Spaces:
Running
on
Zero
Running
on
Zero
SunderAli17
commited on
Commit
•
512dfd4
1
Parent(s):
c244c09
Delete toonmage/fluxpipeline.py
Browse files- toonmage/fluxpipeline.py +0 -188
toonmage/fluxpipeline.py
DELETED
@@ -1,188 +0,0 @@
|
|
1 |
-
import gc
|
2 |
-
|
3 |
-
import cv2
|
4 |
-
import insightface
|
5 |
-
import torch
|
6 |
-
import torch.nn as nn
|
7 |
-
from basicsr.utils import img2tensor, tensor2img
|
8 |
-
from facexlib.parsing import init_parsing_model
|
9 |
-
from facexlib.utils.face_restoration_helper import FaceRestoreHelper
|
10 |
-
from huggingface_hub import hf_hub_download, snapshot_download
|
11 |
-
from insightface.app import FaceAnalysis
|
12 |
-
from safetensors.torch import load_file
|
13 |
-
from torchvision.transforms import InterpolationMode
|
14 |
-
from torchvision.transforms.functional import normalize, resize
|
15 |
-
|
16 |
-
from eva_clip import create_model_and_transforms
|
17 |
-
from eva_clip.constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
|
18 |
-
from toonmage.fluxencoders import IDFormer, PerceiverAttentionCA
|
19 |
-
|
20 |
-
|
21 |
-
class ToonMagePipeline(nn.Module):
|
22 |
-
def __init__(self, dit, device, weight_dtype=torch.bfloat16, *args, **kwargs):
|
23 |
-
super().__init__()
|
24 |
-
self.device = device
|
25 |
-
self.weight_dtype = weight_dtype
|
26 |
-
double_interval = 2
|
27 |
-
single_interval = 4
|
28 |
-
|
29 |
-
# init encoder
|
30 |
-
self.pulid_encoder = IDFormer().to(self.device, self.weight_dtype)
|
31 |
-
|
32 |
-
num_ca = 19 // double_interval + 38 // single_interval
|
33 |
-
if 19 % double_interval != 0:
|
34 |
-
num_ca += 1
|
35 |
-
if 38 % single_interval != 0:
|
36 |
-
num_ca += 1
|
37 |
-
self.pulid_ca = nn.ModuleList([
|
38 |
-
PerceiverAttentionCA().to(self.device, self.weight_dtype) for _ in range(num_ca)
|
39 |
-
])
|
40 |
-
|
41 |
-
dit.pulid_ca = self.pulid_ca
|
42 |
-
dit.toonmage_double_interval = double_interval
|
43 |
-
dit.toonmage_single_interval = single_interval
|
44 |
-
|
45 |
-
# preprocessors
|
46 |
-
# face align and parsing
|
47 |
-
self.face_helper = FaceRestoreHelper(
|
48 |
-
upscale_factor=1,
|
49 |
-
face_size=512,
|
50 |
-
crop_ratio=(1, 1),
|
51 |
-
det_model='retinaface_resnet50',
|
52 |
-
save_ext='png',
|
53 |
-
device=self.device,
|
54 |
-
)
|
55 |
-
self.face_helper.face_parse = None
|
56 |
-
self.face_helper.face_parse = init_parsing_model(model_name='bisenet', device=self.device)
|
57 |
-
# clip-vit backbone
|
58 |
-
model, _, _ = create_model_and_transforms('EVA02-CLIP-L-14-336', 'eva_clip', force_custom_clip=True)
|
59 |
-
model = model.visual
|
60 |
-
self.clip_vision_model = model.to(self.device, dtype=self.weight_dtype)
|
61 |
-
eva_transform_mean = getattr(self.clip_vision_model, 'image_mean', OPENAI_DATASET_MEAN)
|
62 |
-
eva_transform_std = getattr(self.clip_vision_model, 'image_std', OPENAI_DATASET_STD)
|
63 |
-
if not isinstance(eva_transform_mean, (list, tuple)):
