model_lib.py 36.4 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 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.
# ==============================================================================
15
r"""Constructs model, inputs, and training environment."""
16
17
18
19
20

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

21
import copy
22
import functools
23
import os
24
25
26
27

import tensorflow as tf

from object_detection import eval_util
28
from object_detection import exporter as exporter_lib
29
from object_detection import inputs
30
from object_detection.builders import graph_rewriter_builder
31
32
33
34
35
from object_detection.builders import model_builder
from object_detection.builders import optimizer_builder
from object_detection.core import standard_fields as fields
from object_detection.utils import config_util
from object_detection.utils import label_map_util
36
from object_detection.utils import ops
37
38
39
40
from object_detection.utils import shape_utils
from object_detection.utils import variables_helper
from object_detection.utils import visualization_utils as vis_utils

41
42
43
44
45
46
47
48
# A map of names to methods that help build the model.
MODEL_BUILD_UTIL_MAP = {
    'get_configs_from_pipeline_file':
        config_util.get_configs_from_pipeline_file,
    'create_pipeline_proto_from_configs':
        config_util.create_pipeline_proto_from_configs,
    'merge_external_params_with_configs':
        config_util.merge_external_params_with_configs,
49
50
51
52
53
54
    'create_train_input_fn':
        inputs.create_train_input_fn,
    'create_eval_input_fn':
        inputs.create_eval_input_fn,
    'create_predict_input_fn':
        inputs.create_predict_input_fn,
55
    'detection_model_fn_base': model_builder.build,
56
57
58
}


59
60
def _prepare_groundtruth_for_eval(detection_model, class_agnostic,
                                  max_number_of_boxes):
61
  """Extracts groundtruth data from detection_model and prepares it for eval.
62
63
64
65

  Args:
    detection_model: A `DetectionModel` object.
    class_agnostic: Whether the detections are class_agnostic.
66
    max_number_of_boxes: Max number of groundtruth boxes.
67
68
69
70

  Returns:
    A tuple of:
    groundtruth: Dictionary with the following fields:
71
72
73
74
75
      'groundtruth_boxes': [batch_size, num_boxes, 4] float32 tensor of boxes,
        in normalized coordinates.
      'groundtruth_classes': [batch_size, num_boxes] int64 tensor of 1-indexed
        classes.
      'groundtruth_masks': 4D float32 tensor of instance masks (if provided in
76
        groundtruth)
77
78
79
80
      'groundtruth_is_crowd': [batch_size, num_boxes] bool tensor indicating
        is_crowd annotations (if provided in groundtruth).
      'num_groundtruth_boxes': [batch_size] tensor containing the maximum number
        of groundtruth boxes per image..
81
82
83
    class_agnostic: Boolean indicating whether detections are class agnostic.
  """
  input_data_fields = fields.InputDataFields()
84
85
86
  groundtruth_boxes = tf.stack(
      detection_model.groundtruth_lists(fields.BoxListFields.boxes))
  groundtruth_boxes_shape = tf.shape(groundtruth_boxes)
87
88
89
  # For class-agnostic models, groundtruth one-hot encodings collapse to all
  # ones.
  if class_agnostic:
90
91
    groundtruth_classes_one_hot = tf.ones(
        [groundtruth_boxes_shape[0], groundtruth_boxes_shape[1], 1])
92
  else:
93
94
    groundtruth_classes_one_hot = tf.stack(
        detection_model.groundtruth_lists(fields.BoxListFields.classes))
95
96
  label_id_offset = 1  # Applying label id offset (b/63711816)
  groundtruth_classes = (
97
      tf.argmax(groundtruth_classes_one_hot, axis=2) + label_id_offset)
98
99
100
101
102
  groundtruth = {
      input_data_fields.groundtruth_boxes: groundtruth_boxes,
      input_data_fields.groundtruth_classes: groundtruth_classes
  }
  if detection_model.groundtruth_has_field(fields.BoxListFields.masks):
103
104
105
    groundtruth[input_data_fields.groundtruth_instance_masks] = tf.stack(
        detection_model.groundtruth_lists(fields.BoxListFields.masks))

106
  if detection_model.groundtruth_has_field(fields.BoxListFields.is_crowd):
107
108
109
110
111
    groundtruth[input_data_fields.groundtruth_is_crowd] = tf.stack(
        detection_model.groundtruth_lists(fields.BoxListFields.is_crowd))

  groundtruth[input_data_fields.num_groundtruth_boxes] = (
      tf.tile([max_number_of_boxes], multiples=[groundtruth_boxes_shape[0]]))
112
113
114
115
116
117
118
  return groundtruth


def unstack_batch(tensor_dict, unpad_groundtruth_tensors=True):
  """Unstacks all tensors in `tensor_dict` along 0th dimension.

