-
Notifications
You must be signed in to change notification settings - Fork 6
/
SAR_train.py
200 lines (167 loc) · 7.2 KB
/
SAR_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
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
import os
import time
import itertools
from torch.autograd import Variable
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import utils
import numpy as np
import json
from torch.optim.lr_scheduler import MultiStepLR
# standard cross-entropy loss
def instance_bce(logits, labels):
assert logits.dim() == 2
cross_entropy_loss = nn.CrossEntropyLoss()
prediction_ans_k, top_ans_ind = torch.topk(F.softmax(labels, dim=-1), k=1, dim=-1, sorted=False)
ce_loss = cross_entropy_loss(logits, top_ans_ind.squeeze(-1))
return ce_loss
# multi-label soft loss
def instance_bce_with_logits(logits, labels, reduction='mean'):
assert logits.dim() == 2
loss = torch.nn.functional.binary_cross_entropy_with_logits(logits, labels, reduction=reduction)
if reduction == 'mean':
loss *= labels.size(1)
return loss
def compute_score_with_logits(logits, labels):
logits = torch.max(logits, 1)[1].data # argmax
one_hots = torch.zeros(*labels.size()).cuda()
one_hots.scatter_(1, logits.view(-1, 1), 1)
scores = (one_hots * labels)
return scores
def compute_TopKscore_with_logits(logits, labels,n):
prediction_ans_k, top_ans_ind = torch.topk(F.softmax(logits, dim=-1), k=n, dim=-1, sorted=True)
logits_ind = top_ans_ind
one_hots = torch.zeros(*labels.size()).cuda()
one_hots.scatter_(1, logits_ind.view(-1, n), 1)
scores = (one_hots * labels)
scores = torch.max(scores, 1)[0].data
return scores
def compute_self_loss(logits_neg, a):
prediction_ans_k, top_ans_ind = torch.topk(F.softmax(a, dim=-1), k=1, dim=-1, sorted=False)
neg_top_k = torch.gather(F.softmax(logits_neg, dim=-1), 1, top_ans_ind).sum(1)
qice_loss = neg_top_k.mean()
return qice_loss
def train(model, train_loader, eval_loader, opt):
utils.create_dir(opt.output)
optim = torch.optim.Adam(model.parameters(), lr=opt.learning_rate, betas=(0.9, 0.999), eps=1e-08,
weight_decay=opt.weight_decay)
logger = utils.Logger(os.path.join(opt.output, 'log.txt'))
utils.print_model(model, logger)
for param_group in optim.param_groups:
param_group['lr'] = opt.learning_rate
scheduler = MultiStepLR(optim, milestones=[100], gamma=0.8)
scheduler.last_epoch = opt.s_epoch
best_eval_score = 0
for epoch in range(opt.s_epoch, opt.num_epochs):
total_loss = 0
total_norm = 0
count_norm = 0
train_score = 0
t = time.time()
N = len(train_loader.dataset)
scheduler.step()
for i, (v, b, a, _, qa_text, _, _, q_t, bias) in enumerate(train_loader):
v = v.cuda()
b = b.cuda()
a = a.cuda()
bias = bias.cuda()
qa_text = qa_text.cuda()
rand_index = random.sample(range(0, opt.train_candi_ans_num), opt.train_candi_ans_num)
qa_text = qa_text[:,rand_index,:]
a = a[:,rand_index]
bias = bias[:,rand_index]
if opt.lp == 0:
logits = model(qa_text, v, b, epoch, 'train')
loss = instance_bce_with_logits(logits, a, reduction='mean')
elif opt.lp == 1:
logits = model(qa_text, v, b, epoch, 'train')
loss_pos = instance_bce_with_logits(logits, a, reduction='mean')
index = random.sample(range(0, v.shape[0]), v.shape[0])
v_neg = v[index]
b_neg = b[index]
logits_neg = model(qa_text, v_neg, b_neg, epoch, 'train')
self_loss = compute_self_loss(logits_neg, a)
loss = loss_pos + opt.self_loss_weight * self_loss
elif opt.lp == 2:
logits, loss = model(qa_text, v, b, epoch, 'train', bias, a)
else:
assert 1==2
loss.backward()
total_norm += nn.utils.clip_grad_norm_(model.parameters(), opt.grad_clip)
count_norm += 1
optim.step()
optim.zero_grad()
score = compute_score_with_logits(logits, a.data).sum()
train_score += score.item()
total_loss += loss.item() * v.size(0)
if i != 0 and i % 100 == 0:
print(
'training: %d/%d, train_loss: %.6f, train_acc: %.6f' %
(i, len(train_loader), total_loss / (i * v.size(0)),
100 * train_score / (i * v.size(0))))
total_loss /= N
if None != eval_loader:
model.train(False)
eval_score, bound = evaluate(model, eval_loader, opt)
model.train(True)
logger.write('\nlr: %.7f' % optim.param_groups[0]['lr'])
logger.write('epoch %d, time: %.2f' % (epoch, time.time() - t))
logger.write(
'\ttrain_loss: %.2f, norm: %.4f, score: %.2f' % (total_loss, total_norm / count_norm, train_score))
if eval_loader is not None:
logger.write('\teval score: %.2f (%.2f)' % (100 * eval_score, 100 * bound))
if (eval_loader is not None and eval_score > best_eval_score):
if opt.lp == 0:
model_path = os.path.join(opt.output, 'SAR_top'+str(opt.train_candi_ans_num)+'_best_model.pth')
elif opt.lp == 1:
model_path = os.path.join(opt.output, 'SAR_SSL_top'+str(opt.train_candi_ans_num)+'_best_model.pth')
elif opt.lp == 2:
model_path = os.path.join(opt.output, 'SAR_LMH_top'+str(opt.train_candi_ans_num)+'_best_model.pth')
utils.save_model(model_path, model, epoch, optim)
if eval_loader is not None:
best_eval_score = eval_score
@torch.no_grad()
def evaluate(model, dataloader, opt):
score = 0
score_ini_num_list=[]
for num in range(opt.test_candi_ans_num):
score_ini_num_list.append(0)
upper_bound = 0
num_data = 0
entropy = 0
for i, (v, b, a, q_id, qa_text, _, _, q_t, bias) in enumerate(dataloader):
v = v.cuda()
b = b.cuda()
bias = bias.cuda()
a = a.cuda()
q_id = q_id.cuda()
qa_text = qa_text.cuda()
if opt.lp == 0:
logits = model(qa_text, v, b, 0, 'test')
elif opt.lp == 1:
logits = model(qa_text, v, b, 0, 'test')
elif opt.lp == 2:
logits, _ = model(qa_text, v, b, 0, 'test', bias, a)
pred = logits
batch_score = compute_score_with_logits(pred, a.cuda()).sum()
score += batch_score.item()
for num in range(opt.test_candi_ans_num):
batch_score_num = compute_TopKscore_with_logits(pred, a.cuda(), num+1).sum()
score_ini_num_list[num] += batch_score_num.item()
upper_bound += (a.max(1)[0]).sum().item()
num_data += pred.size(0)
score = score / len(dataloader.dataset)
score_num_list = []
for score_num in score_ini_num_list:
score_num = score_num / len(dataloader.dataset)
score_num_list.append(score_num)
upper_bound = upper_bound / len(dataloader.dataset)
return score, upper_bound#, entropy
def calc_entropy(att): # size(att) = [b x v x q]
sizes = att.size()
eps = 1e-8
# att = att.unsqueeze(-1)
p = att.view(-1, sizes[1] * sizes[2])
return (-p * (p + eps).log()).sum(1).sum(0) # g