add app
Browse files- app.py +105 -0
- demo.png +0 -0
- demo.tif +0 -0
- model.pth +3 -0
- models/__init__.py +9 -0
- models/unicell_modules.py +912 -0
- requirements.txt +12 -0
- utils/__init__.py +7 -0
- utils/multi_task_sliding_window_inference.py +187 -0
- utils/postprocess.py +125 -0
app.py
ADDED
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding=utf-8
|
3 |
+
# Author: Jun Ma
|
4 |
+
|
5 |
+
import os
|
6 |
+
join = os.path.join
|
7 |
+
import argparse
|
8 |
+
import numpy as np
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import tifffile as tif
|
12 |
+
import monai
|
13 |
+
from tqdm import tqdm
|
14 |
+
from utils.postprocess import mask_overlay
|
15 |
+
from monai.transforms import Activations, AddChanneld, AsChannelFirstd, AsDiscrete, Compose, EnsureTyped, EnsureType
|
16 |
+
from models.unicell_modules import MiT_B2_UNet_MultiHead, MiT_B3_UNet_MultiHead
|
17 |
+
import matplotlib.pyplot as plt
|
18 |
+
from skimage import io, exposure, segmentation, morphology
|
19 |
+
from utils.postprocess import watershed_post
|
20 |
+
from utils.multi_task_sliding_window_inference import multi_task_sliding_window_inference
|
21 |
+
import gradio as gr
|
22 |
+
|
23 |
+
def normalize_channel(img, lower=0.1, upper=99.9):
|
24 |
+
non_zero_vals = img[np.nonzero(img)]
|
25 |
+
percentiles = np.percentile(non_zero_vals, [lower, upper])
|
26 |
+
if percentiles[1] - percentiles[0] > 0.001:
|
27 |
+
img_norm = exposure.rescale_intensity(img, in_range=(percentiles[0], percentiles[1]), out_range='uint8')
|
28 |
+
else:
|
29 |
+
img_norm = img
|
30 |
+
return img_norm
|
31 |
+
|
32 |
+
def preprocess(img_data):
|
33 |
+
if len(img_data.shape) == 2:
|
34 |
+
img_data = np.repeat(np.expand_dims(img_data, axis=-1), 3, axis=-1)
|
35 |
+
elif len(img_data.shape) == 3 and img_data.shape[-1] > 3:
|
36 |
+
img_data = img_data[:,:, :3]
|
37 |
+
else:
|
38 |
+
pass
|
39 |
+
pre_img_data = np.zeros(img_data.shape, dtype=np.uint8)
|
40 |
+
for i in range(3):
|
41 |
+
img_channel_i = img_data[:,:,i]
|
42 |
+
if len(img_channel_i[np.nonzero(img_channel_i)])>0:
|
43 |
+
pre_img_data[:,:,i] = normalize_channel(img_channel_i, lower=1, upper=99)
|
44 |
+
return pre_img_data
|
45 |
+
|
46 |
+
|
47 |
+
def inference(pre_img_data):
|
48 |
+
test_npy = pre_img_data/np.max(pre_img_data)
|
49 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
50 |
+
model = MiT_B2_UNet_MultiHead(in_channels=3, out_channels=3, regress_class=1, img_size=256).to(device)
|
51 |
+
checkpoint = torch.load('./model.pth', map_location=torch.device(device))
|
52 |
+
model.load_state_dict(checkpoint['model_state_dict'])
|
53 |
+
model.eval()
|
54 |
+
with torch.no_grad():
|
55 |
+
test_tensor = torch.from_numpy(np.expand_dims(test_npy, 0)).permute(0,3,1,2).type(torch.FloatTensor).to(device)
|
56 |
+
|
57 |
+
val_pred, val_pred_dist = multi_task_sliding_window_inference(inputs=test_tensor, roi_size=(256, 256), sw_batch_size=8, predictor=model)
|
58 |
+
|
59 |
+
# watershed postprocessing
|
60 |
+
val_seg_inst = watershed_post(val_pred_dist.squeeze(1).cpu().numpy(), val_pred.squeeze(1).cpu().numpy()[:,1])
|
61 |
+
test_pred_mask = val_seg_inst.squeeze().astype(np.uint16)
|
62 |
+
|
63 |
+
# overlay
|
64 |
+
boundary = segmentation.find_boundaries(test_pred_mask, connectivity=1, mode='inner')
|
65 |
+
boundary = morphology.binary_dilation(boundary, morphology.disk(1))
|
66 |
+
pre_img_data[boundary, 0] = 0
|
67 |
+
pre_img_data[boundary, 1] = 255
|
68 |
+
pre_img_data[boundary, 2] = 0
|
69 |
+
|
70 |
+
return test_pred_mask, pre_img_data
|
71 |
+
|
72 |
+
|
73 |
+
def predict(img):
|
74 |
+
print('##########', img.name)
|
75 |
+
img_name = img.name
|
76 |
+
if img_name.endswith('.tif') or img_name.endswith('.tiff'):
|
77 |
+
img_data = tif.imread(img_name)
|
78 |
+
else:
|
79 |
+
img_data = io.imread(img_name)
|
80 |
+
if len(img_data.shape)==2:
|
81 |
+
pre_img_data = normalize_channel(img_data, lower=0.1, upper=99.9)
|
82 |
+
pre_img_data = np.repeat(np.expand_dims(pre_img_data, -1), repeats=3, axis=-1)
|
83 |
+
else:
|
84 |
+
pre_img_data = np.zeros((img_data.shape[0], img_data.shape[1], 3), dtype=np.uint8)
|
85 |
+
for i in range(3):
|
86 |
+
img_channel_i = img_data[:,:,i]
|
87 |
+
if len(img_channel_i[np.nonzero(img_channel_i)])>0:
|
88 |
+
pre_img_data[:,:,i] = normalize_channel(img_channel_i, lower=0.1, upper=99.9)
|
89 |
+
|
90 |
+
seg_labels, seg_overlay = inference(pre_img_data)
|
91 |
+
|
92 |
+
tif.imwrite(join(os.getcwd(), 'segmentation.tiff'), seg_labels, compression='zlib')
|
93 |
+
|
94 |
+
return seg_overlay, join(os.getcwd(), 'segmentation.tiff')
|
95 |
+
|
96 |
+
unicell_api = gr.Interface(
|
97 |
+
predict,
|
98 |
+
inputs = gr.File(label="Input image (png, bmp, jpg, tif, tiff)"),
|
99 |
+
outputs = [gr.Image(label="Segmentation overlay"), gr.File(label="Download segmentation")],
|
100 |
+
title = "UniCell Online Demo",
|
101 |
+
examples=['demo.png', 'demo.tif']
|
102 |
+
)
|
103 |
+
|
104 |
+
unicell_api.launch()
|
105 |
+
|
demo.png
ADDED
demo.tif
ADDED
model.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1a6849ec1969d4abc37b8eb915d03f7b6d6eb3092fc3f1ac5060d1310ddf89f9
|
3 |
+
size 90440917
|
models/__init__.py
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
"""
|
4 |
+
Created on Sun Mar 20 14:23:55 2022
|
5 |
+
|
6 |
+
@author: jma
|
7 |
+
"""
|
8 |
+
|
9 |
+
from .unicell_modules import *
|
models/unicell_modules.py
ADDED
@@ -0,0 +1,912 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# ---------------------------------------------------------------
|
2 |
+
# Copyright (c) 2021, NVIDIA Corporation. All rights reserved.
|
3 |
+
#
|
4 |
+
# This work is licensed under the NVIDIA Source Code License
|
5 |
+
# ---------------------------------------------------------------
|
6 |
+
import torch
|
7 |
+
import torch.nn as nn
|
8 |
+
from functools import partial
|
9 |
+
import math
|
10 |
+
from itertools import repeat
|
11 |
+
import collections.abc
|
12 |
+
from typing import Tuple, Union
|
13 |
+
from monai.networks.blocks import PatchEmbed, UnetOutBlock, UnetrBasicBlock, UnetrUpBlock, UnetrPrUpBlock
|
14 |
+
from monai.networks.blocks.dynunet_block import get_conv_layer
|
15 |
+
|
16 |
+
# From PyTorch internals
|
17 |
+
def _ntuple(n):
|
18 |
+
def parse(x):
|
19 |
+
if isinstance(x, collections.abc.Iterable):
|
20 |
+
return x
|
21 |
+
return tuple(repeat(x, n))
|
22 |
+
return parse
|
23 |
+
|
24 |
+
to_1tuple = _ntuple(1)
|
25 |
+
to_2tuple = _ntuple(2)
|
26 |
+
to_3tuple = _ntuple(3)
|
27 |
+
to_4tuple = _ntuple(4)
|
28 |
+
to_ntuple = _ntuple
|
29 |
+
|
30 |
+
def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
|
31 |
+
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
32 |
+
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
33 |
+
def norm_cdf(x):
|
34 |
+
# Computes standard normal cumulative distribution function
|
35 |
+
return (1. + math.erf(x / math.sqrt(2.))) / 2.