  Unstacks tensor from the tensor dict along 0th dimension and returns a
119
  tensor_dict containing values that are lists of unstacked, unpadded tensors.
120
121
122
123
124
125

  Tensors in the `tensor_dict` are expected to be of one of the three shapes:
  1. [batch_size]
  2. [batch_size, height, width, channels]
  3. [batch_size, num_boxes, d1, d2, ... dn]

126
127
  When unpad_groundtruth_tensors is set to true, unstacked tensors of form 3
  above are sliced along the `num_boxes` dimension using the value in tensor
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
  field.InputDataFields.num_groundtruth_boxes.

  Note that this function has a static list of input data fields and has to be
  kept in sync with the InputDataFields defined in core/standard_fields.py

  Args:
    tensor_dict: A dictionary of batched groundtruth tensors.
    unpad_groundtruth_tensors: Whether to remove padding along `num_boxes`
      dimension of the groundtruth tensors.

  Returns:
    A dictionary where the keys are from fields.InputDataFields and values are
    a list of unstacked (optionally unpadded) tensors.

  Raises:
    ValueError: If unpad_tensors is True and `tensor_dict` does not contain
      `num_groundtruth_boxes` tensor.
  """
146
147
148
  unbatched_tensor_dict = {
      key: tf.unstack(tensor) for key, tensor in tensor_dict.items()
  }
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
  if unpad_groundtruth_tensors:
    if (fields.InputDataFields.num_groundtruth_boxes not in
        unbatched_tensor_dict):
      raise ValueError('`num_groundtruth_boxes` not found in tensor_dict. '
                       'Keys available: {}'.format(
                           unbatched_tensor_dict.keys()))
    unbatched_unpadded_tensor_dict = {}
    unpad_keys = set([
        # List of input data fields that are padded along the num_boxes
        # dimension. This list has to be kept in sync with InputDataFields in
        # standard_fields.py.
        fields.InputDataFields.groundtruth_instance_masks,
        fields.InputDataFields.groundtruth_classes,
        fields.InputDataFields.groundtruth_boxes,
        fields.InputDataFields.groundtruth_keypoints,
        fields.InputDataFields.groundtruth_group_of,
        fields.InputDataFields.groundtruth_difficult,
        fields.InputDataFields.groundtruth_is_crowd,
        fields.InputDataFields.groundtruth_area,
        fields.InputDataFields.groundtruth_weights
    ]).intersection(set(unbatched_tensor_dict.keys()))

    for key in unpad_keys:
      unpadded_tensor_list = []
      for num_gt, padded_tensor in zip(
          unbatched_tensor_dict[fields.InputDataFields.num_groundtruth_boxes],
          unbatched_tensor_dict[key]):
        tensor_shape = shape_utils.combined_static_and_dynamic_shape(
            padded_tensor)
        slice_begin = tf.zeros([len(tensor_shape)], dtype=tf.int32)
        slice_size = tf.stack(
            [num_gt] + [-1 if dim is None else dim for dim in tensor_shape[1:]])
        unpadded_tensor = tf.slice(padded_tensor, slice_begin, slice_size)
        unpadded_tensor_list.append(unpadded_tensor)
      unbatched_unpadded_tensor_dict[key] = unpadded_tensor_list
    unbatched_tensor_dict.update(unbatched_unpadded_tensor_dict)

  return unbatched_tensor_dict


pkulzc's avatar
pkulzc committed
189
def provide_groundtruth(model, labels):
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
  """Provides the labels to a model as groundtruth.

  This helper function extracts the corresponding boxes, classes,
  keypoints, weights, masks, etc. from the labels, and provides it
  as groundtruth to the models.

  Args:
    model: The detection model to provide groundtruth to.
    labels: The labels for the training or evaluation inputs.
  """
  gt_boxes_list = labels[fields.InputDataFields.groundtruth_boxes]
  gt_classes_list = labels[fields.InputDataFields.groundtruth_classes]
  gt_masks_list = None
  if fields.InputDataFields.groundtruth_instance_masks in labels:
    gt_masks_list = labels[
        fields.InputDataFields.groundtruth_instance_masks]
  gt_keypoints_list = None
  if fields.InputDataFields.groundtruth_keypoints in labels:
    gt_keypoints_list = labels[fields.InputDataFields.groundtruth_keypoints]
  gt_weights_list = None
  if fields.InputDataFields.groundtruth_weights in labels:
    gt_weights_list = labels[fields.InputDataFields.groundtruth_weights]
  gt_confidences_list = None
  if fields.InputDataFields.groundtruth_confidences in labels:
    gt_confidences_list = labels[
        fields.InputDataFields.groundtruth_confidences]
  gt_is_crowd_list = None
  if fields.InputDataFields.groundtruth_is_crowd in labels:
    gt_is_crowd_list = labels[fields.InputDataFields.groundtruth_is_crowd]
  model.provide_groundtruth(
      groundtruth_boxes_list=gt_boxes_list,
      groundtruth_classes_list=gt_classes_list,
      groundtruth_confidences_list=gt_confidences_list,
      groundtruth_masks_list=gt_masks_list,
      groundtruth_keypoints_list=gt_keypoints_list,
      groundtruth_weights_list=gt_weights_list,
      groundtruth_is_crowd_list=gt_is_crowd_list)


