/
tf_hub.py
327 lines (277 loc) · 12.5 KB
/
tf_hub.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
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
r"""Provide functionalities to export and eval tf_hub module.
Example use of export_to_hub:
tf_hub.py --tf_hub_mode='export_to_hub' --cell_name=amoeba_net_a \
--reduction_size=448 --num_cells=18 --image_size=331 \
--drop_connect_keep_prob=0.7 \
--export_path=/tmp/module_export \
--model_dir=/ADD_PATH_WITH_1001_CLASSES_HERE \
--alsologtostderr
Example use of eval_from_hub
tf_hub.py --tf_hub_mode='eval_from_hub' --cell_name=amoeba_net_a \
--reduction_size=448 --num_cells=18 --image_size=331 \
--export_path=/tmp/module_export \
--data_dir=/ADD_DATA_PATH_HERE \
--model_dir=/ADD_PATH_WITH_1001_CLASSES_HERE \
--eval_batch_size=40 --alsologtostderr
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import re
# Standard Imports
from absl import app
from absl import flags
import tensorflow.compat.v1 as tf
from tensorflow.compat.v1 import estimator as tf_estimator
import tensorflow_hub as hub
import amoeba_net
import amoeba_net_model as model_lib
import model_builder
from tensorflow.contrib import slim
flags.DEFINE_string('tf_hub_mode', 'export_to_hub',
'export_to_hub|eval_from_hub')
flags.DEFINE_string('export_path', None,
'Directory where export output is stored')
flags.DEFINE_bool(
'export_feature_vector', False,
'If true, network builder only returns feature vector after global_pool '
'without the fully connected layer.')
flags.DEFINE_bool(
'dryrun_with_untrained_weights', None,
'FOR TESTING USE ONLY. If set, export_to_hub is done without restoring '
'the model\'s trained weights. This helps test the Python code quickly but '
'makes the resulting module useless.')
FLAGS = flags.FLAGS
def _make_module_fn(hparams, num_classes):
"""Returns a module_fn for use with hub.create_module_spec()."""
def _module_fn(is_training):
"""A module_fn for use with hub.create_module_spec().
Args:
is_training: a boolean, passed to the config.network_fn.
This is meant to control whether batch norm, dropout etc. are built
in training or inference mode for this graph version.
Raises:
ValueError: if network_fn outputs are not as expected.
"""
# Set up the module input, and attach an ImageModuleInfo about it.
with tf.name_scope('hub_input'):
default_size = (hparams.image_size,) * 2
image_module_info = hub.ImageModuleInfo()
size_info = image_module_info.default_image_size
size_info.height, size_info.width = default_size
# TODO(b/72731449): Support variable input size.
shape = (None,) + default_size + (3,)
images = tf.placeholder(dtype=tf.float32, shape=shape, name='images')
hub.attach_image_module_info(image_module_info)
# The input is expected to have RGB color values in the range [0,1]
# and gets converted for AmoebaNet to the Inception-style range [-1,+1].
scaled_images = tf.multiply(images, 2.0)
scaled_images = tf.subtract(scaled_images, 1.0)
# Build the net.
logits, end_points = model_builder.build_network(scaled_images, num_classes,
is_training, hparams)
with tf.name_scope('hub_output'):
# Extract the feature_vectors output.
try:
feature_vectors = end_points['global_pool']
except KeyError:
tf.logging.error('Valid keys of end_points are:', ', '.join(end_points))
raise
with tf.name_scope('feature_vector'):
if feature_vectors.shape.ndims != 2:
raise ValueError(
'Wrong rank (expected 2 after squeeze) '
'in feature_vectors:', feature_vectors)
# Extract the logits output (if applicable).
if num_classes:
with tf.name_scope('classification'):
if logits.shape.ndims != 2:
raise ValueError('Wrong rank (expected 2) in logits:', logits)
# Add named signatures.
hub.add_signature('image_feature_vector', dict(images=images),
dict(end_points, default=feature_vectors))
if num_classes:
hub.add_signature('image_classification', dict(images=images),
dict(end_points, default=logits))
# Add the default signature.
if num_classes:
hub.add_signature('default', dict(images=images), dict(default=logits))
else:
hub.add_signature('default', dict(images=images),
dict(default=feature_vectors))
return _module_fn
def export_to_hub(checkpoint_path, export_path, num_classes, hparams):
"""Exports the network as a TF-Hub Module.
If a positive integer num_classes is given, a module for image classification
is exported. If num_classes is 0 or None, a module for feature vector
extraction is exported. In both cases, the default signature returns
a default output that matches the Python slim API net, _ = network_fn(...).
Args:
checkpoint_path: a string with the file name of the checkpoint from which
the trained weights are copied into the Module.
FOR TESTING USE ONLY, this can be set to empty or None, to skip
restoring weights, which ignores the checkpoint and copies the random
initializer values of the weights instead.
export_path: a string with the directory to pass to hub.Module.export().
num_classes: an integer with the number of classes for which the given
checkpoint has been trained. If 0 or None, the classification layer
is omitted.
hparams: hyper parameters.