|
36 |
+
|
37 |
+
if (mean < a - 2 * std) or (mean > b + 2 * std):
|
38 |
+
print("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
|
39 |
+
"The distribution of values may be incorrect.",
|
40 |
+
stacklevel=2)
|
41 |
+
|
42 |
+
with torch.no_grad():
|
43 |
+
# Values are generated by using a truncated uniform distribution and
|
44 |
+
# then using the inverse CDF for the normal distribution.
|
45 |
+
# Get upper and lower cdf values
|
46 |
+
l = norm_cdf((a - mean) / std)
|
47 |
+
u = norm_cdf((b - mean) / std)
|
48 |
+
|
49 |
+
# Uniformly fill tensor with values from [l, u], then translate to
|
50 |
+
# [2l-1, 2u-1].
|
51 |
+
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
52 |
+
|
53 |
+
# Use inverse cdf transform for normal distribution to get truncated
|
54 |
+
# standard normal
|
55 |
+
tensor.erfinv_()
|
56 |
+
|
57 |
+
# Transform to proper mean, std
|
58 |
+
tensor.mul_(std * math.sqrt(2.))
|
59 |
+
tensor.add_(mean)
|
60 |
+
|
61 |
+
# Clamp to ensure it's in the proper range
|
62 |
+
tensor.clamp_(min=a, max=b)
|
63 |
+
return tensor
|
64 |
+
|
65 |
+
#%%
|
66 |
+
class Mlp(nn.Module):
|
67 |
+
def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
|
68 |
+
super().__init__()
|
69 |
+
out_features = out_features or in_features
|
70 |
+
hidden_features = hidden_features or in_features
|
71 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
72 |
+
self.dwconv = DWConv(hidden_features)
|
73 |
+
self.act = act_layer()
|
74 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
75 |
+
self.drop = nn.Dropout(drop)
|
76 |
+
|
77 |
+
self.apply(self._init_weights)
|
78 |
+
|
79 |
+
def _init_weights(self, m):
|
80 |
+
if isinstance(m, nn.Linear):
|
81 |
+
trunc_normal_(m.weight, std=.02)
|
82 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
83 |
+
nn.init.constant_(m.bias, 0)
|
84 |
+
elif isinstance(m, nn.LayerNorm):
|
85 |
+
nn.init.constant_(m.bias, 0)
|
86 |
+
nn.init.constant_(m.weight, 1.0)
|
87 |
+
elif isinstance(m, nn.Conv2d):
|
88 |
+
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
89 |
+
fan_out //= m.groups
|
90 |
+
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
|
91 |
+
if m.bias is not None:
|
92 |
+
m.bias.data.zero_()
|
93 |
+
|
94 |
+
def forward(self, x, H, W):
|
95 |
+
x = self.fc1(x)
|
96 |
+
x = self.dwconv(x, H, W)
|
97 |
+
x = self.act(x)
|
98 |
+
x = self.drop(x)
|
99 |
+
x = self.fc2(x)
|
100 |
+
x = self.drop(x)
|
101 |
+
return x
|
102 |
+
|
103 |
+
|
104 |
+
class Attention(nn.Module):
|
105 |
+
def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):
|
106 |
+
super().__init__()
|
107 |
+
assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
|
108 |
+
|
109 |
+
self.dim = dim
|
110 |
+
self.num_heads = num_heads
|
111 |
+
head_dim = dim // num_heads
|
112 |
+
self.scale = qk_scale or head_dim ** -0.5
|
113 |
+
|
114 |
+
self.q = nn.Linear(dim, dim, bias=qkv_bias)
|
115 |
+
self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)
|
116 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
117 |
+
self.proj = nn.Linear(dim, dim)
|
118 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
119 |
+
|
120 |
+
self.sr_ratio = sr_ratio
|
121 |
+
if sr_ratio > 1:
|
122 |
+
self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)
|
123 |
+
self.norm = nn.LayerNorm(dim)
|
124 |
+
|
125 |
+
self.apply(self._init_weights)
|
126 |
+
|
127 |
+
def _init_weights(self, m):
|
128 |
+
if isinstance(m, nn.Linear):
|
129 |
+
trunc_normal_(m.weight, std=.02)
|
130 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
131 |
+
nn.init.constant_(m.bias, 0)
|
132 |
+
elif isinstance(m, nn.LayerNorm):
|
133 |
+
nn.init.constant_(m.bias, 0)
|
134 |
+
nn.init.constant_(m.weight, 1.0)
|
135 |
+
elif isinstance(m, nn.Conv2d):
|
136 |
+
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
137 |
+
fan_out //= m.groups
|
138 |
+
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
|
139 |
+
if m.bias is not None:
|
140 |
+
m.bias.data.zero_()
|
141 |
+
|
142 |
+
def forward(self, x, H, W):
|
143 |
+
B, N, C = x.shape
|
144 |
+
q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
145 |
+
|
146 |
+
if self.sr_ratio > 1:
|
147 |
+
x_ = x.permute(0, 2, 1).reshape(B, C, H, W)
|
148 |
+
x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)
|
149 |
+
x_ = self.norm(x_)
|
150 |
+
kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
151 |
+
else:
|
152 |
+
kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
|
153 |
+
k, v = kv[0], kv[1]
|
154 |
+
|
155 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
156 |
+
attn = attn.softmax(dim=-1)
|
157 |
+
attn = self.attn_drop(attn)
|
158 |
+
|
159 |
+
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
|
160 |
+
x = self.proj(x)
|
161 |
+
x = self.proj_drop(x)
|
162 |
+
|
163 |
+
return x
|
164 |
+
|
165 |
+
|
166 |
+
class Block(nn.Module):
|
167 |
+
|
168 |
+
def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
|
169 |
+
drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):
|
170 |
+
super().__init__()
|
171 |
+
self.norm1 = norm_layer(dim)
|
172 |
+
self.attn = Attention(
|
173 |
+
dim,
|
174 |
+
num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
|
175 |
+
attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)
|
176 |
+
# NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
|
177 |
+
# self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
178 |
+
self.drop_path = nn.Identity()
|
179 |
+
self.norm2 = norm_layer(dim)
|
180 |
+
mlp_hidden_dim = int(dim * mlp_ratio)
|
181 |
+
self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
|
182 |
+
|
183 |
+
self.apply(self._init_weights)
|
184 |
+
|
185 |
+
def _init_weights(self, m):
|
186 |
+
if isinstance(m, nn.Linear):
|
187 |
+
trunc_normal_(m.weight, std=.02)
|
188 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
189 |
+
nn.init.constant_(m.bias, 0)
|
190 |
+
elif isinstance(m, nn.LayerNorm):
|
191 |
+
nn.init.constant_(m.bias, 0)
|
192 |
+
nn.init.constant_(m.weight, 1.0)
|
193 |
+
elif isinstance(m, nn.Conv2d):
|
194 |
+
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
195 |
+
fan_out //= m.groups
|
196 |
+
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
|
197 |
+
if m.bias is not None:
|
198 |
+
m.bias.data.zero_()
|
199 |
+
|
200 |
+
def forward(self, x, H, W):
|
201 |
+
x = x + self.drop_path(self.attn(self.norm1(x), H, W))
|
202 |
+
x = x + self.drop_path(self.mlp(self.norm2(x), H, W))
|
203 |
+
|
204 |
+
return x
|
205 |
+
#%%
|
206 |
+
|
207 |
+
class OverlapPatchEmbed(nn.Module):
|
208 |
+
""" Image to Patch Embedding
|
209 |
+
"""
|
210 |
+
|
211 |
+
def __init__(self, img_size=224, patch_size=7, stride=4, in_chans=3, embed_dim=768):
|
212 |
+
super().__init__()
|
213 |
+
img_size = to_2tuple(img_size)
|
214 |
+
patch_size = to_2tuple(patch_size)
|
215 |
+
|
216 |
+
self.img_size = img_size
|
217 |
+
self.patch_size = patch_size
|
218 |
+
self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]
|
219 |
+
self.num_patches = self.H * self.W
|
220 |
+
