |
+
recall: The recall curve (list)
|
85 |
+
precision: The precision curve (list)
|
86 |
+
# Returns
|
87 |
+
Average precision, precision curve, recall curve
|
88 |
+
"""
|
89 |
+
|
90 |
+
# Append sentinel values to beginning and end
|
91 |
+
mrec = np.concatenate(([0.], recall, [recall[-1] + 0.01]))
|
92 |
+
mpre = np.concatenate(([1.], precision, [0.]))
|
93 |
+
|
94 |
+
# Compute the precision envelope
|
95 |
+
mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
|
96 |
+
|
97 |
+
# Integrate area under curve
|
98 |
+
method = 'interp' # methods: 'continuous', 'interp'
|
99 |
+
if method == 'interp':
|
100 |
+
x = np.linspace(0, 1, 101) # 101-point interp (COCO)
|
101 |
+
ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate
|
102 |
+
else: # 'continuous'
|
103 |
+
i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes
|
104 |
+
ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve
|
105 |
+
|
106 |
+
return ap, mpre, mrec
|
107 |
+
|
108 |
+
|
109 |
+
class ConfusionMatrix:
|
110 |
+
# Updated version of https://github.com/kaanakan/object_detection_confusion_matrix
|
111 |
+
def __init__(self, nc, conf=0.25, iou_thres=0.45):
|
112 |
+
self.matrix = np.zeros((nc + 1, nc + 1))
|
113 |
+
self.nc = nc # number of classes
|
114 |
+
self.conf = conf
|
115 |
+
self.iou_thres = iou_thres
|
116 |
+
|
117 |
+
def process_batch(self, detections, labels):
|
118 |
+
"""
|
119 |
+
Return intersection-over-union (Jaccard index) of boxes.
|
120 |
+
Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
|
121 |
+
Arguments:
|
122 |
+
detections (Array[N, 6]), x1, y1, x2, y2, conf, class
|
123 |
+
labels (Array[M, 5]), class, x1, y1, x2, y2
|
124 |
+
Returns:
|
125 |
+
None, updates confusion matrix accordingly
|
126 |
+
"""
|
127 |
+
detections = detections[detections[:, 4] > self.conf]
|
128 |
+
gt_classes = labels[:, 0].int()
|
129 |
+
detection_classes = detections[:, 5].int()
|
130 |
+
iou = general.box_iou(labels[:, 1:], detections[:, :4])
|
131 |
+
|
132 |
+
x = torch.where(iou > self.iou_thres)
|
133 |
+
if x[0].shape[0]:
|
134 |
+
matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
|
135 |
+
if x[0].shape[0] > 1:
|
136 |
+
matches = matches[matches[:, 2].argsort()[::-1]]
|
137 |
+
matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
|
138 |
+
matches = matches[matches[:, 2].argsort()[::-1]]
|
139 |
+
matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
|
140 |
+
else:
|
141 |
+
matches = np.zeros((0, 3))
|
142 |
+
|
143 |
+
n = matches.shape[0] > 0
|
144 |
+
m0, m1, _ = matches.transpose().astype(np.int16)
|
145 |
+
for i, gc in enumerate(gt_classes):
|
146 |
+
j = m0 == i
|
147 |
+
if n and sum(j) == 1:
|
148 |
+
self.matrix[gc, detection_classes[m1[j]]] += 1 # correct
|
149 |
+
else:
|
150 |
+
self.matrix[self.nc, gc] += 1 # background FP
|
151 |
+
|
152 |
+
if n:
|
153 |
+
for i, dc in enumerate(detection_classes):
|
154 |
+
if not any(m1 == i):
|
155 |
+
self.matrix[dc, self.nc] += 1 # background FN
|
156 |
+
|
157 |
+
def matrix(self):
|
158 |
+
return self.matrix
|
159 |
+
|
160 |
+
def plot(self, save_dir='', names=()):
|
161 |
+
try:
|
162 |
+
import seaborn as sn
|
163 |
+
|
164 |
+
array = self.matrix / (self.matrix.sum(0).reshape(1, self.nc + 1) + 1E-6) # normalize
|
165 |
+
array[array < 0.005] = np.nan # don't annotate (would appear as 0.00)
|
166 |
+
|
167 |
+
fig = plt.figure(figsize=(12, 9), tight_layout=True)
|
168 |
+
sn.set(font_scale=1.0 if self.nc < 50 else 0.8) # for label size
|
169 |
+
labels = (0 < len(names) < 99) and len(names) == self.nc # apply names to ticklabels
|
170 |
+
sn.heatmap(array, annot=self.nc < 30, annot_kws={"size": 8}, cmap='Blues', fmt='.2f', square=True,
|
171 |
+
xticklabels=names + ['background FP'] if labels else "auto",
|
172 |
+
yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1))
|
173 |
+
fig.axes[0].set_xlabel('True')
|
174 |
+
fig.axes[0].set_ylabel('Predicted')
|
175 |
+
fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250)
|
176 |
+
except Exception as e:
|
177 |
+
pass
|
178 |
+
|
179 |
+
def print(self):
|
180 |
+
for i in range(self.nc + 1):
|
181 |
+
print(' '.join(map(str, self.matrix[i])))
|
182 |
+
|
183 |
+
|
184 |
+
# Plots ----------------------------------------------------------------------------------------------------------------
|
185 |
+
|
186 |
+
def plot_pr_curve(px, py, ap, save_dir='pr_curve.png', names=()):
|
187 |
+
# Precision-recall curve
|
188 |
+
fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
|
189 |
+
py = np.stack(py, axis=1)
|
190 |
+
|
191 |
+
if 0 < len(names) < 21: # display per-class legend if < 21 classes
|
192 |
+
for i, y in enumerate(py.T):
|
193 |
+
ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision)
|
194 |
+
else:
|
195 |
+
ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision)
|
196 |
+
|
197 |
+
ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f [email protected]' % ap[:, 0].mean())
|
198 |
+
ax.set_xlabel('Recall')
|
199 |
+
ax.set_ylabel('Precision')
|
200 |
+
ax.set_xlim(0, 1)
|
201 |
+
ax.set_ylim(0, 1)
|
202 |
+
plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
|
203 |
+
fig.savefig(Path(save_dir), dpi=250)
|
204 |
+
|
205 |
+
|
206 |
+
def plot_mc_curve(px, py, save_dir='mc_curve.png', names=(), xlabel='Confidence', ylabel='Metric'):
|
207 |
+
# Metric-confidence curve
|
208 |
+
fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
|
209 |
+
|
210 |
+
if 0 < len(names) < 21: # display per-class legend if < 21 classes
|
211 |
+
for i, y in enumerate(py):
|
212 |
+
ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric)
