-
Notifications
You must be signed in to change notification settings - Fork 8
/
training_allexpro10k_s10.py
451 lines (353 loc) · 14.2 KB
/
training_allexpro10k_s10.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
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
import pandas as pd
import numpy as np
import os
from sklearn.model_selection import StratifiedShuffleSplit
from tqdm import tqdm
from glob import glob
#os.environ['CUDA_VISIBLE_DEVICES']='3'
from torch.utils.data import Dataset
import matplotlib.pyplot as plt
from torchvision import transforms
import torch
from torch import nn
from PIL import Image
import open_clip
from collections.abc import Iterable
import logging
import datetime
from tqdm import tqdm
import torch.nn.functional as F
import glob
import os
import pandas as pd
from tqdm import tqdm
import torchvision
import albumentations as A
RESULT_DIR = 'result/'
RANDOM_SEED = 0
BATCH_SIZE = 32
EPOCHS = 3
INIT_LR = 0.02 / 80.
WEIGHT_DECAY = 1e-4
MILESTONES = [1]#[1]#
COMMENT = 'stage10_vith_290_overlap_all'
def get_files_from_dir(dir):
if not os.path.exists(dir):
return ''
file_paths = []
for root, directories, files in os.walk(dir):
for filename in files:
filepath = os.path.join(root, filename)
file_paths.append(filepath)
return file_paths
############# initial DDP ###############
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim
from torch.nn.parallel import DistributedDataParallel as DDP
dist.init_process_group("nccl")
rank = dist.get_rank()
print(f"Start running basic DDP example on rank {rank}.")
torch.cuda.set_device(rank)
file_list = get_files_from_dir('data_preprocess/gldv2full/train')
train_csv = pd.read_csv('data_preprocess/gldv2full/train.csv')
id_ = train_csv['id'].values
file_list = []
for i in tqdm(id_):
file_list.append( os.path.join('data_preprocess/gldv2full/train',i[0],i[1],i[2],i +'.jpg') )
label_list = train_csv['landmark_id'].values
classes_num = len(label_list)
train_shopee_pre = np.load(os.path.join('data_preprocess/shopee','image_list.npy'))
label_shopee = np.load(os.path.join('data_preprocess/shopee','label_list.npy'))
train_shopee = []
label_shopee += classes_num
label_list = list(label_list)
file_list.extend(list(train_shopee))
label_list.extend(list(label_shopee))
classes_num += np.unique(label_list).shape[0]
# add met data
dataset_met = torchvision.datasets.ImageFolder(
'data_preprocess/MET',
transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((224,224)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711]
)
]))
train_met = []
label_met = []
for i in tqdm(dataset_met.samples):
train_met.append(i[0])
label_met.append(i[1] + classes_num)
file_list.extend(list(train_met))
label_list.extend(list(label_met))
classes_num += np.unique(label_list).shape[0]
# add alibaba product data
dataset_ali = torchvision.datasets.ImageFolder(
'data_preprocess/Alibaba Goods/goods_categories',
transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((224,224)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711]
)
]))
train_ali = []
label_ali = []
for i in tqdm(dataset_ali.samples):
train_ali.append(i[0])
label_ali.append(i[1] + classes_num)
file_list.extend(list(train_ali))
label_list.extend(list(label_ali))
classes_num += np.unique(label_list).shape[0]
# GPR1200
dataset_gpr = glob('data_preprocess/gpr1200/*')
for i in dataset_gpr:
file_list.append(i)
label_list.append(int(i.split('_')[0]) + classes_num)
classes_num += 1200
# DeepFashion
dataset_deepfashion = torchvision.datasets.ImageFolder(
'data_preprocess/deepfashion/customerandshop/img/imgfolder',
transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((224,224)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711]
)
]))
file_list = []
label_list = []
for i in tqdm(dataset_deepfashion.samples):
file_list.append(i[0])
label_list.append(i[1] + classes_num)
classes_num += 467
##############################################
# datasets and augmentation #
##############################################
class shopeeDataset(Dataset):
def __init__(self, img_path, label_list, transform, train_aug):
self.img_path = img_path
self.label_list = label_list
self.transform = transform
self.train_aug = train_aug
def __len__(self,):
return len(self.img_path)
def __getitem__(self, index):
img = self.img_path[index]
label = self.label_list[index]
img = plt.imread(img) #/ 255.
