model.py 36.2 KB
Newer Older
1
# Lint as: python2, python3
yukun's avatar
yukun committed
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
# 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"""Provides DeepLab model definition and helper functions.

DeepLab is a deep learning system for semantic image segmentation with
the following features:

(1) Atrous convolution to explicitly control the resolution at which
feature responses are computed within Deep Convolutional Neural Networks.

(2) Atrous spatial pyramid pooling (ASPP) to robustly segment objects at
multiple scales with filters at multiple sampling rates and effective
fields-of-views.

(3) ASPP module augmented with image-level feature and batch normalization.

(4) A simple yet effective decoder module to recover the object boundaries.

See the following papers for more details:

"Encoder-Decoder with Atrous Separable Convolution for Semantic Image
Segmentation"
Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, Hartwig Adam.
hsm207's avatar
hsm207 committed
37
(https://arxiv.org/abs/1802.02611)
yukun's avatar
yukun committed
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55

"Rethinking Atrous Convolution for Semantic Image Segmentation,"
Liang-Chieh Chen, George Papandreou, Florian Schroff, Hartwig Adam
(https://arxiv.org/abs/1706.05587)

"DeepLab: Semantic Image Segmentation with Deep Convolutional Nets,
Atrous Convolution, and Fully Connected CRFs",
Liang-Chieh Chen*, George Papandreou*, Iasonas Kokkinos, Kevin Murphy,
Alan L Yuille (* equal contribution)
(https://arxiv.org/abs/1606.00915)

"Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected
CRFs"
Liang-Chieh Chen*, George Papandreou*, Iasonas Kokkinos, Kevin Murphy,
Alan L. Yuille (* equal contribution)
(https://arxiv.org/abs/1412.7062)
"""
import tensorflow as tf
56
from tensorflow.contrib import slim as contrib_slim
57
from deeplab.core import dense_prediction_cell
yukun's avatar
yukun committed
58
from deeplab.core import feature_extractor
59
60
from deeplab.core import utils

61
slim = contrib_slim
yukun's avatar
yukun committed
62

63
64
65
66
67
68
LOGITS_SCOPE_NAME = 'logits'
MERGED_LOGITS_SCOPE = 'merged_logits'
IMAGE_POOLING_SCOPE = 'image_pooling'
ASPP_SCOPE = 'aspp'
CONCAT_PROJECTION_SCOPE = 'concat_projection'
DECODER_SCOPE = 'decoder'
69
META_ARCHITECTURE_SCOPE = 'meta_architecture'
70

71
72
PROB_SUFFIX = '_prob'

73
_resize_bilinear = utils.resize_bilinear
74
75
scale_dimension = utils.scale_dimension
split_separable_conv2d = utils.split_separable_conv2d
76

77

78
def get_extra_layer_scopes(last_layers_contain_logits_only=False):
yukun's avatar
yukun committed
79
80
  """Gets the scopes for extra layers.

81
82
83
84
  Args:
    last_layers_contain_logits_only: Boolean, True if only consider logits as
    the last layer (i.e., exclude ASPP module, decoder module and so on)

yukun's avatar
yukun committed
85
86
87
  Returns:
    A list of scopes for extra layers.
  """
88
  if last_layers_contain_logits_only:
89
    return [LOGITS_SCOPE_NAME]
90
91
  else:
    return [
92
93
94
95
96
        LOGITS_SCOPE_NAME,
        IMAGE_POOLING_SCOPE,
        ASPP_SCOPE,
        CONCAT_PROJECTION_SCOPE,
        DECODER_SCOPE,
97
        META_ARCHITECTURE_SCOPE,
98
    ]
yukun's avatar
yukun committed
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


def predict_labels_multi_scale(images,
                               model_options,
                               eval_scales=(1.0,),
                               add_flipped_images=False):
  """Predicts segmentation labels.

  Args:
    images: A tensor of size [batch, height, width, channels].
    model_options: A ModelOptions instance to configure models.
    eval_scales: The scales to resize images for evaluation.
    add_flipped_images: Add flipped images for evaluation or not.

  Returns:
    A dictionary with keys specifying the output_type (e.g., semantic
      prediction) and values storing Tensors representing predictions (argmax
      over channels). Each prediction has size [batch, height, width].
  """
  outputs_to_predictions = {
      output: []
      for output in model_options.outputs_to_num_classes
  }

  for i, image_scale in enumerate(eval_scales):
    with tf.variable_scope(tf.get_variable_scope(), reuse=True if i else None):
      outputs_to_scales_to_logits = multi_scale_logits(
          images,
          model_options=model_options,
          image_pyramid=[image_scale],
          is_training=False,
          fine_tune_batch_norm=False)

    if add_flipped_images:
      with tf.variable_scope(tf.get_variable_scope(), reuse=True):
        outputs_to_scales_to_logits_reversed = multi_scale_logits(
            tf.reverse_v2(images, [2]),
            model_options=model_options,
            image_pyramid=[image_scale],
            is_training=False,
            fine_tune_batch_norm=False)