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride,
|
221 |
+
padding=(patch_size[0] // 2, patch_size[1] // 2))
|
222 |
+
self.norm = nn.LayerNorm(embed_dim)
|
223 |
+
|
224 |
+
self.apply(self._init_weights)
|
225 |
+
|
226 |
+
def _init_weights(self, m):
|
227 |
+
if isinstance(m, nn.Linear):
|
228 |
+
trunc_normal_(m.weight, std=.02)
|
229 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
230 |
+
nn.init.constant_(m.bias, 0)
|
231 |
+
elif isinstance(m, nn.LayerNorm):
|
232 |
+
nn.init.constant_(m.bias, 0)
|
233 |
+
nn.init.constant_(m.weight, 1.0)
|
234 |
+
elif isinstance(m, nn.Conv2d):
|
235 |
+
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
236 |
+
fan_out //= m.groups
|
237 |
+
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
|
238 |
+
if m.bias is not None:
|
239 |
+
m.bias.data.zero_()
|
240 |
+
|
241 |
+
def forward(self, x):
|
242 |
+
x = self.proj(x) # [2, 3, 224, 224]-> [2, 64, 56, 56]
|
243 |
+
# print(f"{x.shape=}")
|
244 |
+
_, _, H, W = x.shape
|
245 |
+
x = x.flatten(2).transpose(1, 2) # [2, 64, 56, 56]-> [2, 3136, 64]
|
246 |
+
# print(f"{x.shape=}")
|
247 |
+
x = self.norm(x) # [2, 3136, 64]-> [2, 3136, 64]
|
248 |
+
# print(f"{x.shape=}")
|
249 |
+
|
250 |
+
return x, H, W
|
251 |
+
|
252 |
+
# embed_dims=[64, 128, 256, 512]
|
253 |
+
# patch_embed1 = OverlapPatchEmbed(img_size=224,patch_size=7,stride=4,in_chans=in_chans, embed_dim=64)
|
254 |
+
# x1, H, W = patch_embed1(input_img)
|
255 |
+
# x1 = x1.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
256 |
+
# patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0],
|
257 |
+
# embed_dim=embed_dims[1])
|
258 |
+
# x2, H, W = patch_embed2(x1)
|
259 |
+
# x2 = x2.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
260 |
+
|
261 |
+
# patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1],
|
262 |
+
# embed_dim=embed_dims[2])
|
263 |
+
# x3, H, W = patch_embed3(x2)
|
264 |
+
# x3 = x3.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
265 |
+
|
266 |
+
# patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2],embed_dim=embed_dims[3])
|
267 |
+
# x4, H, W = patch_embed4(x3)
|
268 |
+
# x4 = x4.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
269 |
+
#%%
|
270 |
+
|
271 |
+
class MixVisionTransformer(nn.Module):
|
272 |
+
def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dims=[64, 128, 256, 512],
|
273 |
+
num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,
|
274 |
+
attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,
|
275 |
+
depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):
|
276 |
+
super().__init__()
|
277 |
+
# self.num_classes = num_classes
|
278 |
+
self.depths = depths
|
279 |
+
|
280 |
+
# patch_embed
|
281 |
+
self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_chans=in_chans,
|
282 |
+
embed_dim=embed_dims[0])
|
283 |
+
self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0],
|
284 |
+
embed_dim=embed_dims[1])
|
285 |
+
self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1],
|
286 |
+
embed_dim=embed_dims[2])
|
287 |
+
self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2],
|
288 |
+
embed_dim=embed_dims[3])
|
289 |
+
|
290 |
+
# transformer encoder
|
291 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
|
292 |
+
cur = 0
|
293 |
+
self.block1 = nn.ModuleList([Block(
|
294 |
+
dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
|
295 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
|
296 |
+
sr_ratio=sr_ratios[0])
|
297 |
+
for i in range(depths[0])])
|
298 |
+
self.norm1 = norm_layer(embed_dims[0])
|
299 |
+
|
300 |
+
cur += depths[0]
|
301 |
+
self.block2 = nn.ModuleList([Block(
|
302 |
+
dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
|
303 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
|
304 |
+
sr_ratio=sr_ratios[1])
|
305 |
+
for i in range(depths[1])])
|
306 |
+
self.norm2 = norm_layer(embed_dims[1])
|
307 |
+
|
308 |
+
cur += depths[1]
|
309 |
+
self.block3 = nn.ModuleList([Block(
|
310 |
+
dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,
|
311 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
|
312 |
+
sr_ratio=sr_ratios[2])
|
313 |
+
for i in range(depths[2])])
|
314 |
+
self.norm3 = norm_layer(embed_dims[2])
|
315 |
+
|
316 |
+
cur += depths[2]
|
317 |
+
self.block4 = nn.ModuleList([Block(
|
318 |
+
dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
|
319 |
+
drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,
|
320 |
+
sr_ratio=sr_ratios[3])
|
321 |
+
for i in range(depths[3])])
|
322 |
+
self.norm4 = norm_layer(embed_dims[3])
|
323 |
+
|
324 |
+
# classification head
|
325 |
+
# self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()
|
326 |
+
|
327 |
+
self.apply(self._init_weights)
|
328 |
+
|
329 |
+
def _init_weights(self, m):
|
330 |
+
if isinstance(m, nn.Linear):
|
331 |
+
trunc_normal_(m.weight, std=.02)
|
332 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
333 |
+
nn.init.constant_(m.bias, 0)
|
334 |
+
elif isinstance(m, nn.LayerNorm):
|
335 |
+
nn.init.constant_(m.bias, 0)
|
336 |
+
nn.init.constant_(m.weight, 1.0)
|
337 |
+
elif isinstance(m, nn.Conv2d):
|
338 |
+
fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
|
339 |
+
fan_out //= m.groups
|
340 |
+
m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))
|
341 |
+
if m.bias is not None:
|
342 |
+
m.bias.data.zero_()
|
343 |
+
|
344 |
+
def init_weights(self, pretrained=None):
|
345 |
+
if isinstance(pretrained, str):
|
346 |
+
# logger = get_root_logger()
|
347 |
+
# load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)
|
348 |
+
# load_checkpoint(self, pretrained, map_location='cpu', strict=False)
|
349 |
+
torch.load(pretrained, map_location='cpu')
|
350 |
+
|
351 |
+
def reset_drop_path(self, drop_path_rate):
|
352 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]
|
353 |
+
cur = 0
|
354 |
+
for i in range(self.depths[0]):
|
355 |
+
self.block1[i].drop_path.drop_prob = dpr[cur + i]
|
356 |
+
|
357 |
+
cur += self.depths[0]
|
358 |
+
for i in range(self.depths[1]):
|
359 |
+
self.block2[i].drop_path.drop_prob = dpr[cur + i]
|
360 |
+
|
361 |
+
cur += self.depths[1]
|
362 |
+
for i in range(self.depths[2]):
|
363 |
+
self.block3[i].drop_path.drop_prob = dpr[cur + i]
|
364 |
+
|
365 |
+
cur += self.depths[2]
|
366 |
+
for i in range(self.depths[3]):
|
367 |
+
self.block4[i].drop_path.drop_prob = dpr[cur + i]
|
368 |
+
|
369 |
+
def freeze_patch_emb(self):
|
370 |
+
self.patch_embed1.requires_grad = False
|
371 |
+
|
372 |
+
@torch.jit.ignore
|
373 |
+
def no_weight_decay(self):
|
374 |
+
return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'} # has pos_embed may be better
|
375 |
+
|
376 |
+
def get_classifier(self):
|
377 |
+
return self.head
|
378 |
+
|
379 |
+
# def reset_classifier(self, num_classes, global_pool=''):
|
380 |
+
# self.num_classes = num_classes
|
381 |
+
# self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
|
382 |
+
|
383 |
+
def forward_features(self, x):
|
384 |
+
B = x.shape[0]
|
385 |
+
outs = []
|
386 |
+
|
387 |
+
# stage 1
|
388 |