if img.dtype == 'float32' or img.dtype == 'float64' or img.dtype == 'float16':
img = img * 255
img = img.astype('uint8')
if img.shape[2] == 4:
img = img[:,:,:3]
#img = img.transpose(2,0,1)
#print(img.shape)
#img = Image.fromarray(img)
img = self.train_aug(image=img)["image"]
img = self.transform(img)
return img, label
transform_train = transforms.Compose([
transforms.ToPILImage(),
#transforms.Resize((280,280)),
#transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711])
])
transform_val = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((224,224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
std=[0.26862954, 0.26130258, 0.27577711])
])
train_aug = A.Compose([
A.HorizontalFlip(p=0.5),
A.ImageCompression(quality_lower=99, quality_upper=100),
A.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.2, rotate_limit=10, border_mode=0, p=0.7),
A.Resize(290, 290),
A.Cutout(max_h_size=int(290 * 0.4), max_w_size=int(290 * 0.4), num_holes=1, p=0.5),
])
datasets_train = shopeeDataset(file_list, label_list, transform_train, train_aug)
train_sampler = torch.utils.data.distributed.DistributedSampler(datasets_train)
dataloader_train = torch.utils.data.DataLoader(datasets_train, batch_size=BATCH_SIZE,\
shuffle=False, num_workers=32, pin_memory=True,\
drop_last=False, sampler = train_sampler)
print(len(dataloader_train))
##############################################
# Model Preparation #
##############################################
def set_freeze_by_idxs(model, idxs, freeze=True):
if not isinstance(idxs, Iterable):
idxs = [idxs]
num_child = len(list(model.children()))
idxs = tuple(map(lambda idx: num_child + idx if idx < 0 else idx, idxs))
for idx, child in enumerate(model.children()):
if idx not in idxs:
continue
for param in child.parameters():
param.requires_grad = not freeze
class AddMarginProduct(nn.Module):
r"""Implement of large margin cosine distance: :
Args:
in_features: size of each input sample
out_features: size of each output sample
s: norm of input feature
m: margin
cos(theta) - m
"""
def __init__(self, in_features, out_features, s=30.0, m=0.40):
super(AddMarginProduct, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.s = s
self.m = m
self.weight = nn.Parameter(torch.FloatTensor(out_features, in_features))
nn.init.xavier_uniform_(self.weight)
def forward(self, input, label):
# --------------------------- cos(theta) & phi(theta) ---------------------------
cosine = F.linear(F.normalize(input), F.normalize(self.weight))
phi = cosine - self.m
# --------------------------- convert label to one-hot ---------------------------
one_hot = torch.zeros(cosine.size(), device='cuda')
# one_hot = one_hot.cuda() if cosine.is_cuda else one_hot
one_hot.scatter_(1, label.view(-1, 1).long(), 1)
# -------------torch.where(out_i = {x_i if condition_i else y_i) -------------
output = (one_hot * phi) + ((1.0 - one_hot) * cosine) # you can use torch.where if your torch.__version__ is 0.4
output *= self.s
# print(output)
return output
def __repr__(self):
return self.__class__.__name__ + '(' \
+ 'in_features=' + str(self.in_features) \
+ ', out_features=' + str(self.out_features) \
+ ', s=' + str(self.s) \
+ ', m=' + str(self.m) + ')'
class add_layer_model(nn.Module):
def __init__(self,backbone):
super(add_layer_model, self).__init__()
self.backbone = backbone
self.fc1 = nn.Linear(1024,64)
#self.bn = nn.BatchNorm1d(64)
self.fc2 = AddMarginProduct(64, classes_num, s=30, m=0.55)# nn.Linear(64, classes_num, bias = False) #AddMarginProduct(64, classes_num, s=30, m=0.35)
self.drpout = nn.Dropout(0.2)
def forward(self, x, label):
x = self.backbone(x)
x = self.drpout(x)
x = self.fc1(x)
#x = self.bn(x)
x = self.fc2(x, label)
return x
backbone, _, preprocess = open_clip.create_model_and_transforms('ViT-H-14-290')
#weight_backbone = torch.load( 'vit-h-14-laion2b_s32b_b79k.pth' )
'''
weight_clear = weight_backbone
