-
Notifications
You must be signed in to change notification settings - Fork 113
/
train.py
executable file
·291 lines (246 loc) · 11.3 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
# -----------------------------------------------------------
# Stacked Cross Attention Network implementation based on
# https://arxiv.org/abs/1803.08024.
# "Stacked Cross Attention for Image-Text Matching"
# Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, Xiaodong He
#
# Writen by Kuang-Huei Lee, 2018
# ---------------------------------------------------------------
"""Training script"""
import os
import time
import shutil
import torch
import numpy
import data
from vocab import Vocabulary, deserialize_vocab
from model import SCAN
from evaluation import i2t, t2i, AverageMeter, LogCollector, encode_data, shard_xattn_t2i, shard_xattn_i2t
from torch.autograd import Variable
import logging
import tensorboard_logger as tb_logger
import argparse
def main():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('--data_path', default='./data/',
help='path to datasets')
parser.add_argument('--data_name', default='precomp',
help='{coco,f30k}_precomp')
parser.add_argument('--vocab_path', default='./vocab/',
help='Path to saved vocabulary json files.')
parser.add_argument('--margin', default=0.2, type=float,
help='Rank loss margin.')
parser.add_argument('--num_epochs', default=30, type=int,
help='Number of training epochs.')
parser.add_argument('--batch_size', default=128, type=int,
help='Size of a training mini-batch.')
parser.add_argument('--word_dim', default=300, type=int,
help='Dimensionality of the word embedding.')
parser.add_argument('--embed_size', default=1024, type=int,
help='Dimensionality of the joint embedding.')
parser.add_argument('--grad_clip', default=2., type=float,
help='Gradient clipping threshold.')
parser.add_argument('--num_layers', default=1, type=int,
help='Number of GRU layers.')
parser.add_argument('--learning_rate', default=.0002, type=float,
help='Initial learning rate.')
parser.add_argument('--lr_update', default=15, type=int,
help='Number of epochs to update the learning rate.')
parser.add_argument('--workers', default=10, type=int,
help='Number of data loader workers.')
parser.add_argument('--log_step', default=10, type=int,
help='Number of steps to print and record the log.')
parser.add_argument('--val_step', default=500, type=int,
help='Number of steps to run validation.')
parser.add_argument('--logger_name', default='./runs/runX/log',
help='Path to save Tensorboard log.')
parser.add_argument('--model_name', default='./runs/runX/checkpoint',
help='Path to save the model.')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('--max_violation', action='store_true',
help='Use max instead of sum in the rank loss.')
parser.add_argument('--img_dim', default=2048, type=int,
help='Dimensionality of the image embedding.')
parser.add_argument('--no_imgnorm', action='store_true',
help='Do not normalize the image embeddings.')
parser.add_argument('--no_txtnorm', action='store_true',
help='Do not normalize the text embeddings.')
parser.add_argument('--raw_feature_norm', default="clipped_l2norm",
help='clipped_l2norm|l2norm|clipped_l1norm|l1norm|no_norm|softmax')
parser.add_argument('--agg_func', default="LogSumExp",
help='LogSumExp|Mean|Max|Sum')
parser.add_argument('--cross_attn', default="t2i",
help='t2i|i2t')
parser.add_argument('--precomp_enc_type', default="basic",
help='basic|weight_norm')
parser.add_argument('--bi_gru', action='store_true',
help='Use bidirectional GRU.')
parser.add_argument('--lambda_lse', default=6., type=float,
help='LogSumExp temp.')
parser.add_argument('--lambda_softmax', default=9., type=float,
help='Attention softmax temperature.')
