-
Notifications
You must be signed in to change notification settings - Fork 148
/
extract_features.py
337 lines (276 loc) · 13 KB
/
extract_features.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
import os
from pathlib import Path
import sys
current_file_path = Path(__file__).resolve()
sys.path.insert(0, str(current_file_path.parent.parent))
from PIL import Image
import torch
from torchvision import transforms as T
import numpy as np
import json
from tqdm import tqdm
import argparse
import threading
from queue import Queue
from pathlib import Path
from torch.utils.data import DataLoader, RandomSampler
from accelerate import Accelerator
from torchvision.transforms.functional import InterpolationMode
from torchvision.datasets.folder import default_loader
from diffusion.model.t5 import T5Embedder
from diffusers.models import AutoencoderKL
from diffusion.data.datasets.InternalData import InternalData
from diffusion.utils.misc import SimpleTimer
from diffusion.utils.data_sampler import AspectRatioBatchSampler
from diffusion.data.builder import DATASETS
from diffusion.data import ASPECT_RATIO_512, ASPECT_RATIO_1024
def get_closest_ratio(height: float, width: float, ratios: dict):
aspect_ratio = height / width
closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - aspect_ratio))
return ratios[closest_ratio], float(closest_ratio)
@DATASETS.register_module()
class DatasetMS(InternalData):
def __init__(self,
root, # Notice: need absolute path here
image_list_json=['data_info.json'],
transform=None,
resolution=1024,
load_vae_feat=False,
aspect_ratio_type=None,
start_index=0,
end_index=100000000,
**kwargs):
assert os.path.isabs(root), 'root must be a absolute path'
self.root = root
self.img_dir_name = 'InternalImgs' # need to change to according to your data structure
self.json_dir_name = 'InternalData' # need to change to according to your data structure
self.transform = transform
self.load_vae_feat = load_vae_feat
self.resolution = resolution
self.meta_data_clean = []
self.img_samples = []
self.txt_feat_samples = []
self.aspect_ratio = aspect_ratio_type
assert self.aspect_ratio in [ASPECT_RATIO_1024, ASPECT_RATIO_512]
self.ratio_index = {}
self.ratio_nums = {}
for k, v in self.aspect_ratio.items():
self.ratio_index[float(k)] = [] # used for self.getitem
self.ratio_nums[float(k)] = 0 # used for batch-sampler
image_list_json = image_list_json if isinstance(image_list_json, list) else [image_list_json]
for json_file in image_list_json:
meta_data = self.load_json(os.path.join(self.root, 'partition', json_file))
meta_data_clean = [item for item in meta_data if item['ratio'] <= 4]
self.meta_data_clean.extend(meta_data_clean)
self.img_samples.extend([os.path.join(self.root.replace(self.json_dir_name, self.img_dir_name), item['path']) for item in meta_data_clean])
self.img_samples = self.img_samples[start_index: end_index]
# scan the dataset for ratio static
for i, info in enumerate(self.meta_data_clean[:len(self.meta_data_clean)//3]):
ori_h, ori_w = info['height'], info['width']
closest_size, closest_ratio = get_closest_ratio(ori_h, ori_w, self.aspect_ratio)
self.ratio_nums[closest_ratio] += 1
if len(self.ratio_index[closest_ratio]) == 0:
self.ratio_index[closest_ratio].append(i)
# Set loader and extensions
if self.load_vae_feat:
raise ValueError("No VAE loader here")
self.loader = default_loader
def __getitem__(self, idx):
data_info = {}
for i in range(20):
try:
img_path = self.img_samples[idx]
img = self.loader(img_path)
if self.transform:
img = self.transform(img)
# Calculate closest aspect ratio and resize & crop image[w, h]
if isinstance(img, Image.Image):
h, w = (img.size[1], img.size[0])
assert h, w == (self.meta_data_clean[idx]['height'], self.meta_data_clean[idx]['width'])
closest_size, closest_ratio = get_closest_ratio(h, w, self.aspect_ratio)
closest_size = list(map(lambda x: int(x), closest_size))
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB')),
T.Resize(closest_size, interpolation=InterpolationMode.BICUBIC), # Image.BICUBIC
T.CenterCrop(closest_size),
T.ToTensor(),
T.Normalize([.5], [.5]),
])
img = transform(img)
data_info['img_hw'] = torch.tensor([h, w], dtype=torch.float32)
data_info['aspect_ratio'] = closest_ratio
# change the path according to your data structure
return img, '_'.join(self.img_samples[idx].rsplit('/', 2)[-2:]) # change from 'serial-number-of-dir/serial-number-of-image.png' ---> 'serial-number-of-dir_serial-number-of-image.png'
except Exception as e:
print(f"Error details: {str(e)}")
idx = np.random.randint(len(self))
raise RuntimeError('Too many bad data.')