+
x, H, W = self.patch_embed1(x)
|
389 |
+
for i, blk in enumerate(self.block1):
|
390 |
+
x = blk(x, H, W)
|
391 |
+
x = self.norm1(x)
|
392 |
+
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
393 |
+
outs.append(x)
|
394 |
+
|
395 |
+
# stage 2
|
396 |
+
x, H, W = self.patch_embed2(x)
|
397 |
+
for i, blk in enumerate(self.block2):
|
398 |
+
x = blk(x, H, W)
|
399 |
+
x = self.norm2(x)
|
400 |
+
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
401 |
+
outs.append(x)
|
402 |
+
|
403 |
+
# stage 3
|
404 |
+
x, H, W = self.patch_embed3(x)
|
405 |
+
for i, blk in enumerate(self.block3):
|
406 |
+
x = blk(x, H, W)
|
407 |
+
x = self.norm3(x)
|
408 |
+
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
409 |
+
outs.append(x)
|
410 |
+
|
411 |
+
# stage 4
|
412 |
+
x, H, W = self.patch_embed4(x)
|
413 |
+
for i, blk in enumerate(self.block4):
|
414 |
+
x = blk(x, H, W)
|
415 |
+
x = self.norm4(x)
|
416 |
+
x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()
|
417 |
+
outs.append(x)
|
418 |
+
|
419 |
+
return outs
|
420 |
+
|
421 |
+
def forward(self, x):
|
422 |
+
x = self.forward_features(x)
|
423 |
+
# x = self.head(x)
|
424 |
+
|
425 |
+
return x
|
426 |
+
|
427 |
+
|
428 |
+
class DWConv(nn.Module):
|
429 |
+
def __init__(self, dim=768):
|
430 |
+
super(DWConv, self).__init__()
|
431 |
+
self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)
|
432 |
+
|
433 |
+
def forward(self, x, H, W):
|
434 |
+
B, N, C = x.shape
|
435 |
+
x = x.transpose(1, 2).view(B, C, H, W)
|
436 |
+
x = self.dwconv(x)
|
437 |
+
x = x.flatten(2).transpose(1, 2)
|
438 |
+
return x
|
439 |
+
|
440 |
+
|
441 |
+
|
442 |
+
|
443 |
+
class mit_b0(MixVisionTransformer):
|
444 |
+
def __init__(self, **kwargs):
|
445 |
+
super(mit_b0, self).__init__(
|
446 |
+
patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
447 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
|
448 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
449 |
+
|
450 |
+
|
451 |
+
|
452 |
+
class mit_b1(MixVisionTransformer):
|
453 |
+
def __init__(self, **kwargs):
|
454 |
+
super(mit_b1, self).__init__(
|
455 |
+
patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
456 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],
|
457 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
458 |
+
|
459 |
+
|
460 |
+
class mit_b2(MixVisionTransformer):
|
461 |
+
def __init__(self, **kwargs):
|
462 |
+
super(mit_b2, self).__init__(
|
463 |
+
patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
464 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],
|
465 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
466 |
+
|
467 |
+
|
468 |
+
|
469 |
+
class mit_b3(MixVisionTransformer):
|
470 |
+
def __init__(self, **kwargs):
|
471 |
+
super(mit_b3, self).__init__(
|
472 |
+
patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
473 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
|
474 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
475 |
+
|
476 |
+
|
477 |
+
|
478 |
+
class mit_b4(MixVisionTransformer):
|
479 |
+
def __init__(self, **kwargs):
|
480 |
+
super(mit_b4, self).__init__(
|
481 |
+
patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
482 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],
|
483 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
484 |
+
|
485 |
+
|
486 |
+
|
487 |
+
class mit_b5(MixVisionTransformer):
|
488 |
+
def __init__(self, **kwargs):
|
489 |
+
super(mit_b5, self).__init__(
|
490 |
+
patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],
|
491 |
+
qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],
|
492 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
493 |
+
|
494 |
+
|
495 |
+
#%% B2
|
496 |
+
class MiT_B2_UNet_MultiHead(nn.Module):
|
497 |
+
def __init__(self,
|
498 |
+
in_channels: int,
|
499 |
+
out_channels: int,
|
500 |
+
regress_class: int = 1,
|
501 |
+
img_size: Tuple[int, int] = (256,256),
|
502 |
+
|
503 |
+
feature_size: int = 16,
|
504 |
+
spatial_dims: int = 2,
|
505 |
+
# hidden_size: int = 768,
|
506 |
+
# mlp_dim: int = 3072,
|
507 |
+
num_heads = [1, 2, 4, 8],
|
508 |
+
# pos_embed: str = "perceptron",
|
509 |
+
norm_name: Union[Tuple, str] = "instance",
|
510 |
+
conv_block: bool = False,
|
511 |
+
res_block: bool = True,
|
512 |
+
dropout_rate: float = 0.0,
|
513 |
+
debug: bool = False
|
514 |
+
):
|
515 |
+
super().__init__()
|
516 |
+
self.debug = debug
|
517 |
+
self.mit_b3 = MixVisionTransformer(img_size=img_size, patch_size=4, embed_dims=[feature_size*2, feature_size*4, feature_size*8, feature_size*16],
|
518 |
+
num_heads=num_heads, mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6),
|
519 |
+
depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1], drop_rate=0.0, drop_path_rate=0.1)
|
520 |
+
|
521 |
+
self.encoder1 = UnetrBasicBlock(
|
522 |
+
spatial_dims=spatial_dims,
|
523 |
+
in_channels=in_channels,
|
524 |
+
out_channels=feature_size,
|
525 |
+
kernel_size=3,
|
526 |
+
stride=1,
|
527 |
+
norm_name=norm_name,
|
528 |
+
res_block=True,
|
529 |
+
)
|
530 |
+
|
531 |
+
self.encoder2 = UnetrBasicBlock(
|
532 |
+
spatial_dims=spatial_dims,
|
533 |
+
in_channels=2 * feature_size,
|
534 |
+
out_channels=2 * feature_size,
|
535 |
+
kernel_size=3,
|
536 |
+
stride=1,
|
537 |
+
norm_name=norm_name,
|
538 |
+
res_block=True,
|
539 |
+
)
|
540 |
+
|
541 |
+
self.encoder3 = UnetrBasicBlock(
|
542 |
+
spatial_dims=spatial_dims,
|
543 |
+
in_channels=4 * feature_size,
|
544 |
+
out_channels=4 * feature_size,
|
545 |
+
kernel_size=3,
|
546 |
+
stride=1,
|
547 |
+
norm_name=norm_name,
|
548 |
+
res_block=True,
|
549 |
+
)
|
550 |
+
|
551 |
+
self.encoder4 = UnetrBasicBlock(
|
552 |
+
spatial_dims=spatial_dims,
|
553 |
+
in_channels=8 * feature_size,
|
554 |
+
out_channels=8 * feature_size,
|
555 |
+
kernel_size=3,
|
556 |
+
stride=1,
|
557 |
+
norm_name=norm_name,
|
558 |
+
res_block=True,
|
559 |
+
)
|
560 |
+
|
561 |
+
self.encoder5 = UnetrBasicBlock(
|
562 |
+
spatial_dims=spatial_dims,
|
563 |
+
in_channels=16 * feature_size,
|
564 |
+
out_channels=16 * feature_size,
|
565 |
+
kernel_size=3,
|
566 |
+
stride=1,
|
567 |
+
norm_name=norm_name,
|
568 |
+
res_block=True,
|
569 |
+
)
|
570 |
+
|
571 |
+
self.decoder4 = UnetrUpBlock(
|
572 |
+
spatial_dims=2,
|
573 |
+
in_channels=feature_size * 16,
|
574 |
+
out_channels=feature_size * 8,
|
575 |
+
kernel_size=3,
|
576 |
+
upsample_kernel_size=2,
|
577 |
+
norm_name=norm_name,
|
578 |
+
res_block=res_block,
|
579 |
+
)
|
580 |
+
self.decoder3 = UnetrUpBlock(
|
581 |
+
spatial_dims=2,
|
582 |
+
in_channels=feature_size * 8,
|
583 |
+
out_channels=feature_size * 4,
|
584 |
+
kernel_size=3,
|
585 |
+
upsample_kernel_size=2,
|
586 |
+
norm_name=norm_name,
|
587 |
+
res_block=res_block,
|
588 |
+
)
|
589 |
+
self.decoder2 = UnetrUpBlock(
|
590 |
+
spatial_dims=2,
|
591 |
+
in_channels=feature_size * 4,
|
592 |
+
out_channels=feature_size * 2,
|
593 |
+
kernel_size=3,
|
594 |
+
upsample_kernel_size=2,
|
595 |
+
norm_name=norm_name,
|
596 |
+
res_block=res_block,