    for output in sorted(outputs_to_scales_to_logits):
      scales_to_logits = outputs_to_scales_to_logits[output]
143
      logits = _resize_bilinear(
144
          scales_to_logits[MERGED_LOGITS_SCOPE],
yukun's avatar
yukun committed
145
          tf.shape(images)[1:3],
146
          scales_to_logits[MERGED_LOGITS_SCOPE].dtype)
yukun's avatar
yukun committed
147
148
149
150
151
152
      outputs_to_predictions[output].append(
          tf.expand_dims(tf.nn.softmax(logits), 4))

      if add_flipped_images:
        scales_to_logits_reversed = (
            outputs_to_scales_to_logits_reversed[output])
153
        logits_reversed = _resize_bilinear(
154
            tf.reverse_v2(scales_to_logits_reversed[MERGED_LOGITS_SCOPE], [2]),
yukun's avatar
yukun committed
155
            tf.shape(images)[1:3],
156
            scales_to_logits_reversed[MERGED_LOGITS_SCOPE].dtype)
yukun's avatar
yukun committed
157
158
159
160
161
162
163
164
        outputs_to_predictions[output].append(
            tf.expand_dims(tf.nn.softmax(logits_reversed), 4))

  for output in sorted(outputs_to_predictions):
    predictions = outputs_to_predictions[output]
    # Compute average prediction across different scales and flipped images.
    predictions = tf.reduce_mean(tf.concat(predictions, 4), axis=4)
    outputs_to_predictions[output] = tf.argmax(predictions, 3)
165
    outputs_to_predictions[output + PROB_SUFFIX] = tf.nn.softmax(predictions)
yukun's avatar
yukun committed
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

  return outputs_to_predictions


def predict_labels(images, model_options, image_pyramid=None):
  """Predicts segmentation labels.

  Args:
    images: A tensor of size [batch, height, width, channels].
    model_options: A ModelOptions instance to configure models.
    image_pyramid: Input image scales for multi-scale feature extraction.

  Returns:
    A dictionary with keys specifying the output_type (e.g., semantic
      prediction) and values storing Tensors representing predictions (argmax
      over channels). Each prediction has size [batch, height, width].
  """
  outputs_to_scales_to_logits = multi_scale_logits(
      images,
      model_options=model_options,
      image_pyramid=image_pyramid,
      is_training=False,
      fine_tune_batch_norm=False)

  predictions = {}
  for output in sorted(outputs_to_scales_to_logits):
    scales_to_logits = outputs_to_scales_to_logits[output]
193
194
195
196
197
198
199
200
201
202
    logits = scales_to_logits[MERGED_LOGITS_SCOPE]
    # There are two ways to obtain the final prediction results: (1) bilinear
    # upsampling the logits followed by argmax, or (2) argmax followed by
    # nearest neighbor upsampling. The second option may introduce the "blocking
    # effect" but is computationally efficient.
    if model_options.prediction_with_upsampled_logits:
      logits = _resize_bilinear(logits,
                                tf.shape(images)[1:3],
                                scales_to_logits[MERGED_LOGITS_SCOPE].dtype)
      predictions[output] = tf.argmax(logits, 3)
203
      predictions[output + PROB_SUFFIX] = tf.nn.softmax(logits)
204
205
206
207
208
209
210
211
    else:
      argmax_results = tf.argmax(logits, 3)
      argmax_results = tf.image.resize_nearest_neighbor(
          tf.expand_dims(argmax_results, 3),
          tf.shape(images)[1:3],
          align_corners=True,
          name='resize_prediction')
      predictions[output] = tf.squeeze(argmax_results, 3)
212
213
214
215
216
      predictions[output + PROB_SUFFIX] = tf.image.resize_bilinear(
          tf.nn.softmax(logits),
          tf.shape(images)[1:3],
          align_corners=True,
          name='resize_prob')
yukun's avatar
yukun committed
217
218
219
220
221
222
223
224
  return predictions


def multi_scale_logits(images,
                       model_options,
                       image_pyramid,
                       weight_decay=0.0001,
                       is_training=False,
225
226
                       fine_tune_batch_norm=False,
                       nas_training_hyper_parameters=None):
yukun's avatar
yukun committed
227
228
229
230
231
232
233
234
235
236
237
238
  """Gets the logits for multi-scale inputs.

  The returned logits are all downsampled (due to max-pooling layers)
  for both training and evaluation.

