-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_fn.py
327 lines (248 loc) · 16.2 KB
/
model_fn.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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
"""Define the model."""
import os
import tensorflow as tf
from pathlib import Path
from tensorflow.contrib.rnn import LSTMStateTuple
from tensorflow.contrib import layers
# Upgrade to 2.0
from tensorflow.python.ops import lookup_ops, array_ops
START_TOKEN = 0
END_TOKEN = 1
UNK_TOKEN = 2
PAD_TOKEN = 3
def emb_initialize():
def intialize(shape, dtype, partition_info=None):
embedding = tf.random_uniform(shape, minval=-1.0, maxval=1.0, dtype=dtype)
embedding = tf.nn.l2_normalize(embedding, -1)
return embedding
return intialize
def conv2d(layer, name, n_filters, trainable, k_size=3):
return tf.layers.conv2d(layer, n_filters, kernel_size=(k_size, k_size),
activation=tf.nn.relu, padding=self.conv_padding, name=name, trainable=trainable,
kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(),
use_bias=True)
def vgg_16(inputs,
dropout_keep_prob=0.5,
spatial_squeeze=True,
scope='vgg_16'):
end_points = {}
with tf.variable_scope(scope):
# Conv1
with tf.variable_scope('conv1'):
net = tf.layers.conv2d(inputs, 64, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv1_1')
net = tf.layers.conv2d(net, 64, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv1_2')
# Pool 1
net = tf.layers.max_pooling2d(net, (2, 2), (2, 2), padding='SAME')
# Conv2
with tf.variable_scope('conv2'):
net = tf.layers.conv2d(net, 128, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv2_1')
net = tf.layers.conv2d(net, 128, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv2_2')
end_points['conv2_2'] = net
# Pool 2
net = tf.layers.max_pooling2d(net, (2, 2), (2, 2), padding='SAME')
# Conv3
with tf.variable_scope('conv3'):
net = tf.layers.conv2d(net, 256, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv3_1')
net = tf.layers.conv2d(net, 256, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv3_2')
net = tf.layers.conv2d(net, 256, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv3_3')
end_points['conv3_3'] = net
# Pool 3
net = tf.layers.max_pooling2d(net, (2, 2), (2, 2), padding='SAME')
# Conv 4
with tf.variable_scope('conv4'):
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv4_1')
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv4_2')
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv4_3')
end_points['conv4_3'] = net
# Pool 4
net = tf.layers.max_pooling2d(net, (2, 2), (2, 2), padding='SAME')
# Conv 5
with tf.variable_scope('conv5'):
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv5_1')
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv5_2')
net = tf.layers.conv2d(net, 512, [3,3], activation=tf.nn.relu, padding='SAME', kernel_initializer=tf.glorot_uniform_initializer(),
bias_initializer=tf.zeros_initializer(), use_bias=True, name='conv5_3')
end_points['conv5_3'] = net
# Pool 5
net = tf.layers.max_pooling2d(net, (2, 2), (2, 2), padding='SAME')
return net, end_points
def build_encoder(images, is_training, params):
with tf.variable_scope('encoder'):
# Conv1
W1 = tf.get_variable(dtype=tf.float32, shape=[3,3,1,64], name='weights_conv1', initializer=tf.glorot_uniform_initializer())
b1 = tf.get_variable(dtype=tf.float32, shape=[64], name='biases_conv1', initializer=tf.zeros_initializer())
conv1 = tf.nn.bias_add(tf.nn.conv2d(images, W1, strides=[1,1,1,1], padding='SAME'), b1)
conv1 = tf.nn.relu(conv1)
conv1 = tf.layers.batch_normalization(conv1, training=is_training)
pool1 = tf.nn.max_pool(conv1, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
# Conv2
W2 = tf.get_variable(dtype=tf.float32, shape=[3,3,64,128], name='weights_conv2', initializer=tf.glorot_uniform_initializer())
b2 = tf.get_variable(dtype=tf.float32, shape=[128], name='biases_conv2', initializer=tf.zeros_initializer())
conv2 = tf.nn.bias_add(tf.nn.conv2d(pool1, W2, strides=[1,1,1,1], padding='SAME'), b2)
conv2 = tf.nn.relu(conv2)
conv2 = tf.layers.batch_normalization(conv2, training=is_training)
pool2 = tf.nn.max_pool(conv2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='VALID')
# Conv3
W3 = tf.get_variable(dtype=tf.float32, shape=[3,3,128,256], name='weights_conv3', initializer=tf.glorot_uniform_initializer())