"""
module_fn = _make_module_fn(hparams, num_classes)
tags_and_args = [
# The default graph is built with batch_norm, dropout etc. in inference
# mode. This graph version is good for inference, not training.
([], {
'is_training': False
}),
# A separate 'train' graph builds batch_norm, dropout etc. in training
# mode.
(['train'], {
'is_training': True
}),
]
drop_collections = [
'moving_vars', tf.GraphKeys.GLOBAL_STEP,
tf.GraphKeys.MOVING_AVERAGE_VARIABLES
]
spec = hub.create_module_spec(module_fn, tags_and_args, drop_collections)
with tf.Graph().as_default():
module = hub.Module(spec)
init_fn = _get_init_fn(
module,
checkpoint_path,
hparams.moving_average_decay > 0,
moving_averages_blacklist_regex='global_step')
with tf.Session() as session:
init_fn(session)
module.export(export_path, session=session)
tf.logging.info('Export to {} succeeded.'.format(export_path))
def _get_init_fn(module,
checkpoint_path,
export_moving_averages=False,
moving_averages_blacklist_regex=None):
"""Returns init_fn for the session that calls hub.Module.export()."""
if not checkpoint_path:
tf.logging.warn('DRYRUN: using random weight initializers, no checkpoint')
return lambda session: session.run(tf.global_variables_initializer())
# Build `variables_to_restore` as a map from names in the checkpoint to the
# variable in the instantiated module.
if export_moving_averages:
variables_to_restore = {}
num_averaged = num_blacklisted = 0
for variable_name, variable in module.variable_map.items():
if (moving_averages_blacklist_regex and
re.match(moving_averages_blacklist_regex, variable_name)):
num_blacklisted += 1
else:
variable_name += '/ExponentialMovingAverage'
num_averaged += 1
variables_to_restore[variable_name] = variable
tf.logging.info('Export of moving averages is applied to %d variables '
'with %d exempted by matching the blacklist_regex' %
(num_averaged, num_blacklisted))
else:
variables_to_restore = module.variable_map
tf.logging.info('Export of moving averages is disabled')
unchecked_init_fn = slim.assign_from_checkpoint_fn(checkpoint_path,
variables_to_restore)
def init_fn(session):
unchecked_init_fn(session)
_check_shapes_of_restored_variables(session, variables_to_restore)
return init_fn
def _check_shapes_of_restored_variables(session, variables_to_restore):
"""Raises TypeError if restored variables have unexpected shapes."""
num_errors = 0
for variable_name, variable in variables_to_restore.items():
graph_shape = variable.value().shape
# Values are big, but tf.shape(..) whould echo graph_shape if fully defined.
checkpoint_shape = session.run(variable.value()).shape
if not graph_shape.is_compatible_with(checkpoint_shape):
tf.logging.error('Shape mismatch for variable %s: '
'graph expects %s but checkpoint has %s' %
(variable_name, graph_shape, checkpoint_shape))
num_errors += 1
if num_errors:
raise TypeError(
'Shape mismatch for %d variables, see error log for list.' % num_errors)
def _make_model_fn(hub_module_spec):
"""Returns a model_fn for estimator using hub_module."""
def _model_fn(features, labels, mode, params):
"""model_fn for estimator."""
del params
features = tf.transpose(features, [3, 0, 1, 2]) # HWCN to NHWC
hub_module = hub.Module(spec=hub_module_spec, trainable=False)
logits = hub_module(features)
labels_onehot = tf.one_hot(labels, logits.shape[1])
loss = tf.losses.softmax_cross_entropy(labels_onehot, logits)
eval_metric_ops = None
def metric_fn(labels, logits):
"""Evaluation metric fn. Performed on CPU, do not reference TPU ops."""
predictions = tf.argmax(logits, axis=1)
top_1_accuracy = tf.metrics.accuracy(labels, predictions)
in_top_5 = tf.cast(tf.nn.in_top_k(logits, labels, 5), tf.float32)
top_5_accuracy = tf.metrics.mean(in_top_5)
return {
'top_1_accuracy': top_1_accuracy,
'top_5_accuracy': top_5_accuracy,
}
eval_metric_ops = metric_fn(labels, logits)
return tf_estimator.EstimatorSpec(
mode=mode, loss=loss, train_op=None, eval_metric_ops=eval_metric_ops)
return _model_fn
def eval_from_hub(model_dir, input_fn, eval_steps):
"""Eval using hub module."""
hub_module_spec = hub.load_module_spec(model_dir)
run_config = tf_estimator.RunConfig(model_dir=model_dir)
image_classifier = tf_estimator.Estimator(
model_fn=_make_model_fn(hub_module_spec), config=run_config, params={})
eval_results = image_classifier.evaluate(input_fn=input_fn, steps=eval_steps)
tf.logging.info('Evaluation results: %s' % eval_results)
def main(_):
mode = FLAGS.tf_hub_mode
data_dir = amoeba_net.FLAGS.data_dir
model_dir = amoeba_net.FLAGS.model_dir
hparams = amoeba_net.build_hparams()
hparams.add_hparam('drop_path_burn_in_steps', 0)
hparams.set_hparam('use_tpu', False)
# So far, there is no standardized way of exposing aux heads for
# fine-tuning Hub image modules. Disable aux heads to avoid putting unused
# variables and ops into the module.
hparams.set_hparam('use_aux_head', False)
eval_steps = FLAGS.num_eval_images // FLAGS.eval_batch_size
export_path = FLAGS.export_path or (model_dir + '/export')
input_pipeline = model_lib.InputPipeline(
is_training=False, data_dir=data_dir, hparams=hparams, eval_from_hub=True)
if mode == 'eval_from_hub':
eval_from_hub(export_path, input_pipeline.input_fn, eval_steps=eval_steps)
elif mode == 'export_to_hub':
num_classes = (None if FLAGS.export_feature_vector else
input_pipeline.num_classes)
if FLAGS.dryrun_with_untrained_weights:
checkpoint_path = None
else:
checkpoint_path = tf.train.latest_checkpoint(model_dir)
if not checkpoint_path:
raise IOError('No checkpoint found.')
export_to_hub(
checkpoint_path, export_path, num_classes, hparams)
else:
raise ValueError('Unsupported tf_hub_mode = {}'.format(mode))
if __name__ == '__main__':
tf.logging.set_verbosity(tf.logging.INFO)
app.run(main)