  Args:
    images: A tensor of size [batch, height, width, channels].
    model_options: A ModelOptions instance to configure models.
    image_pyramid: Input image scales for multi-scale feature extraction.
    weight_decay: The weight decay for model variables.
    is_training: Is training or not.
    fine_tune_batch_norm: Fine-tune the batch norm parameters or not.
239
240
241
242
243
244
    nas_training_hyper_parameters: A dictionary storing hyper-parameters for
      training nas models. Its keys are:
      - `drop_path_keep_prob`: Probability to keep each path in the cell when
        training.
      - `total_training_steps`: Total training steps to help drop path
        probability calculation.
yukun's avatar
yukun committed
245
246
247
248
249
250
251
252
253
254
255
256

  Returns:
    outputs_to_scales_to_logits: A map of maps from output_type (e.g.,
      semantic prediction) to a dictionary of multi-scale logits names to
      logits. For each output_type, the dictionary has keys which
      correspond to the scales and values which correspond to the logits.
      For example, if `scales` equals [1.0, 1.5], then the keys would
      include 'merged_logits', 'logits_1.00' and 'logits_1.50'.

  Raises:
    ValueError: If model_options doesn't specify crop_size and its
      add_image_level_feature = True, since add_image_level_feature requires
257
      crop_size information.
yukun's avatar
yukun committed
258
259
260
261
262
263
264
265
266
267
  """
  # Setup default values.
  if not image_pyramid:
    image_pyramid = [1.0]
  crop_height = (
      model_options.crop_size[0]
      if model_options.crop_size else tf.shape(images)[1])
  crop_width = (
      model_options.crop_size[1]
      if model_options.crop_size else tf.shape(images)[2])
268
269
270
  if model_options.image_pooling_crop_size:
    image_pooling_crop_height = model_options.image_pooling_crop_size[0]
    image_pooling_crop_width = model_options.image_pooling_crop_size[1]
yukun's avatar
yukun committed
271
272

  # Compute the height, width for the output logits.
273
274
275
276
  if model_options.decoder_output_stride:
    logits_output_stride = min(model_options.decoder_output_stride)
  else:
    logits_output_stride = model_options.output_stride
yukun's avatar
yukun committed
277
278
279
280
281
282
283
284
285
286
287
288
289
290

  logits_height = scale_dimension(
      crop_height,
      max(1.0, max(image_pyramid)) / logits_output_stride)
  logits_width = scale_dimension(
      crop_width,
      max(1.0, max(image_pyramid)) / logits_output_stride)

  # Compute the logits for each scale in the image pyramid.
  outputs_to_scales_to_logits = {
      k: {}
      for k in model_options.outputs_to_num_classes
  }

291
292
  num_channels = images.get_shape().as_list()[-1]

293
  for image_scale in image_pyramid:
yukun's avatar
yukun committed
294
295
296
297
    if image_scale != 1.0:
      scaled_height = scale_dimension(crop_height, image_scale)
      scaled_width = scale_dimension(crop_width, image_scale)
      scaled_crop_size = [scaled_height, scaled_width]
298
      scaled_images = _resize_bilinear(images, scaled_crop_size, images.dtype)
yukun's avatar
yukun committed
299
      if model_options.crop_size:
300
301
302
303
304
305
306
307
        scaled_images.set_shape(
            [None, scaled_height, scaled_width, num_channels])
      # Adjust image_pooling_crop_size accordingly.
      scaled_image_pooling_crop_size = None
      if model_options.image_pooling_crop_size:
        scaled_image_pooling_crop_size = [
            scale_dimension(image_pooling_crop_height, image_scale),
            scale_dimension(image_pooling_crop_width, image_scale)]
yukun's avatar
yukun committed
308
309
310
    else:
      scaled_crop_size = model_options.crop_size
      scaled_images = images
311
      scaled_image_pooling_crop_size = model_options.image_pooling_crop_size
yukun's avatar
yukun committed
312

313
314
315
    updated_options = model_options._replace(
        crop_size=scaled_crop_size,
        image_pooling_crop_size=scaled_image_pooling_crop_size)
yukun's avatar
yukun committed
316
317
318
319
    outputs_to_logits = _get_logits(
        scaled_images,
        updated_options,
        weight_decay=weight_decay,
320
        reuse=tf.AUTO_REUSE,
yukun's avatar
yukun committed
321
        is_training=is_training,
322
323
        fine_tune_batch_norm=fine_tune_batch_norm,
        nas_training_hyper_parameters=nas_training_hyper_parameters)
yukun's avatar
yukun committed
324
325
326

    # Resize the logits to have the same dimension before merging.
    for output in sorted(outputs_to_logits):
327
      outputs_to_logits[output] = _resize_bilinear(
yukun's avatar
yukun committed
328
          outputs_to_logits[output], [logits_height, logits_width],
329
          outputs_to_logits[output].dtype)
yukun's avatar
yukun committed
330
331
332
333
334

    # Return when only one input scale.
    if len(image_pyramid) == 1:
      for output in sorted(model_options.outputs_to_num_classes):
        outputs_to_scales_to_logits[output][
335
            MERGED_LOGITS_SCOPE] = outputs_to_logits[output]
yukun's avatar
yukun committed
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
      return outputs_to_scales_to_logits

    # Save logits to the output map.
    for output in sorted(model_options.outputs_to_num_classes):
      outputs_to_scales_to_logits[output][
          'logits_%.2f' % image_scale] = outputs_to_logits[output]