|
597 |
+
)
|
598 |
+
|
599 |
+
self.transp_conv = get_conv_layer(
|
600 |
+
spatial_dims=2,
|
601 |
+
in_channels=feature_size*2,
|
602 |
+
out_channels=feature_size*2,
|
603 |
+
kernel_size=3,
|
604 |
+
stride=2,
|
605 |
+
conv_only=True,
|
606 |
+
is_transposed=True,
|
607 |
+
)
|
608 |
+
self.decoder1 = UnetrUpBlock(
|
609 |
+
spatial_dims=2,
|
610 |
+
in_channels=feature_size * 2,
|
611 |
+
out_channels=feature_size,
|
612 |
+
kernel_size=3,
|
613 |
+
upsample_kernel_size=2,
|
614 |
+
norm_name=norm_name,
|
615 |
+
res_block=res_block,
|
616 |
+
)
|
617 |
+
|
618 |
+
self.out_interior = UnetOutBlock(spatial_dims=2, in_channels=feature_size, out_channels=out_channels) # type: ignore
|
619 |
+
self.out_dist = UnetOutBlock(spatial_dims=2, in_channels=feature_size, out_channels=1) # type: ignore
|
620 |
+
|
621 |
+
def forward(self, x_in):
|
622 |
+
hidden_states_out = self.mit_b3(x_in) # x: (B, 256,768), hidden_states_out: list, 12 elements, (B,256,768)
|
623 |
+
enc1 = self.encoder1(x_in) # (B, 16, 256, 256)
|
624 |
+
x1 = hidden_states_out[0] # (B, 32, 64, 64)
|
625 |
+
enc2 = self.encoder2(x1) # (B, 64, 32, 32)
|
626 |
+
x2 = hidden_states_out[1] # (B, 64, 32, 32)
|
627 |
+
enc3 = self.encoder3(x2) # (B, 128, 16, 16)
|
628 |
+
x3 = hidden_states_out[2] # (B, 128, 16,16)
|
629 |
+
enc4 = self.encoder4(x3) # (B, 256, 8, 8)
|
630 |
+
x4 = hidden_states_out[3] # (B, 256, 8, 8)
|
631 |
+
enc5 = self.encoder5(x4) # (B, 256, 8, 8)
|
632 |
+
# print(f"{enc1.shape=}, {enc2.shape=}, {enc3.shape=}, {enc4.shape=}, {enc5.shape=}")
|
633 |
+
|
634 |
+
dec4 = self.decoder4(enc5, enc4) # (B, 128, 16, 16); up -> cat -> ResConv; (B, 128, 16, 16)
|
635 |
+
dec3 = self.decoder3(dec4, enc3) # (B, 64, 32, 32)
|
636 |
+
dec2 = self.decoder2(dec3, enc2) # (B, 32, 64, 64)
|
637 |
+
dec2_up = self.transp_conv(dec2) # [B, 32, 128, 128]
|
638 |
+
dec1 = self.decoder1(dec2_up, enc1) # (B, 16, 256, 256)
|
639 |
+
logits = self.out_interior(dec1)
|
640 |
+
dist = self.out_dist(dec1)
|
641 |
+
|
642 |
+
if self.debug:
|
643 |
+
return hidden_states_out, enc1, enc2, enc3, enc4, dec4, dec3, dec2, dec1, logits
|
644 |
+
else:
|
645 |
+
return logits, dist
|
646 |
+
|
647 |
+
# print(f"{dec1.shape=}, {dec2.shape=}, {dec3.shape=}, {dec4.shape=}, {logits.shape=}")
|
648 |
+
|
649 |
+
img_size = 256
|
650 |
+
in_chans = 3
|
651 |
+
B = 2
|
652 |
+
input_img = torch.randn((B,in_chans,img_size,img_size))
|
653 |
+
|
654 |
+
b2 = MiT_B2_UNet_MultiHead(3, 3, img_size=img_size)
|
655 |
+
logits, dist = b2(input_img)
|
656 |
+
|
657 |
+
|
658 |
+
#%% B3
|
659 |
+
class MiT_B3_UNet_MultiHead(nn.Module):
|
660 |
+
def __init__(self,
|
661 |
+
in_channels: int,
|
662 |
+
out_channels: int,
|
663 |
+
regress_class: int = 1,
|
664 |
+
img_size: Tuple[int, int] = (256,256),
|
665 |
+
|
666 |
+
feature_size: int = 16,
|
667 |
+
spatial_dims: int = 2,
|
668 |
+
# hidden_size: int = 768,
|
669 |
+
# mlp_dim: int = 3072,
|
670 |
+
num_heads = [1, 2, 4, 8],
|
671 |
+
# pos_embed: str = "perceptron",
|
672 |
+
norm_name: Union[Tuple, str] = "instance",
|
673 |
+
conv_block: bool = False,
|
674 |
+
res_block: bool = True,
|
675 |
+
dropout_rate: float = 0.0,
|
676 |
+
debug: bool = False
|
677 |
+
):
|
678 |
+
super().__init__()
|
679 |
+
self.debug = debug
|
680 |
+
self.mit_b3 = MixVisionTransformer(img_size=img_size, patch_size=4, embed_dims=[feature_size*2, feature_size*4, feature_size*8, feature_size*16],
|
681 |
+
num_heads=num_heads, mlp_ratios=[4, 4, 4, 4], qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],
|
682 |
+
drop_rate=0.0, drop_path_rate=0.1)
|
683 |
+
|
684 |
+
self.encoder1 = UnetrBasicBlock(
|
685 |
+
spatial_dims=spatial_dims,
|
686 |
+
in_channels=in_channels,
|
687 |
+
out_channels=feature_size,
|
688 |
+
kernel_size=3,
|
689 |
+
stride=1,
|
690 |
+
norm_name=norm_name,
|
691 |
+
res_block=True,
|
692 |
+
)
|
693 |
+
|
694 |
+
self.encoder2 = UnetrBasicBlock(
|
695 |
+
spatial_dims=spatial_dims,
|
696 |
+
in_channels=2 * feature_size,
|
697 |
+
out_channels=2 * feature_size,
|
698 |
+
kernel_size=3,
|
699 |
+
stride=1,
|
700 |
+
norm_name=norm_name,
|
701 |
+
res_block=True,
|
702 |
+
)
|
703 |
+
|
704 |
+
self.encoder3 = UnetrBasicBlock(
|
705 |
+
spatial_dims=spatial_dims,
|
706 |
+
in_channels=4 * feature_size,
|
707 |
+
out_channels=4 * feature_size,
|
708 |
+
kernel_size=3,
|
709 |
+
stride=1,
|
710 |
+
norm_name=norm_name,
|
711 |
+
res_block=True,
|
712 |
+
)
|
713 |
+
|
714 |
+
self.encoder4 = UnetrBasicBlock(
|
715 |
+
spatial_dims=spatial_dims,
|
716 |
+
in_channels=8 * feature_size,
|
717 |
+
out_channels=8 * feature_size,
|
718 |
+
kernel_size=3,
|
719 |
+
stride=1,
|
720 |
+
norm_name=norm_name,
|
721 |
+
res_block=True,
|
722 |
+
)
|
723 |
+
|
724 |
+
self.encoder5 = UnetrBasicBlock(
|
725 |
+
spatial_dims=spatial_dims,
|
726 |
+
in_channels=16 * feature_size,
|
727 |
+
out_channels=16 * feature_size,
|
728 |
+
kernel_size=3,
|
729 |
+
stride=1,
|
730 |
+
norm_name=norm_name,
|
731 |
+
res_block=True,
|
732 |
+
)
|
733 |
+
|
734 |
+
self.decoder4 = UnetrUpBlock(
|
735 |
+
spatial_dims=2,
|
736 |
+
in_channels=feature_size * 16,
|
737 |
+
out_channels=feature_size * 8,
|
738 |
+
kernel_size=3,
|
739 |
+
upsample_kernel_size=2,
|
740 |
+
norm_name=norm_name,
|
741 |
+
res_block=res_block,
|
742 |
+
)
|
743 |
+
self.decoder3 = UnetrUpBlock(
|
744 |
+
spatial_dims=2,
|
745 |
+
in_channels=feature_size * 8,
|
746 |
+
out_channels=feature_size * 4,
|
747 |
+
kernel_size=3,
|
748 |
+
upsample_kernel_size=2,
|
749 |
+
norm_name=norm_name,
|
750 |
+
res_block=res_block,
|
751 |
+
)
|
752 |
+
self.decoder2 = UnetrUpBlock(
|
753 |
+
spatial_dims=2,
|
754 |
+
in_channels=feature_size * 4,
|
755 |
+
out_channels=feature_size * 2,
|
756 |
+
kernel_size=3,
|
757 |
+
upsample_kernel_size=2,
|
758 |
+
norm_name=norm_name,
|
759 |
+
res_block=res_block,
|
760 |
+
)
|
761 |
+
|
762 |
+
self.transp_conv = get_conv_layer(
|
763 |
+
spatial_dims=2,
|
764 |
+
in_channels=feature_size*2,
|
765 |
+
out_channels=feature_size*2,
|
766 |
+
kernel_size=3,
|
767 |
+
stride=2,
|
768 |
+
conv_only=True,
|
769 |
+
is_transposed=True,
|
770 |
+
)
|
771 |
+
self.decoder1 = UnetrUpBlock(
|
772 |
+
spatial_dims=2,
|
773 |
+
in_channels=feature_size * 2,
|
774 |
+
out_channels=feature_size,
|
775 |
+
kernel_size=3,
|
776 |
+
upsample_kernel_size=2,
|
777 |
+
norm_name=norm_name,
|
778 |
+
res_block=res_block,
|
779 |
+
)
|
780 |
+
|
781 |
+
self.out_interior = UnetOutBlock(spatial_dims=2, in_channels=feature_size, out_channels=out_channels) # type: ignore
|
782 |
+
self.out_dist = UnetOutBlock(spatial_dims=2, in_channels=feature_size, out_channels=1) # type: ignore
|
783 |
+
|
784 |
+
def forward(self, x_in):
|
785 |
+
hidden_states_out = self.mit_b3(x_in) # x: (B, 256,768), hidden_states_out: list, 12 elements, (B,256,768)
|
786 |
+
enc1 = self.encoder1(x_in) # (B, 16, 256, 256)
|
787 |
+
x1 = hidden_states_out[0] # (B, 32, 64, 64)
|
788 |
+
enc2 = self.encoder2(x1) # (B, 64, 32, 32)
|
789 |
+
