factory.py 14.3 KB
Newer Older
Yeqing Li's avatar
Yeqing Li committed
1
# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
Abdullah Rashwan's avatar
Abdullah Rashwan committed
2
3
4
5
6
7
8
9
10
11
12
13
#
# 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.
Yeqing Li's avatar
Yeqing Li committed
14

Abdullah Rashwan's avatar
Abdullah Rashwan committed
15
16
17
"""Factory methods to build models."""

# Import libraries
Abdullah Rashwan's avatar
Abdullah Rashwan committed
18

Abdullah Rashwan's avatar
Abdullah Rashwan committed
19
20
21
22
23
import tensorflow as tf

from official.vision.beta.configs import image_classification as classification_cfg
from official.vision.beta.configs import maskrcnn as maskrcnn_cfg
from official.vision.beta.configs import retinanet as retinanet_cfg
Abdullah Rashwan's avatar
Abdullah Rashwan committed
24
from official.vision.beta.configs import semantic_segmentation as segmentation_cfg
Yeqing Li's avatar
Yeqing Li committed
25
from official.vision.beta.modeling import backbones
Abdullah Rashwan's avatar
Abdullah Rashwan committed
26
27
28
from official.vision.beta.modeling import classification_model
from official.vision.beta.modeling import maskrcnn_model
from official.vision.beta.modeling import retinanet_model
Abdullah Rashwan's avatar
Abdullah Rashwan committed
29
from official.vision.beta.modeling import segmentation_model
Abdullah Rashwan's avatar
Abdullah Rashwan committed
30
31
32
from official.vision.beta.modeling.decoders import factory as decoder_factory
from official.vision.beta.modeling.heads import dense_prediction_heads
from official.vision.beta.modeling.heads import instance_heads
Abdullah Rashwan's avatar
Abdullah Rashwan committed
33
from official.vision.beta.modeling.heads import segmentation_heads
Abdullah Rashwan's avatar
Abdullah Rashwan committed
34
35
36
37
38
39
40
41
42
43
from official.vision.beta.modeling.layers import detection_generator
from official.vision.beta.modeling.layers import mask_sampler
from official.vision.beta.modeling.layers import roi_aligner
from official.vision.beta.modeling.layers import roi_generator
from official.vision.beta.modeling.layers import roi_sampler


def build_classification_model(
    input_specs: tf.keras.layers.InputSpec,
    model_config: classification_cfg.ImageClassificationModel,
Pengchong Jin's avatar
Pengchong Jin committed
44
    l2_regularizer: tf.keras.regularizers.Regularizer = None,
Fan Yang's avatar
Fan Yang committed
45
    skip_logits_layer: bool = False) -> tf.keras.Model:
Abdullah Rashwan's avatar
Abdullah Rashwan committed
46
  """Builds the classification model."""
Yeqing Li's avatar
Yeqing Li committed
47
  backbone = backbones.factory.build_backbone(
Abdullah Rashwan's avatar
Abdullah Rashwan committed
48
49
50
51
52
53
54
55
56
57
58
59
60
61
      input_specs=input_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  norm_activation_config = model_config.norm_activation
  model = classification_model.ClassificationModel(
      backbone=backbone,
      num_classes=model_config.num_classes,
      input_specs=input_specs,
      dropout_rate=model_config.dropout_rate,
      kernel_regularizer=l2_regularizer,
      add_head_batch_norm=model_config.add_head_batch_norm,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
Pengchong Jin's avatar
Pengchong Jin committed
62
63
      norm_epsilon=norm_activation_config.norm_epsilon,
      skip_logits_layer=skip_logits_layer)
Abdullah Rashwan's avatar
Abdullah Rashwan committed
64
65
66
  return model


Fan Yang's avatar
Fan Yang committed
67
68
69
70
def build_maskrcnn(
    input_specs: tf.keras.layers.InputSpec,
    model_config: maskrcnn_cfg.MaskRCNN,
    l2_regularizer: tf.keras.regularizers.Regularizer = None) -> tf.keras.Model:
Abdullah Rashwan's avatar
Abdullah Rashwan committed
71
  """Builds Mask R-CNN model."""
Yeqing Li's avatar
Yeqing Li committed
72
  backbone = backbones.factory.build_backbone(
Abdullah Rashwan's avatar
Abdullah Rashwan committed
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
      input_specs=input_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  decoder = decoder_factory.build_decoder(
      input_specs=backbone.output_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  rpn_head_config = model_config.rpn_head
  roi_generator_config = model_config.roi_generator
  roi_sampler_config = model_config.roi_sampler
  roi_aligner_config = model_config.roi_aligner
  detection_head_config = model_config.detection_head
  generator_config = model_config.detection_generator
  norm_activation_config = model_config.norm_activation
  num_anchors_per_location = (
      len(model_config.anchor.aspect_ratios) * model_config.anchor.num_scales)