|
213 |
+
else:
|
214 |
+
ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric)
|
215 |
+
|
216 |
+
y = py.mean(0)
|
217 |
+
ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}')
|
218 |
+
ax.set_xlabel(xlabel)
|
219 |
+
ax.set_ylabel(ylabel)
|
220 |
+
ax.set_xlim(0, 1)
|
221 |
+
ax.set_ylim(0, 1)
|
222 |
+
plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
|
223 |
+
fig.savefig(Path(save_dir), dpi=250)
|
utils/plots.py
ADDED
@@ -0,0 +1,489 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Plotting utils
|
2 |
+
|
3 |
+
import glob
|
4 |
+
import math
|
5 |
+
import os
|
6 |
+
import random
|
7 |
+
from copy import copy
|
8 |
+
from pathlib import Path
|
9 |
+
|
10 |
+
import cv2
|
11 |
+
import matplotlib
|
12 |
+
import matplotlib.pyplot as plt
|
13 |
+
import numpy as np
|
14 |
+
import pandas as pd
|
15 |
+
import seaborn as sns
|
16 |
+
import torch
|
17 |
+
import yaml
|
18 |
+
from PIL import Image, ImageDraw, ImageFont
|
19 |
+
from scipy.signal import butter, filtfilt
|
20 |
+
|
21 |
+
from utils.general import xywh2xyxy, xyxy2xywh
|
22 |
+
from utils.metrics import fitness
|
23 |
+
|
24 |
+
# Settings
|
25 |
+
matplotlib.rc('font', **{'size': 11})
|
26 |
+
matplotlib.use('Agg') # for writing to files only
|
27 |
+
|
28 |
+
|
29 |
+
def color_list():
|
30 |
+
# Return first 10 plt colors as (r,g,b) https://stackoverflow.com/questions/51350872/python-from-color-name-to-rgb
|
31 |
+
def hex2rgb(h):
|
32 |
+
return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
|
33 |
+
|
34 |
+
return [hex2rgb(h) for h in matplotlib.colors.TABLEAU_COLORS.values()] # or BASE_ (8), CSS4_ (148), XKCD_ (949)
|
35 |
+
|
36 |
+
|
37 |
+
def hist2d(x, y, n=100):
|
38 |
+
# 2d histogram used in labels.png and evolve.png
|
39 |
+
xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n)
|
40 |
+
hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges))
|
41 |
+
xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1)
|
42 |
+
yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1)
|
43 |
+
return np.log(hist[xidx, yidx])
|
44 |
+
|
45 |
+
|
46 |
+
def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5):
|
47 |
+
# https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy
|
48 |
+
def butter_lowpass(cutoff, fs, order):
|
49 |
+
nyq = 0.5 * fs
|
50 |
+
normal_cutoff = cutoff / nyq
|
51 |
+
return butter(order, normal_cutoff, btype='low', analog=False)
|
52 |
+
|
53 |
+
b, a = butter_lowpass(cutoff, fs, order=order)
|
54 |
+
return filtfilt(b, a, data) # forward-backward filter
|
55 |
+
|
56 |
+
|
57 |
+
def plot_one_box(x, img, color=None, label=None, line_thickness=3):
|
58 |
+
# Plots one bounding box on image img
|
59 |
+
tl = line_thickness or round(0.002 * (img.shape[0] + img.shape[1]) / 2) + 1 # line/font thickness
|
60 |
+
color = color or [random.randint(0, 255) for _ in range(3)]
|
61 |
+
c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
|
62 |
+
cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
|
63 |
+
if label:
|
64 |
+
tf = max(tl - 1, 1) # font thickness
|
65 |
+
t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
|
66 |
+
c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
|
67 |
+
cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA) # filled
|
68 |
+
cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3, [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)
|
69 |
+
|
70 |
+
|
71 |
+
def plot_one_box_PIL(box, img, color=None, label=None, line_thickness=None):
|
72 |
+
img = Image.fromarray(img)
|
73 |
+
draw = ImageDraw.Draw(img)
|
74 |
+
line_thickness = line_thickness or max(int(min(img.size) / 200), 2)
|
75 |
+
draw.rectangle(box, width=line_thickness, outline=tuple(color)) # plot
|
76 |
+
if label:
|
77 |
+
fontsize = max(round(max(img.size) / 40), 12)
|
78 |
+
font = ImageFont.truetype("Arial.ttf", fontsize)
|
79 |
+
txt_width, txt_height = font.getsize(label)
|
80 |
+
draw.rectangle([box[0], box[1] - txt_height + 4, box[0] + txt_width, box[1]], fill=tuple(color))
|
81 |
+
draw.text((box[0], box[1] - txt_height + 1), label, fill=(255, 255, 255), font=font)
|
82 |
+
return np.asarray(img)
|
83 |
+
|
84 |
+
|
85 |
+
def plot_wh_methods(): # from utils.plots import *; plot_wh_methods()
|
86 |
+
# Compares the two methods for width-height anchor multiplication
|
87 |
+
# https://github.com/ultralytics/yolov3/issues/168
|
88 |
+
x = np.arange(-4.0, 4.0, .1)
|
89 |
+
ya = np.exp(x)
|
90 |
+
yb = torch.sigmoid(torch.from_numpy(x)).numpy() * 2
|
91 |
+
|
92 |
+
fig = plt.figure(figsize=(6, 3), tight_layout=True)
|
93 |
+
plt.plot(x, ya, '.-', label='YOLOv3')
|
94 |
+
plt.plot(x, yb ** 2, '.-', label='YOLOR ^2')
|
95 |
+
plt.plot(x, yb ** 1.6, '.-', label='YOLOR ^1.6')
|
96 |
+
plt.xlim(left=-4, right=4)
|
97 |
+
plt.ylim(bottom=0, top=6)
|
98 |
+
plt.xlabel('input')
|
99 |
+
plt.ylabel('output')
|
100 |
+
plt.grid()
|
101 |
+
plt.legend()
|
102 |
+
fig.savefig('comparison.png', dpi=200)
|
103 |
+
|
104 |
+
|
105 |
+
def output_to_target(output):
|
106 |
+
# Convert model output to target format [batch_id, class_id, x, y, w, h, conf]
|
107 |
+
targets = []
|
108 |
+
for i, o in enumerate(output):
|
109 |
+
for *box, conf, cls in o.cpu().numpy():
|
110 |
+
targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf])
|
111 |
+
return np.array(targets)
|
112 |
+
|
113 |
+
|
114 |
+
def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=640, max_subplots=16):
|
115 |
+
# Plot image grid with labels
|
116 |
+
|
117 |
+
if isinstance(images, torch.Tensor):
|
118 |
+
images = images.cpu().float().numpy()
|
119 |
+
