exporter.py 19.9 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
# Copyright 2017 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.
# ==============================================================================

"""Functions to export object detection inference graph."""
import os
Vivek Rathod's avatar
Vivek Rathod committed
18
import tempfile
19
import tensorflow as tf
20
from tensorflow.contrib.quantize.python import graph_matcher
21
from tensorflow.core.protobuf import saver_pb2
22
from tensorflow.python.tools import freeze_graph  # pylint: disable=g-direct-tensorflow-import
23
from object_detection.builders import graph_rewriter_builder
24
25
26
from object_detection.builders import model_builder
from object_detection.core import standard_fields as fields
from object_detection.data_decoders import tf_example_decoder
27
from object_detection.utils import config_util
28
from object_detection.utils import shape_utils
29
30
31

slim = tf.contrib.slim

32
freeze_graph_with_def_protos = freeze_graph.freeze_graph_with_def_protos
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
def rewrite_nn_resize_op(is_quantized=False):
  """Replaces a custom nearest-neighbor resize op with the Tensorflow version.

  Some graphs use this custom version for TPU-compatibility.

  Args:
    is_quantized: True if the default graph is quantized.
  """
  input_pattern = graph_matcher.OpTypePattern(
      'FakeQuantWithMinMaxVars' if is_quantized else '*')
  reshape_1_pattern = graph_matcher.OpTypePattern(
      'Reshape', inputs=[input_pattern, 'Const'], ordered_inputs=False)
  mul_pattern = graph_matcher.OpTypePattern(
      'Mul', inputs=[reshape_1_pattern, 'Const'], ordered_inputs=False)
  # The quantization script may or may not insert a fake quant op after the
  # Mul. In either case, these min/max vars are not needed once replaced with
  # the TF version of NN resize.
  fake_quant_pattern = graph_matcher.OpTypePattern(
      'FakeQuantWithMinMaxVars',
      inputs=[mul_pattern, 'Identity', 'Identity'],
      ordered_inputs=False)
  reshape_2_pattern = graph_matcher.OpTypePattern(
      'Reshape',
      inputs=[graph_matcher.OneofPattern([fake_quant_pattern, mul_pattern]),
              'Const'],
      ordered_inputs=False)
  add_pattern = graph_matcher.OpTypePattern(
      'Add', inputs=[reshape_2_pattern, '*'], ordered_inputs=False)

  matcher = graph_matcher.GraphMatcher(add_pattern)
  for match in matcher.match_graph(tf.get_default_graph()):
    projection_op = match.get_op(input_pattern)
    reshape_2_op = match.get_op(reshape_2_pattern)
    add_op = match.get_op(add_pattern)
    nn_resize = tf.image.resize_nearest_neighbor(
        projection_op.outputs[0],
        add_op.outputs[0].shape.dims[1:3],
72
73
        align_corners=False,
        name=os.path.split(reshape_2_op.name)[0] + '/resize_nearest_neighbor')
74
75
76
77
78
79
80

    for index, op_input in enumerate(add_op.inputs):
      if op_input == reshape_2_op.outputs[0]:
        add_op._update_input(index, nn_resize)  # pylint: disable=protected-access
        break


Vivek Rathod's avatar
Vivek Rathod committed
81
82
83
84
def replace_variable_values_with_moving_averages(graph,
                                                 current_checkpoint_file,
                                                 new_checkpoint_file):
  """Replaces variable values in the checkpoint with their moving averages.
85

Vivek Rathod's avatar
Vivek Rathod committed
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
  If the current checkpoint has shadow variables maintaining moving averages of
  the variables defined in the graph, this function generates a new checkpoint
  where the variables contain the values of their moving averages.

  Args:
    graph: a tf.Graph object.
    current_checkpoint_file: a checkpoint containing both original variables and
      their moving averages.
    new_checkpoint_file: file path to write a new checkpoint.
  """
  with graph.as_default():
    variable_averages = tf.train.ExponentialMovingAverage(0.0)
    ema_variables_to_restore = variable_averages.variables_to_restore()
    with tf.Session() as sess:
      read_saver = tf.train.Saver(ema_variables_to_restore)
      read_saver.restore(sess, current_checkpoint_file)
      write_saver = tf.train.Saver()
      write_saver.save(sess, new_checkpoint_file)


def _image_tensor_input_placeholder(input_shape=None):
  """Returns input placeholder and a 4-D uint8 image tensor."""
  if input_shape is None:
    input_shape = (None, None, None, 3)
  input_tensor = tf.placeholder(
      dtype=tf.uint8, shape=input_shape, name='image_tensor')
Derek Chow's avatar
Derek Chow committed
112
  return input_tensor, input_tensor
113

114

115
def _tf_example_input_placeholder():
Derek Chow's avatar
Derek Chow committed
116
117
118
  """Returns input that accepts a batch of strings with tf examples.