  # Merge the logits from all the multi-scale inputs.
  for output in sorted(model_options.outputs_to_num_classes):
    # Concatenate the multi-scale logits for each output type.
    all_logits = [
        tf.expand_dims(logits, axis=4)
        for logits in outputs_to_scales_to_logits[output].values()
    ]
    all_logits = tf.concat(all_logits, 4)
    merge_fn = (
        tf.reduce_max
        if model_options.merge_method == 'max' else tf.reduce_mean)
354
    outputs_to_scales_to_logits[output][MERGED_LOGITS_SCOPE] = merge_fn(
yukun's avatar
yukun committed
355
356
357
358
359
        all_logits, axis=4)

  return outputs_to_scales_to_logits


360
361
362
363
364
def extract_features(images,
                     model_options,
                     weight_decay=0.0001,
                     reuse=None,
                     is_training=False,
365
366
                     fine_tune_batch_norm=False,
                     nas_training_hyper_parameters=None):
yukun's avatar
yukun committed
367
368
369
370
371
372
373
374
375
  """Extracts features by the particular model_variant.

  Args:
    images: A tensor of size [batch, height, width, channels].
    model_options: A ModelOptions instance to configure models.
    weight_decay: The weight decay for model variables.
    reuse: Reuse the model variables or not.
    is_training: Is training or not.
    fine_tune_batch_norm: Fine-tune the batch norm parameters or not.
376
377
378
379
380
381
    nas_training_hyper_parameters: A dictionary storing hyper-parameters for
      training nas models. Its keys are:
      - `drop_path_keep_prob`: Probability to keep each path in the cell when
        training.
      - `total_training_steps`: Total training steps to help drop path
        probability calculation.
yukun's avatar
yukun committed
382
383
384
385
386
387
388
389
390
391
392
393
394

  Returns:
    concat_logits: A tensor of size [batch, feature_height, feature_width,
      feature_channels], where feature_height/feature_width are determined by
      the images height/width and output_stride.
    end_points: A dictionary from components of the network to the corresponding
      activation.
  """
  features, end_points = feature_extractor.extract_features(
      images,
      output_stride=model_options.output_stride,
      multi_grid=model_options.multi_grid,
      model_variant=model_options.model_variant,
395
      depth_multiplier=model_options.depth_multiplier,
396
      divisible_by=model_options.divisible_by,
yukun's avatar
yukun committed
397
398
399
      weight_decay=weight_decay,
      reuse=reuse,
      is_training=is_training,
400
401
      preprocessed_images_dtype=model_options.preprocessed_images_dtype,
      fine_tune_batch_norm=fine_tune_batch_norm,
402
      nas_architecture_options=model_options.nas_architecture_options,
403
404
      nas_training_hyper_parameters=nas_training_hyper_parameters,
      use_bounded_activation=model_options.use_bounded_activation)
yukun's avatar
yukun committed
405
406
407
408

  if not model_options.aspp_with_batch_norm:
    return features, end_points
  else:
409
410
411
412
413
    if model_options.dense_prediction_cell_config is not None:
      tf.logging.info('Using dense prediction cell config.')
      dense_prediction_layer = dense_prediction_cell.DensePredictionCell(
          config=model_options.dense_prediction_cell_config,
          hparams={
414
              'conv_rate_multiplier': 16 // model_options.output_stride,
415
416
417
418
419
420
421
422
423
424
425
426
          })
      concat_logits = dense_prediction_layer.build_cell(
          features,
          output_stride=model_options.output_stride,
          crop_size=model_options.crop_size,
          image_pooling_crop_size=model_options.image_pooling_crop_size,
          weight_decay=weight_decay,
          reuse=reuse,
          is_training=is_training,
          fine_tune_batch_norm=fine_tune_batch_norm)
      return concat_logits, end_points
    else:
427
      # The following codes employ the DeepLabv3 ASPP module. Note that we
428
      # could express the ASPP module as one particular dense prediction
429
430
      # cell architecture. We do not do so but leave the following codes
      # for backward compatibility.
431
432
433
434
435
436
437
438
      batch_norm_params = utils.get_batch_norm_params(
          decay=0.9997,
          epsilon=1e-5,
          scale=True,
          is_training=(is_training and fine_tune_batch_norm),
          sync_batch_norm_method=model_options.sync_batch_norm_method)
      batch_norm = utils.get_batch_norm_fn(
          model_options.sync_batch_norm_method)
439
440
      activation_fn = (
          tf.nn.relu6 if model_options.use_bounded_activation else tf.nn.relu)
441
442
443
      with slim.arg_scope(
          [slim.conv2d, slim.separable_conv2d],
          weights_regularizer=slim.l2_regularizer(weight_decay),
444
          activation_fn=activation_fn,
445
          normalizer_fn=batch_norm,
446
447
448
          padding='SAME',
          stride=1,
          reuse=reuse):
449
450
        with slim.arg_scope([batch_norm], **batch_norm_params):
          depth = model_options.aspp_convs_filters
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
          branch_logits = []