if isinstance(targets, torch.Tensor):
|
120 |
+
targets = targets.cpu().numpy()
|
121 |
+
|
122 |
+
# un-normalise
|
123 |
+
if np.max(images[0]) <= 1:
|
124 |
+
images *= 255
|
125 |
+
|
126 |
+
tl = 3 # line thickness
|
127 |
+
tf = max(tl - 1, 1) # font thickness
|
128 |
+
bs, _, h, w = images.shape # batch size, _, height, width
|
129 |
+
bs = min(bs, max_subplots) # limit plot images
|
130 |
+
ns = np.ceil(bs ** 0.5) # number of subplots (square)
|
131 |
+
|
132 |
+
# Check if we should resize
|
133 |
+
scale_factor = max_size / max(h, w)
|
134 |
+
if scale_factor < 1:
|
135 |
+
h = math.ceil(scale_factor * h)
|
136 |
+
w = math.ceil(scale_factor * w)
|
137 |
+
|
138 |
+
colors = color_list() # list of colors
|
139 |
+
mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init
|
140 |
+
for i, img in enumerate(images):
|
141 |
+
if i == max_subplots: # if last batch has fewer images than we expect
|
142 |
+
break
|
143 |
+
|
144 |
+
block_x = int(w * (i // ns))
|
145 |
+
block_y = int(h * (i % ns))
|
146 |
+
|
147 |
+
img = img.transpose(1, 2, 0)
|
148 |
+
if scale_factor < 1:
|
149 |
+
img = cv2.resize(img, (w, h))
|
150 |
+
|
151 |
+
mosaic[block_y:block_y + h, block_x:block_x + w, :] = img
|
152 |
+
if len(targets) > 0:
|
153 |
+
image_targets = targets[targets[:, 0] == i]
|
154 |
+
boxes = xywh2xyxy(image_targets[:, 2:6]).T
|
155 |
+
classes = image_targets[:, 1].astype('int')
|
156 |
+
labels = image_targets.shape[1] == 6 # labels if no conf column
|
157 |
+
conf = None if labels else image_targets[:, 6] # check for confidence presence (label vs pred)
|
158 |
+
|
159 |
+
if boxes.shape[1]:
|
160 |
+
if boxes.max() <= 1.01: # if normalized with tolerance 0.01
|
161 |
+
boxes[[0, 2]] *= w # scale to pixels
|
162 |
+
boxes[[1, 3]] *= h
|
163 |
+
elif scale_factor < 1: # absolute coords need scale if image scales
|
164 |
+
boxes *= scale_factor
|
165 |
+
boxes[[0, 2]] += block_x
|
166 |
+
boxes[[1, 3]] += block_y
|
167 |
+
for j, box in enumerate(boxes.T):
|
168 |
+
cls = int(classes[j])
|
169 |
+
color = colors[cls % len(colors)]
|
170 |
+
cls = names[cls] if names else cls
|
171 |
+
if labels or conf[j] > 0.25: # 0.25 conf thresh
|
172 |
+
label = '%s' % cls if labels else '%s %.1f' % (cls, conf[j])
|
173 |
+
plot_one_box(box, mosaic, label=label, color=color, line_thickness=tl)
|
174 |
+
|
175 |
+
# Draw image filename labels
|
176 |
+
if paths:
|
177 |
+
label = Path(paths[i]).name[:40] # trim to 40 char
|
178 |
+
t_size = cv2.getTextSize(label, 0, fontScale=tl / 3, thickness=tf)[0]
|
179 |
+
cv2.putText(mosaic, label, (block_x + 5, block_y + t_size[1] + 5), 0, tl / 3, [220, 220, 220], thickness=tf,
|
180 |
+
lineType=cv2.LINE_AA)
|
181 |
+
|
182 |
+
# Image border
|
183 |
+
cv2.rectangle(mosaic, (block_x, block_y), (block_x + w, block_y + h), (255, 255, 255), thickness=3)
|
184 |
+
|
185 |
+
if fname:
|
186 |
+
r = min(1280. / max(h, w) / ns, 1.0) # ratio to limit image size
|
187 |
+
mosaic = cv2.resize(mosaic, (int(ns * w * r), int(ns * h * r)), interpolation=cv2.INTER_AREA)
|
188 |
+
# cv2.imwrite(fname, cv2.cvtColor(mosaic, cv2.COLOR_BGR2RGB)) # cv2 save
|
189 |
+
Image.fromarray(mosaic).save(fname) # PIL save
|
190 |
+
return mosaic
|
191 |
+
|
192 |
+
|
193 |
+
def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''):
|
194 |
+
# Plot LR simulating training for full epochs
|
195 |
+
optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals
|
196 |
+
y = []
|
197 |
+
for _ in range(epochs):
|
198 |
+
scheduler.step()
|
199 |
+
y.append(optimizer.param_groups[0]['lr'])
|
200 |
+
plt.plot(y, '.-', label='LR')
|
201 |
+
plt.xlabel('epoch')
|
202 |
+
plt.ylabel('LR')
|
203 |
+
plt.grid()
|
204 |
+
plt.xlim(0, epochs)
|
205 |
+
plt.ylim(0)
|
206 |
+
plt.savefig(Path(save_dir) / 'LR.png', dpi=200)
|
207 |
+
plt.close()
|
208 |
+
|
209 |
+
|
210 |
+
def plot_test_txt(): # from utils.plots import *; plot_test()
|
211 |
+
# Plot test.txt histograms
|
212 |
+
x = np.loadtxt('test.txt', dtype=np.float32)
|
213 |
+
box = xyxy2xywh(x[:, :4])
|
214 |
+
cx, cy = box[:, 0], box[:, 1]
|
215 |
+
|
216 |
+
fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True)
|
217 |
+
ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0)
|
218 |
+
ax.set_aspect('equal')
|
219 |
+
plt.savefig('hist2d.png', dpi=300)
|
220 |
+
|
221 |
+
fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True)
|
222 |
+
ax[0].hist(cx, bins=600)
|
223 |
+
ax[1].hist(cy, bins=600)
|
224 |
+
plt.savefig('hist1d.png', dpi=200)
|
225 |
+
|
226 |
+
|
227 |
+
def plot_targets_txt(): # from utils.plots import *; plot_targets_txt()
|
228 |
+
# Plot targets.txt histograms
|
229 |
+
x = np.loadtxt('targets.txt', dtype=np.float32).T
|
230 |
+
s = ['x targets', 'y targets', 'width targets', 'height targets']
|
231 |
+
fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)
|
232 |
+
ax = ax.ravel()
|
233 |
+
for i in range(4):
|
234 |
+
ax[i].hist(x[i], bins=100, label='%.3g +/- %.3g' % (x[i].mean(), x[i].std()))
|
235 |
+
ax[i].legend()
|
236 |
+
ax[i].set_title(s[i])
|
237 |
+
plt.savefig('targets.jpg', dpi=200)
|
238 |
+
|
239 |
+
|
240 |
+
def plot_study_txt(path='', x=None): # from utils.plots import *; plot_study_txt()
|
241 |
+
# Plot study.txt generated by test.py
|
242 |
+
fig, ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)
|
243 |
+
# ax = ax.ravel()
|
244 |
+
|
245 |
+
fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
|
246 |
+
# for f in [Path(path) / f'study_coco_{x}.txt' for x in ['yolor-p6', 'yolor-w6', 'yolor-e6', 'yolor-d6']]:
|
247 |
+
for f in sorted(Path(path).glob('study*.txt')):
|