x2 = hidden_states_out[1] # (B, 64, 32, 32)
|
790 |
+
enc3 = self.encoder3(x2) # (B, 128, 16, 16)
|
791 |
+
x3 = hidden_states_out[2] # (B, 128, 16,16)
|
792 |
+
enc4 = self.encoder4(x3) # (B, 256, 8, 8)
|
793 |
+
x4 = hidden_states_out[3] # (B, 256, 8, 8)
|
794 |
+
enc5 = self.encoder5(x4) # (B, 256, 8, 8)
|
795 |
+
# print(f"{enc1.shape=}, {enc2.shape=}, {enc3.shape=}, {enc4.shape=}, {enc5.shape=}")
|
796 |
+
|
797 |
+
dec4 = self.decoder4(enc5, enc4) # (B, 128, 16, 16); up -> cat -> ResConv; (B, 128, 16, 16)
|
798 |
+
dec3 = self.decoder3(dec4, enc3) # (B, 64, 32, 32)
|
799 |
+
dec2 = self.decoder2(dec3, enc2) # (B, 32, 64, 64)
|
800 |
+
dec2_up = self.transp_conv(dec2) # [B, 32, 128, 128]
|
801 |
+
dec1 = self.decoder1(dec2_up, enc1) # (B, 16, 256, 256)
|
802 |
+
logits = self.out_interior(dec1)
|
803 |
+
dist = self.out_dist(dec1)
|
804 |
+
|
805 |
+
if self.debug:
|
806 |
+
return hidden_states_out, enc1, enc2, enc3, enc4, dec4, dec3, dec2, dec1, logits
|
807 |
+
else:
|
808 |
+
return logits, dist
|
809 |
+
|
810 |
+
# print(f"{dec1.shape=}, {dec2.shape=}, {dec3.shape=}, {dec4.shape=}, {logits.shape=}")
|
811 |
+
|
812 |
+
|
813 |
+
|
814 |
+
#%% head
|
815 |
+
class MLPEmbedding(nn.Module):
|
816 |
+
"""
|
817 |
+
Linear Embedding
|
818 |
+
used in head
|
819 |
+
"""
|
820 |
+
def __init__(self, input_dim=2048, embed_dim=768):
|
821 |
+
super().__init__()
|
822 |
+
self.proj = nn.Linear(input_dim, embed_dim)
|
823 |
+
|
824 |
+
def forward(self, x):
|
825 |
+
x = x.flatten(2).transpose(1, 2)
|
826 |
+
x = self.proj(x)
|
827 |
+
return x
|
828 |
+
|
829 |
+
class All_MLP_Head(nn.Module):
|
830 |
+
"""
|
831 |
+
All MLP head in segformer
|
832 |
+
Simple and Efficient Design for Semantic Segmentation with Transformers
|
833 |
+
"""
|
834 |
+
def __init__(self, in_channels=[64,128,320,512], # channel number of multi-scale features
|
835 |
+
in_index=[0,1,2,3],
|
836 |
+
feature_strides=[4,8,16,32],
|
837 |
+
dropout_ratio=0.1,
|
838 |
+
num_classes=3,
|
839 |
+
embedding_dim=768,
|
840 |
+
output_hidden_states=False):
|
841 |
+
super().__init__()
|
842 |
+
self.in_channels = in_channels
|
843 |
+
assert len(feature_strides) == len(self.in_channels)
|
844 |
+
assert min(feature_strides) == feature_strides[0]
|
845 |
+
self.in_index = in_index
|
846 |
+
self.feature_strides = feature_strides
|
847 |
+
self.dropout_ratio = dropout_ratio
|
848 |
+
self.num_classes = num_classes
|
849 |
+
self.output_hidden_states = output_hidden_states
|
850 |
+
|
851 |
+
c1_in_channels, c2_in_channels, c3_in_channels, c4_in_channels = self.in_channels
|
852 |
+
|
853 |
+
# unify channel number to 768
|
854 |
+
self.linear_c4 = MLPEmbedding(input_dim=c4_in_channels, embed_dim=embedding_dim)
|
855 |
+
self.linear_c3 = MLPEmbedding(input_dim=c3_in_channels, embed_dim=embedding_dim)
|
856 |
+
self.linear_c2 = MLPEmbedding(input_dim=c2_in_channels, embed_dim=embedding_dim)
|
857 |
+
self.linear_c1 = MLPEmbedding(input_dim=c1_in_channels, embed_dim=embedding_dim)
|
858 |
+
|
859 |
+
self.linear_fuse = nn.Conv2d(in_channels=embedding_dim*4, out_channels=embedding_dim, kernel_size=1,bias=False)
|
860 |
+
self.batch_norm = nn.BatchNorm2d(embedding_dim) # 4: number of blocks
|
861 |
+
self.activation = nn.ReLU()
|
862 |
+
if dropout_ratio>0:
|
863 |
+
self.dropout = nn.Dropout2d(self.dropout_ratio)
|
864 |
+
self.linear_pred = nn.Conv2d(embedding_dim, self.num_classes, kernel_size=1)
|
865 |
+
|
866 |
+
def forward(self, inputs):
|
867 |
+
# x = self._transform_inputs(inputs) # len=4, 1/4,1/8,1/16,1/32
|
868 |
+
c1, c2, c3, c4 = inputs
|
869 |
+
|
870 |
+
############## MLP decoder on C1-C4 ###########
|
871 |
+
n, _, h, w = c4.shape
|
872 |
+
# normalize channel number and resample to 1/4 HxW
|
873 |
+
_c4 = self.linear_c4(c4).permute(0,2,1).reshape(n, -1, c4.shape[2], c4.shape[3])
|
874 |
+
_c4 = nn.functional.interpolate(_c4, size=c1.size()[2:], mode='bilinear',align_corners=False)
|
875 |
+
|
876 |
+
_c3 = self.linear_c3(c3).permute(0,2,1).reshape(n, -1, c3.shape[2], c3.shape[3])
|
877 |
+
_c3 = nn.functional.interpolate(_c3, size=c1.size()[2:], mode='bilinear',align_corners=False)
|
878 |
+
|
879 |
+
_c2 = self.linear_c2(c2).permute(0,2,1).reshape(n, -1, c2.shape[2], c2.shape[3])
|
880 |
+
_c2 = nn.functional.interpolate(_c2, size=c1.size()[2:], mode='bilinear',align_corners=False)
|
881 |
+
|
882 |
+
_c1 = self.linear_c1(c1).permute(0,2,1).reshape(n, -1, c1.shape[2], c1.shape[3])
|
883 |
+
|
884 |
+
# concatenate features
|
885 |
+
hidden_states = self.linear_fuse(torch.cat([_c4, _c3, _c2, _c1], dim=1))
|
886 |
+
hidden_states = self.batch_norm(hidden_states)
|
887 |
+
hidden_states = self.activation(hidden_states)
|
888 |
+
hidden_states = self.dropout(hidden_states)
|
889 |
+
# predict results
|
890 |
+
x = self.linear_pred(hidden_states)
|
891 |
+
if self.output_hidden_states:
|
892 |
+
return x, hidden_states
|
893 |
+
else:
|
894 |
+
return x
|
895 |
+
|
896 |
+
|
897 |
+
|
898 |
+
#%% test different networks
|
899 |
+
# img_size = 256
|
900 |
+
# in_chans = 3
|
901 |
+
# B = 2
|
902 |
+
# input_img = torch.randn((B,in_chans,img_size,img_size))
|
903 |
+
|
904 |
+
# b3 = mit_b3_demo(img_size=img_size)
|
905 |
+
# b3_out = b3(input_img)
|
906 |
+
# for feature in b3_out:
|
907 |
+
# print(f"{feature.shape=}")
|
908 |
+
# head = All_MLP_Head()
|
909 |
+
# outputs = head(b3_out)
|
910 |
+
# print(f"{outputs.shape = }")
|
911 |
+
|
912 |
+
|
requirements.txt
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
numpy
|
2 |
+
scipy
|
3 |
+
numba
|
4 |
+
einops
|
5 |
+
imagecodecs
|
6 |
+
matplotlib
|
7 |
+
monai
|
8 |
+
pandas
|
9 |
+
pillow
|
10 |
+
scikit-image
|
11 |
+
torch
|
12 |
+
torchvision
|
utils/__init__.py
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
"""
|
4 |
+
Created on Thu Apr 7 10:53:23 2022
|
5 |
+
|
6 |
+
@author: jma
|
7 |
+
"""
|
utils/multi_task_sliding_window_inference.py
ADDED
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
"""
|
4 |
+
Created on Fri Apr 1 19:18:58 2022
|
5 |
+
|
6 |
+
@author: jma
|
7 |
+
"""
|
8 |
+
|
9 |
+
from typing import Any, Callable, List, Sequence, Tuple, Union
|
10 |
+
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
from monai.data.utils import compute_importance_map, dense_patch_slices, get_valid_patch_size
|
13 |
+
from monai.utils import BlendMode, PytorchPadMode, fall_back_tuple, look_up_option
|
14 |
+
|
15 |
+
|
16 |
+
__all__ = ["multi_task_sliding_window_inference"]
|
17 |
+
|
18 |
+
def multi_task_sliding_window_inference(
|
19 |
+
inputs: torch.Tensor,
|
20 |
+
roi_size: Union[Sequence[int], int],
|
21 |
+
sw_batch_size: int,
|
22 |
+
predictor: Callable[..., torch.Tensor],
|
23 |
+
overlap = 0.25,
|
24 |
+
mode = "constant",
|
25 |
+
sigma_scale = 0.125,
|
26 |
+
padding_mode = "constant",
|
27 |
+
cval = 0.0,
|
28 |
+
sw_device = None,
|
29 |
+
device = None,
|
30 |
+
*args: Any,
|
31 |
+
**kwargs: Any,
|
32 |
+
) -> torch.Tensor:
|
33 |
+
"""
|
34 |
+
Sliding window inference on `inputs` with `predictor`.