          if model_options.add_image_level_feature:
            if model_options.crop_size is not None:
              image_pooling_crop_size = model_options.image_pooling_crop_size
              # If image_pooling_crop_size is not specified, use crop_size.
              if image_pooling_crop_size is None:
                image_pooling_crop_size = model_options.crop_size
              pool_height = scale_dimension(
                  image_pooling_crop_size[0],
                  1. / model_options.output_stride)
              pool_width = scale_dimension(
                  image_pooling_crop_size[1],
                  1. / model_options.output_stride)
              image_feature = slim.avg_pool2d(
466
467
                  features, [pool_height, pool_width],
                  model_options.image_pooling_stride, padding='VALID')
468
469
470
471
472
473
              resize_height = scale_dimension(
                  model_options.crop_size[0],
                  1. / model_options.output_stride)
              resize_width = scale_dimension(
                  model_options.crop_size[1],
                  1. / model_options.output_stride)
yukun's avatar
yukun committed
474
            else:
475
476
477
478
479
480
481
              # If crop_size is None, we simply do global pooling.
              pool_height = tf.shape(features)[1]
              pool_width = tf.shape(features)[2]
              image_feature = tf.reduce_mean(
                  features, axis=[1, 2], keepdims=True)
              resize_height = pool_height
              resize_width = pool_width
482
483
484
485
486
487
488
            image_feature_activation_fn = tf.nn.relu
            image_feature_normalizer_fn = batch_norm
            if model_options.aspp_with_squeeze_and_excitation:
              image_feature_activation_fn = tf.nn.sigmoid
              if model_options.image_se_uses_qsigmoid:
                image_feature_activation_fn = utils.q_sigmoid
              image_feature_normalizer_fn = None
489
            image_feature = slim.conv2d(
490
491
492
493
                image_feature, depth, 1,
                activation_fn=image_feature_activation_fn,
                normalizer_fn=image_feature_normalizer_fn,
                scope=IMAGE_POOLING_SCOPE)
494
495
496
497
498
499
500
501
502
503
            image_feature = _resize_bilinear(
                image_feature,
                [resize_height, resize_width],
                image_feature.dtype)
            # Set shape for resize_height/resize_width if they are not Tensor.
            if isinstance(resize_height, tf.Tensor):
              resize_height = None
            if isinstance(resize_width, tf.Tensor):
              resize_width = None
            image_feature.set_shape([None, resize_height, resize_width, depth])
504
505
            if not model_options.aspp_with_squeeze_and_excitation:
              branch_logits.append(image_feature)
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528

          # Employ a 1x1 convolution.
          branch_logits.append(slim.conv2d(features, depth, 1,
                                           scope=ASPP_SCOPE + str(0)))

          if model_options.atrous_rates:
            # Employ 3x3 convolutions with different atrous rates.
            for i, rate in enumerate(model_options.atrous_rates, 1):
              scope = ASPP_SCOPE + str(i)
              if model_options.aspp_with_separable_conv:
                aspp_features = split_separable_conv2d(
                    features,
                    filters=depth,
                    rate=rate,
                    weight_decay=weight_decay,
                    scope=scope)
              else:
                aspp_features = slim.conv2d(
                    features, depth, 3, rate=rate, scope=scope)
              branch_logits.append(aspp_features)

          # Merge branch logits.
          concat_logits = tf.concat(branch_logits, 3)
529
530
531
532
533
534
535
536
537
538
539
          if model_options.aspp_with_concat_projection:
            concat_logits = slim.conv2d(
                concat_logits, depth, 1, scope=CONCAT_PROJECTION_SCOPE)
            concat_logits = slim.dropout(
                concat_logits,
                keep_prob=0.9,
                is_training=is_training,
                scope=CONCAT_PROJECTION_SCOPE + '_dropout')
          if (model_options.add_image_level_feature and
              model_options.aspp_with_squeeze_and_excitation):
            concat_logits *= image_feature
540
541

          return concat_logits, end_points
yukun's avatar
yukun committed
542
543
544
545
546
547
548


def _get_logits(images,
                model_options,
                weight_decay=0.0001,
                reuse=None,
                is_training=False,
549
550
                fine_tune_batch_norm=False,
                nas_training_hyper_parameters=None):
yukun's avatar
yukun committed
551
552
553
554
555
556
557
558
559
  """Gets the logits by atrous/image spatial pyramid pooling.