248 |
+
y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T
|
249 |
+
x = np.arange(y.shape[1]) if x is None else np.array(x)
|
250 |
+
s = ['P', 'R', '[email protected]', '[email protected]:.95', 't_inference (ms/img)', 't_NMS (ms/img)', 't_total (ms/img)']
|
251 |
+
# for i in range(7):
|
252 |
+
# ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8)
|
253 |
+
# ax[i].set_title(s[i])
|
254 |
+
|
255 |
+
j = y[3].argmax() + 1
|
256 |
+
ax2.plot(y[6, 1:j], y[3, 1:j] * 1E2, '.-', linewidth=2, markersize=8,
|
257 |
+
label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
|
258 |
+
|
259 |
+
ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5],
|
260 |
+
'k.-', linewidth=2, markersize=8, alpha=.25, label='EfficientDet')
|
261 |
+
|
262 |
+
ax2.grid(alpha=0.2)
|
263 |
+
ax2.set_yticks(np.arange(20, 60, 5))
|
264 |
+
ax2.set_xlim(0, 57)
|
265 |
+
ax2.set_ylim(30, 55)
|
266 |
+
ax2.set_xlabel('GPU Speed (ms/img)')
|
267 |
+
ax2.set_ylabel('COCO AP val')
|
268 |
+
ax2.legend(loc='lower right')
|
269 |
+
plt.savefig(str(Path(path).name) + '.png', dpi=300)
|
270 |
+
|
271 |
+
|
272 |
+
def plot_labels(labels, names=(), save_dir=Path(''), loggers=None):
|
273 |
+
# plot dataset labels
|
274 |
+
print('Plotting labels... ')
|
275 |
+
c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes
|
276 |
+
nc = int(c.max() + 1) # number of classes
|
277 |
+
colors = color_list()
|
278 |
+
x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height'])
|
279 |
+
|
280 |
+
# seaborn correlogram
|
281 |
+
sns.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9))
|
282 |
+
plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200)
|
283 |
+
plt.close()
|
284 |
+
|
285 |
+
# matplotlib labels
|
286 |
+
matplotlib.use('svg') # faster
|
287 |
+
ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel()
|
288 |
+
ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8)
|
289 |
+
ax[0].set_ylabel('instances')
|
290 |
+
if 0 < len(names) < 30:
|
291 |
+
ax[0].set_xticks(range(len(names)))
|
292 |
+
ax[0].set_xticklabels(names, rotation=90, fontsize=10)
|
293 |
+
else:
|
294 |
+
ax[0].set_xlabel('classes')
|
295 |
+
sns.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9)
|
296 |
+
sns.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9)
|
297 |
+
|
298 |
+
# rectangles
|
299 |
+
labels[:, 1:3] = 0.5 # center
|
300 |
+
labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000
|
301 |
+
img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255)
|
302 |
+
for cls, *box in labels[:1000]:
|
303 |
+
ImageDraw.Draw(img).rectangle(box, width=1, outline=colors[int(cls) % 10]) # plot
|
304 |
+
ax[1].imshow(img)
|
305 |
+
ax[1].axis('off')
|
306 |
+
|
307 |
+
for a in [0, 1, 2, 3]:
|
308 |
+
for s in ['top', 'right', 'left', 'bottom']:
|
309 |
+
ax[a].spines[s].set_visible(False)
|
310 |
+
|
311 |
+
plt.savefig(save_dir / 'labels.jpg', dpi=200)
|
312 |
+
matplotlib.use('Agg')
|
313 |
+
plt.close()
|
314 |
+
|
315 |
+
# loggers
|
316 |
+
for k, v in loggers.items() or {}:
|
317 |
+
if k == 'wandb' and v:
|
318 |
+
v.log({"Labels": [v.Image(str(x), caption=x.name) for x in save_dir.glob('*labels*.jpg')]}, commit=False)
|
319 |
+
|
320 |
+
|
321 |
+
def plot_evolution(yaml_file='data/hyp.finetune.yaml'): # from utils.plots import *; plot_evolution()
|
322 |
+
# Plot hyperparameter evolution results in evolve.txt
|
323 |
+
with open(yaml_file) as f:
|
324 |
+
hyp = yaml.load(f, Loader=yaml.SafeLoader)
|
325 |
+
x = np.loadtxt('evolve.txt', ndmin=2)
|
326 |
+
f = fitness(x)
|
327 |
+
# weights = (f - f.min()) ** 2 # for weighted results
|
328 |
+
plt.figure(figsize=(10, 12), tight_layout=True)
|
329 |
+
matplotlib.rc('font', **{'size': 8})
|
330 |
+
for i, (k, v) in enumerate(hyp.items()):
|
331 |
+
y = x[:, i + 7]
|
332 |
+
# mu = (y * weights).sum() / weights.sum() # best weighted result
|
333 |
+
mu = y[f.argmax()] # best single result
|
334 |
+
plt.subplot(6, 5, i + 1)
|
335 |
+
plt.scatter(y, f, c=hist2d(y, f, 20), cmap='viridis', alpha=.8, edgecolors='none')
|
336 |
+
plt.plot(mu, f.max(), 'k+', markersize=15)
|
337 |
+
plt.title('%s = %.3g' % (k, mu), fontdict={'size': 9}) # limit to 40 characters
|
338 |
+
if i % 5 != 0:
|
339 |
+
plt.yticks([])
|
340 |
+
print('%15s: %.3g' % (k, mu))
|
341 |
+
plt.savefig('evolve.png', dpi=200)
|
342 |
+
print('\nPlot saved as evolve.png')
|
343 |
+
|
344 |
+
|
345 |
+
def profile_idetection(start=0, stop=0, labels=(), save_dir=''):
|
346 |
+
# Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection()
|
347 |
+
ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel()
|
348 |
+
s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS']
|
349 |
+
files = list(Path(save_dir).glob('frames*.txt'))
|
350 |
+
for fi, f in enumerate(files):
|
351 |
+
try:
|
352 |
+
results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows
|
353 |
+
n = results.shape[1] # number of rows
|
354 |
+
x = np.arange(start, min(stop, n) if stop else n)
|
355 |
+
results = results[:, x]
|
356 |
+
t = (results[0] - results[0].min()) # set t0=0s
|
357 |
+
results[0] = x
|
358 |
+
for i, a in enumerate(ax):
|
359 |
+
if i < len(results):
|
360 |
+
label = labels[fi] if len(labels) else f.stem.replace('frames_', '')
|
361 |
+
a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5)
|
362 |
+
a.set_title(s[i])
|
363 |
+
a.set_xlabel('time (s)')
|
364 |
+
# if fi == len(files) - 1:
|
365 |
+
# a.set_ylim(bottom=0)
|
366 |
+
for side in ['top', 'right']:
|
367 |
+
a.spines[side].set_visible(False)
|
368 |
+
else:
|
369 |
+
a.remove()
|
370 |
+
except Exception as e:
|
371 |
+