|
64 |
-
eva_transform_mean = (eva_transform_mean,) * 3
|
65 |
-
if not isinstance(eva_transform_std, (list, tuple)):
|
66 |
-
eva_transform_std = (eva_transform_std,) * 3
|
67 |
-
self.eva_transform_mean = eva_transform_mean
|
68 |
-
self.eva_transform_std = eva_transform_std
|
69 |
-
# antelopev2
|
70 |
-
snapshot_download('DIAMONIK7777/antelopev2', local_dir='models/antelopev2')
|
71 |
-
self.app = FaceAnalysis(
|
72 |
-
name='antelopev2', root='.', providers=['CPUExecutionProvider']
|
73 |
-
)
|
74 |
-
self.app.prepare(ctx_id=0, det_size=(640, 640))
|
75 |
-
self.handler_ante = insightface.model_zoo.get_model('models/antelopev2/glintr100.onnx', providers=['CPUExecutionProvider'])
|
76 |
-
self.handler_ante.prepare(ctx_id=0)
|
77 |
-
|
78 |
-
gc.collect()
|
79 |
-
torch.cuda.empty_cache()
|
80 |
-
|
81 |
-
# self.load_pretrain()
|
82 |
-
|
83 |
-
# other configs
|
84 |
-
self.debug_img_list = []
|
85 |
-
|
86 |
-
def load_pretrain(self, pretrain_path=None):
|
87 |
-
hf_hub_download('SunderAli17/SAK', 'toonmage_flux_v2.safetensors', local_dir ='models')
|
88 |
-
ckpt_path = 'models/toonmage_flux_v2.safetensors'
|
89 |
-
if pretrain_path is not None:
|
90 |
-
ckpt_path = pretrain_path
|
91 |
-
state_dict = load_file(ckpt_path)
|
92 |
-
state_dict_dict = {}
|
93 |
-
for k, v in state_dict.items():
|
94 |
-
module = k.split('.')[0]
|
95 |
-
state_dict_dict.setdefault(module, {})
|
96 |
-
new_k = k[len(module) + 1:]
|
97 |
-
state_dict_dict[module][new_k] = v
|
98 |
-
|
99 |
-
for module in state_dict_dict:
|
100 |
-
print(f'loading from {module}')
|
101 |
-
getattr(self, module).load_state_dict(state_dict_dict[module], strict=True)
|
102 |
-
|
103 |
-
del state_dict
|
104 |
-
del state_dict_dict
|
105 |
-
|
106 |
-
def to_gray(self, img):
|
107 |
-
x = 0.299 * img[:, 0:1] + 0.587 * img[:, 1:2] + 0.114 * img[:, 2:3]
|
108 |
-
x = x.repeat(1, 3, 1, 1)
|
109 |
-
return x
|
110 |
-
|
111 |
-
def get_id_embedding(self, image, cal_uncond=False):
|
112 |
-
"""
|
113 |
-
Args:
|
114 |
-
image: numpy rgb image, range [0, 255]
|
115 |
-
"""
|
116 |
-
self.face_helper.clean_all()
|
117 |
-
self.debug_img_list = []
|
118 |
-
image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
|
119 |
-
# get antelopev2 embedding
|
120 |
-
# for k in self.app.models.keys():
|
121 |
-
# self.app.models[k].session.set_providers(['CUDAExecutionProvider'])
|
122 |
-
face_info = self.app.get(image_bgr)
|
123 |
-
if len(face_info) > 0:
|
124 |
-
face_info = sorted(face_info, key=lambda x: (x['bbox'][2] - x['bbox'][0]) * (x['bbox'][3] - x['bbox'][1]))[
|
125 |
-
-1
|
126 |
-
] # only use the maximum face
|
127 |
-
id_ante_embedding = face_info['embedding']
|
128 |
-
self.debug_img_list.append(
|
129 |
-
image[
|
130 |
-
int(face_info['bbox'][1]) : int(face_info['bbox'][3]),
|
131 |
-
int(face_info['bbox'][0]) : int(face_info['bbox'][2]),
|
132 |
-
]
|
133 |
-
)
|
134 |
-
else:
|
135 |
-
id_ante_embedding = None
