-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_GGE.py
232 lines (196 loc) · 8.53 KB
/
train_GGE.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
import json
import os
import pickle
import time
from os.path import join
import torch
import torch.nn as nn
import utils
from torch.autograd import Variable
import numpy as np
from tqdm import tqdm
import random
import copy
import math
import time
def compute_score_with_logits(logits, labels):
logits = torch.argmax(logits, 1)
one_hots = torch.zeros(*labels.size()).cuda()
one_hots.scatter_(1, logits.view(-1, 1), 1)
scores = (one_hots * labels)
return scores
def train(model, train_loader, eval_loader,args,qid2type):
dataset=args.dataset
num_epochs=args.epochs
mode=args.mode
run_eval=args.eval_each_epoch
output=args.output
optim = torch.optim.Adamax(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001)
logger = utils.Logger(os.path.join(output, 'log.txt'))
total_step = 0
best_eval_score = 0
if mode=='q_debias':
topq=args.topq
keep_qtype=args.keep_qtype
elif mode=='v_debias':
topv=args.topv
top_hint=args.top_hint
elif mode=='q_v_debias':
topv=args.topv
top_hint=args.top_hint
topq=args.topq
keep_qtype=args.keep_qtype
qvp=args.qvp
for epoch in range(num_epochs):
total_loss = 0
train_score = 0
t = time.time()
for i, (v, q, a, b, hintscore,type_mask,notype_mask,q_mask) in tqdm(enumerate(train_loader), ncols=100,
desc="Epoch %d" % (epoch + 1), total=len(train_loader)):
total_step += 1
scale = math.sin(math.pi/2 * (epoch+30) / (num_epochs+30))
# scale = 1 / (1 + math.exp(-10 * epoch / num_epochs))
#########################################
v = Variable(v).cuda().requires_grad_()
q = Variable(q).cuda()
q_mask=Variable(q_mask).cuda()
a = Variable(a).cuda()
b = Variable(b).cuda().requires_grad_()
hintscore = Variable(hintscore).cuda()
type_mask=Variable(type_mask).float().cuda()
notype_mask=Variable(notype_mask).float().cuda()
#########################################
assert mode in ['base', 'gge_iter', 'gge_tog', 'gge_d_bias', 'gge_q_bias'], " %s not in modes. Please \'import train_ab as train\' in main.py" % mode
if mode == 'gge_iter':
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = 'q', weight=scale)
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
# optim.zero_grad()
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = 'joint', weight=scale)
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
optim.zero_grad()
total_loss += loss.item() * q.size(0)
batch_score = compute_score_with_logits(pred, a.data).sum()
train_score += batch_score
elif mode =='gge_tog':
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = 'tog')
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
optim.zero_grad()
total_loss += loss.item() * q.size(0)
batch_score = compute_score_with_logits(pred, a.data).sum()
train_score += batch_score
elif mode =='gge_d_bias':
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = 'd_bias')
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
optim.zero_grad()
total_loss += loss.item() * q.size(0)
batch_score = compute_score_with_logits(pred, a.data).sum()
train_score += batch_score
elif mode =='gge_q_bias':
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = 'q_bias')
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
optim.zero_grad()
total_loss += loss.item() * q.size(0)
batch_score = compute_score_with_logits(pred, a.data).sum()
train_score += batch_score
elif mode == 'base':
pred, loss, _ = model(v, q, a, b, None, q_mask, loss_type = None)
if (loss != loss).any():
raise ValueError("NaN loss")
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 0.25)
optim.step()
optim.zero_grad()
total_loss += loss.item() * q.size(0)
batch_score = compute_score_with_logits(pred, a.data).sum()
train_score += batch_score
total_loss /= len(train_loader.dataset)
train_score = 100 * train_score / len(train_loader.dataset)
logger.write('epoch %d, time: %.2f' % (epoch, time.time() - t))
logger.write('\ttrain_loss: %.2f, score: %.2f' % (total_loss, train_score))
if run_eval:
# if epoch % 2 == 0:
model.train(False)
results = evaluate(model, eval_loader, qid2type)
results["epoch"] = epoch
results["step"] = total_step
results["train_loss"] = total_loss
results["train_score"] = train_score
model.train(True)
eval_score = results["score"]
bound = results["upper_bound"]
yn = results['score_yesno']
other = results['score_other']
num = results['score_number']
logger.write('\teval score: %.2f (%.2f)' % (100 * eval_score, 100 * bound))
logger.write('\tyn score: %.2f other score: %.2f num score: %.2f' % (100 * yn, 100 * other, 100 * num))
if eval_score > best_eval_score:
model_path = os.path.join(output, 'model.pth')
torch.save(model.state_dict(), model_path)
best_eval_score = eval_score
model_path = os.path.join(output, 'model_final.pth')
torch.save(model.state_dict(), model_path)
def evaluate(model, dataloader, qid2type):
score = 0
upper_bound = 0
score_yesno = 0
score_number = 0
score_other = 0
total_yesno = 0
total_number = 0
total_other = 0
for v, q, a, b, qids, _, q_mask in tqdm(dataloader, ncols=100, total=len(dataloader), desc="eval"):
v = Variable(v, requires_grad=False).cuda()
q = Variable(q, requires_grad=False).cuda()
q_mask=Variable(q_mask).cuda()
pred, _, _ = model(v, q, None, None, None, q_mask, loss_type = None)
batch_score = compute_score_with_logits(pred, a.cuda()).cpu().numpy().sum(1)
score += batch_score.sum()
upper_bound += (a.max(1)[0]).sum()
qids = qids.detach().cpu().int().numpy()
for j in range(len(qids)):
qid = qids[j]
typ = qid2type[str(qid)]
if typ == 'yes/no':
score_yesno += batch_score[j]
total_yesno += 1
elif typ == 'other':
score_other += batch_score[j]
total_other += 1
elif typ == 'number':
score_number += batch_score[j]
total_number += 1
else:
print('Hahahahahahahahahahaha')
score = score / len(dataloader.dataset)
upper_bound = upper_bound / len(dataloader.dataset)
score_yesno /= total_yesno
score_other /= total_other
score_number /= total_number
results = dict(
score=score,
upper_bound=upper_bound,
score_yesno=score_yesno,
score_other=score_other,
score_number=score_number,
)
return results