-
Notifications
You must be signed in to change notification settings - Fork 2
/
VSRN.py
153 lines (123 loc) · 5.14 KB
/
VSRN.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
import numpy as np
import torch
from transformers import BatchEncoding
class VSRN:
"""
"""
def __init__(
self,
image_encoder,
text_encoder,
caption_model,
retrieval_criterion,
caption_criterion,
weight_retrieval_loss,
weight_caption_loss,
learning_rate,
grad_clip,
device = 'cpu',
):
self.device = device
### makes fusion of bboxes + ocr-encoded
self.image_encoder = image_encoder.to(self.device)
### encodes caption inputs
self.text_encoder = text_encoder.to(self.device)
### generates caption from hidden states
self.caption_model = caption_model.to(self.device)
# losses and optimizer
self.retrieval_criterion = retrieval_criterion.to(self.device)
self.caption_criterion = caption_criterion
self.weight_retrieval_loss = weight_retrieval_loss
self.weight_caption_loss = weight_caption_loss
# self.params = params
self.optimizer = torch.optim.Adam(self.get_params(), lr=learning_rate)
### scalar on gradient vector norm
self.grad_clip = grad_clip
### they're already in caption_model - that was causing below warning
### "UserWarning: optimizer contains a parameter group with duplicate parameters;
### in future, this will cause an error; see github.com/pytorch/pytorch/issues/40967 for more information"
# params += list(self.decoder.parameters())
# params += list(self.encoder.parameters())
def get_params(self):
params = []
params += list(self.image_encoder.parameters())
params += list(self.text_encoder.parameters())
params += list(self.caption_model.parameters())
return params
def state_dict(self):
state_dict = [
self.image_encoder.state_dict(),
self.text_encoder.state_dict(),
self.caption_model.state_dict(),
]
return state_dict
def load_state_dict(self, state_dict):
self.image_encoder.load_state_dict(state_dict[0])
self.text_encoder.load_state_dict(state_dict[1])
if len(state_dict) > 2:
self.caption_model.load_state_dict(state_dict[2])
def train(self):
"""switch to train mode"""
self.image_encoder.train()
self.text_encoder.train()
self.caption_model.train()
def eval(self):
"""switch to evaluate mode"""
self.image_encoder.eval()
self.text_encoder.eval()
self.caption_model.eval()
def to(self, device):
self.image_encoder.to(device)
self.text_encoder.to(device)
self.caption_model.to(device)
### !!!
self.retrieval_criterion.to(device)
self.device = device
return self
def calculate_caption_loss(self, fc_feats, labels, masks):
labels = labels.to(self.device)
masks = masks.to(self.device)
seq_probs, _ = self.caption_model(fc_feats, labels, 'train')
loss = self.caption_criterion(seq_probs, labels[:, 1:], masks[:, 1:])
return loss
def forward_emb(self, images, ocr_features, tokenizer_outputs, captions, lengths):
"""
returns
fusion of image and ocr features
caption embeddings
pure image embeddings
"""
images = images.to(self.device)
ocr_features = ocr_features.to(self.device)
# print(type(tokenizer_outputs))
# tokenizer_outputs = BatchEncoding(tokenizer_outputs).to(self.device)
# tokenizer_outputs = tokenizer_outputs.to(self.device)
captions = captions.to(self.device)
# cap_emb = self.text_encoder(captions, lengths)
cap_emb = self.text_encoder(tokenizer_outputs)
img_emb, GCN_img_emd = self.image_encoder(images, ocr_features)
return img_emb, cap_emb, GCN_img_emd
def make_train_step(self, ids, images, ocr_features, tokenizer_outputs, captions, lengths, caption_labels, caption_masks):
"""training step"""
# print(type(tokenizer_outputs))
# compute the embeddings
img_emb, cap_emb, GCN_img_emb = self.forward_emb(images, ocr_features, tokenizer_outputs, captions, lengths)
caption_loss = self.calculate_caption_loss(GCN_img_emb, caption_labels, caption_masks)
# todo Тут бы выяснить, почему одно, а не другое
# caption_loss = self.calculate_caption_loss(img_emb, caption_labels, caption_masks)
retrieval_loss = self.retrieval_criterion(img_emb, cap_emb)
loss = (
self.weight_retrieval_loss * retrieval_loss
+
self.weight_caption_loss * caption_loss
)
print(f"Loss: {loss}, caption loss: {caption_loss}, retrieval loss: {retrieval_loss}")
# compute gradient and make optimizer step
self.optimizer.zero_grad()
loss.backward()
if self.grad_clip > 0:
params = self.get_params()
# for p in params:
# print(p.device)
torch.nn.utils.clip_grad.clip_grad_norm_(params, self.grad_clip)
self.optimizer.step()