-
Notifications
You must be signed in to change notification settings - Fork 0
/
Frozen_resnet.py
392 lines (326 loc) · 13.2 KB
/
Frozen_resnet.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
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
import torchvision
import matplotlib.pyplot as plt
import cv2
import math
from copy import deepcopy
from tqdm import tqdm
import json
import os
import requests
import cv2
import numpy as np
import torch
import torchvision.transforms
from datasets import load_dataset
from transformers import T5Tokenizer
from torch.utils.tensorboard import SummaryWriter
log_dir = './data'
writer = SummaryWriter(log_dir)
# import T5
from transformers import T5Tokenizer, T5ForConditionalGeneration
# forward hook for reading resnet penultimate layer logits
def forward_hook(module, input, output):
global resnet_avgpool_output
resnet_avgpool_output = output
# class for image embeddings - obtained from resnet
class ImageEmbeddings(nn.Module):
def __init__(self, hook_fn, d_model):
super().__init__()
self.resnet = torchvision.models.resnet50()
self.resnet.avgpool.register_forward_hook(hook_fn)
self.proj = nn.Linear(2048, d_model * 2, bias=False) # d_model * 2 because each image is supposed to constitute embeddings of seq_len = 2 (according to the paper)
def forward(self, imgs): # imgs.shape: [b,c,w,h]
batch_size = imgs.shape[0]
_ = self.resnet(imgs)
emb = resnet_avgpool_output # emb.shape: [b, 2048, 1, 1]
emb = emb.flatten(start_dim=1, end_dim=-1) # emb.shape: [b, 2048]
emb = self.proj(emb) # emb.shape: [b, d_model * 2]
emb = emb.reshape(batch_size, d_model, 2)
emb = emb.permute(0, 2, 1) # emb.shape: [batch_size, 2, d_model]
return emb
def slice_datadict(dct, start_idx, end_idx):
slice_dict = {}
keys = list(dct.keys())
for key in keys:
slice_dict[key] = dct[key][start_idx:end_idx]
return slice_dict
def process_batch_(minibatch, img_size):
"""process the url
Parameters
----------
minibatch: Dict
key: ['URL','text']
value: list[URL],list[text]
img_size: int
the size of image
Returns
-------
augmented_imgs: List
length of augmented_imgs: batch
captions: List
length of caption: batch
"""
value_list = list(minibatch.values())
url_list = value_list[0]
captions = value_list[1]
augmented_imgs = []
#processing
for url,cap in zip(url_list,captions):
print(f"processing url: {url}")
# print(f"caption: {cap}")
response = requests.get(url)
if response.status_code == 200:
img_data = response.content
else:
print(f"Failed to fetch image from URL. Status code: {response.status_code}")
continue
# img_data = response.content
img = cv2.imdecode(np.frombuffer(img_data, np.uint8), -1)
resize_shape = (img_size, img_size)
img = cv2.resize(img, resize_shape, interpolation=cv2.INTER_LINEAR)
img = np.float32(img) / 255
img = torch.tensor(img)
img = img.permute(2, 1, 0) # [w, h, c] -> [c, h, w]
transforms = torchvision.transforms.Compose([
torchvision.transforms.Resize(int(1.25 * img_size)), # image_size + 1/4 * image_size
torchvision.transforms.RandomResizedCrop(resize_shape, scale=(0.8, 1.0)),
torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) # zero mean, unit std
])
img = transforms(img)
augmented_imgs.append(img)
return augmented_imgs, captions
def load_img2cap(batch_size,dataset,tokenizer,img_size, device):
"""load the image-caption dataset and return torch.Tensor
Parameters
----------
batch_size: int
dataset: DataDict
tokenizer: T5
device: string -- torch.device
cuda or cpu
Returns
-------
img_tenosr: torch.Tensor
caption_tenosr: torch.Tensor
"""
img_list = []
caption_list = []
n = len(list(dataset.values())[0])
for i in range(0, n, batch_size):
minibatch = slice_datadict(dataset, i, i+batch_size)
augmented_imgs, captions = process_batch_(minibatch, img_size)
img_list.extend(augmented_imgs)
caption_list.extend(captions)
print("-----------------------")