  rpn_head = dense_prediction_heads.RPNHead(
      min_level=model_config.min_level,
      max_level=model_config.max_level,
      num_anchors_per_location=num_anchors_per_location,
      num_convs=rpn_head_config.num_convs,
      num_filters=rpn_head_config.num_filters,
      use_separable_conv=rpn_head_config.use_separable_conv,
      activation=norm_activation_config.activation,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
      norm_epsilon=norm_activation_config.norm_epsilon,
      kernel_regularizer=l2_regularizer)

  detection_head = instance_heads.DetectionHead(
      num_classes=model_config.num_classes,
      num_convs=detection_head_config.num_convs,
      num_filters=detection_head_config.num_filters,
      use_separable_conv=detection_head_config.use_separable_conv,
      num_fcs=detection_head_config.num_fcs,
      fc_dims=detection_head_config.fc_dims,
Xianzhi Du's avatar
Xianzhi Du committed
112
      class_agnostic_bbox_pred=detection_head_config.class_agnostic_bbox_pred,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
113
114
115
116
      activation=norm_activation_config.activation,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
      norm_epsilon=norm_activation_config.norm_epsilon,
Xianzhi Du's avatar
Xianzhi Du committed
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
      kernel_regularizer=l2_regularizer,
      name='detection_head')
  if roi_sampler_config.cascade_iou_thresholds:
    detection_head_cascade = [detection_head]
    for cascade_num in range(len(roi_sampler_config.cascade_iou_thresholds)):
      detection_head = instance_heads.DetectionHead(
          num_classes=model_config.num_classes,
          num_convs=detection_head_config.num_convs,
          num_filters=detection_head_config.num_filters,
          use_separable_conv=detection_head_config.use_separable_conv,
          num_fcs=detection_head_config.num_fcs,
          fc_dims=detection_head_config.fc_dims,
          class_agnostic_bbox_pred=detection_head_config
          .class_agnostic_bbox_pred,
          activation=norm_activation_config.activation,
          use_sync_bn=norm_activation_config.use_sync_bn,
          norm_momentum=norm_activation_config.norm_momentum,
          norm_epsilon=norm_activation_config.norm_epsilon,
          kernel_regularizer=l2_regularizer,
          name='detection_head_{}'.format(cascade_num + 1))
      detection_head_cascade.append(detection_head)
    detection_head = detection_head_cascade
Abdullah Rashwan's avatar
Abdullah Rashwan committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

  roi_generator_obj = roi_generator.MultilevelROIGenerator(
      pre_nms_top_k=roi_generator_config.pre_nms_top_k,
      pre_nms_score_threshold=roi_generator_config.pre_nms_score_threshold,
      pre_nms_min_size_threshold=(
          roi_generator_config.pre_nms_min_size_threshold),
      nms_iou_threshold=roi_generator_config.nms_iou_threshold,
      num_proposals=roi_generator_config.num_proposals,
      test_pre_nms_top_k=roi_generator_config.test_pre_nms_top_k,
      test_pre_nms_score_threshold=(
          roi_generator_config.test_pre_nms_score_threshold),
      test_pre_nms_min_size_threshold=(
          roi_generator_config.test_pre_nms_min_size_threshold),
      test_nms_iou_threshold=roi_generator_config.test_nms_iou_threshold,
      test_num_proposals=roi_generator_config.test_num_proposals,
      use_batched_nms=roi_generator_config.use_batched_nms)