|
136 |
-
|
137 |
-
# using facexlib to detect and align face
|
138 |
-
self.face_helper.read_image(image_bgr)
|
139 |
-
self.face_helper.get_face_landmarks_5(only_center_face=True)
|
140 |
-
self.face_helper.align_warp_face()
|
141 |
-
if len(self.face_helper.cropped_faces) == 0:
|
142 |
-
raise RuntimeError('facexlib align face fail')
|
143 |
-
align_face = self.face_helper.cropped_faces[0]
|
144 |
-
# incase insightface didn't detect face
|
145 |
-
if id_ante_embedding is None:
|
146 |
-
print('fail to detect face using insightface, extract embedding on align face')
|
147 |
-
# self.handler_ante.session.set_providers(['CUDAExecutionProvider'])
|
148 |
-
id_ante_embedding = self.handler_ante.get_feat(align_face)
|
149 |
-
|
150 |
-
id_ante_embedding = torch.from_numpy(id_ante_embedding).to(self.device, self.weight_dtype)
|
151 |
-
if id_ante_embedding.ndim == 1:
|
152 |
-
id_ante_embedding = id_ante_embedding.unsqueeze(0)
|
153 |
-
|
154 |
-
# parsing
|
155 |
-
input = img2tensor(align_face, bgr2rgb=True).unsqueeze(0) / 255.0
|
156 |
-
input = input.to(self.device)
|
157 |
-
parsing_out = self.face_helper.face_parse(normalize(input, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225]))[0]
|
158 |
-
parsing_out = parsing_out.argmax(dim=1, keepdim=True)
|
159 |
-
bg_label = [0, 16, 18, 7, 8, 9, 14, 15]
|
160 |
-
bg = sum(parsing_out == i for i in bg_label).bool()
|
161 |
-
white_image = torch.ones_like(input)
|
162 |
-
# only keep the face features
|
163 |
-
face_features_image = torch.where(bg, white_image, self.to_gray(input))
|
164 |
-
self.debug_img_list.append(tensor2img(face_features_image, rgb2bgr=False))
|
165 |
-
|
166 |
-
# transform img before sending to eva-clip-vit
|
167 |
-
face_features_image = resize(face_features_image, self.clip_vision_model.image_size, InterpolationMode.BICUBIC)
|
168 |
-
face_features_image = normalize(face_features_image, self.eva_transform_mean, self.eva_transform_std)
|
169 |
-
id_cond_vit, id_vit_hidden = self.clip_vision_model(
|
170 |
-
face_features_image.to(self.weight_dtype), return_all_features=False, return_hidden=True, shuffle=False
|
171 |
-
)
|
172 |
-
id_cond_vit_norm = torch.norm(id_cond_vit, 2, 1, True)
|
173 |
-
id_cond_vit = torch.div(id_cond_vit, id_cond_vit_norm)
|
174 |
-
|
175 |
-
id_cond = torch.cat([id_ante_embedding, id_cond_vit], dim=-1)
|
176 |
-
|
177 |
-
id_embedding = self.pulid_encoder(id_cond, id_vit_hidden)
|
178 |
-
|
179 |
-
if not cal_uncond:
|
180 |
-
return id_embedding, None
|
181 |
-
|
182 |
-
id_uncond = torch.zeros_like(id_cond)
|
183 |
-
id_vit_hidden_uncond = []
|
184 |
-
for layer_idx in range(0, len(id_vit_hidden)):
|
185 |
-
id_vit_hidden_uncond.append(torch.zeros_like(id_vit_hidden[layer_idx]))
|
186 |
-
uncond_id_embedding = self.pulid_encoder(id_uncond, id_vit_hidden_uncond)
|
187 |
-
|
188 |
-
return id_embedding, uncond_id_embedding
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|