229
230
def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False,
                    postprocess_on_cpu=False):
231
232
233
234
235
236
237
238
  """Creates a model function for `Estimator`.

  Args:
    detection_model_fn: Function that returns a `DetectionModel` instance.
    configs: Dictionary of pipeline config objects.
    hparams: `HParams` object.
    use_tpu: Boolean indicating whether model should be constructed for
        use on TPU.
239
240
    postprocess_on_cpu: When use_tpu and postprocess_on_cpu is true, postprocess
        is scheduled on the host cpu.
241
242
243
244
245
246

  Returns:
    `model_fn` for `Estimator`.
  """
  train_config = configs['train_config']
  eval_input_config = configs['eval_input_config']
247
  eval_config = configs['eval_config']
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265

  def model_fn(features, labels, mode, params=None):
    """Constructs the object detection model.

    Args:
      features: Dictionary of feature tensors, returned from `input_fn`.
      labels: Dictionary of groundtruth tensors if mode is TRAIN or EVAL,
        otherwise None.
      mode: Mode key from tf.estimator.ModeKeys.
      params: Parameter dictionary passed from the estimator.

    Returns:
      An `EstimatorSpec` that encapsulates the model and its serving
        configurations.
    """
    params = params or {}
    total_loss, train_op, detections, export_outputs = None, None, None, None
    is_training = mode == tf.estimator.ModeKeys.TRAIN
266
267
268
269

    # Make sure to set the Keras learning phase. True during training,
    # False for inference.
    tf.keras.backend.set_learning_phase(is_training)
270
271
272
273
274
275
276
    # Set policy for mixed-precision training with Keras-based models.
    if use_tpu and train_config.use_bfloat16:
      from tensorflow.python.keras.engine import base_layer_utils  # pylint: disable=g-import-not-at-top
      # Enable v2 behavior, as `mixed_bfloat16` is only supported in TF 2.0.
      base_layer_utils.enable_v2_dtype_behavior()
      tf.compat.v2.keras.mixed_precision.experimental.set_policy(
          'mixed_bfloat16')
277
278
    detection_model = detection_model_fn(
        is_training=is_training, add_summaries=(not use_tpu))
279
280
281
282
283
284
285
    scaffold_fn = None

    if mode == tf.estimator.ModeKeys.TRAIN:
      labels = unstack_batch(
          labels,
          unpad_groundtruth_tensors=train_config.unpad_groundtruth_tensors)
    elif mode == tf.estimator.ModeKeys.EVAL:
286
287
288
289
290
      # For evaling on train data, it is necessary to check whether groundtruth
      # must be unpadded.
      boxes_shape = (
          labels[fields.InputDataFields.groundtruth_boxes].get_shape()
          .as_list())
291
      unpad_groundtruth_tensors = boxes_shape[1] is not None and not use_tpu
292
293
      labels = unstack_batch(
          labels, unpad_groundtruth_tensors=unpad_groundtruth_tensors)
294
295

    if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
pkulzc's avatar
pkulzc committed
296
      provide_groundtruth(detection_model, labels)
297
298

    preprocessed_images = features[fields.InputDataFields.image]
299
300
301
302
303
    if use_tpu and train_config.use_bfloat16:
      with tf.contrib.tpu.bfloat16_scope():
        prediction_dict = detection_model.predict(
            preprocessed_images,
            features[fields.InputDataFields.true_image_shape])
304
        prediction_dict = ops.bfloat16_to_float32_nested(prediction_dict)
305
306
307
308
    else:
      prediction_dict = detection_model.predict(
          preprocessed_images,
          features[fields.InputDataFields.true_image_shape])
309
310
311
312

    def postprocess_wrapper(args):
      return detection_model.postprocess(args[0], args[1])