Xianzhi Du's avatar
Xianzhi Du committed
156
  roi_sampler_cascade = []
Abdullah Rashwan's avatar
Abdullah Rashwan committed
157
158
159
160
161
162
163
164
165
  roi_sampler_obj = roi_sampler.ROISampler(
      mix_gt_boxes=roi_sampler_config.mix_gt_boxes,
      num_sampled_rois=roi_sampler_config.num_sampled_rois,
      foreground_fraction=roi_sampler_config.foreground_fraction,
      foreground_iou_threshold=roi_sampler_config.foreground_iou_threshold,
      background_iou_high_threshold=(
          roi_sampler_config.background_iou_high_threshold),
      background_iou_low_threshold=(
          roi_sampler_config.background_iou_low_threshold))
Xianzhi Du's avatar
Xianzhi Du committed
166
167
168
169
170
171
172
173
174
175
176
177
  roi_sampler_cascade.append(roi_sampler_obj)
  # Initialize addtional roi simplers for cascade heads.
  if roi_sampler_config.cascade_iou_thresholds:
    for iou in roi_sampler_config.cascade_iou_thresholds:
      roi_sampler_obj = roi_sampler.ROISampler(
          mix_gt_boxes=False,
          num_sampled_rois=roi_sampler_config.num_sampled_rois,
          foreground_iou_threshold=iou,
          background_iou_high_threshold=iou,
          background_iou_low_threshold=0.0,
          skip_subsampling=True)
      roi_sampler_cascade.append(roi_sampler_obj)
Abdullah Rashwan's avatar
Abdullah Rashwan committed
178
179
180
181
182
183

  roi_aligner_obj = roi_aligner.MultilevelROIAligner(
      crop_size=roi_aligner_config.crop_size,
      sample_offset=roi_aligner_config.sample_offset)

  detection_generator_obj = detection_generator.DetectionGenerator(
Fan Yang's avatar
Fan Yang committed
184
      apply_nms=generator_config.apply_nms,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
      pre_nms_top_k=generator_config.pre_nms_top_k,
      pre_nms_score_threshold=generator_config.pre_nms_score_threshold,
      nms_iou_threshold=generator_config.nms_iou_threshold,
      max_num_detections=generator_config.max_num_detections,
      use_batched_nms=generator_config.use_batched_nms)

  if model_config.include_mask:
    mask_head = instance_heads.MaskHead(
        num_classes=model_config.num_classes,
        upsample_factor=model_config.mask_head.upsample_factor,
        num_convs=model_config.mask_head.num_convs,
        num_filters=model_config.mask_head.num_filters,
        use_separable_conv=model_config.mask_head.use_separable_conv,
        activation=model_config.norm_activation.activation,
        norm_momentum=model_config.norm_activation.norm_momentum,
        norm_epsilon=model_config.norm_activation.norm_epsilon,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
201
202
        kernel_regularizer=l2_regularizer,
        class_agnostic=model_config.mask_head.class_agnostic)
Abdullah Rashwan's avatar
Abdullah Rashwan committed
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223

    mask_sampler_obj = mask_sampler.MaskSampler(
        mask_target_size=(
            model_config.mask_roi_aligner.crop_size *
            model_config.mask_head.upsample_factor),
        num_sampled_masks=model_config.mask_sampler.num_sampled_masks)

    mask_roi_aligner_obj = roi_aligner.MultilevelROIAligner(
        crop_size=model_config.mask_roi_aligner.crop_size,
        sample_offset=model_config.mask_roi_aligner.sample_offset)
  else:
    mask_head = None
    mask_sampler_obj = None
    mask_roi_aligner_obj = None

  model = maskrcnn_model.MaskRCNNModel(
      backbone=backbone,
      decoder=decoder,
      rpn_head=rpn_head,
      detection_head=detection_head,
      roi_generator=roi_generator_obj,
Xianzhi Du's avatar
Xianzhi Du committed
224
      roi_sampler=roi_sampler_cascade,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
225
226
227
228
      roi_aligner=roi_aligner_obj,
      detection_generator=detection_generator_obj,
      mask_head=mask_head,
      mask_sampler=mask_sampler_obj,
Xianzhi Du's avatar
Xianzhi Du committed
229
230
      mask_roi_aligner=mask_roi_aligner_obj,
      class_agnostic_bbox_pred=detection_head_config.class_agnostic_bbox_pred,
231
232
233
234
235
236
      cascade_class_ensemble=detection_head_config.cascade_class_ensemble,
      min_level=model_config.min_level,
      max_level=model_config.max_level,
      num_scales=model_config.anchor.num_scales,
      aspect_ratios=model_config.anchor.aspect_ratios,
      anchor_size=model_config.anchor.anchor_size)
Abdullah Rashwan's avatar
Abdullah Rashwan committed
237
238
239
  return model