print('Warning: Plotting error for %s; %s' % (f, e))
|
372 |
+
|
373 |
+
ax[1].legend()
|
374 |
+
plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200)
|
375 |
+
|
376 |
+
|
377 |
+
def plot_results_overlay(start=0, stop=0): # from utils.plots import *; plot_results_overlay()
|
378 |
+
# Plot training 'results*.txt', overlaying train and val losses
|
379 |
+
s = ['train', 'train', 'train', 'Precision', '[email protected]', 'val', 'val', 'val', 'Recall', '[email protected]:0.95'] # legends
|
380 |
+
t = ['Box', 'Objectness', 'Classification', 'P-R', 'mAP-F1'] # titles
|
381 |
+
for f in sorted(glob.glob('results*.txt') + glob.glob('../../Downloads/results*.txt')):
|
382 |
+
results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T
|
383 |
+
n = results.shape[1] # number of rows
|
384 |
+
x = range(start, min(stop, n) if stop else n)
|
385 |
+
fig, ax = plt.subplots(1, 5, figsize=(14, 3.5), tight_layout=True)
|
386 |
+
ax = ax.ravel()
|
387 |
+
for i in range(5):
|
388 |
+
for j in [i, i + 5]:
|
389 |
+
y = results[j, x]
|
390 |
+
ax[i].plot(x, y, marker='.', label=s[j])
|
391 |
+
# y_smooth = butter_lowpass_filtfilt(y)
|
392 |
+
# ax[i].plot(x, np.gradient(y_smooth), marker='.', label=s[j])
|
393 |
+
|
394 |
+
ax[i].set_title(t[i])
|
395 |
+
ax[i].legend()
|
396 |
+
ax[i].set_ylabel(f) if i == 0 else None # add filename
|
397 |
+
fig.savefig(f.replace('.txt', '.png'), dpi=200)
|
398 |
+
|
399 |
+
|
400 |
+
def plot_results(start=0, stop=0, bucket='', id=(), labels=(), save_dir=''):
|
401 |
+
# Plot training 'results*.txt'. from utils.plots import *; plot_results(save_dir='runs/train/exp')
|
402 |
+
fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True)
|
403 |
+
ax = ax.ravel()
|
404 |
+
s = ['Box', 'Objectness', 'Classification', 'Precision', 'Recall',
|
405 |
+
'val Box', 'val Objectness', 'val Classification', '[email protected]', '[email protected]:0.95']
|
406 |
+
if bucket:
|
407 |
+
# files = ['https://storage.googleapis.com/%s/results%g.txt' % (bucket, x) for x in id]
|
408 |
+
files = ['results%g.txt' % x for x in id]
|
409 |
+
c = ('gsutil cp ' + '%s ' * len(files) + '.') % tuple('gs://%s/results%g.txt' % (bucket, x) for x in id)
|
410 |
+
os.system(c)
|
411 |
+
else:
|
412 |
+
files = list(Path(save_dir).glob('results*.txt'))
|
413 |
+
assert len(files), 'No results.txt files found in %s, nothing to plot.' % os.path.abspath(save_dir)
|
414 |
+
for fi, f in enumerate(files):
|
415 |
+
try:
|
416 |
+
results = np.loadtxt(f, usecols=[2, 3, 4, 8, 9, 12, 13, 14, 10, 11], ndmin=2).T
|
417 |
+
n = results.shape[1] # number of rows
|
418 |
+
x = range(start, min(stop, n) if stop else n)
|
419 |
+
for i in range(10):
|
420 |
+
y = results[i, x]
|
421 |
+
if i in [0, 1, 2, 5, 6, 7]:
|
422 |
+
y[y == 0] = np.nan # don't show zero loss values
|
423 |
+
# y /= y[0] # normalize
|
424 |
+
label = labels[fi] if len(labels) else f.stem
|
425 |
+
ax[i].plot(x, y, marker='.', label=label, linewidth=2, markersize=8)
|
426 |
+
ax[i].set_title(s[i])
|
427 |
+
# if i in [5, 6, 7]: # share train and val loss y axes
|
428 |
+
# ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
|
429 |
+
except Exception as e:
|
430 |
+
print('Warning: Plotting error for %s; %s' % (f, e))
|
431 |
+
|
432 |
+
ax[1].legend()
|
433 |
+
fig.savefig(Path(save_dir) / 'results.png', dpi=200)
|
434 |
+
|
435 |
+
|
436 |
+
def output_to_keypoint(output):
|
437 |
+
# Convert model output to target format [batch_id, class_id, x, y, w, h, conf]
|
438 |
+
targets = []
|
439 |
+
for i, o in enumerate(output):
|
440 |
+
kpts = o[:,6:]
|
441 |
+
o = o[:,:6]
|
442 |
+
for index, (*box, conf, cls) in enumerate(o.detach().cpu().numpy()):
|
443 |
+
targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf, *list(kpts.detach().cpu().numpy()[index])])
|
444 |
+
return np.array(targets)
|
445 |
+
|
446 |
+
|
447 |
+
def plot_skeleton_kpts(im, kpts, steps, orig_shape=None):
|
448 |
+
#Plot the skeleton and keypointsfor coco datatset
|
449 |
+
palette = np.array([[255, 128, 0], [255, 153, 51], [255, 178, 102],
|
450 |
+
[230, 230, 0], [255, 153, 255], [153, 204, 255],
|
451 |
+
[255, 102, 255], [255, 51, 255], [102, 178, 255],
|
452 |
+
[51, 153, 255], [255, 153, 153], [255, 102, 102],
|
453 |
+
[255, 51, 51], [153, 255, 153], [102, 255, 102],
|
454 |
+
[51, 255, 51], [0, 255, 0], [0, 0, 255], [255, 0, 0],
|
455 |
+
[255, 255, 255]])
|
456 |
+
|
457 |
+
skeleton = [[16, 14], [14, 12], [17, 15], [15, 13], [12, 13], [6, 12],
|
458 |
+
[7, 13], [6, 7], [6, 8], [7, 9], [8, 10], [9, 11], [2, 3],
|
459 |
+
[1, 2], [1, 3], [2, 4], [3, 5], [4, 6], [5, 7]]
|
460 |
+
|
461 |
+
pose_limb_color = palette[[9, 9, 9, 9, 7, 7, 7, 0, 0, 0, 0, 0, 16, 16, 16, 16, 16, 16, 16]]
|
462 |
+
pose_kpt_color = palette[[16, 16, 16, 16, 16, 0, 0, 0, 0, 0, 0, 9, 9, 9, 9, 9, 9]]
|
463 |
+
radius = 5
|
464 |
+
num_kpts = len(kpts) // steps
|
465 |
+
|
466 |
+
for kid in range(num_kpts):
|
467 |
+
r, g, b = pose_kpt_color[kid]
|
468 |
+
x_coord, y_coord = kpts[steps * kid], kpts[steps * kid + 1]
|
469 |
+
if not (x_coord % 640 == 0 or y_coord % 640 == 0):
|
470 |
+
if steps == 3:
|
471 |
+
conf = kpts[steps * kid + 2]
|
472 |
+
if conf < 0.5:
|
473 |
+
continue
|
474 |
+
cv2.circle(im, (int(x_coord), int(y_coord)), radius, (int(r), int(g), int(b)), -1)
|
475 |
+
|
476 |
+
for sk_id, sk in enumerate(skeleton):
|
477 |
+
r, g, b = pose_limb_color[sk_id]
|
478 |
+
pos1 = (int(kpts[(sk[0]-1)*steps]), int(kpts[(sk[0]-1)*steps+1]))
|
479 |
+
pos2 = (int(kpts[(sk[1]-1)*steps]), int(kpts[(sk[1]-1)*steps+1]))
|
480 |
+
if steps == 3:
|
481 |
+
