-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathtrain.py
165 lines (140 loc) · 6.53 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
import os
import pickle
import argparse
import torch
from torch.autograd import Variable
from build_vocab import Vocab
from data_loader import get_data_loader
from data_loader import get_styled_data_loader
from models import EncoderCNN
from models import FactoredLSTM
from loss import masked_cross_entropy
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
def to_var(x, volatile=False):
if torch.cuda.is_available():
x = x.cuda()
return Variable(x, volatile=volatile)
def eval_outputs(outputs, vocab):
# outputs: [batch, max_len - 1, vocab_size]
indices = torch.topk(outputs, 1)[1]
indices = indices.squeeze(2)
indices = indices.data
for i in range(len(indices)):
caption = [vocab.i2w[x] for x in indices[i]]
print(caption)
def main(args):
model_path = args.model_path
if not os.path.exists(model_path):
os.makedirs(model_path)
# load vocablary
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
img_path = args.img_path
factual_cap_path = args.factual_caption_path
humorous_cap_path = args.humorous_caption_path
# import data_loader
data_loader = get_data_loader(img_path, factual_cap_path, vocab,
args.caption_batch_size, shuffle=True)
styled_data_loader = get_styled_data_loader(humorous_cap_path, vocab,
args.language_batch_size,
shuffle=True)
# import models
emb_dim = args.emb_dim
hidden_dim = args.hidden_dim
factored_dim = args.factored_dim
vocab_size = len(vocab)
encoder = EncoderCNN(emb_dim)
decoder = FactoredLSTM(emb_dim, hidden_dim, factored_dim, vocab_size)
if torch.cuda.is_available():
encoder = encoder.cuda()
decoder = decoder.cuda()
# loss and optimizer
criterion = masked_cross_entropy
cap_params = list(decoder.parameters()) + list(encoder.A.parameters())
lang_params = list(decoder.parameters())
optimizer_cap = torch.optim.Adam(cap_params, lr=args.lr_caption)
optimizer_lang = torch.optim.Adam(lang_params, lr=args.lr_language)
# train
total_cap_step = len(data_loader)
total_lang_step = len(styled_data_loader)
epoch_num = args.epoch_num
for epoch in range(epoch_num):
# caption
for i, (images, captions, lengths) in enumerate(data_loader):
images = to_var(images, volatile=True)
captions = to_var(captions.long())
# forward, backward and optimize
decoder.zero_grad()
encoder.zero_grad()
features = encoder(images)
outputs = decoder(captions, features, mode="factual")
loss = criterion(outputs[:, 1:, :].contiguous(),
captions[:, 1:].contiguous(), lengths - 1)
loss.backward()
optimizer_cap.step()
# print log
if i % args.log_step_caption == 0:
print("Epoch [%d/%d], CAP, Step [%d/%d], Loss: %.4f"
% (epoch+1, epoch_num, i, total_cap_step,
loss.data.mean()))
eval_outputs(outputs, vocab)
# language
for i, (captions, lengths) in enumerate(styled_data_loader):
captions = to_var(captions.long())
# forward, backward and optimize
decoder.zero_grad()
outputs = decoder(captions, mode='humorous')
loss = criterion(outputs, captions[:, 1:].contiguous(), lengths-1)
loss.backward()
optimizer_lang.step()
# print log
if i % args.log_step_language == 0:
print("Epoch [%d/%d], LANG, Step [%d/%d], Loss: %.4f"
% (epoch+1, epoch_num, i, total_lang_step,
loss.data.mean()))
# save models
torch.save(decoder.state_dict(),
os.path.join(model_path, 'decoder-%d.pkl' % (epoch + 1,)))
torch.save(encoder.state_dict(),
os.path.join(model_path, 'encoder-%d.pkl' % (epoch + 1,)))
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='StyleNet: Generating Attractive Visual Captions \
with Styles')
parser.add_argument('--model_path', type=str, default='pretrained_models',
help='path for saving trained models')
parser.add_argument('--vocab_path', type=str, default='./data/vocab.pkl',
help='path for vocabrary')
parser.add_argument('--img_path', type=str,
default='./data/flickr7k_images',
help='path for train images directory')
parser.add_argument('--factual_caption_path', type=str,
default='./data/factual_train.txt',
help='path for factual caption file')
parser.add_argument('--humorous_caption_path', type=str,
default='./data/humor/funny_train.txt',
help='path for humorous caption file')
parser.add_argument('--romantic_caption_path', type=str,
default='./data/romantic/romanntic_train.txt',
help='path for romantic caption file')
parser.add_argument('--caption_batch_size', type=int, default=64,
help='mini batch size for caption model training')
parser.add_argument('--language_batch_size', type=int, default=96,
help='mini batch size for language model training')
parser.add_argument('--emb_dim', type=int, default=300,
help='embedding size of word, image')
parser.add_argument('--hidden_dim', type=int, default=512,
help='hidden state size of factored LSTM')
parser.add_argument('--factored_dim', type=int, default=512,
help='size of factored matrix')
parser.add_argument('--lr_caption', type=int, default=0.0002,
help='learning rate for caption model training')
parser.add_argument('--lr_language', type=int, default=0.0005,
help='learning rate for language model training')
parser.add_argument('--epoch_num', type=int, default=30)
parser.add_argument('--log_step_caption', type=int, default=50,
help='steps for print log while train caption model')
parser.add_argument('--log_step_language', type=int, default=10,
help='steps for print log while train language model')
args = parser.parse_args()
main(args)