opt = parser.parse_args()
print(opt)
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
tb_logger.configure(opt.logger_name, flush_secs=5)
# Load Vocabulary Wrapper
vocab = deserialize_vocab(os.path.join(opt.vocab_path, '%s_vocab.json' % opt.data_name))
opt.vocab_size = len(vocab)
# Load data loaders
train_loader, val_loader = data.get_loaders(
opt.data_name, vocab, opt.batch_size, opt.workers, opt)
# Construct the model
model = SCAN(opt)
best_rsum = 0
start_epoch = 0
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
start_epoch = checkpoint['epoch'] + 1
best_rsum = checkpoint['best_rsum']
model.load_state_dict(checkpoint['model'])
# Eiters is used to show logs as the continuation of another
# training
model.Eiters = checkpoint['Eiters']
print("=> loaded checkpoint '{}' (epoch {}, best_rsum {})"
.format(opt.resume, start_epoch, best_rsum))
validate(opt, val_loader, model)
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# Train the Model
for epoch in range(start_epoch, opt.num_epochs):
print(opt.logger_name)
print(opt.model_name)
adjust_learning_rate(opt, model.optimizer, epoch)
# train for one epoch
train(opt, train_loader, model, epoch, val_loader)
# evaluate on validation set
rsum = validate(opt, val_loader, model)
# remember best R@ sum and save checkpoint
is_best = rsum > best_rsum
best_rsum = max(rsum, best_rsum)
if not os.path.exists(opt.model_name):
os.mkdir(opt.model_name)
save_checkpoint({
'epoch': epoch,
'model': model.state_dict(),
'best_rsum': best_rsum,
'opt': opt,
'Eiters': model.Eiters,
}, is_best, filename='checkpoint_{}.pth.tar'.format(epoch), prefix=opt.model_name + '/')
def train(opt, train_loader, model, epoch, val_loader):
# average meters to record the training statistics
batch_time = AverageMeter()
data_time = AverageMeter()
train_logger = LogCollector()
end = time.time()
for i, train_data in enumerate(train_loader):
# switch to train mode
model.train_start()
# measure data loading time
data_time.update(time.time() - end)
# make sure train logger is used
model.logger = train_logger
# Update the model
model.train_emb(*train_data)
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# Print log info
if model.Eiters % opt.log_step == 0:
logging.info(
'Epoch: [{0}][{1}/{2}]\t'
'{e_log}\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
.format(
epoch, i, len(train_loader), batch_time=batch_time,
data_time=data_time, e_log=str(model.logger)))
# Record logs in tensorboard
tb_logger.log_value('epoch', epoch, step=model.Eiters)
tb_logger.log_value('step', i, step=model.Eiters)
tb_logger.log_value('batch_time', batch_time.val, step=model.Eiters)
tb_logger.log_value('data_time', data_time.val, step=model.Eiters)
model.logger.tb_log(tb_logger, step=model.Eiters)
# validate at every val_step
if model.Eiters % opt.val_step == 0:
validate(opt, val_loader, model)
def validate(opt, val_loader, model):
# compute the encoding for all the validation images and captions
img_embs, cap_embs, cap_lens = encode_data(
model, val_loader, opt.log_step, logging.info)
img_embs = numpy.array([img_embs[i] for i in range(0, len(img_embs), 5)])
start = time.time()
if opt.cross_attn == 't2i':
sims = shard_xattn_t2i(img_embs, cap_embs, cap_lens, opt, shard_size=128)
elif opt.cross_attn == 'i2t':
sims = shard_xattn_i2t(img_embs, cap_embs, cap_lens, opt, shard_size=128)
else:
raise NotImplementedError
end = time.time()
print("calculate similarity time:", end-start)
# caption retrieval
(r1, r5, r10, medr, meanr) = i2t(img_embs, cap_embs, cap_lens, sims)
logging.info("Image to text: %.1f, %.1f, %.1f, %.1f, %.1f" %
(r1, r5, r10, medr, meanr))
# image retrieval
(r1i, r5i, r10i, medri, meanr) = t2i(
img_embs, cap_embs, cap_lens, sims)
logging.info("Text to image: %.1f, %.1f, %.1f, %.1f, %.1f" %
(r1i, r5i, r10i, medri, meanr))
# sum of recalls to be used for early stopping
currscore = r1 + r5 + r10 + r1i + r5i + r10i
# record metrics in tensorboard
tb_logger.log_value('r1', r1, step=model.Eiters)
tb_logger.log_value('r5', r5, step=model.Eiters)
tb_logger.log_value('r10', r10, step=model.Eiters)
tb_logger.log_value('medr', medr, step=model.Eiters)
tb_logger.log_value('meanr', meanr, step=model.Eiters)
tb_logger.log_value('r1i', r1i, step=model.Eiters)
tb_logger.log_value('r5i', r5i, step=model.Eiters)
tb_logger.log_value('r10i', r10i, step=model.Eiters)
tb_logger.log_value('medri', medri, step=model.Eiters)
tb_logger.log_value('meanr', meanr, step=model.Eiters)
tb_logger.log_value('rsum', currscore, step=model.Eiters)
return currscore
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar', prefix=''):
tries = 15
error = None
# deal with unstable I/O. Usually not necessary.
while tries:
try:
torch.save(state, prefix + filename)
if is_best:
shutil.copyfile(prefix + filename, prefix + 'model_best.pth.tar')
except IOError as e:
error = e
tries -= 1
else:
break
print('model save {} failed, remaining {} trials'.format(filename, tries))
if not tries:
raise error
def adjust_learning_rate(opt, optimizer, epoch):
"""Sets the learning rate to the initial LR
decayed by 10 every 30 epochs"""
lr = opt.learning_rate * (0.1 ** (epoch // opt.lr_update))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def accuracy(output, target, topk=(1,)):
"""Computes the precision@k for the specified values of k"""
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
res = []
for k in topk:
correct_k = correct[:k].view(-1).float().sum(0)
res.append(correct_k.mul_(100.0 / batch_size))
return res
if __name__ == '__main__':
main()