conf1 = kpts[(sk[0]-1)*steps+2]
|
482 |
+
conf2 = kpts[(sk[1]-1)*steps+2]
|
483 |
+
if conf1<0.5 or conf2<0.5:
|
484 |
+
continue
|
485 |
+
if pos1[0]%640 == 0 or pos1[1]%640==0 or pos1[0]<0 or pos1[1]<0:
|
486 |
+
continue
|
487 |
+
if pos2[0] % 640 == 0 or pos2[1] % 640 == 0 or pos2[0]<0 or pos2[1]<0:
|
488 |
+
continue
|
489 |
+
cv2.line(im, pos1, pos2, (int(r), int(g), int(b)), thickness=2)
|
utils/torch_utils.py
ADDED
@@ -0,0 +1,374 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# YOLOR PyTorch utils
|
2 |
+
|
3 |
+
import datetime
|
4 |
+
import logging
|
5 |
+
import math
|
6 |
+
import os
|
7 |
+
import platform
|
8 |
+
import subprocess
|
9 |
+
import time
|
10 |
+
from contextlib import contextmanager
|
11 |
+
from copy import deepcopy
|
12 |
+
from pathlib import Path
|
13 |
+
|
14 |
+
import torch
|
15 |
+
import torch.backends.cudnn as cudnn
|
16 |
+
import torch.nn as nn
|
17 |
+
import torch.nn.functional as F
|
18 |
+
import torchvision
|
19 |
+
|
20 |
+
try:
|
21 |
+
import thop # for FLOPS computation
|
22 |
+
except ImportError:
|
23 |
+
thop = None
|
24 |
+
logger = logging.getLogger(__name__)
|
25 |
+
|
26 |
+
|
27 |
+
@contextmanager
|
28 |
+
def torch_distributed_zero_first(local_rank: int):
|
29 |
+
"""
|
30 |
+
Decorator to make all processes in distributed training wait for each local_master to do something.
|
31 |
+
"""
|
32 |
+
if local_rank not in [-1, 0]:
|
33 |
+
torch.distributed.barrier()
|
34 |
+
yield
|
35 |
+
if local_rank == 0:
|
36 |
+
torch.distributed.barrier()
|
37 |
+
|
38 |
+
|
39 |
+
def init_torch_seeds(seed=0):
|
40 |
+
# Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html
|
41 |
+
torch.manual_seed(seed)
|
42 |
+
if seed == 0: # slower, more reproducible
|
43 |
+
cudnn.benchmark, cudnn.deterministic = False, True
|
44 |
+
else: # faster, less reproducible
|
45 |
+
cudnn.benchmark, cudnn.deterministic = True, False
|
46 |
+
|
47 |
+
|
48 |
+
def date_modified(path=__file__):
|
49 |
+
# return human-readable file modification date, i.e. '2021-3-26'
|
50 |
+
t = datetime.datetime.fromtimestamp(Path(path).stat().st_mtime)
|
51 |
+
return f'{t.year}-{t.month}-{t.day}'
|
52 |
+
|
53 |
+
|
54 |
+
def git_describe(path=Path(__file__).parent): # path must be a directory
|
55 |
+
# return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe
|
56 |
+
s = f'git -C {path} describe --tags --long --always'
|
57 |
+
try:
|
58 |
+
return subprocess.check_output(s, shell=True, stderr=subprocess.STDOUT).decode()[:-1]
|
59 |
+
except subprocess.CalledProcessError as e:
|
60 |
+
return '' # not a git repository
|
61 |
+
|
62 |
+
|
63 |
+
def select_device(device='', batch_size=None):
|
64 |
+
# device = 'cpu' or '0' or '0,1,2,3'
|
65 |
+
s = f'YOLOR 🚀 {git_describe() or date_modified()} torch {torch.__version__} ' # string
|
66 |
+
cpu = device.lower() == 'cpu'
|
67 |
+
if cpu:
|
68 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False
|
69 |
+
elif device: # non-cpu device requested
|
70 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable
|
71 |
+
assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested' # check availability
|
72 |
+
|
73 |
+
cuda = not cpu and torch.cuda.is_available()
|
74 |
+
if cuda:
|
75 |
+
n = torch.cuda.device_count()
|
76 |
+
if n > 1 and batch_size: # check that batch_size is compatible with device_count
|
77 |
+
assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
|
78 |
+
space = ' ' * len(s)
|
79 |
+
for i, d in enumerate(device.split(',') if device else range(n)):
|
80 |
+
p = torch.cuda.get_device_properties(i)
|
81 |
+
s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n" # bytes to MB
|
82 |
+
else:
|
83 |
+
s += 'CPU\n'
|
84 |
+
|
85 |
+
logger.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s) # emoji-safe
|
86 |
+
return torch.device('cuda:0' if cuda else 'cpu')
|
87 |
+
|
88 |
+
|
89 |
+
def time_synchronized():
|
90 |
+
# pytorch-accurate time
|
91 |
+
if torch.cuda.is_available():
|
92 |
+
torch.cuda.synchronize()
|
93 |
+
return time.time()
|
94 |
+
|
95 |
+
|
96 |
+
def profile(x, ops, n=100, device=None):
|
97 |
+
# profile a pytorch module or list of modules. Example usage:
|
98 |
+
# x = torch.randn(16, 3, 640, 640) # input
|
99 |
+
# m1 = lambda x: x * torch.sigmoid(x)
|
100 |
+
# m2 = nn.SiLU()
|
101 |
+
# profile(x, [m1, m2], n=100) # profile speed over 100 iterations
|
102 |
+
|
103 |
+
device = device or torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
|
104 |
+
x = x.to(device)
|
105 |
+
x.requires_grad = True
|
106 |
+
print(torch.__version__, device.type, torch.cuda.get_device_properties(0) if device.type == 'cuda' else '')
|
107 |
+
print(f"\n{'Params':>12s}{'GFLOPS':>12s}{'forward (ms)':>16s}{'backward (ms)':>16s}{'input':>24s}{'output':>24s}")
|
108 |
+
for m in ops if isinstance(ops, list) else [ops]:
|
109 |
+
m = m.to(device) if hasattr(m, 'to') else m # device
|
110 |
+
m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m # type
|
111 |
+
dtf, dtb, t = 0., 0., [0., 0., 0.] # dt forward, backward
|
112 |
+
try:
|
113 |
+
flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPS
|
114 |
+
except:
|
115 |
+
flops = 0
|
116 |
+
|
117 |
+
for _ in range(n):
|
118 |
+
t[0] = time_synchronized()
|
119 |
+
y = m(x)
|
120 |
+
t[1] = time_synchronized()
|
121 |
+
try:
|
122 |
+
_ = y.sum().backward()
|
123 |
+
t[2] = time_synchronized()
|
124 |
+
except: # no backward method
|
125 |
+
t[2] = float('nan')
|
126 |
+
dtf += (t[1] - t[0]) * 1000 / n # ms per op forward
|
127 |
+
dtb += (t[2] - t[1]) * 1000 / n # ms per op backward
|
128 |
+
|
129 |
+
