-
Notifications
You must be signed in to change notification settings - Fork 0
/
dynamic_model.py
187 lines (165 loc) · 9.68 KB
/
dynamic_model.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import numpy as np
import tensorflow as tf
import time
import encoder
import decoder
from data_util import collect_variables, modify_variables
import sys
sys.path.append("..")
from model import id_to_arg_scope, id_to_scope, id_to_saverscope, id_to_model, id_to_checkpoint
slim = tf.contrib.slim
def get_train_ops(image_emb_train, encoder_train_input, encoder_train_target, decoder_train_target, params,
reuse=tf.AUTO_REUSE):
with tf.variable_scope('EPD', reuse=reuse):
my_encoder = encoder.Model(image_emb_train, encoder_train_input, encoder_train_target, params, tf.estimator.ModeKeys.TRAIN,
'Encoder', reuse)
encoder_outputs = my_encoder.arch_emb
encoder_outputs.set_shape([None, params['encoder_hidden_size']])
my_decoder = decoder.Model(encoder_outputs, decoder_train_target, params, tf.estimator.ModeKeys.TRAIN, 'Decoder', reuse)
encoder_loss = my_encoder.loss
decoder_loss = my_decoder.loss
decoder_correct_rate = my_decoder.correct_rate
mse = encoder_loss
cross_entropy = decoder_loss
total_loss = params['trade_off'] * encoder_loss + (1 - params['trade_off']) * decoder_loss + params['weight_decay'] * tf.add_n(
[tf.nn.l2_loss(v) for v in tf.trainable_variables()])
tf.summary.scalar('training_loss', total_loss)
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.constant(params['lr'])
if params['optimizer'] == "sgd":
learning_rate = tf.cond(
global_step < params['start_decay_step'],
lambda: learning_rate,
lambda: tf.train.exponential_decay(
learning_rate,
(global_step - params['start_decay_step']),
params['decay_steps'],
params['decay_factor'],
staircase=True),
name="calc_learning_rate")
opt = tf.train.GradientDescentOptimizer(learning_rate)
elif params['optimizer'] == "adam":
assert float(params['lr']) <= 0.001, "! High Adam learning rate %g" % params['lr']
opt = tf.train.AdamOptimizer(learning_rate)
elif params['optimizer'] == 'adadelta':
opt = tf.train.AdadeltaOptimizer(learning_rate=learning_rate)
tf.summary.scalar("learning_rate", learning_rate)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
var_list = collect_variables(underired_scope=id_to_saverscope[params["model_id"]])
gradients, variables = zip(*opt.compute_gradients(total_loss, var_list))
grad_norm = tf.global_norm(gradients)
clipped_gradients, _ = tf.clip_by_global_norm(gradients, params['max_gradient_norm'])
train_op = opt.apply_gradients(
zip(clipped_gradients, variables), global_step=global_step)
debug_dict = {
'predict_value': my_encoder.predict_value,
'arch_emb': my_encoder.arch_emb,
'image_emb': my_encoder.image_emb,
'logits': my_decoder.logits,
'encoder_input': my_encoder.x,
'decoder_target': my_decoder.target,
}
return mse, cross_entropy, total_loss, learning_rate, train_op, global_step, grad_norm, decoder_correct_rate, debug_dict
def get_evaluate_ops(image_emb_eval, encoder_eval_input, encoder_eval_target, decoder_eval_target, params,
reuse=tf.AUTO_REUSE):
with tf.variable_scope('EPD', reuse=reuse):
my_encoder = encoder.Model(image_emb_eval, encoder_eval_input, encoder_eval_target, params, tf.estimator.ModeKeys.EVAL, 'Encoder', reuse)
encoder_outputs = my_encoder.arch_emb
encoder_outputs.set_shape([None, params['decoder_hidden_size']])
my_decoder = decoder.Model(encoder_outputs, decoder_eval_target, params, tf.estimator.ModeKeys.EVAL, 'Decoder', reuse)
encoder_loss = my_encoder.loss
decoder_loss = my_decoder.loss
decoder_correct_rate = my_decoder.correct_rate
mse = encoder_loss
cross_entropy = decoder_loss
total_loss = params['trade_off'] * encoder_loss + (1 - params['trade_off']) * decoder_loss + params['weight_decay'] * tf.add_n(
[tf.nn.l2_loss(v) for v in tf.trainable_variables()])
debug_dict = {
'predict_value': my_encoder.predict_value,
'encoder_input': my_encoder.x,
'target_value': my_encoder.y,
}
return mse, cross_entropy, total_loss, decoder_correct_rate, debug_dict
