-
Notifications
You must be signed in to change notification settings - Fork 42
/
train.py
238 lines (183 loc) · 9.81 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
import math
import json
import argparse
import torch
import torch.nn as nn
import numpy as np
import os
import pickle
from utils import coco_eval, to_var
from data_loader import get_loader
from adaptive import Encoder2Decoder
from build_vocab import Vocabulary
from torch.autograd import Variable
from torchvision import transforms
from torch.nn.utils.rnn import pack_padded_sequence
def main(args):
# To reproduce training results
torch.manual_seed( args.seed )
if torch.cuda.is_available():
torch.cuda.manual_seed( args.seed )
# Create model directory
if not os.path.exists( args.model_path ):
os.makedirs(args.model_path)
# Image Preprocessing
# For normalization, see https://github.com/pytorch/vision#models
transform = transforms.Compose([
transforms.RandomCrop( args.crop_size ),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(( 0.485, 0.456, 0.406 ),
( 0.229, 0.224, 0.225 ))])
# Load vocabulary wrapper.
with open( args.vocab_path, 'rb') as f:
vocab = pickle.load( f )
# Build training data loader
data_loader = get_loader( args.image_dir, args.caption_path, vocab,
transform, args.batch_size,
shuffle=True, num_workers=args.num_workers )
# Load pretrained model or build from scratch
adaptive = Encoder2Decoder( args.embed_size, len(vocab), args.hidden_size )
if args.pretrained:
adaptive.load_state_dict( torch.load( args.pretrained ) )
# Get starting epoch #, note that model is named as '...your path to model/algoname-epoch#.pkl'
# A little messy here.
start_epoch = int( args.pretrained.split('/')[-1].split('-')[1].split('.')[0] ) + 1
else:
start_epoch = 1
# Constructing CNN parameters for optimization, only fine-tuning higher layers
cnn_subs = list( adaptive.encoder.resnet_conv.children() )[ args.fine_tune_start_layer: ]
cnn_params = [ list( sub_module.parameters() ) for sub_module in cnn_subs ]
cnn_params = [ item for sublist in cnn_params for item in sublist ]
cnn_optimizer = torch.optim.Adam( cnn_params, lr=args.learning_rate_cnn,
betas=( args.alpha, args.beta ) )
# Other parameter optimization
params = list( adaptive.encoder.affine_a.parameters() ) + list( adaptive.encoder.affine_b.parameters() ) \
+ list( adaptive.decoder.parameters() )
# Will decay later
learning_rate = args.learning_rate
# Language Modeling Loss
LMcriterion = nn.CrossEntropyLoss()
# Change to GPU mode if available
if torch.cuda.is_available():
adaptive.cuda()
LMcriterion.cuda()
# Train the Models
total_step = len( data_loader )
cider_scores = []
best_cider = 0.0
best_epoch = 0
# Start Training
for epoch in range( start_epoch, args.num_epochs + 1 ):
# Start Learning Rate Decay
if epoch > args.lr_decay:
frac = float( epoch - args.lr_decay ) / args.learning_rate_decay_every
decay_factor = math.pow( 0.5, frac )
# Decay the learning rate
learning_rate = args.learning_rate * decay_factor
print 'Learning Rate for Epoch %d: %.6f'%( epoch, learning_rate )
optimizer = torch.optim.Adam( params, lr=learning_rate, betas=( args.alpha, args.beta ) )
# Language Modeling Training
print '------------------Training for Epoch %d----------------'%( epoch )
for i, (images, captions, lengths, _, _ ) in enumerate( data_loader ):
# Set mini-batch dataset
images = to_var( images )
captions = to_var( captions )
lengths = [ cap_len - 1 for cap_len in lengths ]
targets = pack_padded_sequence( captions[:,1:], lengths, batch_first=True )[0]
# Forward, Backward and Optimize
adaptive.train()
adaptive.zero_grad()
packed_scores = adaptive( images, captions, lengths )
# Compute loss and backprop
loss = LMcriterion( packed_scores[0], targets )
loss.backward()
# Gradient clipping for gradient exploding problem in LSTM
for p in adaptive.decoder.LSTM.parameters():
p.data.clamp_( -args.clip, args.clip )
optimizer.step()
# Start CNN fine-tuning
if epoch > args.cnn_epoch:
cnn_optimizer.step()
# Print log info
if i % args.log_step == 0:
print 'Epoch [%d/%d], Step [%d/%d], CrossEntropy Loss: %.4f, Perplexity: %5.4f'%( epoch,
args.num_epochs,
i, total_step,
loss.data[0],
np.exp( loss.data[0] ) )
# Save the Adaptive Attention model after each epoch
torch.save( adaptive.state_dict(),
os.path.join( args.model_path,
'adaptive-%d.pkl'%( epoch ) ) )
# Evaluation on validation set
cider = coco_eval( adaptive, args, epoch )
cider_scores.append( cider )
if cider > best_cider:
best_cider = cider
best_epoch = epoch
if len( cider_scores ) > 5:
last_6 = cider_scores[-6:]
last_6_max = max( last_6 )
# Test if there is improvement, if not do early stopping
if last_6_max != best_cider:
print 'No improvement with CIDEr in the last 6 epochs...Early stopping triggered.'