  Args:
    images: A tensor of size [batch, height, width, channels].
    model_options: A ModelOptions instance to configure models.
    weight_decay: The weight decay for model variables.
    reuse: Reuse the model variables or not.
    is_training: Is training or not.
    fine_tune_batch_norm: Fine-tune the batch norm parameters or not.
560
561
562
563
564
565
    nas_training_hyper_parameters: A dictionary storing hyper-parameters for
      training nas models. Its keys are:
      - `drop_path_keep_prob`: Probability to keep each path in the cell when
        training.
      - `total_training_steps`: Total training steps to help drop path
        probability calculation.
yukun's avatar
yukun committed
566
567
568
569

  Returns:
    outputs_to_logits: A map from output_type to logits.
  """
570
  features, end_points = extract_features(
yukun's avatar
yukun committed
571
572
573
574
575
      images,
      model_options,
      weight_decay=weight_decay,
      reuse=reuse,
      is_training=is_training,
576
577
      fine_tune_batch_norm=fine_tune_batch_norm,
      nas_training_hyper_parameters=nas_training_hyper_parameters)
yukun's avatar
yukun committed
578

579
580
581
582
  if model_options.decoder_output_stride:
    crop_size = model_options.crop_size
    if crop_size is None:
      crop_size = [tf.shape(images)[1], tf.shape(images)[2]]
yukun's avatar
yukun committed
583
584
585
    features = refine_by_decoder(
        features,
        end_points,
586
        crop_size=crop_size,
587
        decoder_output_stride=model_options.decoder_output_stride,
yukun's avatar
yukun committed
588
        decoder_use_separable_conv=model_options.decoder_use_separable_conv,
589
590
591
        decoder_use_sum_merge=model_options.decoder_use_sum_merge,
        decoder_filters=model_options.decoder_filters,
        decoder_output_is_logits=model_options.decoder_output_is_logits,
yukun's avatar
yukun committed
592
593
594
595
        model_variant=model_options.model_variant,
        weight_decay=weight_decay,
        reuse=reuse,
        is_training=is_training,
596
597
        fine_tune_batch_norm=fine_tune_batch_norm,
        use_bounded_activation=model_options.use_bounded_activation)
yukun's avatar
yukun committed
598
599
600

  outputs_to_logits = {}
  for output in sorted(model_options.outputs_to_num_classes):
601
602
603
604
605
606
607
608
609
610
611
612
613
    if model_options.decoder_output_is_logits:
      outputs_to_logits[output] = tf.identity(features,
                                              name=output)
    else:
      outputs_to_logits[output] = get_branch_logits(
          features,
          model_options.outputs_to_num_classes[output],
          model_options.atrous_rates,
          aspp_with_batch_norm=model_options.aspp_with_batch_norm,
          kernel_size=model_options.logits_kernel_size,
          weight_decay=weight_decay,
          reuse=reuse,
          scope_suffix=output)
yukun's avatar
yukun committed
614
615
616
617
618
619

  return outputs_to_logits


def refine_by_decoder(features,
                      end_points,
620
621
                      crop_size=None,
                      decoder_output_stride=None,
yukun's avatar
yukun committed
622
                      decoder_use_separable_conv=False,
623
624
625
                      decoder_use_sum_merge=False,
                      decoder_filters=256,
                      decoder_output_is_logits=False,
yukun's avatar
yukun committed
626
627
628
629
                      model_variant=None,
                      weight_decay=0.0001,
                      reuse=None,
                      is_training=False,
630
                      fine_tune_batch_norm=False,
631
632
                      use_bounded_activation=False,
                      sync_batch_norm_method='None'):
yukun's avatar
yukun committed
633
634
635
636
637
638
639
  """Adds the decoder to obtain sharper segmentation results.

  Args:
    features: A tensor of size [batch, features_height, features_width,
      features_channels].
    end_points: A dictionary from components of the network to the corresponding
      activation.
640
641
642
643
    crop_size: A tuple [crop_height, crop_width] specifying whole patch crop
      size.
    decoder_output_stride: A list of integers specifying the output stride of
      low-level features used in the decoder module.
yukun's avatar
yukun committed
644
    decoder_use_separable_conv: Employ separable convolution for decoder or not.
645
646
647
    decoder_use_sum_merge: Boolean, decoder uses simple sum merge or not.
    decoder_filters: Integer, decoder filter size.
    decoder_output_is_logits: Boolean, using decoder output as logits or not.
yukun's avatar
yukun committed
648
649
650
651
652
    model_variant: Model variant for feature extraction.
    weight_decay: The weight decay for model variables.
    reuse: Reuse the model variables or not.
    is_training: Is training or not.
    fine_tune_batch_norm: Fine-tune the batch norm parameters or not.
653
654
    use_bounded_activation: Whether or not to use bounded activations. Bounded
      activations better lend themselves to quantized inference.
655
656
657
    sync_batch_norm_method: String, method used to sync batch norm. Currently
     only support `None` (no sync batch norm) and `tpu` (use tpu code to
     sync batch norm).
yukun's avatar
yukun committed
658
659
660
661

  Returns:
    Decoder output with size [batch, decoder_height, decoder_width,
      decoder_channels].
662
663
664