313
    if mode in (tf.estimator.ModeKeys.EVAL, tf.estimator.ModeKeys.PREDICT):
314
315
316
317
318
319
320
321
322
      if use_tpu and postprocess_on_cpu:
        detections = tf.contrib.tpu.outside_compilation(
            postprocess_wrapper,
            (prediction_dict,
             features[fields.InputDataFields.true_image_shape]))
      else:
        detections = postprocess_wrapper((
            prediction_dict,
            features[fields.InputDataFields.true_image_shape]))
323
324

    if mode == tf.estimator.ModeKeys.TRAIN:
325
326
      load_pretrained = hparams.load_pretrained if hparams else False
      if train_config.fine_tune_checkpoint and load_pretrained:
327
328
329
330
331
332
333
334
        if not train_config.fine_tune_checkpoint_type:
          # train_config.from_detection_checkpoint field is deprecated. For
          # backward compatibility, set train_config.fine_tune_checkpoint_type
          # based on train_config.from_detection_checkpoint.
          if train_config.from_detection_checkpoint:
            train_config.fine_tune_checkpoint_type = 'detection'
          else:
            train_config.fine_tune_checkpoint_type = 'classification'
335
        asg_map = detection_model.restore_map(
336
            fine_tune_checkpoint_type=train_config.fine_tune_checkpoint_type,
337
338
339
340
            load_all_detection_checkpoint_vars=(
                train_config.load_all_detection_checkpoint_vars))
        available_var_map = (
            variables_helper.get_variables_available_in_checkpoint(
341
342
                asg_map,
                train_config.fine_tune_checkpoint,
343
344
                include_global_step=False))
        if use_tpu:
345

346
347
348
349
          def tpu_scaffold():
            tf.train.init_from_checkpoint(train_config.fine_tune_checkpoint,
                                          available_var_map)
            return tf.train.Scaffold()
350

351
352
353
354
355
356
357
358
          scaffold_fn = tpu_scaffold
        else:
          tf.train.init_from_checkpoint(train_config.fine_tune_checkpoint,
                                        available_var_map)

    if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
      losses_dict = detection_model.loss(
          prediction_dict, features[fields.InputDataFields.true_image_shape])
359
      losses = [loss_tensor for loss_tensor in losses_dict.values()]
360
      if train_config.add_regularization_loss:
361
        regularization_losses = detection_model.regularization_losses()
362
363
364
        if use_tpu and train_config.use_bfloat16:
          regularization_losses = ops.bfloat16_to_float32_nested(
              regularization_losses)
365
        if regularization_losses:
366
367
          regularization_loss = tf.add_n(
              regularization_losses, name='regularization_loss')
368
          losses.append(regularization_loss)
369
          losses_dict['Loss/regularization_loss'] = regularization_loss
370
      total_loss = tf.add_n(losses, name='total_loss')
371
      losses_dict['Loss/total_loss'] = total_loss
372

373
374
375
376
377
      if 'graph_rewriter_config' in configs:
        graph_rewriter_fn = graph_rewriter_builder.build(
            configs['graph_rewriter_config'], is_training=is_training)
        graph_rewriter_fn()

378
379
      # TODO(rathodv): Stop creating optimizer summary vars in EVAL mode once we
      # can write learning rate summaries on TPU without host calls.
380
381
382
383
      global_step = tf.train.get_or_create_global_step()
      training_optimizer, optimizer_summary_vars = optimizer_builder.build(
          train_config.optimizer)

384
    if mode == tf.estimator.ModeKeys.TRAIN:
385
      if use_tpu:
386
        training_optimizer = tf.contrib.tpu.CrossShardOptimizer(
387
388
389
390
            training_optimizer)

      # Optionally freeze some layers by setting their gradients to be zero.
      trainable_variables = None
391
392
393
394
395
396
397
398
399
400
      include_variables = (
          train_config.update_trainable_variables
          if train_config.update_trainable_variables else None)
      exclude_variables = (
          train_config.freeze_variables
          if train_config.freeze_variables else None)
      trainable_variables = tf.contrib.framework.filter_variables(
          tf.trainable_variables(),
          include_patterns=include_variables,
          exclude_patterns=exclude_variables)
401
402
403
404
405
406
407
408
409

      clip_gradients_value = None
      if train_config.gradient_clipping_by_norm > 0:
        clip_gradients_value = train_config.gradient_clipping_by_norm

      if not use_tpu:
        for var in optimizer_summary_vars:
          tf.summary.scalar(var.op.name, var)
      summaries = [] if use_tpu else None
410
411
      if train_config.summarize_gradients:
        summaries = ['gradients', 'gradient_norm', 'global_gradient_norm']
412
413
414
415
416
417
      train_op = tf.contrib.layers.optimize_loss(
          loss=total_loss,
          global_step=global_step,
          learning_rate=None,
          clip_gradients=clip_gradients_value,
          optimizer=training_optimizer,
418
          update_ops=detection_model.updates(),
419
420
421
422
423
          variables=trainable_variables,
          summaries=summaries,
          name='')  # Preventing scope prefix on all variables.