s_in = tuple(x.shape) if isinstance(x, torch.Tensor) else 'list'
|
130 |
+
s_out = tuple(y.shape) if isinstance(y, torch.Tensor) else 'list'
|
131 |
+
p = sum(list(x.numel() for x in m.parameters())) if isinstance(m, nn.Module) else 0 # parameters
|
132 |
+
print(f'{p:12}{flops:12.4g}{dtf:16.4g}{dtb:16.4g}{str(s_in):>24s}{str(s_out):>24s}')
|
133 |
+
|
134 |
+
|
135 |
+
def is_parallel(model):
|
136 |
+
return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
|
137 |
+
|
138 |
+
|
139 |
+
def intersect_dicts(da, db, exclude=()):
|
140 |
+
# Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values
|
141 |
+
return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape}
|
142 |
+
|
143 |
+
|
144 |
+
def initialize_weights(model):
|
145 |
+
for m in model.modules():
|
146 |
+
t = type(m)
|
147 |
+
if t is nn.Conv2d:
|
148 |
+
pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
|
149 |
+
elif t is nn.BatchNorm2d:
|
150 |
+
m.eps = 1e-3
|
151 |
+
m.momentum = 0.03
|
152 |
+
elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6]:
|
153 |
+
m.inplace = True
|
154 |
+
|
155 |
+
|
156 |
+
def find_modules(model, mclass=nn.Conv2d):
|
157 |
+
# Finds layer indices matching module class 'mclass'
|
158 |
+
return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
|
159 |
+
|
160 |
+
|
161 |
+
def sparsity(model):
|
162 |
+
# Return global model sparsity
|
163 |
+
a, b = 0., 0.
|
164 |
+
for p in model.parameters():
|
165 |
+
a += p.numel()
|
166 |
+
b += (p == 0).sum()
|
167 |
+
return b / a
|
168 |
+
|
169 |
+
|
170 |
+
def prune(model, amount=0.3):
|
171 |
+
# Prune model to requested global sparsity
|
172 |
+
import torch.nn.utils.prune as prune
|
173 |
+
print('Pruning model... ', end='')
|
174 |
+
for name, m in model.named_modules():
|
175 |
+
if isinstance(m, nn.Conv2d):
|
176 |
+
prune.l1_unstructured(m, name='weight', amount=amount) # prune
|
177 |
+
prune.remove(m, 'weight') # make permanent
|
178 |
+
print(' %.3g global sparsity' % sparsity(model))
|
179 |
+
|
180 |
+
|
181 |
+
def fuse_conv_and_bn(conv, bn):
|
182 |
+
# Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
|
183 |
+
fusedconv = nn.Conv2d(conv.in_channels,
|
184 |
+
conv.out_channels,
|
185 |
+
kernel_size=conv.kernel_size,
|
186 |
+
stride=conv.stride,
|
187 |
+
padding=conv.padding,
|
188 |
+
groups=conv.groups,
|
189 |
+
bias=True).requires_grad_(False).to(conv.weight.device)
|
190 |
+
|
191 |
+
# prepare filters
|
192 |
+
w_conv = conv.weight.clone().view(conv.out_channels, -1)
|
193 |
+
w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
|
194 |
+
fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
|
195 |
+
|
196 |
+
# prepare spatial bias
|
197 |
+
b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
|
198 |
+
b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
|
199 |
+
fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
|
200 |
+
|
201 |
+
return fusedconv
|
202 |
+
|
203 |
+
|
204 |
+
def model_info(model, verbose=False, img_size=640):
|
205 |
+
# Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
|
206 |
+
n_p = sum(x.numel() for x in model.parameters()) # number parameters
|
207 |
+
n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients
|
208 |
+
if verbose:
|
209 |
+
print('%5s %40s %9s %12s %20s %10s %10s' % ('layer', 'name', 'gradient', 'parameters', 'shape', 'mu', 'sigma'))
|
210 |
+
for i, (name, p) in enumerate(model.named_parameters()):
|
211 |
+
name = name.replace('module_list.', '')
|
212 |
+
print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
|
213 |
+
(i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
|
214 |
+
|
215 |
+
try: # FLOPS
|
216 |
+
from thop import profile
|
217 |
+
stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32
|
218 |
+
img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input
|
219 |
+
flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPS
|
220 |
+
img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float
|
221 |
+
fs = ', %.1f GFLOPS' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPS
|
222 |
+
except (ImportError, Exception):
|
223 |
+
fs = ''
|
224 |
+
|
225 |
+
logger.info(f"Model Summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
|
226 |
+
|
227 |
+
|
228 |
+
def load_classifier(name='resnet101', n=2):
|
229 |
+
# Loads a pretrained model reshaped to n-class output
|
230 |
+
model = torchvision.models.__dict__[name](pretrained=True)
|
231 |
+
|
232 |
+
# ResNet model properties
|
233 |
+
# input_size = [3, 224, 224]
|
234 |
+
# input_space = 'RGB'
|
235 |
+
# input_range = [0, 1]
|
236 |
+
# mean = [0.485, 0.456, 0.406]
|
237 |
+
# std = [0.229, 0.224, 0.225]
|
238 |
+
|
239 |
+
# Reshape output to n classes
|
240 |
+
filters = model.fc.weight.shape[1]
|
241 |
+
model.fc.bias = nn.Parameter(torch.zeros(n), requires_grad=True)
|
242 |
+
model.fc.weight = nn.Parameter(torch.zeros(n, filters), requires_grad=True)
|
243 |
+
model.fc.out_features = n
|
244 |
+
return model
|
245 |
+
|
246 |
+
|
247 |
+
def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416)
|
248 |
+
# scales img(bs,3,y,x) by ratio constrained to gs-multiple
|
249 |
+
if ratio == 1.0:
|
250 |
+
return img
|
251 |
+
else:
|
252 |
+
h, w = img.shape[2:]
|
253 |
+
s = (int(h * ratio), int(w * ratio)) # new size
|
254 |
+
img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize
|
255 |
+
if not same_shape: # pad/crop img
|
256 |
+
h, w = [math.ceil(x * ratio / gs) * gs for x in (h, w)]
|
257 |
+
return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean
|
258 |
+
|
259 |
+
|
260 |
+
def copy_attr(a, b, include=(), exclude=()):