def get_predict_ops(image_emb_pred, encoder_pred_input, params, reuse=tf.AUTO_REUSE):
encoder_pred_target = None
decoder_pred_target = None
with tf.variable_scope('EPD', reuse=reuse):
my_encoder = encoder.Model(image_emb_pred, encoder_pred_input, encoder_pred_target, params, tf.estimator.ModeKeys.PREDICT, 'Encoder', reuse)
encoder_outputs = my_encoder.arch_emb
my_decoder = decoder.Model(encoder_outputs, decoder_pred_target, params, tf.estimator.ModeKeys.PREDICT, 'Decoder', reuse)
arch_emb, predict_value, new_arch_emb = my_encoder.infer()
sample_id = my_decoder.decode()
return predict_value, sample_id, arch_emb, new_arch_emb
def build_model_train_graph(params):
with tf.Graph().as_default() as g:
with tf.device('/gpu:0'):
tf.logging.error('Building model train graph')
image_ph = tf.placeholder(shape=[None, 299, 299, 3], dtype=tf.float32)
encoder_input_ph = tf.placeholder(shape=[None, params['encoder_length']], dtype=tf.int32)
encoder_target_ph = tf.placeholder(shape=[None, 1], dtype=tf.float32)
decoder_target_ph = tf.placeholder(shape=[None, params['decoder_length']], dtype=tf.int32)
model_id = params['model_id']
with slim.arg_scope(id_to_arg_scope[model_id]):
logits, end_points = id_to_model[model_id](
image_ph, num_classes=1001, is_training=False, scope=id_to_scope[model_id], reuse=tf.AUTO_REUSE)
image = end_points['PreLogits']
image_emb = tf.squeeze(image, [1, 2], name='SpatialSqueeze')
train_mse, train_cross_entropy, train_loss, learning_rate, train_op, global_step, grad_norm, decoder_correct_rate, debug_dict = get_train_ops(
image_emb, encoder_input_ph, encoder_target_ph, decoder_target_ph, params)
merged_summary = tf.summary.merge_all()
run_ops = [
train_mse,
train_cross_entropy,
train_loss,
learning_rate,
train_op,
global_step,
grad_norm,
decoder_correct_rate,
merged_summary,
]
return g, image_ph, encoder_input_ph, encoder_target_ph, decoder_target_ph, run_ops
def build_model_evaluate_graph(params):
with tf.Graph().as_default() as g:
with tf.device('/gpu:0'):
tf.logging.error('Building model evaluate graph')
image_ph = tf.placeholder(shape=[None, 299, 299, 3], dtype=tf.float32)
encoder_input_ph = tf.placeholder(shape=[None, params['encoder_length']], dtype=tf.int32)
encoder_target_ph = tf.placeholder(shape=[None, 1], dtype=tf.float32)
decoder_target_ph = tf.placeholder(shape=[None, params['decoder_length']], dtype=tf.int32)
model_id = params['model_id']
with slim.arg_scope(id_to_arg_scope[model_id]):
logits, end_points = id_to_model[model_id](
image_ph, num_classes=1001, is_training=False, scope=id_to_scope[model_id], reuse=tf.AUTO_REUSE)
image = end_points['PreLogits']
image_emb = tf.squeeze(image, [1, 2], name='SpatialSqueeze')
eval_mse, eval_cross_entropy, eval_loss, decoder_correct_rate, debug_dict = get_evaluate_ops(
image_emb, encoder_input_ph, encoder_target_ph, decoder_target_ph, params)
run_ops = [
eval_mse,
eval_cross_entropy,
eval_loss,
decoder_correct_rate,
]
return g, image_ph, encoder_input_ph, encoder_target_ph, decoder_target_ph, run_ops, debug_dict
def build_model_optim_graph(params):
with tf.Graph().as_default() as g:
with tf.device('/gpu:0'):
tf.logging.error('Building model optimizer graph')
image_ph = tf.placeholder(shape=[None, 299, 299, 3], dtype=tf.float32)
encoder_input_ph = tf.placeholder(shape=[None, params['encoder_length']], dtype=tf.int32)
encoder_target_ph = tf.placeholder(shape=[None, 1], dtype=tf.float32)
decoder_target_ph = tf.placeholder(shape=[None, params['decoder_length']], dtype=tf.int32)
model_id = params['model_id']
with slim.arg_scope(id_to_arg_scope[model_id]):
logits, end_points = id_to_model[model_id](
image_ph, num_classes=1001, is_training=False, scope=id_to_scope[model_id], reuse=tf.AUTO_REUSE)
image = end_points['PreLogits']
image_emb = tf.squeeze(image, [1, 2], name='SpatialSqueeze')
predict_value, sample_id, arch_emb, new_arch_emb = get_predict_ops(image_emb, encoder_input_ph, params)
return g, image_ph, encoder_input_ph, encoder_target_ph, decoder_target_ph, predict_value, sample_id, image_emb, arch_emb, new_arch_emb