-
Notifications
You must be signed in to change notification settings - Fork 16
/
train.py
331 lines (287 loc) · 14.3 KB
/
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
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
#!user/bin/env python
# -*- coding:utf-8 -*-
import argparse
import json
import os
import pickle
import random
import torch
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from bisect import bisect
from math import fabs
from torch.optim import lr_scheduler
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import LxmertTokenizer
from config import args
from contrastive_loss import ContrastiveLoss, l2_sim
from dataset import KgDataset, my_collate_pretrain, PretrainDataset, my_collate
from dataset import vocab_num
from dataset_val import KgDatasetVal
from model import KgPreModel, tokenizer
from transformers import get_linear_schedule_with_warmup
# dist.init_process_group(backend='nccl')
# torch.cuda.set_device(args.local_rank)
# torch.manual_seed(10)
# torch.cuda.manual_seed(10)
# cudnn.benchmark = False
# cudnn.deterministic = True
torch.multiprocessing.set_sharing_strategy('file_system')
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
def generate_tripleid(batch_anchor, candidate):
# cos distance
similarity = batch_anchor.mm(candidate.t()) # b * v
# l2 distance
# similarity = l2_sim(batch_anchor, candidate) #b * v
# cos largest:True l2 largest:False
prob, idx_1 = torch.topk(similarity, k=1, dim=1, largest=True)
prob3, idx_3 = torch.topk(similarity, k=3, dim=1, largest=True)
return idx_1.squeeze(), idx_3.squeeze()
def cal_batch_loss(target, target_true, criterion):
target = target.view(-1, 2)
target_true = target_true.view(-1, 1).squeeze()
batch_loss = criterion(target, target_true)
return batch_loss
def cal_acc_multi(ground_truth, preds, return_id = False):
all_num = len(ground_truth)
acc_num = 0
ids = []
temp = []
for i, answer_id in enumerate(ground_truth):
pred = preds[i]
# ids.append([i, int(pred)])
cnt = 0
for aid in answer_id:
if pred == aid:
cnt += 1
if cnt ==1:
acc_num += 0.3
# ids.append([int(pred), 1])
elif cnt == 2:
acc_num += 0.6
# ids.append([int(pred), 1])
elif cnt > 2:
acc_num += 1
# ids.append([int(pred), 1])
# else:
# ids.append([int(pred), 0])
if return_id:
return acc_num / all_num, ids
else:
return acc_num / all_num
def cal_acc(ground_truth, preds, return_id = False):
all_num = len(ground_truth)
acc_num = 0
ids = []
temp = []
for i, answer_id in enumerate(ground_truth):
pred = preds[i]
ids.append([i, int(pred)])
cnt = 0
for aid in answer_id:
if pred == aid:
acc_num += 1
if return_id:
return acc_num / all_num, ids
else:
return acc_num / all_num
def train():
if not args.pretrain:
train_dataset = KgDataset(val=False)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=4, collate_fn=my_collate)
if args.validate:
test_dataset = KgDatasetVal(val=False)
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=4, collate_fn=my_collate)
else:
train_dataset = PretrainDataset(val=False)
# train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_dataloader = DataLoader(dataset=train_dataset, batch_size=args.batch_size,
num_workers=8, collate_fn=my_collate_pretrain, shuffle=True)#sampler=train_sampler)
if args.validate:
test_dataset = KgDatasetVal(val=False)
# test_sampler = torch.utils.data.distributed.DistributedSampler(test_dataset)
test_dataloader = DataLoader(dataset=test_dataset, batch_size=args.batch_size,
num_workers=8, collate_fn=my_collate, shuffle=False)#sampler=test_sampler)
model = KgPreModel(vocab_num)
model = model.to(device)
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