|
261 |
+
# Copy attributes from b to a, options to only include [...] and to exclude [...]
|
262 |
+
for k, v in b.__dict__.items():
|
263 |
+
if (len(include) and k not in include) or k.startswith('_') or k in exclude:
|
264 |
+
continue
|
265 |
+
else:
|
266 |
+
setattr(a, k, v)
|
267 |
+
|
268 |
+
|
269 |
+
class ModelEMA:
|
270 |
+
""" Model Exponential Moving Average from https://github.com/rwightman/pytorch-image-models
|
271 |
+
Keep a moving average of everything in the model state_dict (parameters and buffers).
|
272 |
+
This is intended to allow functionality like
|
273 |
+
https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
|
274 |
+
A smoothed version of the weights is necessary for some training schemes to perform well.
|
275 |
+
This class is sensitive where it is initialized in the sequence of model init,
|
276 |
+
GPU assignment and distributed training wrappers.
|
277 |
+
"""
|
278 |
+
|
279 |
+
def __init__(self, model, decay=0.9999, updates=0):
|
280 |
+
# Create EMA
|
281 |
+
self.ema = deepcopy(model.module if is_parallel(model) else model).eval() # FP32 EMA
|
282 |
+
# if next(model.parameters()).device.type != 'cpu':
|
283 |
+
# self.ema.half() # FP16 EMA
|
284 |
+
self.updates = updates # number of EMA updates
|
285 |
+
self.decay = lambda x: decay * (1 - math.exp(-x / 2000)) # decay exponential ramp (to help early epochs)
|
286 |
+
for p in self.ema.parameters():
|
287 |
+
p.requires_grad_(False)
|
288 |
+
|
289 |
+
def update(self, model):
|
290 |
+
# Update EMA parameters
|
291 |
+
with torch.no_grad():
|
292 |
+
self.updates += 1
|
293 |
+
d = self.decay(self.updates)
|
294 |
+
|
295 |
+
msd = model.module.state_dict() if is_parallel(model) else model.state_dict() # model state_dict
|
296 |
+
for k, v in self.ema.state_dict().items():
|
297 |
+
if v.dtype.is_floating_point:
|
298 |
+
v *= d
|
299 |
+
v += (1. - d) * msd[k].detach()
|
300 |
+
|
301 |
+
def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
|
302 |
+
# Update EMA attributes
|
303 |
+
copy_attr(self.ema, model, include, exclude)
|
304 |
+
|
305 |
+
|
306 |
+
class BatchNormXd(torch.nn.modules.batchnorm._BatchNorm):
|
307 |
+
def _check_input_dim(self, input):
|
308 |
+
# The only difference between BatchNorm1d, BatchNorm2d, BatchNorm3d, etc
|
309 |
+
# is this method that is overwritten by the sub-class
|
310 |
+
# This original goal of this method was for tensor sanity checks
|
311 |
+
# If you're ok bypassing those sanity checks (eg. if you trust your inference
|
312 |
+
# to provide the right dimensional inputs), then you can just use this method
|
313 |
+
# for easy conversion from SyncBatchNorm
|
314 |
+
# (unfortunately, SyncBatchNorm does not store the original class - if it did
|
315 |
+
# we could return the one that was originally created)
|
316 |
+
return
|
317 |
+
|
318 |
+
def revert_sync_batchnorm(module):
|
319 |
+
# this is very similar to the function that it is trying to revert:
|
320 |
+
# https://github.com/pytorch/pytorch/blob/c8b3686a3e4ba63dc59e5dcfe5db3430df256833/torch/nn/modules/batchnorm.py#L679
|
321 |
+
module_output = module
|
322 |
+
if isinstance(module, torch.nn.modules.batchnorm.SyncBatchNorm):
|
323 |
+
new_cls = BatchNormXd
|
324 |
+
module_output = BatchNormXd(module.num_features,
|
325 |
+
module.eps, module.momentum,
|
326 |
+
module.affine,
|
327 |
+
module.track_running_stats)
|
328 |
+
if module.affine:
|
329 |
+
with torch.no_grad():
|
330 |
+
module_output.weight = module.weight
|
331 |
+
module_output.bias = module.bias
|
332 |
+
module_output.running_mean = module.running_mean
|
333 |
+
module_output.running_var = module.running_var
|
334 |
+
module_output.num_batches_tracked = module.num_batches_tracked
|
335 |
+
if hasattr(module, "qconfig"):
|
336 |
+
module_output.qconfig = module.qconfig
|
337 |
+
for name, child in module.named_children():
|
338 |
+
module_output.add_module(name, revert_sync_batchnorm(child))
|
339 |
+
del module
|
340 |
+
return module_output
|
341 |
+
|
342 |
+
|
343 |
+
class TracedModel(nn.Module):
|
344 |
+
|
345 |
+
def __init__(self, model=None, device=None, img_size=(640,640)):
|
346 |
+
super(TracedModel, self).__init__()
|
347 |
+
|
348 |
+
print(" Convert model to Traced-model... ")
|
349 |
+
self.stride = model.stride
|
350 |
+
self.names = model.names
|
351 |
+
self.model = model
|
352 |
+
|
353 |
+
self.model = revert_sync_batchnorm(self.model)
|
354 |
+
self.model.to('cpu')
|
355 |
+
self.model.eval()
|
356 |
+
|
357 |
+
self.detect_layer = self.model.model[-1]
|
358 |
+
self.model.traced = True
|
359 |
+
|
360 |
+
rand_example = torch.rand(1, 3, img_size, img_size)
|
361 |
+
|
362 |
+
traced_script_module = torch.jit.trace(self.model, rand_example, strict=False)
|
363 |
+
#traced_script_module = torch.jit.script(self.model)
|
364 |
+
traced_script_module.save("traced_model.pt")
|
365 |
+
print(" traced_script_module saved! ")
|
366 |
+
self.model = traced_script_module
|
367 |
+
self.model.to(device)
|
368 |
+
self.detect_layer.to(device)
|
369 |
+
print(" model is traced! \n")
|
370 |
+
|
371 |
+
def forward(self, x, augment=False, profile=False):
|
372 |
+
out = self.model(x)
|
373 |
+
out = self.detect_layer(out)
|
374 |
+
return out
|