b3 = tf.get_variable(dtype=tf.float32, shape=[256], name='biases_conv3', initializer=tf.zeros_initializer())
conv3 = tf.nn.bias_add(tf.nn.conv2d(pool2, W3, strides=[1,1,1,1], padding='SAME'), b3)
conv3 = tf.nn.relu(conv3)
conv3 = tf.layers.batch_normalization(conv3, training=is_training)
# Conv4
W4 = tf.get_variable(dtype=tf.float32, shape=[3,3,256,256], name='weights_conv4', initializer=tf.glorot_uniform_initializer())
b4 = tf.get_variable(dtype=tf.float32, shape=[256], name='biases_conv4', initializer=tf.zeros_initializer())
conv4 = tf.nn.bias_add(tf.nn.conv2d(conv3, W4, strides=[1,1,1,1], padding='SAME'), b4)
conv4 = tf.nn.relu(conv4)
conv4 = tf.layers.batch_normalization(conv4, training=is_training)
pool3 = tf.nn.max_pool(conv4, ksize=[1, 2, 1, 1], strides=[1, 2, 1, 1], padding='VALID')
# Conv5
W5 = tf.get_variable(dtype=tf.float32, shape=[3,3,256,512], name='weights_conv5', initializer=tf.glorot_uniform_initializer())
b5 = tf.get_variable(dtype=tf.float32, shape=[512], name='biases_conv5', initializer=tf.zeros_initializer())
conv5 = tf.nn.bias_add(tf.nn.conv2d(pool3, W5, strides=[1,1,1,1], padding='SAME'), b5)
conv5 = tf.nn.relu(conv5)
conv5 = tf.layers.batch_normalization(conv5, training=is_training)
# Conv6
W6 = tf.get_variable(dtype=tf.float32, shape=[3,3,512,512], name='weights_conv6', initializer=tf.glorot_uniform_initializer())
b6 = tf.get_variable(dtype=tf.float32, shape=[512], name='biases_conv6', initializer=tf.zeros_initializer())
conv6 = tf.nn.bias_add(tf.nn.conv2d(conv5, W6, strides=[1,1,1,1], padding='SAME'), b6)
conv6 = tf.nn.relu(conv6)
conv6 = tf.layers.batch_normalization(conv6, training=is_training)
pool5 = tf.nn.max_pool(conv6, ksize=[1, 2, 1, 1], strides=[1, 2, 1, 1], padding='VALID')
# Conv7
W7 = tf.get_variable(dtype=tf.float32, shape=[2,2,512,512], name='weights_conv7', initializer=tf.glorot_uniform_initializer())
b7 = tf.get_variable(dtype=tf.float32, shape=[512], name='biases_conv7', initializer=tf.zeros_initializer())
conv7 = tf.nn.bias_add(tf.nn.conv2d(pool5, W7, strides=[1,1,1,1], padding='VALID'), b7)
conv7 = tf.nn.relu(conv7)
out_conv = tf.squeeze(conv7, axis=1)
cell_fw = tf.nn.rnn_cell.LSTMCell(params.encoder_lstm_hidden)
cell_bw = tf.nn.rnn_cell.LSTMCell(params.encoder_lstm_hidden)
out, (state_fw, state_bw) = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, out_conv, dtype=tf.float32)
out = tf.concat(out, axis=-1)
final_state_c = tf.concat((state_fw.c, state_bw.c), 1)
final_state_h = tf.concat((state_fw.h, state_bw.h), 1)
state = LSTMStateTuple(c=final_state_c, h=final_state_h)
return out, state
def build_encoder_v1(images, is_training, params):
# Experimental is w x h = 150 x 120
vgg_res, end_points = vgg_16(images)
if is_training == True:
batch_size = params.batch_size
else:
batch_size = 1
# Add positional embedding
_, h, w, _ = vgg_res.shape.as_list()
x, y = tf.meshgrid(tf.range(w), tf.range(h))
w_loc = tf.one_hot(x, depth=w)
h_loc = tf.one_hot(y, depth=h)
loc = tf.concat([h_loc, w_loc], 2)
# Batch size cheat
loc = tf.tile(tf.expand_dims(loc, 0), [batch_size, 1, 1, 1])
out = tf.concat([vgg_res, loc], 3)
# Flatten positional
features_h = array_ops.shape(out)[1]
feature_size = out.get_shape().dims[3].value
out = tf.reshape(out, [batch_size, -1, feature_size])
return out, None
def build_train_decoder(decoder_in, encoder_out, state, seq_length, params):
# Declare embedding:
# with tf.device('/cpu:0'):
with tf.variable_scope('embedding', reuse=tf.AUTO_REUSE):
embedding = tf.get_variable('embedding', initializer=emb_initialize(), shape=[params.max_char_vocab, params.char_embedding_dim], dtype=tf.float32)
batch_size = tf.shape(encoder_out)[0]
start_token = tf.ones([batch_size], dtype=tf.int32) * START_TOKEN
# Create training params
# Add start token and update seq length
train_inp = tf.concat([tf.expand_dims(start_token, 1), decoder_in], 1)
max_length = tf.reduce_max((seq_length+1))
# Create embedding lookup for training
train_embedding = tf.nn.embedding_lookup(embedding, train_inp)
train_helper = tf.contrib.seq2seq.TrainingHelper(train_embedding, seq_length+1)
with tf.variable_scope('decode', reuse=tf.AUTO_REUSE):
# Build attention
attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(num_units=params.attention_units, memory=encoder_out)