  Raises:
    ValueError: If crop_size is None.
yukun's avatar
yukun committed
665
  """
666
667
  if crop_size is None:
    raise ValueError('crop_size must be provided when using decoder.')
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
  batch_norm_params = utils.get_batch_norm_params(
      decay=0.9997,
      epsilon=1e-5,
      scale=True,
      is_training=(is_training and fine_tune_batch_norm),
      sync_batch_norm_method=sync_batch_norm_method)
  batch_norm = utils.get_batch_norm_fn(sync_batch_norm_method)
  decoder_depth = decoder_filters
  projected_filters = 48
  if decoder_use_sum_merge:
    # When using sum merge, the projected filters must be equal to decoder
    # filters.
    projected_filters = decoder_filters
  if decoder_output_is_logits:
    # Overwrite the setting when decoder output is logits.
    activation_fn = None
    normalizer_fn = None
    conv2d_kernel = 1
    # Use original conv instead of separable conv.
    decoder_use_separable_conv = False
  else:
    # Default setting when decoder output is not logits.
    activation_fn = tf.nn.relu6 if use_bounded_activation else tf.nn.relu
    normalizer_fn = batch_norm
    conv2d_kernel = 3
yukun's avatar
yukun committed
693
694
695
  with slim.arg_scope(
      [slim.conv2d, slim.separable_conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
696
697
      activation_fn=activation_fn,
      normalizer_fn=normalizer_fn,
yukun's avatar
yukun committed
698
699
700
      padding='SAME',
      stride=1,
      reuse=reuse):
701
    with slim.arg_scope([batch_norm], **batch_norm_params):
702
      with tf.variable_scope(DECODER_SCOPE, DECODER_SCOPE, [features]):
703
704
705
706
707
708
709
710
711
712
713
        decoder_features = features
        decoder_stage = 0
        scope_suffix = ''
        for output_stride in decoder_output_stride:
          feature_list = feature_extractor.networks_to_feature_maps[
              model_variant][
                  feature_extractor.DECODER_END_POINTS][output_stride]
          # If only one decoder stage, we do not change the scope name in
          # order for backward compactibility.
          if decoder_stage:
            scope_suffix = '_{}'.format(decoder_stage)
yukun's avatar
yukun committed
714
715
          for i, name in enumerate(feature_list):
            decoder_features_list = [decoder_features]
716
            # MobileNet and NAS variants use different naming convention.
717
718
719
            if ('mobilenet' in model_variant or
                model_variant.startswith('mnas') or
                model_variant.startswith('nas')):
720
721
722
723
              feature_name = name
            else:
              feature_name = '{}/{}'.format(
                  feature_extractor.name_scope[model_variant], name)
yukun's avatar
yukun committed
724
725
726
            decoder_features_list.append(
                slim.conv2d(
                    end_points[feature_name],
727
                    projected_filters,
yukun's avatar
yukun committed
728
                    1,
729
730
731
732
                    scope='feature_projection' + str(i) + scope_suffix))
            # Determine the output size.
            decoder_height = scale_dimension(crop_size[0], 1.0 / output_stride)
            decoder_width = scale_dimension(crop_size[1], 1.0 / output_stride)
yukun's avatar
yukun committed
733
734
            # Resize to decoder_height/decoder_width.
            for j, feature in enumerate(decoder_features_list):
735
736
              decoder_features_list[j] = _resize_bilinear(
                  feature, [decoder_height, decoder_width], feature.dtype)
737
738
739
740
741
              h = (None if isinstance(decoder_height, tf.Tensor)
                   else decoder_height)
              w = (None if isinstance(decoder_width, tf.Tensor)
                   else decoder_width)
              decoder_features_list[j].set_shape([None, h, w, None])
742
743
744
745
746
747
            if decoder_use_sum_merge:
              decoder_features = _decoder_with_sum_merge(
                  decoder_features_list,
                  decoder_depth,
                  conv2d_kernel=conv2d_kernel,
                  decoder_use_separable_conv=decoder_use_separable_conv,
yukun's avatar
yukun committed
748
                  weight_decay=weight_decay,
749
                  scope_suffix=scope_suffix)
yukun's avatar
yukun committed
750
            else:
751
752
753
754
              if not decoder_use_separable_conv:
                scope_suffix = str(i) + scope_suffix
              decoder_features = _decoder_with_concat_merge(
                  decoder_features_list,
yukun's avatar
yukun committed
755
                  decoder_depth,
756
757
758
                  decoder_use_separable_conv=decoder_use_separable_conv,
                  weight_decay=weight_decay,
                  scope_suffix=scope_suffix)
759
760
          decoder_stage += 1
        return decoder_features
yukun's avatar
yukun committed
761
762


763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
def _decoder_with_sum_merge(decoder_features_list,
                            decoder_depth,
                            conv2d_kernel=3,
                            decoder_use_separable_conv=True,
                            weight_decay=0.0001,
                            scope_suffix=''):
  """Decoder with sum to merge features.