|
35 |
+
|
36 |
+
When roi_size is larger than the inputs' spatial size, the input image are padded during inference.
|
37 |
+
To maintain the same spatial sizes, the output image will be cropped to the original input size.
|
38 |
+
|
39 |
+
Args:
|
40 |
+
inputs: input image to be processed (assuming NCHW[D])
|
41 |
+
roi_size: the spatial window size for inferences.
|
42 |
+
When its components have None or non-positives, the corresponding inputs dimension will be used.
|
43 |
+
if the components of the `roi_size` are non-positive values, the transform will use the
|
44 |
+
corresponding components of img size. For example, `roi_size=(32, -1)` will be adapted
|
45 |
+
to `(32, 64)` if the second spatial dimension size of img is `64`.
|
46 |
+
sw_batch_size: the batch size to run window slices.
|
47 |
+
predictor: given input tensor `patch_data` in shape NCHW[D], `predictor(patch_data)`
|
48 |
+
should return a prediction with the same spatial shape and batch_size, i.e. NMHW[D];
|
49 |
+
where HW[D] represents the patch spatial size, M is the number of output channels, N is `sw_batch_size`.
|
50 |
+
overlap: Amount of overlap between scans.
|
51 |
+
mode: {``"constant"``, ``"gaussian"``}
|
52 |
+
How to blend output of overlapping windows. Defaults to ``"constant"``.
|
53 |
+
|
54 |
+
- ``"constant``": gives equal weight to all predictions.
|
55 |
+
- ``"gaussian``": gives less weight to predictions on edges of windows.
|
56 |
+
|
57 |
+
sigma_scale: the standard deviation coefficient of the Gaussian window when `mode` is ``"gaussian"``.
|
58 |
+
Default: 0.125. Actual window sigma is ``sigma_scale`` * ``dim_size``.
|
59 |
+
When sigma_scale is a sequence of floats, the values denote sigma_scale at the corresponding
|
60 |
+
spatial dimensions.
|
61 |
+
padding_mode: {``"constant"``, ``"reflect"``, ``"replicate"``, ``"circular"``}
|
62 |
+
Padding mode for ``inputs``, when ``roi_size`` is larger than inputs. Defaults to ``"constant"``
|
63 |
+
See also: https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html
|
64 |
+
cval: fill value for 'constant' padding mode. Default: 0
|
65 |
+
sw_device: device for the window data.
|
66 |
+
By default the device (and accordingly the memory) of the `inputs` is used.
|
67 |
+
Normally `sw_device` should be consistent with the device where `predictor` is defined.
|
68 |
+
device: device for the stitched output prediction.
|
69 |
+
By default the device (and accordingly the memory) of the `inputs` is used. If for example
|
70 |
+
set to device=torch.device('cpu') the gpu memory consumption is less and independent of the
|
71 |
+
`inputs` and `roi_size`. Output is on the `device`.
|
72 |
+
args: optional args to be passed to ``predictor``.
|
73 |
+
kwargs: optional keyword args to be passed to ``predictor``.
|
74 |
+
|
75 |
+
Note:
|
76 |
+
- input must be channel-first and have a batch dim, supports N-D sliding window.
|
77 |
+
|
78 |
+
"""
|
79 |
+
num_spatial_dims = len(inputs.shape) - 2
|
80 |
+
if overlap < 0 or overlap >= 1:
|
81 |
+
raise AssertionError("overlap must be >= 0 and < 1.")
|
82 |
+
|
83 |
+
# determine image spatial size and batch size
|
84 |
+
# Note: all input images must have the same image size and batch size
|
85 |
+
image_size_ = list(inputs.shape[2:])
|
86 |
+
batch_size = inputs.shape[0]
|
87 |
+
|
88 |
+
if device is None:
|
89 |
+
device = inputs.device
|
90 |
+
if sw_device is None:
|
91 |
+
sw_device = inputs.device
|
92 |
+
|
93 |
+
roi_size = fall_back_tuple(roi_size, image_size_)
|
94 |
+
# in case that image size is smaller than roi size
|
95 |
+
image_size = tuple(max(image_size_[i], roi_size[i]) for i in range(num_spatial_dims))
|
96 |
+
pad_size = []
|
97 |
+
for k in range(len(inputs.shape) - 1, 1, -1):
|
98 |
+
diff = max(roi_size[k - 2] - inputs.shape[k], 0)
|
99 |
+
half = diff // 2
|
100 |
+
pad_size.extend([half, diff - half])
|
101 |
+
inputs = F.pad(inputs, pad=pad_size, mode=mode, value=cval)
|
102 |
+
|
103 |
+
scan_interval = _get_scan_interval(image_size, roi_size, num_spatial_dims, overlap)
|
104 |
+
|
105 |
+
# Store all slices in list
|
106 |
+
slices = dense_patch_slices(image_size, roi_size, scan_interval)
|
107 |
+
num_win = len(slices) # number of windows per image
|
108 |
+
total_slices = num_win * batch_size # total number of windows
|
109 |
+
|
110 |
+
# Create window-level importance map
|
111 |
+
importance_map = compute_importance_map(
|
112 |
+
get_valid_patch_size(image_size, roi_size), mode="gaussian", sigma_scale=sigma_scale, device=device
|
113 |
+
)
|
114 |
+
|
115 |
+
# Perform predictions
|
116 |
+
output_image, count_map = torch.tensor(0.0, device=device), torch.tensor(0.0, device=device)
|
117 |
+
output_dist = torch.tensor(0.0, device=device)
|
118 |
+
_initialized = False
|
119 |
+
for slice_g in range(0, total_slices, sw_batch_size):
|
120 |
+
slice_range = range(slice_g, min(slice_g + sw_batch_size, total_slices))
|
121 |
+
unravel_slice = [
|
122 |
+
[slice(int(idx / num_win), int(idx / num_win) + 1), slice(None)] + list(slices[idx % num_win])
|
123 |
+
for idx in slice_range
|
124 |
+
]
|
125 |
+
window_data = torch.cat([inputs[win_slice] for win_slice in unravel_slice]).to(sw_device)
|
126 |
+
seg_logit, seg_dist = predictor(window_data)# .to(device) # batched patch segmentation
|
127 |
+
seg_logit = torch.nn.functional.interpolate(seg_logit, size=roi_size, mode="bilinear", align_corners=False)
|
128 |
+
seg_logit = torch.softmax(seg_logit, dim=1)
|
129 |
+
seg_dist = torch.nn.functional.interpolate(seg_dist, size=roi_size, mode="bilinear", align_corners=False)
|
130 |
+
seg_dist = torch.sigmoid(seg_dist)
|
131 |
+
|
132 |
+
if not _initialized: # init. buffer at the first iteration
|
133 |
+
output_classes = seg_logit.shape[1]
|
134 |
+
dist_class = seg_dist.shape[1]
|
135 |
+
output_shape = [batch_size, output_classes] + list(image_size)
|
136 |
+
output_dist_shape = [batch_size, dist_class] + list(image_size)
|
137 |
+
# allocate memory to store the full output and the count for overlapping parts
|
138 |
+
output_image = torch.zeros(output_shape, dtype=torch.float32, device=device)
|
139 |
+
output_dist = torch.zeros(output_dist_shape, dtype=torch.float32, device=device)
|
140 |
+
count_map = torch.zeros(output_shape, dtype=torch.float32, device=device)
|
141 |
+
count_dist_map = torch.zeros(output_dist_shape, dtype=torch.float32, device=device)
