-
Notifications
You must be signed in to change notification settings - Fork 0
/
ve_train.py
173 lines (126 loc) · 6.02 KB
/
ve_train.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
import json
import os
import zarr
from tqdm import tqdm
import torch
import numpy as np
from torch.nn import BCELoss
from torch.utils.data import Dataset
from torch.optim import Adam
from transformers import LxmertTokenizer, LxmertModel
import utils
from argparse import ArgumentParser
from ve_dataset import VEQADataset
# Accuracy function for metrics
def compute_score_with_logits(outputs, scores):
return (outputs>0.5).int()==scores.int()
# return s>0
# Evaluate branch
def evaluate(model, loader, cfg):
model.eval()
eval_score = 0
total_loss = 0
total_size = 0
loss_fn = BCELoss(reduction="sum")
for i, (features, boxes, qa_tokens_item, scores) in tqdm(enumerate(loader)):
qa_tokens, qa_tokens_padded, qa_tokens_ids = qa_tokens_item["input_ids"].cuda(), qa_tokens_item["attention_mask"].cuda(), qa_tokens_item["token_type_ids"].cuda()
# qa_tokens = qa_tokens.cuda()
features = features.cuda()
boxes = boxes.cuda()
scores = scores.cuda()
# Question-answer tokenized units
qa_tokens_item["input_ids"] = qa_tokens.squeeze(1) if len(qa_tokens.shape)>2 else qa_tokens
qa_tokens_item["attention_mask"] = qa_tokens_padded.squeeze(1) if len(qa_tokens_padded.shape)>2 else qa_tokens_padded
qa_tokens_item["token_type_ids"] = qa_tokens_ids.squeeze(1) if len(qa_tokens_ids.shape)>2 else qa_tokens_ids
# Visual attention masks
visual_attention_mask = (features.sum(2)>0).int()
# Model outputs. Unlike in train.py, we dont reshape the output since the training is a direct entailment objective.
outputs = model(qa_tokens_item, features, boxes, visual_attention_mask).squeeze(1)
loss = loss_fn(outputs, scores.float()) #* scores.size(1)
i+=1
score = compute_score_with_logits(outputs, scores.data).sum()
total_size += features.size(0)
eval_score += score.item()
total_loss += loss.item()
final_loss = total_loss / total_size
final_score = eval_score / total_size
return final_loss, final_score
if __name__=="__main__":
from torch.utils.data import DataLoader
from model import VEQ
class AttributeDict(dict):
__getattr__ = dict.__getitem__
__setattr__ = dict.__setitem__
__delattr__ = dict.__delitem__
parser = ArgumentParser()
parser.add_argument("--config")
args = parser.parse_args()
cfg = AttributeDict(json.load(open(args.config)))
results = {"train":{"loss":[], "accuracy":[]}, "eval":{"loss":[], "accuracy":[]}}
# VE dataset for train and validation
dataset = VEQADataset("train")
eval_dataset = VEQADataset("eval")
# Dataloaders for train and validation datasets
loader = DataLoader(dataset, cfg.batch_size, shuffle=True, num_workers=1, collate_fn=utils.trim_collate)
eval_loader = DataLoader(eval_dataset, cfg.batch_size, shuffle=True, num_workers=1, collate_fn=utils.trim_collate)
# VEQA model for SNLI-VE configuration
model = VEQ(cfg)
model = model.cuda()
optim = Adam(model.parameters(), lr=1e-5)
optim.zero_grad()
loss_fn = BCELoss(reduction="sum")
best_eval_score = 0
eval_loss, eval_score = 0, 0
for epoch in range(cfg.epochs):
print(f"Epoch {epoch}/{cfg.epochs}:\n")
model.train()
train_score = 0
total_loss = 0
total_size = 0
for i, (features, boxes, qa_tokens_item, scores) in tqdm(enumerate(loader)):
qa_tokens, qa_tokens_padded, qa_tokens_ids = qa_tokens_item["input_ids"].cuda(), qa_tokens_item["attention_mask"].cuda(), qa_tokens_item["token_type_ids"].cuda()
features = features.cuda()
boxes = boxes.cuda()
scores = scores.cuda()
# Question answer pairs
qa_tokens_item["input_ids"] = qa_tokens.squeeze(1) if len(qa_tokens.shape)>2 else qa_tokens
qa_tokens_item["attention_mask"] = qa_tokens_padded.squeeze(1) if len(qa_tokens_padded.shape)>2 else qa_tokens_padded
qa_tokens_item["token_type_ids"] = qa_tokens_ids.squeeze(1) if len(qa_tokens_ids.shape)>2 else qa_tokens_ids
# Visual attention maks creation
visual_attention_mask = (features.sum(2)>0).int()
outputs = model(qa_tokens_item, features, boxes, visual_attention_mask).squeeze()
loss = loss_fn(outputs, scores.float())
i+=1
loss.backward()
optim.step()
optim.zero_grad()
# Compute scores for entailment prediction
score = compute_score_with_logits(outputs, scores.data).sum()
train_score += score.item()
total_loss += loss.item()
total_size += features.size(0)
if i != 0 and i % 1000 == 0:
print(
'training: %d/%d, train_loss: %.3f, train_acc: %.3f%%' %
(i, len(loader), total_loss / total_size,
100 * train_score / total_size))
# computing loss
total_loss /= total_size
model.eval()
eval_loss, eval_score = evaluate(model, eval_loader, cfg)
model.train()
print('eval loss: %.3f eval score: %.3f%%' % (eval_loss, 100 * eval_score))
# Save the best model
if (eval_score > best_eval_score):
print("Saving best model")
model_path = os.path.join(cfg.output_dir, "VE_best_model.pth")
utils.save_model(model_path, model, epoch, optim)
best_eval_score = eval_score
results["eval"]["loss"].append(eval_loss)
results["eval"]["accuracy"].append(eval_score)
results["train"]["loss"].append(total_loss)
results["train"]["accuracy"].append(train_score/total_size)
# Save every epoch's results
import json
with open("/home/meghana/meg/VEQA/Trained/ve_train_results.json", "w") as f:
json.dump(results, f)