  Returns:
Vivek Rathod's avatar
Vivek Rathod committed
119
    a tuple of input placeholder and the output decoded images.
Derek Chow's avatar
Derek Chow committed
120
  """
121
122
123
124
125
126
127
  batch_tf_example_placeholder = tf.placeholder(
      tf.string, shape=[None], name='tf_example')
  def decode(tf_example_string_tensor):
    tensor_dict = tf_example_decoder.TfExampleDecoder().decode(
        tf_example_string_tensor)
    image_tensor = tensor_dict[fields.InputDataFields.image]
    return image_tensor
Derek Chow's avatar
Derek Chow committed
128
  return (batch_tf_example_placeholder,
129
130
131
132
133
134
          shape_utils.static_or_dynamic_map_fn(
              decode,
              elems=batch_tf_example_placeholder,
              dtype=tf.uint8,
              parallel_iterations=32,
              back_prop=False))
135
136


137
def _encoded_image_string_tensor_input_placeholder():
Derek Chow's avatar
Derek Chow committed
138
139
140
  """Returns input that accepts a batch of PNG or JPEG strings.

  Returns:
Vivek Rathod's avatar
Vivek Rathod committed
141
    a tuple of input placeholder and the output decoded images.
Derek Chow's avatar
Derek Chow committed
142
  """
143
144
145
146
147
148
149
150
151
  batch_image_str_placeholder = tf.placeholder(
      dtype=tf.string,
      shape=[None],
      name='encoded_image_string_tensor')
  def decode(encoded_image_string_tensor):
    image_tensor = tf.image.decode_image(encoded_image_string_tensor,
                                         channels=3)
    image_tensor.set_shape((None, None, 3))
    return image_tensor
Derek Chow's avatar
Derek Chow committed
152
153
154
155
156
157
158
  return (batch_image_str_placeholder,
          tf.map_fn(
              decode,
              elems=batch_image_str_placeholder,
              dtype=tf.uint8,
              parallel_iterations=32,
              back_prop=False))
159
160


161
input_placeholder_fn_map = {
162
163
164
    'image_tensor': _image_tensor_input_placeholder,
    'encoded_image_string_tensor':
    _encoded_image_string_tensor_input_placeholder,
165
166
167
168
    'tf_example': _tf_example_input_placeholder,
}


169
170
def add_output_tensor_nodes(postprocessed_tensors,
                            output_collection_name='inference_op'):
171
172
173
174
175
176
177
178
179
180
  """Adds output nodes for detection boxes and scores.

  Adds the following nodes for output tensors -
    * num_detections: float32 tensor of shape [batch_size].
    * detection_boxes: float32 tensor of shape [batch_size, num_boxes, 4]
      containing detected boxes.
    * detection_scores: float32 tensor of shape [batch_size, num_boxes]
      containing scores for the detected boxes.
    * detection_classes: float32 tensor of shape [batch_size, num_boxes]
      containing class predictions for the detected boxes.
181
182
183
    * detection_keypoints: (Optional) float32 tensor of shape
      [batch_size, num_boxes, num_keypoints, 2] containing keypoints for each
      detection box.
184
185
186
    * detection_masks: (Optional) float32 tensor of shape
      [batch_size, num_boxes, mask_height, mask_width] containing masks for each
      detection box.
187
188
189
190
191
192

  Args:
    postprocessed_tensors: a dictionary containing the following fields
      'detection_boxes': [batch, max_detections, 4]
      'detection_scores': [batch, max_detections]
      'detection_classes': [batch, max_detections]
193
194
      'detection_masks': [batch, max_detections, mask_height, mask_width]
        (optional).
195
196
      'detection_keypoints': [batch, max_detections, num_keypoints, 2]
        (optional).
197
      'num_detections': [batch]
198
    output_collection_name: Name of collection to add output tensors to.
199
200
201

