This repository has been archived by the owner on May 19, 2018. It is now read-only.
/
main.py
215 lines (183 loc) · 9.38 KB
/
main.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
import time
import numpy as np
import skimage
import tensorflow as tf
import os
import reader
from models.multimodal import MultiModal
flags = tf.flags
flags.DEFINE_integer(
'N', -1, 'The number of samples of the Flickr8k dataset. -1 for all the samples.')
flags.DEFINE_boolean('restore', True, "Wether we restore the graph from the checkpoint or not")
logging = tf.logging
class Configs(object):
learning_rate = 0.001 #0.25
max_grad_norm = 5
num_layers = 2
num_steps = 14
hidden_size = 400
image_features_size = 500
max_epoch = 4
max_max_epoch = 3
keep_prob = 0.5
lr_decay = 0.5
batch_size = 20
vocab_size = 10000
image_size = 227
pad_symbol = 0.
def main(rnn_config, eval_config):
print('Loading data')
train_data, valid_data, vocab = reader.flickr_raw_data(flags.FLAGS.N, rnn_config.num_steps,
rnn_config.image_size)
rnn_config.pad_symbol = vocab['<PAD>']
eval_config.pad_symbol = vocab['<PAD>']
test_image = np.zeros(
(eval_config.batch_size, eval_config.image_size, eval_config.image_size, 3))
im = skimage.img_as_float(skimage.io.imread('data/test_image.jpg')
) #plt.imread('data/test_image.jpg') / 255.
test_image[0, :, :] = skimage.transform.rescale(im, rnn_config.image_size / 227.0)
rnn_config.vocab_size = len(vocab)
eval_config.vocab_size = len(vocab)
print('Vocab size of: {}'.format(rnn_config.vocab_size))
with tf.Graph().as_default(), tf.Session() as sess:
global_step_tensor, m, merged, mtest, mvalid, mgen, gen_image_tensor, image_train_op, saver, writer = init_models(eval_config,
rnn_config, sess)
train_model(eval_config, global_step_tensor, m, merged, mtest, mvalid, saver, sess,
test_image, train_data, valid_data, vocab, writer)
create_image(sess, mgen, gen_image_tensor, image_train_op, ['<BOS>', 'a', 'woman', 'wear', 'a', 'blue', 'coat', '<EOS>'], eval_config, vocab, writer)
def create_image(session, mgen, image_tensor, image_train_op, input_sentence, eval_config, vocab, writer):
input = np.zeros((eval_config.batch_size, eval_config.num_steps))
for i, word in enumerate(input_sentence):
input[:, i] = vocab[word]
target = np.zeros((eval_config.batch_size, eval_config.num_steps))
target[:,:-1] = input[:,:-1]
feed_dict = {mgen.input_data: input, mgen.targets:target}
image_sum = tf.image_summary('picture', image_tensor)
for i in range(10000):
cost, _, summary = session.run([mgen.cost, image_train_op, image_sum], feed_dict=feed_dict)
if i % 10 == 0:
writer.add_summary(summary, global_step=i)
print(cost)
def train_model(eval_config, global_step_tensor, m, merged, mtest, mvalid, saver, sess, test_image,
train_data, valid_data, vocab, writer):
print('Getting the global step : {}'.format(tf.train.global_step(sess, global_step_tensor)))
epoch = get_epoch(global_step_tensor, sess, flags.FLAGS.N, config.batch_size)
while epoch <= config.max_max_epoch:
lr_decay = config.lr_decay ** max(epoch - config.max_epoch, 0.0)
m.assign_lr(sess, config.learning_rate * lr_decay)
print('Running epoch {} with learning rate {}'.format(epoch,
config.learning_rate * lr_decay))
train_loss = run_epoch(sess, train_data, m.train_op, config, m, mtest, test_image, vocab,
global_step_tensor,
summary=merged,
summary_writer=writer,
verbose=True)
print('Training Loss : {:.3f}'.format(train_loss))
valid_loss = run_epoch(sess, valid_data, mvalid.train_op, config, mvalid, None, test_image,
vocab, global_step_tensor,
verbose=False)
print('Valid Loss: {:.3f}'.format(valid_loss))
print('Sampling rnn')
s = mtest.sample(sess, vocab, eval_config, test_image, ['<BOS>'])
print(s)
saver.save(sess, os.path.abspath('ckpts/captionning'), global_step_tensor)
epoch = get_epoch(global_step_tensor, sess, flags.FLAGS.N, config.batch_size)
def init_models(eval_config, rnn_config, sess):