img_tensor = torch.stack(img_list, dim=0).to(device)
caption = tokenizer(caption_list, padding=True, truncation=True, return_tensors="pt")
caption = caption.to(device)
# caption = {key: val.to(device) for key, val in caption.items()}
# caption = {
# key: val.to(device) if isinstance(val, torch.Tensor) else val
# for key, val in tokenizer(caption_list, padding=True, truncation=True, return_tensors="pt").items()
# }
return img_tensor,caption
def calculate_loss(imgs,caption,image_encoder,t5_model,tokenizer,device):
"""Calculate Loss
Parameters
----------
imgs: torch.Tensor
shape of imgs `[batch_size, channels, height, width]
caption: Dict
keys: input_ids, attn_mask
values: torch.Tensor
image_encoder: nn.Module
t5_model: nn.Module
Pre-Trained Language Model
tokenizer: Hugging Face Tokenizer
device: torch.device
cuda or cpu
Returns
-------
loss: float
batch_accuracy: float
"""
batch_size = imgs.shape[0]
# obtain img embeddings
img_embs = image_encoder(imgs) # img_embs.shape: [batch_size,2,d_model]
# feed img_embs to t5 encoder to get encoder output
enc_out = t5_model.encoder(
inputs_embeds = img_embs
).last_hidden_state # enc_out.shape: [batch_size, 2, d_model]
# extract cap tokens and attn_mask
cap_tokens, cap_attn_mask = caption.input_ids, caption.attention_mask
# shift cap_tokens right (pre-pend start token) - as input to decoder is
# expected to be right shifted and starting with pad token (used as start token by T5)
start_token_id = tokenizer(tokenizer.pad_token,
return_tensors='pt',
padding=False, truncation=True).input_ids
start_token_id = start_token_id[:, 0] # trim end token appended by the tokenizer
start_token_id = start_token_id.expand(batch_size, -1).to(device) # start_token_id.shape: [batch_size, 1]
cap_tokens_rshifted = torch.cat((start_token_id, cap_tokens), dim=-1) # cap_tokens_rshifted.shape: [batch_size, seq_len+1]
cap_tokens_rshifted = cap_tokens_rshifted[:, :-1] # cap_tokens_rshifted.shape: [batch_size, seq_len]
# feed cap tokens to t5 decoder to get decoder output
dec_out = t5_model.decoder(input_ids=cap_tokens_rshifted,
attention_mask=cap_attn_mask,
encoder_hidden_states=enc_out).last_hidden_state
# dec_out.shape: [batch_size, seq_len, d_model]
# get scores from dec_out
scores = t5_model.lm_head(dec_out) # scores.shape: [batch_size, seq_len, vocab_size]
scores = scores.permute(0, 2, 1) # scores.shape: [batch_size, vocab_size, seq_len] - required for crossEntropyLoss
# create targets = cap_tokens (unshifted)
targets = cap_tokens # targets.shape: [batch_size, seq_len]
# cross entropy loss
criterion = nn.CrossEntropyLoss(reduction='mean')
loss = criterion(scores, targets)
# calculate batch accuracy
pred_cap_tokens = torch.argmax(scores, dim=1) # shape: [batch_size, seq_len]
batch_accuracy = (pred_cap_tokens == cap_tokens).float().mean() * 100
return loss, batch_accuracy,targets,pred_cap_tokens
def ids2text(targets,pre_cap_tokens,tokenizer):
"""convert input_ids to caption
Parameters
----------
targets: torch.Tensor
shape of targets `[batch_size,seq_len]`
pre_cap_tokens: torch.Tensor
shape of pre_cap_tokens `[batch_size,seq_len]`
tokenizer: Hugging Face Tokenizer
T5
Returns
-------
True_Caption: List
Pred_Caption: List
"""
true_caption = []
# true_caption = tokenizer.decode(
# targets,
# skip_special_tokens=True,
# )
batch = targets.shape[0]
for i in range(batch):
true_caption.append(tokenizer.decode(
targets[i],
skip_special_tokens=True,
))
pred_caption = []
for i in range(batch):
true_caption.append(tokenizer.decode(
pred_cap_tokens[i],
skip_special_tokens=True,
))
return true_caption,pred_caption
# utility function to load model weights from checkpoint - loads to the device passed as 'device' argument
def load_ckpt(checkpoint_path, model, optimizer=None, scheduler=None, device=torch.device('cpu'), mode='eval'):