def get_data_info(self, idx):
data_info = self.meta_data_clean[idx]
return {'height': data_info['height'], 'width': data_info['width']}
def extract_caption_t5_do(q):
while not q.empty():
item = q.get()
extract_caption_t5_job(item)
q.task_done()
def extract_caption_t5_job(item):
global mutex
global t5
global t5_save_dir
with torch.no_grad():
caption = item['prompt'].strip()
if isinstance(caption, str):
caption = [caption]
save_path = os.path.join(t5_save_dir, Path(item['path']).stem)
if os.path.exists(save_path + ".npz"):
return
try:
mutex.acquire()
caption_emb, emb_mask = t5.get_text_embeddings(caption)
mutex.release()
emb_dict = {
'caption_feature': caption_emb.float().cpu().data.numpy(),
'attention_mask': emb_mask.cpu().data.numpy(),
}
np.savez_compressed(save_path, **emb_dict)
except Exception as e:
print(e)
def extract_caption_t5():
global t5
global t5_save_dir
# global images_extension
t5 = T5Embedder(device="cuda", local_cache=True, cache_dir=f'{args.pretrained_models_dir}/t5_ckpts')
t5_save_dir = args.t5_save_root
os.makedirs(t5_save_dir, exist_ok=True)
train_data_json = json.load(open(args.json_path, 'r'))
train_data = train_data_json[args.start_index: args.end_index]
global mutex
mutex = threading.Lock()
jobs = Queue()
for item in tqdm(train_data):
jobs.put(item)
for _ in range(20):
worker = threading.Thread(target=extract_caption_t5_do, args=(jobs,))
worker.start()
jobs.join()
def extract_img_vae_do(q):
while not q.empty():
item = q.get()
extract_img_vae_job(item)
q.task_done()
def extract_img_vae_job(item):
return
def extract_img_vae():
vae = AutoencoderKL.from_pretrained(f'{args.pretrained_models_dir}/sd-vae-ft-ema').to(device)
train_data_json = json.load(open(args.json_path, 'r'))
image_names = set()
vae_save_root = f'{args.vae_save_root}/{image_resize}resolution'
os.umask(0o000) # file permission: 666; dir permission: 777
os.makedirs(vae_save_root, exist_ok=True)
vae_save_dir = os.path.join(vae_save_root, 'noflip')
os.makedirs(vae_save_dir, exist_ok=True)
for item in train_data_json:
image_name = item['path']
if image_name in image_names:
continue
image_names.add(image_name)
lines = list(image_names)
lines.sort()
lines = lines[args.start_index: args.end_index]
_, images_extension = os.path.splitext(lines[0])
transform = T.Compose([
T.Lambda(lambda img: img.convert('RGB')),
T.Resize(image_resize), # Image.BICUBIC
T.CenterCrop(image_resize),
T.ToTensor(),
T.Normalize([.5], [.5]),
])
os.umask(0o000) # file permission: 666; dir permission: 777
for image_name in tqdm(lines):
save_path = os.path.join(vae_save_dir, Path(image_name).stem)
if os.path.exists(save_path + ".npy"):
continue
try:
img = Image.open(f'{args.dataset_root}/{image_name}')
img = transform(img).to(device)[None]
with torch.no_grad():
posterior = vae.encode(img).latent_dist
z = torch.cat([posterior.mean, posterior.std], dim=1).detach().cpu().numpy().squeeze()
np.save(save_path, z)
except Exception as e:
print(e)
print(image_name)
def save_results(results, paths, signature, work_dir):
timer = SimpleTimer(len(results), log_interval=100, desc=f"Saving Results")
# save to npy
new_paths = []
os.umask(0o000) # file permission: 666; dir permission: 777
for res, p in zip(results, paths):
file_name = p.split('.')[0] + '.npy'
new_folder = signature
save_folder = os.path.join(work_dir, new_folder)
if os.path.exists(save_folder):
raise FileExistsError(f"{save_folder} exists. BE careful not to overwrite your files. Comment this error raising for overwriting!!")