optimizer = optim.AdamW(model.parameters(), lr=1e-4)
# warm up
total_steps = (len(train_dataset) // (args.batch_size / torch.cuda.device_count())) * args.num_epochs \
if len(train_dataset) % args.batch_size == 0 \
else (len(train_dataset) // (args.batch_size / torch.cuda.device_count()) + 1) * args.num_epochs
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0.01 * total_steps,
num_training_steps=total_steps)
criterion_cls = nn.CrossEntropyLoss()
criterion_mse = nn.MSELoss()
criterion_graph = ContrastiveLoss(measure='dot', margin=1.0, max_violation=False)
if args.load_pthpath == "":
start_epoch = 0
else:
print('load model')
# "path/to/checkpoint_xx.pth" -> xx
start_epoch = int(args.load_pthpath.split("_")[-1][:-4]) + 1
model.module.load_state_dict(torch.load(args.load_pthpath))
best_acc = 0
best_epoch = 0
best_acc_t = 0
best_epoch_t = 0
best_acc_t3 = 0
if args.embedding:
answer_candidate_tensor = torch.arange(0, vocab_num).view(-1, 1).long().cuda()
# else:
# answer_candidate_tensor = torch.tensor(answer_embedding).float().cuda()
# answer_candidate_tensor = F.normalize(answer_candidate_tensor, dim=1, p=2)
#model.module.load_state_dict(torch.load('contrasloss_check_v3/model_for_epoch_4.pth'))
for epoch in range(start_epoch, args.num_epochs):
train_answers = []
train_preds = []
train_preds_trip = []
train_preds_trip_3 = []
train_answers_trip = []
for batch_data in tqdm(train_dataloader):
visual_faetures = torch.from_numpy(np.array(batch_data['img'], dtype=float)).float().to(device)
source_seq = tokenizer(batch_data['ques'], padding=True, return_tensors="pt",
add_special_tokens=True)
input_id = source_seq['input_ids'].to(device)
attention_mask = source_seq['attention_mask'].to(device)
token_type_ids = source_seq['token_type_ids'].to(device)
spatial_feature = torch.tensor(batch_data['spatial']).float().to(device)
most_id = batch_data['mostid']
most_id_tensor = torch.tensor(most_id).long().cuda()
model.zero_grad()
anchor = model(input_id, attention_mask, token_type_ids, visual_faetures, spatial_feature)
if args.embedding:
most_id_tensor = torch.tensor(most_id).view(anchor.shape[0], -1).long().cuda()
if torch.cuda.device_count() > 1:
most = model.module.decode_tail(most_id_tensor)
else:
most = model.decode_tail(most_id_tensor)
else:
most = torch.tensor(batch_data['most']).float().to(device)
most = F.normalize(most, dim=-1, p=2)
if args.embedding:
if torch.cuda.device_count() > 1:
answer_candidate_tensor_train = model.module.decode_tail(answer_candidate_tensor)
cls = model.module.cal_sim(anchor, answer_candidate_tensor_train)
else:
answer_candidate_tensor_train = model.decode_tail(answer_candidate_tensor)
cls = model.cal_sim(anchor, answer_candidate_tensor_train)
anchor = F.normalize(anchor, dim=1, p=2)
optimizer.zero_grad()
most_id_tensor = most_id_tensor[:,0].squeeze()
loss_cl = criterion_cls(cls, most_id_tensor)
if args.dataset == 'okvqa':
loss = 0
for i in range(10):
most_i = most[:,i,:]
loss_mse = criterion_mse(anchor, most_i)
loss_graph = criterion_graph(anchor, most_i)
loss = loss + loss_mse + loss_graph + loss_cl
else:
loss_mse = criterion_mse(anchor, most)
loss_graph = criterion_graph(anchor, most)
loss = loss_mse + loss_graph + loss_cl
loss_stat = loss.item()
loss.backward()
optimizer.step()
scheduler.step()
with torch.no_grad():
if args.embedding:
if torch.cuda.device_count() > 1:
answer_candidate_tensor_train = model.module.decode_tail(answer_candidate_tensor)
else:
answer_candidate_tensor_train = model.decode_tail(answer_candidate_tensor)
answer_candidate_tensor_train = F.normalize(answer_candidate_tensor_train, dim=1, p=2)