  Args:
    decoder_features_list: A list of decoder features.
    decoder_depth: Integer, the filters used in the convolution.
    conv2d_kernel: Integer, the convolution kernel size.
    decoder_use_separable_conv: Boolean, use separable conv or not.
    weight_decay: Weight decay for the model variables.
    scope_suffix: String, used in the scope suffix.

  Returns:
    decoder features merged with sum.

  Raises:
    RuntimeError: If decoder_features_list have length not equal to 2.
  """
  if len(decoder_features_list) != 2:
    raise RuntimeError('Expect decoder_features has length 2.')
  # Only apply one convolution when decoder use sum merge.
  if decoder_use_separable_conv:
    decoder_features = split_separable_conv2d(
        decoder_features_list[0],
        filters=decoder_depth,
        rate=1,
        weight_decay=weight_decay,
        scope='decoder_split_sep_conv0'+scope_suffix) + decoder_features_list[1]
  else:
    decoder_features = slim.conv2d(
        decoder_features_list[0],
        decoder_depth,
        conv2d_kernel,
        scope='decoder_conv0'+scope_suffix) + decoder_features_list[1]
  return decoder_features


def _decoder_with_concat_merge(decoder_features_list,
                               decoder_depth,
                               decoder_use_separable_conv=True,
                               weight_decay=0.0001,
                               scope_suffix=''):
  """Decoder with concatenation to merge features.

  This decoder method applies two convolutions to smooth the features obtained
  by concatenating the input decoder_features_list.

  This decoder module is proposed in the DeepLabv3+ paper.

  Args:
    decoder_features_list: A list of decoder features.
    decoder_depth: Integer, the filters used in the convolution.
    decoder_use_separable_conv: Boolean, use separable conv or not.
    weight_decay: Weight decay for the model variables.
    scope_suffix: String, used in the scope suffix.

  Returns:
    decoder features merged with concatenation.
  """
  if decoder_use_separable_conv:
    decoder_features = split_separable_conv2d(
        tf.concat(decoder_features_list, 3),
        filters=decoder_depth,
        rate=1,
        weight_decay=weight_decay,
        scope='decoder_conv0'+scope_suffix)
    decoder_features = split_separable_conv2d(
        decoder_features,
        filters=decoder_depth,
        rate=1,
        weight_decay=weight_decay,
        scope='decoder_conv1'+scope_suffix)
  else:
    num_convs = 2
    decoder_features = slim.repeat(
        tf.concat(decoder_features_list, 3),
        num_convs,
        slim.conv2d,
        decoder_depth,
        3,
        scope='decoder_conv'+scope_suffix)
  return decoder_features


851
852
853
854
855
856
857
858
def get_branch_logits(features,
                      num_classes,
                      atrous_rates=None,
                      aspp_with_batch_norm=False,
                      kernel_size=1,
                      weight_decay=0.0001,
                      reuse=None,
                      scope_suffix=''):
yukun's avatar
yukun committed
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
  """Gets the logits from each model's branch.

  The underlying model is branched out in the last layer when atrous
  spatial pyramid pooling is employed, and all branches are sum-merged
  to form the final logits.

  Args:
    features: A float tensor of shape [batch, height, width, channels].
    num_classes: Number of classes to predict.
    atrous_rates: A list of atrous convolution rates for last layer.
    aspp_with_batch_norm: Use batch normalization layers for ASPP.
    kernel_size: Kernel size for convolution.
    weight_decay: Weight decay for the model variables.
    reuse: Reuse model variables or not.
    scope_suffix: Scope suffix for the model variables.

  Returns:
    Merged logits with shape [batch, height, width, num_classes].

  Raises:
    ValueError: Upon invalid input kernel_size value.
  """
  # When using batch normalization with ASPP, ASPP has been applied before
882
  # in extract_features, and thus we simply apply 1x1 convolution here.
yukun's avatar
yukun committed
883
884
885
886
887
888
889
890
891
892
893
  if aspp_with_batch_norm or atrous_rates is None:
    if kernel_size != 1:
      raise ValueError('Kernel size must be 1 when atrous_rates is None or '
                       'using aspp_with_batch_norm. Gets %d.' % kernel_size)
    atrous_rates = [1]

  with slim.arg_scope(
      [slim.conv2d],
      weights_regularizer=slim.l2_regularizer(weight_decay),
      weights_initializer=tf.truncated_normal_initializer(stddev=0.01),
      reuse=reuse):
894
    with tf.variable_scope(LOGITS_SCOPE_NAME, LOGITS_SCOPE_NAME, [features]):
yukun's avatar
yukun committed
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
      branch_logits = []
      for i, rate in enumerate(atrous_rates):
        scope = scope_suffix
        if i:
          scope += '_%d' % i

        branch_logits.append(
            slim.conv2d(
                features,
                num_classes,
                kernel_size=kernel_size,
                rate=rate,
                activation_fn=None,
                normalizer_fn=None,
                scope=scope))

      return tf.add_n(branch_logits)