  Returns:
    A tensor dict containing the added output tensor nodes.
202
  """
203
  detection_fields = fields.DetectionResultFields
204
  label_id_offset = 1
205
206
  boxes = postprocessed_tensors.get(detection_fields.detection_boxes)
  scores = postprocessed_tensors.get(detection_fields.detection_scores)
207
208
  raw_boxes = postprocessed_tensors.get(detection_fields.raw_detection_boxes)
  raw_scores = postprocessed_tensors.get(detection_fields.raw_detection_scores)
209
210
  classes = postprocessed_tensors.get(
      detection_fields.detection_classes) + label_id_offset
211
  keypoints = postprocessed_tensors.get(detection_fields.detection_keypoints)
212
213
  masks = postprocessed_tensors.get(detection_fields.detection_masks)
  num_detections = postprocessed_tensors.get(detection_fields.num_detections)
214
  outputs = {}
215
216
217
218
219
220
221
222
  outputs[detection_fields.detection_boxes] = tf.identity(
      boxes, name=detection_fields.detection_boxes)
  outputs[detection_fields.detection_scores] = tf.identity(
      scores, name=detection_fields.detection_scores)
  outputs[detection_fields.detection_classes] = tf.identity(
      classes, name=detection_fields.detection_classes)
  outputs[detection_fields.num_detections] = tf.identity(
      num_detections, name=detection_fields.num_detections)
223
224
225
226
227
228
  if raw_boxes is not None:
    outputs[detection_fields.raw_detection_boxes] = tf.identity(
        raw_boxes, name=detection_fields.raw_detection_boxes)
  if raw_scores is not None:
    outputs[detection_fields.raw_detection_scores] = tf.identity(
        raw_scores, name=detection_fields.raw_detection_scores)
229
230
231
  if keypoints is not None:
    outputs[detection_fields.detection_keypoints] = tf.identity(
        keypoints, name=detection_fields.detection_keypoints)
232
  if masks is not None:
233
234
    outputs[detection_fields.detection_masks] = tf.identity(
        masks, name=detection_fields.detection_masks)
235
236
  for output_key in outputs:
    tf.add_to_collection(output_collection_name, outputs[output_key])
237

238
  return outputs
239
240


241
242
243
244
def write_saved_model(saved_model_path,
                      frozen_graph_def,
                      inputs,
                      outputs):
245
246
247
248
249
250
251
252
253
  """Writes SavedModel to disk.

  If checkpoint_path is not None bakes the weights into the graph thereby
  eliminating the need of checkpoint files during inference. If the model
  was trained with moving averages, setting use_moving_averages to true
  restores the moving averages, otherwise the original set of variables
  is restored.

  Args:
254
255
    saved_model_path: Path to write SavedModel.
    frozen_graph_def: tf.GraphDef holding frozen graph.
256
    inputs: The input placeholder tensor.
257
258
259
    outputs: A tensor dictionary containing the outputs of a DetectionModel.
  """
  with tf.Graph().as_default():
260
    with tf.Session() as sess:
261

262
      tf.import_graph_def(frozen_graph_def, name='')
263

264
      builder = tf.saved_model.builder.SavedModelBuilder(saved_model_path)
265
266
267
268
269
270
271
272
273
274
275

      tensor_info_inputs = {
          'inputs': tf.saved_model.utils.build_tensor_info(inputs)}
      tensor_info_outputs = {}
      for k, v in outputs.items():
        tensor_info_outputs[k] = tf.saved_model.utils.build_tensor_info(v)

      detection_signature = (
          tf.saved_model.signature_def_utils.build_signature_def(
              inputs=tensor_info_inputs,
              outputs=tensor_info_outputs,
276
277
              method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME
          ))
278
279

      builder.add_meta_graph_and_variables(
280
281
          sess,
          [tf.saved_model.tag_constants.SERVING],
282
          signature_def_map={
283
284
              tf.saved_model.signature_constants
              .DEFAULT_SERVING_SIGNATURE_DEF_KEY:
285
286
287
288
289
290
                  detection_signature,
          },
      )
      builder.save()


291
292
293
294
295
def write_graph_and_checkpoint(inference_graph_def,
                               model_path,
                               input_saver_def,
                               trained_checkpoint_prefix):
  """Writes the graph and the checkpoint into disk."""
296
297
298
299
  for node in inference_graph_def.node:
    node.device = ''
  with tf.Graph().as_default():
    tf.import_graph_def(inference_graph_def, name='')
300
301
302
    with tf.Session() as sess:
      saver = tf.train.Saver(
          saver_def=input_saver_def, save_relative_paths=True)
303
304
305
306
      saver.restore(sess, trained_checkpoint_prefix)
      saver.save(sess, model_path)