os.makedirs(save_folder, exist_ok=True)
new_paths.append(os.path.join(new_folder, file_name))
np.save(os.path.join(save_folder, file_name), res)
timer.log()
# save paths
with open(os.path.join(work_dir, f"VAE-{signature}.txt"), 'w') as f:
f.write('\n'.join(new_paths))
def inference(vae, dataloader, signature, work_dir):
timer = SimpleTimer(len(dataloader), log_interval=100, desc=f"VAE-Inference")
for step, batch in enumerate(dataloader):
with torch.no_grad():
with torch.cuda.amp.autocast(enabled=True):
posterior = vae.encode(batch[0]).latent_dist
results = torch.cat([posterior.mean, posterior.std], dim=1).detach().cpu().numpy()
path = batch[1]
save_results(results, path, signature=signature, work_dir=work_dir)
timer.log()
def extract_img_vae_multiscale(bs=1):
assert image_resize in [512, 1024]
work_dir = os.path.abspath(args.vae_save_root)
os.umask(0o000) # file permission: 666; dir permission: 777
os.makedirs(work_dir, exist_ok=True)
accelerator = Accelerator(mixed_precision='fp16')
vae = AutoencoderKL.from_pretrained(f'{args.pretrained_models_dir}/sd-vae-ft-ema').to(device)
signature = 'ms'
aspect_ratio_type = ASPECT_RATIO_1024 if image_resize == 1024 else ASPECT_RATIO_512
dataset = DatasetMS(args.dataset_root, image_list_json=[args.json_file], transform=None, sample_subset=None,
aspect_ratio_type=aspect_ratio_type, start_index=args.start_index, end_index=args.end_index)
# 创建 AspectRatioBatchSampler
sampler = AspectRatioBatchSampler(sampler=RandomSampler(dataset), dataset=dataset, batch_size=bs, aspect_ratios=dataset.aspect_ratio, ratio_nums=dataset.ratio_nums)
# 创建 DataLoader
dataloader = DataLoader(dataset, batch_sampler=sampler, num_workers=13, pin_memory=True)
dataloader = accelerator.prepare(dataloader, )
inference(vae, dataloader, signature=signature, work_dir=work_dir)
accelerator.wait_for_everyone()
print('done')
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--multi_scale", action='store_true', default=False, help="multi-scale feature extraction")
parser.add_argument("--img_size", default=512, type=int, help="image scale for multi-scale feature extraction")
parser.add_argument('--start_index', default=0, type=int)
parser.add_argument('--end_index', default=1000000, type=int)
parser.add_argument('--json_path', type=str)
parser.add_argument('--t5_save_root', default='data/data_toy/caption_feature_wmask', type=str)
parser.add_argument('--vae_save_root', default='data/data_toy/img_vae_features', type=str)
parser.add_argument('--dataset_root', default='data/data_toy', type=str)
parser.add_argument('--pretrained_models_dir', default='output/pretrained_models', type=str)
### for multi-scale(ms) vae feauture extraction
parser.add_argument('--json_file', type=str)
return parser.parse_args()
if __name__ == '__main__':
args = get_args()
device = "cuda" if torch.cuda.is_available() else "cpu"
image_resize = args.img_size
# prepare extracted caption t5 features for training
extract_caption_t5()
# prepare extracted image vae features for training
if args.multi_scale:
print('Extracting Multi-scale Image Resolution based on %s' % image_resize)
extract_img_vae_multiscale(bs=1) # recommend bs = 1 for AspectRatioBatchSampler
else:
print('Extracting Single Image Resolution %s' % image_resize)
extract_img_vae()