global_step_tensor = tf.Variable(0, trainable=False, name='global_step')
print('Creating rnn model')
if flags.FLAGS.restore:
initializer = None
else:
initializer = tf.uniform_unit_scaling_initializer()
with tf.variable_scope("model", reuse=None, initializer=initializer):
train_image_tensor = tf.placeholder(np.float32,
(rnn_config.batch_size, rnn_config.image_size,
rnn_config.image_size, 3), 'input_image')
m = MultiModal(is_training=True,
config=rnn_config,
image_tensor=train_image_tensor,
global_step_tensor=global_step_tensor)
m.load_alexnet('models/alexnet_weights.npy', sess)
variables_to_save = tf.trainable_variables() + [global_step_tensor]
#print(variables_to_save)
with tf.variable_scope("model", reuse=True, initializer=initializer):
mvalid = MultiModal(is_training=False,
config=rnn_config,
image_tensor=train_image_tensor,
global_step_tensor=global_step_tensor)
test_image_tensor = tf.placeholder(np.float32,
(eval_config.batch_size, eval_config.image_size,
eval_config.image_size, 3), 'test_input_image')
mtest = MultiModal(is_training=False,
config=eval_config,
image_tensor=test_image_tensor,
global_step_tensor=global_step_tensor)
initial_value = np.zeros((eval_config.batch_size, eval_config.image_size, eval_config.image_size, 3)).astype(np.float32)
initial_value[0,:,:,:] = skimage.img_as_float(skimage.io.imread('data/test_image.jpg'))
image_gen = tf.Variable(initial_value, trainable=True)
mgen = MultiModal(is_training=False,
config=eval_config,
image_tensor=image_gen,
global_step_tensor=global_step_tensor)
gradients = tf.gradients(mgen.cost, [image_gen])
print(gradients)
optimizer = tf.train.AdamOptimizer(0.1)
image_train = optimizer.apply_gradients(zip(gradients, [image_gen]))
merged = tf.merge_all_summaries()
writer = tf.train.SummaryWriter("logs/")
saver = tf.train.Saver()
tf.initialize_all_variables().run()
if flags.FLAGS.restore:
checkpoint = tf.train.latest_checkpoint(os.path.abspath('ckpts/'))
if checkpoint:
print('Restoring from checkpoint: {}'.format(checkpoint))
saver.restore(sess, checkpoint)
return global_step_tensor, m, merged, mtest, mvalid, mgen, image_gen, image_train, saver, writer
def get_epoch(global_step_tensor, sess, N, batch_size):
return (tf.train.global_step(sess, global_step_tensor) + 1) // (N // batch_size)
def run_epoch(session, data, eval_op, config, train_model, test_model, test_image, vocab,
global_step,
summary_writer=None,
summary=None,
verbose=False):
"""Runs the model on the given data."""
epoch_size = len(data['dataset']) / config.batch_size
start_time = time.time()
costs = 0.0
iters = 0
state = train_model.initial_state.eval()
image_shape = (config.image_size, config.image_size, 3)
eval_ops = [train_model.cost, train_model.logits, eval_op]
if summary is not None:
eval_ops.append(summary)
flickr_iter = reader.flickr_iterator(data, config.batch_size, config.num_steps, image_shape)
for step, (x, im, y) in enumerate(flickr_iter):
feed_dict = {
train_model.input_data: x,
train_model.targets: y,
train_model.initial_state: state,
train_model.image_input: im
}
results = session.run(eval_ops, feed_dict)
cost = results[0]
if summary is not None:
summary_result = results[3]
costs += cost
iters += config.num_steps
if verbose:
completed = step * 1.0 / epoch_size
perplexity = np.exp(costs / iters)
wps = iters * config.batch_size / (time.time() - start_time)
print("{0:.2%} perplexity: {1:.3f} speed: {2:.0f} wps".format(completed, perplexity,
wps))
summary_writer.add_summary(summary_result,
global_step=tf.train.global_step(session, global_step))
if verbose and step % 25 == 0 and step != 0:
print(test_model.sample(session, vocab, eval_config, test_image, prime=['<BOS>']))
print(test_model.sample(session, vocab, eval_config, test_image, prime=['<BOS>'], sampling_func=np.argmax))
return np.exp(costs / iters)
if __name__ == "__main__":
config = Configs()
eval_config = Configs()
eval_config.batch_size = 1
main(config, eval_config)