Fan Yang's avatar
Fan Yang committed
240
241
242
243
def build_retinanet(
    input_specs: tf.keras.layers.InputSpec,
    model_config: retinanet_cfg.RetinaNet,
    l2_regularizer: tf.keras.regularizers.Regularizer = None) -> tf.keras.Model:
Abdullah Rashwan's avatar
Abdullah Rashwan committed
244
  """Builds RetinaNet model."""
Yeqing Li's avatar
Yeqing Li committed
245
  backbone = backbones.factory.build_backbone(
Abdullah Rashwan's avatar
Abdullah Rashwan committed
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
      input_specs=input_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  decoder = decoder_factory.build_decoder(
      input_specs=backbone.output_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  head_config = model_config.head
  generator_config = model_config.detection_generator
  norm_activation_config = model_config.norm_activation
  num_anchors_per_location = (
      len(model_config.anchor.aspect_ratios) * model_config.anchor.num_scales)

  head = dense_prediction_heads.RetinaNetHead(
      min_level=model_config.min_level,
      max_level=model_config.max_level,
      num_classes=model_config.num_classes,
      num_anchors_per_location=num_anchors_per_location,
      num_convs=head_config.num_convs,
      num_filters=head_config.num_filters,
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
268
269
270
      attribute_heads=[
          cfg.as_dict() for cfg in (head_config.attribute_heads or [])
      ],
Abdullah Rashwan's avatar
Abdullah Rashwan committed
271
272
273
274
275
276
277
278
      use_separable_conv=head_config.use_separable_conv,
      activation=norm_activation_config.activation,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
      norm_epsilon=norm_activation_config.norm_epsilon,
      kernel_regularizer=l2_regularizer)

  detection_generator_obj = detection_generator.MultilevelDetectionGenerator(
Fan Yang's avatar
Fan Yang committed
279
      apply_nms=generator_config.apply_nms,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
280
281
282
283
284
285
286
      pre_nms_top_k=generator_config.pre_nms_top_k,
      pre_nms_score_threshold=generator_config.pre_nms_score_threshold,
      nms_iou_threshold=generator_config.nms_iou_threshold,
      max_num_detections=generator_config.max_num_detections,
      use_batched_nms=generator_config.use_batched_nms)

  model = retinanet_model.RetinaNetModel(
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
287
288
289
290
291
292
293
294
295
      backbone,
      decoder,
      head,
      detection_generator_obj,
      min_level=model_config.min_level,
      max_level=model_config.max_level,
      num_scales=model_config.anchor.num_scales,
      aspect_ratios=model_config.anchor.aspect_ratios,
      anchor_size=model_config.anchor.anchor_size)
Abdullah Rashwan's avatar
Abdullah Rashwan committed
296
  return model
Abdullah Rashwan's avatar
Abdullah Rashwan committed
297
298
299
300


def build_segmentation_model(
    input_specs: tf.keras.layers.InputSpec,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
301
    model_config: segmentation_cfg.SemanticSegmentationModel,
Fan Yang's avatar
Fan Yang committed
302
    l2_regularizer: tf.keras.regularizers.Regularizer = None) -> tf.keras.Model:
Abdullah Rashwan's avatar
Abdullah Rashwan committed
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
  """Builds Segmentation model."""
  backbone = backbones.factory.build_backbone(
      input_specs=input_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  decoder = decoder_factory.build_decoder(
      input_specs=backbone.output_specs,
      model_config=model_config,
      l2_regularizer=l2_regularizer)

  head_config = model_config.head
  norm_activation_config = model_config.norm_activation

  head = segmentation_heads.SegmentationHead(
      num_classes=model_config.num_classes,
      level=head_config.level,
      num_convs=head_config.num_convs,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
321
      prediction_kernel_size=head_config.prediction_kernel_size,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
322
323
      num_filters=head_config.num_filters,
      upsample_factor=head_config.upsample_factor,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
324
325
326
      feature_fusion=head_config.feature_fusion,
      low_level=head_config.low_level,
      low_level_num_filters=head_config.low_level_num_filters,
Abdullah Rashwan's avatar
Abdullah Rashwan committed
327
328
329
330
331
332
333
334
      activation=norm_activation_config.activation,
      use_sync_bn=norm_activation_config.use_sync_bn,
      norm_momentum=norm_activation_config.norm_momentum,
      norm_epsilon=norm_activation_config.norm_epsilon,
      kernel_regularizer=l2_regularizer)

  model = segmentation_model.SegmentationModel(backbone, decoder, head)
  return model