print 'Model of best epoch #: %d with CIDEr score %.2f'%( best_epoch, best_cider )
break
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument( '-f', default='self', help='To make it runnable in jupyter' )
parser.add_argument( '--model_path', type=str, default='./models',
help='path for saving trained models')
parser.add_argument('--crop_size', type=int, default=224 ,
help='size for randomly cropping images')
parser.add_argument('--vocab_path', type=str, default='./data/vocab.pkl',
help='path for vocabulary wrapper')
parser.add_argument('--image_dir', type=str, default='./data/resized' ,
help='directory for resized training images')
parser.add_argument('--caption_path', type=str,
default='./data/annotations/karpathy_split_train.json',
help='path for train annotation json file')
parser.add_argument('--caption_val_path', type=str,
default='./data/annotations/karpathy_split_val.json',
help='path for validation annotation json file')
parser.add_argument('--log_step', type=int, default=10,
help='step size for printing log info')
parser.add_argument('--seed', type=int, default=123,
help='random seed for model reproduction')
# ---------------------------Hyper Parameter Setup------------------------------------
# CNN fine-tuning
parser.add_argument('--fine_tune_start_layer', type=int, default=5,
help='CNN fine-tuning layers from: [0-7]')
parser.add_argument('--cnn_epoch', type=int, default=20,
help='start fine-tuning CNN after')
# Optimizer Adam parameter
parser.add_argument( '--alpha', type=float, default=0.8,
help='alpha in Adam' )
parser.add_argument( '--beta', type=float, default=0.999,
help='beta in Adam' )
parser.add_argument( '--learning_rate', type=float, default=4e-4,
help='learning rate for the whole model' )
parser.add_argument( '--learning_rate_cnn', type=float, default=1e-4,
help='learning rate for fine-tuning CNN' )
# LSTM hyper parameters
parser.add_argument( '--embed_size', type=int, default=256,
help='dimension of word embedding vectors, also dimension of v_g' )
parser.add_argument( '--hidden_size', type=int, default=512,
help='dimension of lstm hidden states' )
# Training details
parser.add_argument( '--pretrained', type=str, default='', help='start from checkpoint or scratch' )
parser.add_argument( '--num_epochs', type=int, default=50 )
parser.add_argument( '--batch_size', type=int, default=60 ) # on cluster setup, 60 each x 4 for Huckle server
# For eval_size > 30, it will cause cuda OOM error on Huckleberry
parser.add_argument( '--eval_size', type=int, default=28 ) # on cluster setup, 30 each x 4
parser.add_argument( '--num_workers', type=int, default=4 )
parser.add_argument( '--clip', type=float, default=0.1 )
parser.add_argument( '--lr_decay', type=int, default=20, help='epoch at which to start lr decay' )
parser.add_argument( '--learning_rate_decay_every', type=int, default=50,
help='decay learning rate at every this number')
args = parser.parse_args()
print '------------------------Model and Training Details--------------------------'
print(args)
# Start training
main( args )