|
142 |
+
_initialized = True
|
143 |
+
|
144 |
+
# store the result in the proper location of the full output. Apply weights from importance map.
|
145 |
+
for idx, original_idx in zip(slice_range, unravel_slice):
|
146 |
+
output_image[original_idx] += importance_map * seg_logit[idx - slice_g]
|
147 |
+
output_dist[original_idx] += importance_map * seg_dist[idx - slice_g]
|
148 |
+
count_map[original_idx] += importance_map
|
149 |
+
count_dist_map[original_idx] += importance_map
|
150 |
+
|
151 |
+
# account for any overlapping sections
|
152 |
+
output_image = output_image / count_map
|
153 |
+
output_dist = output_dist / count_dist_map
|
154 |
+
|
155 |
+
final_slicing: List[slice] = []
|
156 |
+
for sp in range(num_spatial_dims):
|
157 |
+
slice_dim = slice(pad_size[sp * 2], image_size_[num_spatial_dims - sp - 1] + pad_size[sp * 2])
|
158 |
+
final_slicing.insert(0, slice_dim)
|
159 |
+
while len(final_slicing) < len(output_image.shape):
|
160 |
+
final_slicing.insert(0, slice(None))
|
161 |
+
return output_image[final_slicing], output_dist[final_slicing]
|
162 |
+
|
163 |
+
|
164 |
+
|
165 |
+
def _get_scan_interval(
|
166 |
+
image_size: Sequence[int], roi_size: Sequence[int], num_spatial_dims: int, overlap: float
|
167 |
+
) -> Tuple[int, ...]:
|
168 |
+
"""
|
169 |
+
Compute scan interval according to the image size, roi size and overlap.
|
170 |
+
Scan interval will be `int((1 - overlap) * roi_size)`, if interval is 0,
|
171 |
+
use 1 instead to make sure sliding window works.
|
172 |
+
|
173 |
+
"""
|
174 |
+
if len(image_size) != num_spatial_dims:
|
175 |
+
raise ValueError("image coord different from spatial dims.")
|
176 |
+
if len(roi_size) != num_spatial_dims:
|
177 |
+
raise ValueError("roi coord different from spatial dims.")
|
178 |
+
|
179 |
+
scan_interval = []
|
180 |
+
for i in range(num_spatial_dims):
|
181 |
+
if roi_size[i] == image_size[i]:
|
182 |
+
scan_interval.append(int(roi_size[i]))
|
183 |
+
else:
|
184 |
+
interval = int(roi_size[i] * (1 - overlap))
|
185 |
+
scan_interval.append(interval if interval > 0 else 1)
|
186 |
+
return tuple(scan_interval)
|
187 |
+
|
utils/postprocess.py
ADDED
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
# -*- coding: utf-8 -*-
|
3 |
+
"""
|
4 |
+
Created on Thu Apr 7 10:51:48 2022
|
5 |
+
|
6 |
+
@author: jma
|
7 |
+
"""
|
8 |
+
|
9 |
+
import numpy as np
|
10 |
+
from skimage import segmentation, measure, exposure, morphology
|
11 |
+
import scipy.ndimage as nd
|
12 |
+
from tqdm import tqdm
|
13 |
+
import skimage
|
14 |
+
import colorsys
|
15 |
+
|
16 |
+
def fill_holes(label_img, size=10, connectivity=1):
|
17 |
+
output_image = np.copy(label_img)
|
18 |
+
props = measure.regionprops(np.squeeze(label_img.astype('int')), cache=False)
|
19 |
+
for prop in props:
|
20 |
+
if prop.euler_number < 1:
|
21 |
+
|
22 |
+
patch = output_image[prop.slice]
|
23 |
+
|
24 |
+
filled = morphology.remove_small_holes(
|
25 |
+
ar=(patch == prop.label),
|
26 |
+
area_threshold=size,
|
27 |
+
connectivity=connectivity)
|
28 |
+
|
29 |
+
output_image[prop.slice] = np.where(filled, prop.label, patch)
|
30 |
+
|
31 |
+
return output_image
|
32 |
+
|
33 |
+
def watershed_post(distmaps, interiors, dist_thre=0.1, interior_thre=0.2):
|
34 |
+
"""
|
35 |
+
Parameters
|
36 |
+
----------
|
37 |
+
distmaps : float (N, H, W) N is the number of cells
|
38 |
+
distance transform map of cell/nuclear [0,1].
|
39 |
+
interiors : float (N, H, W)
|
40 |
+
interior map of cell/nuclear [0,1].
|
41 |
+
|
42 |
+
Returns
|
43 |
+
-------
|
44 |
+
label_images : uint (N, H, W)
|
45 |
+
cell/nuclear instance segmentation.
|
46 |
+
|
47 |
+
"""
|
48 |
+
|
49 |
+
label_images = []
|
50 |
+
for maxima, interior in zip(distmaps, interiors):# in interiors[0:num]:
|
51 |
+
interior = nd.gaussian_filter(interior.astype(np.float32), 2)
|
52 |
+
# find marker based on distance map
|
53 |
+
if skimage.__version__ > '0.18.2':
|
54 |
+
markers = measure.label(morphology.h_maxima(image=maxima, h=dist_thre, footprint=morphology.disk(2)))
|
55 |
+
else:
|
56 |
+
markers = measure.label(morphology.h_maxima(image=maxima, h=dist_thre, selem=morphology.disk(2)))
|
57 |
+
# print('distmap marker num:', np.max(markers), 'interior marker num:', np.max(makers_interior))
|
58 |
+
|
59 |
+
label_image = segmentation.watershed(-1 * interior, markers,
|
60 |
+
mask=interior > interior_thre, # 0.2/0.3
|
61 |
+
watershed_line=0)
|
62 |
+
|
63 |
+
label_image = morphology.remove_small_objects(label_image, min_size=15)
|
64 |
+
# fill in holes that lie completely within a segmentation label
|
65 |
+
label_image = fill_holes(label_image, size=15)
|
66 |
+
|
67 |
+
# Relabel the label image
|
68 |
+
label_image, _, _ = segmentation.relabel_sequential(label_image)
|
69 |
+
label_images.append(label_image)
|
70 |
+
label_images = np.stack(label_images, axis=0).astype(np.uint)
|
71 |
+
return label_images
|
72 |
+
|
73 |
+
|
74 |
+
|
75 |
+
def hsv_to_rgb(arr):
|
76 |
+
hsv_to_rgb_channels = np.vectorize(colorsys.hsv_to_rgb)
|
77 |
+
h, s, v = np.rollaxis(arr, axis=-1)
|
78 |
+
r, g, b = hsv_to_rgb_channels(h, s, v)
|
79 |
+
rgb = np.stack((r,g,b), axis=-1)
|
80 |
+
return rgb
|
81 |
+
|
82 |
+
def mask_overlay(img, masks):
|
83 |
+
""" overlay masks on image (set image to grayscale)
|
84 |
+
Adapted from https://github.com/MouseLand/cellpose/blob/06df602fbe074be02db3d716e280f0990816c726/cellpose/plot.py#L172
|
85 |
+
Parameters
|
86 |
+
----------------
|
87 |
+
|
88 |
+
img: int or float, 2D or 3D array
|
89 |
+
img is of size [Ly x Lx (x nchan)]
|
90 |
+
|
91 |
+
masks: int, 2D array
|
92 |
+
masks where 0=NO masks; 1,2,...=mask labels
|
93 |
+
|
94 |
+
Returns
|
95 |
+
----------------
|
96 |
+
|
97 |
+
RGB: uint8, 3D array
|
98 |
+
array of masks overlaid on grayscale image
|
99 |
+
|
100 |
+
"""
|
101 |
+
|
102 |
+
if img.ndim>2:
|
103 |
+
img = img.astype(np.float32).mean(axis=-1)
|
104 |
+
else:
|
105 |
+
img = img.astype(np.float32)
|
106 |
+
|
107 |
+
HSV = np.zeros((img.shape[0], img.shape[1], 3), np.float32)
|
108 |
+
HSV[:,:,2] = np.clip((img / 255. if img.max() > 1 else img) * 1.5, 0, 1)
|
109 |
+
hues = np.linspace(0, 1, masks.max()+1)[np.random.permutation(masks.max())]
|
110 |
+
for n in range(int(masks.max())):
|
111 |
+
ipix = (masks==n+1).nonzero()
|
112 |
+
HSV[ipix[0],ipix[1],0] = hues[n]
|
113 |
+
|
114 |
+
HSV[ipix[0],ipix[1],1] = 1.0
|
115 |
+
RGB = (hsv_to_rgb(HSV) * 255).astype(np.uint8)
|
116 |
+
return RGB
|
117 |
+
|
118 |
+
|
119 |
+
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
|
124 |
+
|
125 |
+
|