ckpt = torch.load(checkpoint_path, map_location=device)
model.load_state_dict(ckpt['model_state_dict'])
if mode == 'eval':
model.eval() # clip is only used for inference
return model
else:
model.train()
optimizer.load_state_dict(ckpt['optimizer_state_dict'])
if scheduler is not None:
scheduler.load_state_dict(ckpt['scheduler_state_dict'])
return model, optimizer, scheduler
else:
return model, optimizer
# utility function to save a checkpoint (model_state, optimizer_state, scheduler_state) - saves on whatever device the model was training on
def save_ckpt(checkpoint_path, model, optimizer, scheduler=None):
save_dict = {'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()}
if scheduler is not None:
save_dict['scheduler_state_dict'] = scheduler.state_dict()
torch.save(save_dict, checkpoint_path)
## main ##
if __name__ == '__main__':
# hyperparams
img_size = 224 # resize for resnet
d_model = 768 # d_model for T5 (required for resnet proj head)
max_seq_len = 512 # required to init T5 Tokenizer
batch_size = 16
lr = 3e-4
num_epochs = 10
random_seed = 1010
t5_model_name = 't5-base'
checkpoint_path = 'ckpts_frozen_resnet/latest.pt' # path to a save and load checkpoint of the trained resnet
resume_training_from_ckpt = False
# set random seed
np.random.seed(random_seed)
torch.manual_seed(random_seed)
# cuda
# device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
device = torch.device("mps")
# init image encoder model (resnet)
image_encoder = ImageEmbeddings(forward_hook, d_model).to(device)
# init T5 tokenizer and transformer model
t5_tokenizer = T5Tokenizer.from_pretrained(t5_model_name, model_max_length=max_seq_len)
t5_model = T5ForConditionalGeneration.from_pretrained(t5_model_name).to(device)
dataset = load_dataset("ChristophSchuhmann/MS_COCO_2017_URL_TEXT")
dataset = dataset['train'][:100]
# imgs, caption = load_img2cap(
# batch_size=batch_size,
# dataset=dataset,
# tokenizer=t5_tokenizer,
# img_size=img_size,
# device=device,
# )
# load checkpoint to resume training from
# if resume_training_from_ckpt:
# image_encoder, optimizer = load_ckpt(checkpoint_path, image_encoder, optimizer=optimizer, device=device, mode='train')
if resume_training_from_ckpt:
image_encoder = torch.load(checkpoint_path,map_location=device)
# init optimizer
optimizer = torch.optim.Adam(params=image_encoder.parameters(), lr=lr, betas=(0.9, 0.95))
# train loop
n = len(dataset)
for ep in tqdm(range(num_epochs)):
# Loss = []
#fetch minibatch
caption = []
pre_caption = []
for i in range(0, n, batch_size):
sample = slice_datadict(dataset,i,i+batch_size)
img, cap = load_img2cap(
batch_size=batch_size,
dataset=sample,
tokenizer=t5_tokenizer,
img_size=img_size,
device=device
)
# calculate loss
loss, batch_accuracy,targets,pred_cap_tokens = calculate_loss(
imgs=img,
caption=cap,
image_encoder=image_encoder,
t5_model=t5_model,
tokenizer=t5_tokenizer,
device=device,
)
# update params
optimizer.zero_grad()
loss.backward()
optimizer.step()
true_cap,pred_cap = ids2text(
targets=targets,
pre_cap_tokens=pred_cap_tokens,
tokenizer=t5_tokenizer
)
caption.extend(true_cap)
pre_caption.extend(pred_cap)
# Loss.append(loss.detach().cpu().item())
# if (ep+1) % (num_epochs // 1) == 0:
# # print metrics
# print('ep:{} \t loss:{:.3f} \t batch_accuracy:{:.3f}'.format(ep, loss.item(), batch_accuracy.item()))
# # save model checkpoint
# save_ckpt(checkpoint_path, image_encoder, optimizer)
print("epoch: {} | loss: {:.3f}".format(ep+1, loss.detach().cpu().item()))
print("epoch: {} | acc: {:.3f}".format(ep+1, batch_accuracy.detach().cpu().item()))
writer.add_scalar("loss",loss.detach().cpu().item(),ep)
writer.add_scalar("accuracy",batch_accuracy.detach().cpu().item(),ep)
# print("epoch: {} | batch_loss: {:.3f}".format(ep+1, np.average(np.array(Loss))))
print(f"caption: {caption}")
print(f"pred_caption: {pre_caption}")