307
308
309
310
311
312
313
314
def _get_outputs_from_inputs(input_tensors, detection_model,
                             output_collection_name):
  inputs = tf.to_float(input_tensors)
  preprocessed_inputs, true_image_shapes = detection_model.preprocess(inputs)
  output_tensors = detection_model.predict(
      preprocessed_inputs, true_image_shapes)
  postprocessed_tensors = detection_model.postprocess(
      output_tensors, true_image_shapes)
315
316
  return add_output_tensor_nodes(postprocessed_tensors,
                                 output_collection_name)
317
318


319
320
def build_detection_graph(input_type, detection_model, input_shape,
                          output_collection_name, graph_hook_fn):
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
  """Build the detection graph."""
  if input_type not in input_placeholder_fn_map:
    raise ValueError('Unknown input type: {}'.format(input_type))
  placeholder_args = {}
  if input_shape is not None:
    if input_type != 'image_tensor':
      raise ValueError('Can only specify input shape for `image_tensor` '
                       'inputs.')
    placeholder_args['input_shape'] = input_shape
  placeholder_tensor, input_tensors = input_placeholder_fn_map[input_type](
      **placeholder_args)
  outputs = _get_outputs_from_inputs(
      input_tensors=input_tensors,
      detection_model=detection_model,
      output_collection_name=output_collection_name)

  # Add global step to the graph.
  slim.get_or_create_global_step()

  if graph_hook_fn: graph_hook_fn()

  return outputs, placeholder_tensor


345
346
347
def _export_inference_graph(input_type,
                            detection_model,
                            use_moving_averages,
348
349
                            trained_checkpoint_prefix,
                            output_directory,
Vivek Rathod's avatar
Vivek Rathod committed
350
351
                            additional_output_tensor_names=None,
                            input_shape=None,
352
                            output_collection_name='inference_op',
353
                            graph_hook_fn=None,
354
355
                            write_inference_graph=False,
                            temp_checkpoint_prefix=''):
356
  """Export helper."""
357
358
359
360
361
362
  tf.gfile.MakeDirs(output_directory)
  frozen_graph_path = os.path.join(output_directory,
                                   'frozen_inference_graph.pb')
  saved_model_path = os.path.join(output_directory, 'saved_model')
  model_path = os.path.join(output_directory, 'model.ckpt')

363
  outputs, placeholder_tensor = build_detection_graph(
364
365
366
367
368
      input_type=input_type,
      detection_model=detection_model,
      input_shape=input_shape,
      output_collection_name=output_collection_name,
      graph_hook_fn=graph_hook_fn)
369

370
  profile_inference_graph(tf.get_default_graph())
371
  saver_kwargs = {}
372
  if use_moving_averages:
373
374
375
376
377
378
379
    if not temp_checkpoint_prefix:
      # This check is to be compatible with both version of SaverDef.
      if os.path.isfile(trained_checkpoint_prefix):
        saver_kwargs['write_version'] = saver_pb2.SaverDef.V1
        temp_checkpoint_prefix = tempfile.NamedTemporaryFile().name
      else:
        temp_checkpoint_prefix = tempfile.mkdtemp()
Vivek Rathod's avatar
Vivek Rathod committed
380
381
    replace_variable_values_with_moving_averages(
        tf.get_default_graph(), trained_checkpoint_prefix,
382
383
        temp_checkpoint_prefix)
    checkpoint_to_use = temp_checkpoint_prefix
384
  else:
Vivek Rathod's avatar
Vivek Rathod committed
385
386
    checkpoint_to_use = trained_checkpoint_prefix

387
  saver = tf.train.Saver(**saver_kwargs)
388
389
  input_saver_def = saver.as_saver_def()

390
  write_graph_and_checkpoint(
391
392
393
      inference_graph_def=tf.get_default_graph().as_graph_def(),
      model_path=model_path,
      input_saver_def=input_saver_def,
Vivek Rathod's avatar
Vivek Rathod committed
394
      trained_checkpoint_prefix=checkpoint_to_use)
395
396
397
398
399
400
  if write_inference_graph:
    inference_graph_def = tf.get_default_graph().as_graph_def()
    inference_graph_path = os.path.join(output_directory,
                                        'inference_graph.pbtxt')
    for node in inference_graph_def.node:
      node.device = ''
401
    with tf.gfile.GFile(inference_graph_path, 'wb') as f:
402
      f.write(str(inference_graph_def))
Vivek Rathod's avatar
Vivek Rathod committed
403
404
405
406
407

  if additional_output_tensor_names is not None:
    output_node_names = ','.join(outputs.keys()+additional_output_tensor_names)
  else:
    output_node_names = ','.join(outputs.keys())
408