decoder_cell = tf.nn.rnn_cell.LSTMCell(params.encoder_lstm_hidden*2)
attn_cell = tf.contrib.seq2seq.AttentionWrapper(decoder_cell, attention_mechanism,
attention_layer_size=params.attention_layer_size)
project_cell = tf.contrib.rnn.OutputProjectionWrapper(attn_cell, params.max_char_vocab)
decoder_initial_state = attn_cell.zero_state(dtype=tf.float32, batch_size=batch_size)
if state is not None:
decoder_initial_state = decoder_initial_state.clone(cell_state=state)
decoder = tf.contrib.seq2seq.BasicDecoder(project_cell, train_helper, decoder_initial_state)
train_out, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=decoder, output_time_major=False, impute_finished=True,
maximum_iterations=max_length)
return train_out
def build_infer_decoder(encoder_out, state, params):
# Declare embedding:
with tf.variable_scope('embedding', reuse=tf.AUTO_REUSE):
embedding = tf.get_variable('embedding', initializer=emb_initialize(), shape=[params.max_char_vocab, params.char_embedding_dim], dtype=tf.float32)
batch_size = tf.shape(encoder_out)[0]
start_token = tf.ones([batch_size], dtype=tf.int32) * START_TOKEN
# Create training params
# Add start token and update seq length
pred_helper = tf.contrib.seq2seq.GreedyEmbeddingHelper(embedding, start_tokens=tf.to_int32(start_token), end_token=1)
with tf.variable_scope('decode', reuse=tf.AUTO_REUSE):
attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(num_units=params.attention_units, memory=encoder_out)
decoder_cell = tf.nn.rnn_cell.LSTMCell(params.encoder_lstm_hidden*2)
attn_cell = tf.contrib.seq2seq.AttentionWrapper(decoder_cell, attention_mechanism,
attention_layer_size=params.attention_layer_size)
project_cell = tf.contrib.rnn.OutputProjectionWrapper(attn_cell, params.max_char_vocab)
decoder_initial_state = attn_cell.zero_state(dtype=tf.float32, batch_size=batch_size)
if state is not None:
decoder_initial_state = decoder_initial_state.clone(cell_state=state)
decoder = tf.contrib.seq2seq.BasicDecoder(project_cell, pred_helper, decoder_initial_state)
pred_out, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=decoder, output_time_major=False, impute_finished=True,
maximum_iterations=32)
return pred_out
def model_fn(features, labels, mode, params):
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
if isinstance(features, dict): # For serving
print(features)
images = features['feature']
else:
(images, decoder_in, seq_length) = features
if params.use_vgg == 0:
encoder_out, state = build_encoder(images, is_training, params)
else:
print("Use vgg to build encoder")
encoder_out, state = build_encoder_v1(images, is_training, params)
# Perform main business
if mode == tf.estimator.ModeKeys.PREDICT:
pred_out = build_infer_decoder(encoder_out, state, params)
export_outputs = {'sample_id': tf.estimator.export.PredictOutput(pred_out.sample_id)}
predictions = {'pred_out': pred_out.sample_id}
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions, export_outputs=export_outputs)
else:
if mode == tf.estimator.ModeKeys.TRAIN:
train_out = build_train_decoder(decoder_in, encoder_out, state, seq_length, params)
# Train
global_step = tf.train.get_global_step()
mask = tf.cast(tf.sequence_mask(seq_length+1), tf.float32)
loss = tf.contrib.seq2seq.sequence_loss(train_out.rnn_output, labels, weights=mask,
average_across_timesteps=True, average_across_batch=True)
learning_rate = tf.train.exponential_decay(params.learning_rate, global_step, params.learning_rate_step,
params.learning_rate_decay, staircase=True, name='exponential_learning_rate')
update_global_step = tf.assign(global_step, global_step + 1, name = 'update_global_step')
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
optimizer = tf.train.AdamOptimizer(learning_rate)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss, global_step=global_step)
return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=tf.group(train_op, update_global_step))
else:
pred_out = build_infer_decoder(encoder_out, state, params)
mask = tf.cast(tf.sequence_mask(seq_length+1), tf.float32)
loss = tf.contrib.seq2seq.sequence_loss(pred_out.rnn_output, labels, weights=mask,
average_across_timesteps=True, average_across_batch=True)
return tf.estimator.EstimatorSpec(mode, loss=loss)