trip_predict, trip_predict_3 = generate_tripleid(anchor.float(), answer_candidate_tensor_train)
else:
trip_predict, trip_predict_3 = generate_tripleid(anchor.float(), answer_candidate_tensor)
# _, idx_1 = torch.topk(cls, k=1)
for i, pre in enumerate(most_id):
# train_preds.append(idx_1[i]) # [(num_nodes,)]
train_answers.append(most_id[i])
train_preds_trip.append(trip_predict[i])
train_preds_trip_3.append(trip_predict_3[i])
train_answers_trip.append(most_id[i])
# train_acc_1 = cal_acc_old(train_answers, train_preds)
if args.dataset == 'krvqa':
train_acc_1_trip = cal_acc(train_answers_trip, train_preds_trip)
print('epoch %d train_loss = %.1f, acc_trip = %.4f' % (epoch, loss_stat,
train_acc_1_trip))
else:
# train_acc_1_ce = cal_acc_old(train_answers, train_preds)
train_acc_1_trip = cal_acc_multi(train_answers_trip, train_preds_trip)
print('epoch %d train_loss = %.1f, acc_trip = %.4f' % (epoch, loss_stat,
train_acc_1_trip))
# print('acc_ce = %.4f' % train_acc_1_ce)
if torch.cuda.device_count() > 1:
torch.save(model.module.state_dict(), args.model_dir + 'model_for_epoch_%d.pth' % epoch)
else:
torch.save(model.state_dict(), args.model_dir + 'model_for_epoch_%d.pth' % epoch)
if args.validate:
model.eval()
answers = [] # [batch_answers,...]
preds = [] # [batch_preds,...]
preds_trip = []
preds_trip_3 = []
answers_trip = []
print(f"\nValidation after epoch {epoch}:")
for i, batch_data in enumerate(tqdm(test_dataloader)):
with torch.no_grad():
visual_faetures = torch.tensor(batch_data['img']).float().to(device)
source_seq = tokenizer(batch_data['ques'], padding=True, return_tensors="pt",
add_special_tokens=True).to(device)
input_id = source_seq['input_ids'].to(device)
attention_mask = source_seq['attention_mask'].to(device)
token_type_ids = source_seq['token_type_ids'].to(device)
spatial_feature = torch.tensor(batch_data['spatial']).float().to(device)
# most = torch.tensor(batch_data['most']).to(device)
most_id = batch_data['mostid']
anchor = model(input_id, attention_mask, token_type_ids, visual_faetures, spatial_feature)
# if args.embedding:
# answer_candidate_tensor_test = model.module.decode_tail(answer_candidate_tensor)
# cls = model.module.cal_sim(anchor, answer_candidate_tensor_test)
anchor = F.normalize(anchor, dim=1, p=2)
if args.embedding:
if torch.cuda.device_count() > 1:
answer_candidate_tensor_test = model.module.decode_tail(answer_candidate_tensor)
else:
answer_candidate_tensor_test = model.decode_tail(answer_candidate_tensor)
answer_candidate_tensor_test = F.normalize(answer_candidate_tensor_test, dim=1, p=2)
trip_predict, trip_predict_3 = generate_tripleid(anchor, answer_candidate_tensor_test)
else:
trip_predict, trip_predict_3 = generate_tripleid(anchor, answer_candidate_tensor)
# _, idx_1 = torch.topk(cls, k=1)
for i, pre in enumerate(most_id):
# preds.append(idx_1[i]) # [(num_nodes,)]
answers.append(most_id[i])
preds_trip.append(trip_predict[i])
preds_trip_3.append(trip_predict_3[i])
answers_trip.append(most_id[i])
# acc_1 = cal_acc_old(answers, preds)
if args.dataset == 'krvqa':
acc_1_trip = cal_acc(answers_trip, preds_trip)
print('epoch %d , acc_trip = %.4f' % (
epoch, acc_1_trip))
else:
acc_1_trip = cal_acc_multi(answers_trip, preds_trip)
print('epoch %d , acc_trip = %.4f' % (
epoch, acc_1_trip))
# print('acc_ce = %.4f' % acc_1)
if acc_1_trip > best_acc_t:
best_acc_t = acc_1_trip
best_epoch_t = epoch
print("best_acc@1t={:.2%}, epoch{}".format(best_acc_t, best_epoch_t))
if args.dataset == 'fvqa':
print("best_acc@3t={:.2%}".format(best_acc_t3))
model.train()
if __name__ == "__main__":
train()