409
  frozen_graph_def = freeze_graph.freeze_graph_with_def_protos(
410
411
      input_graph_def=tf.get_default_graph().as_graph_def(),
      input_saver_def=input_saver_def,
Vivek Rathod's avatar
Vivek Rathod committed
412
413
      input_checkpoint=checkpoint_to_use,
      output_node_names=output_node_names,
414
415
      restore_op_name='save/restore_all',
      filename_tensor_name='save/Const:0',
416
      output_graph=frozen_graph_path,
417
418
      clear_devices=True,
      initializer_nodes='')
419

420
421
  write_saved_model(saved_model_path, frozen_graph_def,
                    placeholder_tensor, outputs)
422
423


424
425
426
427
def export_inference_graph(input_type,
                           pipeline_config,
                           trained_checkpoint_prefix,
                           output_directory,
Vivek Rathod's avatar
Vivek Rathod committed
428
429
                           input_shape=None,
                           output_collection_name='inference_op',
430
431
                           additional_output_tensor_names=None,
                           write_inference_graph=False):
432
433
434
  """Exports inference graph for the model specified in the pipeline config.

  Args:
435
436
    input_type: Type of input for the graph. Can be one of ['image_tensor',
      'encoded_image_string_tensor', 'tf_example'].
437
    pipeline_config: pipeline_pb2.TrainAndEvalPipelineConfig proto.
438
439
    trained_checkpoint_prefix: Path to the trained checkpoint file.
    output_directory: Path to write outputs.
Vivek Rathod's avatar
Vivek Rathod committed
440
441
    input_shape: Sets a fixed shape for an `image_tensor` input. If not
      specified, will default to [None, None, None, 3].
442
443
    output_collection_name: Name of collection to add output tensors to.
      If None, does not add output tensors to a collection.
Vivek Rathod's avatar
Vivek Rathod committed
444
    additional_output_tensor_names: list of additional output
445
      tensors to include in the frozen graph.
446
    write_inference_graph: If true, writes inference graph to disk.
447
448
449
  """
  detection_model = model_builder.build(pipeline_config.model,
                                        is_training=False)
450
451
452
453
454
  graph_rewriter_fn = None
  if pipeline_config.HasField('graph_rewriter'):
    graph_rewriter_config = pipeline_config.graph_rewriter
    graph_rewriter_fn = graph_rewriter_builder.build(graph_rewriter_config,
                                                     is_training=False)
455
456
457
458
459
460
461
462
463
  _export_inference_graph(
      input_type,
      detection_model,
      pipeline_config.eval_config.use_moving_averages,
      trained_checkpoint_prefix,
      output_directory,
      additional_output_tensor_names,
      input_shape,
      output_collection_name,
464
      graph_hook_fn=graph_rewriter_fn,
465
      write_inference_graph=write_inference_graph)
466
  pipeline_config.eval_config.use_moving_averages = False
467
  config_util.save_pipeline_config(pipeline_config, output_directory)
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498


def profile_inference_graph(graph):
  """Profiles the inference graph.

  Prints model parameters and computation FLOPs given an inference graph.
  BatchNorms are excluded from the parameter count due to the fact that
  BatchNorms are usually folded. BatchNorm, Initializer, Regularizer
  and BiasAdd are not considered in FLOP count.

  Args:
    graph: the inference graph.
  """
  tfprof_vars_option = (
      tf.contrib.tfprof.model_analyzer.TRAINABLE_VARS_PARAMS_STAT_OPTIONS)
  tfprof_flops_option = tf.contrib.tfprof.model_analyzer.FLOAT_OPS_OPTIONS

  # Batchnorm is usually folded during inference.
  tfprof_vars_option['trim_name_regexes'] = ['.*BatchNorm.*']
  # Initializer and Regularizer are only used in training.
  tfprof_flops_option['trim_name_regexes'] = [
      '.*BatchNorm.*', '.*Initializer.*', '.*Regularizer.*', '.*BiasAdd.*'
  ]

  tf.contrib.tfprof.model_analyzer.print_model_analysis(
      graph,
      tfprof_options=tfprof_vars_option)

  tf.contrib.tfprof.model_analyzer.print_model_analysis(
      graph,
      tfprof_options=tfprof_flops_option)