positional_embedding = weight_clear['backbone.positional_embedding']
pos_embed_before = positional_embedding[:1,:]
pos_embed_after = positional_embedding[1:,:]
print(pos_embed_after.shape)
pos_embed_after = pos_embed_after.view(1,16,16,1024).permute(0,3,1,2)
pos_embed_after = torch.nn.functional.interpolate(pos_embed_after, size=(20,20), mode='bicubic') # 1,1024, 24,24
pos_embed_after = pos_embed_after.permute(0,2,3,1).view(20*20,1024)
pos_embed = torch.cat([pos_embed_before, pos_embed_after])
weight_clear['backbone.positional_embedding'] = pos_embed
'''
#backbone.load_state_dict(weight_backbone)
backbone = backbone.visual
model = add_layer_model(backbone)
### load pretrained weight
weight_try = torch.load( os.path.join('result','stage9_vith_280_pro10k_backbone.pth') )
#weight_try = torch.load( 'vit-h-14-laion2b_s32b_b79k.pth' )
weight_clear = {}
for i in weight_try.items():
weight_clear[i[0].split('module.')[-1]] = i[1]
weight_try.popitem('fc2.weight')
positional_embedding = weight_clear['backbone.positional_embedding']
pos_embed_before = positional_embedding[:1,:]
pos_embed_after = positional_embedding[1:,:]
print(pos_embed_after.shape)
pos_embed_after = pos_embed_after.view(1,20,20,1280).permute(0,3,1,2)
pos_embed_after = torch.nn.functional.interpolate(pos_embed_after, size=(24,24), mode='bicubic') # 1,1024, 24,24
pos_embed_after = pos_embed_after.permute(0,2,3,1).view(24*24,1280)
pos_embed = torch.cat([pos_embed_before, pos_embed_after])
weight_clear['backbone.positional_embedding'] = pos_embed
'''
weight_add_forlabel = torch.randn( (len(dataset_ali.classes),64) )
weight_add_forlabel = weight_add_forlabel.to(weight_clear['fc2.weight'].device)
nn.init.xavier_uniform_(weight_add_forlabel)
weight_clear['fc2.weight'] = torch.cat([weight_clear['fc2.weight'],weight_add_forlabel],0)
'''
model.load_state_dict(weight_clear , strict=True)
model.cuda()
# freeze backbone
#set_freeze_by_idxs(model,[1,2])
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[rank])
##############################################
# Optimizer, Loss, Scheduler,... #
##############################################
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()),\
lr=INIT_LR,\
weight_decay=WEIGHT_DECAY)
criterian = nn.CrossEntropyLoss()
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, MILESTONES, gamma = 0.1, last_epoch = -1, verbose = False)
##############################################
# Training #
##############################################
def train_one_epoch(dataloader, model, criterian): # return loss and acc1
loss_total = 0
for batch, (X, y) in tqdm(enumerate(dataloader)):
# Compute prediction and loss
X = X.cuda()
y = y.cuda()
pred = model(X, y)
loss = criterian(pred, y)
# Backpropagation
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_total += loss.item()
if (batch +1)%100 == 0 and rank == 0:
print(f'Loss : {loss_total/batch}')
return loss_total/batch
# we don't need validation for now
def val_one_epoch(model, criterian):
raise NotImplementedError
ep_global = 0
def train():
logging.basicConfig(format='%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s:%(message)s',
filename= f'logs/{datetime.datetime.now().strftime("%Y_%m_%d_%H_%M_%S")}-{COMMENT}.txt',
level='INFO',
filemode='a'
)
#model.cuda()
model.train()
loss_total_train = 0
acc1_total_train = 0
for ep in tqdm(range(EPOCHS)):
print(f'epoch {ep}')
dataloader_train.sampler.set_epoch(ep)
ep_global = ep
loss = train_one_epoch(dataloader_train, model, criterian)
# we don't need validation for now
# val_one_epoch(model, criterian)
#loss_total_train += loss
#acc1_total_train += acc1
if rank == 0:
logging.info(f"epoch {ep}: train loss {loss}")
scheduler.step()
if rank == 0:
logging.info(f"training done. {EPOCHS} epochs.")
model.eval()
if rank == 0:
torch.save(model.state_dict(), os.path.join(RESULT_DIR, f'{COMMENT}.pth'))
train()