    if mode == tf.estimator.ModeKeys.PREDICT:
424
      exported_output = exporter_lib.add_output_tensor_nodes(detections)
425
426
      export_outputs = {
          tf.saved_model.signature_constants.PREDICT_METHOD_NAME:
427
              tf.estimator.export.PredictOutput(exported_output)
428
429
430
      }

    eval_metric_ops = None
431
    scaffold = None
432
    if mode == tf.estimator.ModeKeys.EVAL:
433
434
      class_agnostic = (
          fields.DetectionResultFields.detection_classes not in detections)
435
436
437
      groundtruth = _prepare_groundtruth_for_eval(
          detection_model, class_agnostic,
          eval_input_config.max_number_of_boxes)
438
      use_original_images = fields.InputDataFields.original_image in features
pkulzc's avatar
pkulzc committed
439
      if use_original_images:
440
441
442
443
444
        eval_images = features[fields.InputDataFields.original_image]
        true_image_shapes = tf.slice(
            features[fields.InputDataFields.true_image_shape], [0, 0], [-1, 3])
        original_image_spatial_shapes = features[fields.InputDataFields
                                                 .original_image_spatial_shape]
pkulzc's avatar
pkulzc committed
445
446
      else:
        eval_images = features[fields.InputDataFields.image]
447
448
        true_image_shapes = None
        original_image_spatial_shapes = None
pkulzc's avatar
pkulzc committed
449

450
451
452
      eval_dict = eval_util.result_dict_for_batched_example(
          eval_images,
          features[inputs.HASH_KEY],
453
454
455
          detections,
          groundtruth,
          class_agnostic=class_agnostic,
456
457
458
          scale_to_absolute=True,
          original_image_spatial_shapes=original_image_spatial_shapes,
          true_image_shapes=true_image_shapes)
459

460
461
462
463
      if fields.InputDataFields.image_additional_channels in features:
        eval_dict[fields.InputDataFields.image_additional_channels] = features[
            fields.InputDataFields.image_additional_channels]

464
465
466
467
468
      if class_agnostic:
        category_index = label_map_util.create_class_agnostic_category_index()
      else:
        category_index = label_map_util.create_category_index_from_labelmap(
            eval_input_config.label_map_path)
469
      vis_metric_ops = None
470
      if not use_tpu and use_original_images:
471
472
473
474
475
476
477
478
        eval_metric_op_vis = vis_utils.VisualizeSingleFrameDetections(
            category_index,
            max_examples_to_draw=eval_config.num_visualizations,
            max_boxes_to_draw=eval_config.max_num_boxes_to_visualize,
            min_score_thresh=eval_config.min_score_threshold,
            use_normalized_coordinates=False)
        vis_metric_ops = eval_metric_op_vis.get_estimator_eval_metric_ops(
            eval_dict)
479

480
481
      # Eval metrics on a single example.
      eval_metric_ops = eval_util.get_eval_metric_ops_for_evaluators(
DefineFC's avatar
DefineFC committed
482
          eval_config, list(category_index.values()), eval_dict)
483
484
485
486
      for loss_key, loss_tensor in iter(losses_dict.items()):
        eval_metric_ops[loss_key] = tf.metrics.mean(loss_tensor)
      for var in optimizer_summary_vars:
        eval_metric_ops[var.op.name] = (var, tf.no_op())
487
488
      if vis_metric_ops is not None:
        eval_metric_ops.update(vis_metric_ops)
489
      eval_metric_ops = {str(k): v for k, v in eval_metric_ops.items()}
490

491
492
493
494
495
496
497
498
499
500
      if eval_config.use_moving_averages:
        variable_averages = tf.train.ExponentialMovingAverage(0.0)
        variables_to_restore = variable_averages.variables_to_restore()
        keep_checkpoint_every_n_hours = (
            train_config.keep_checkpoint_every_n_hours)
        saver = tf.train.Saver(
            variables_to_restore,
            keep_checkpoint_every_n_hours=keep_checkpoint_every_n_hours)
        scaffold = tf.train.Scaffold(saver=saver)

501
502
    # EVAL executes on CPU, so use regular non-TPU EstimatorSpec.
    if use_tpu and mode != tf.estimator.ModeKeys.EVAL:
503
504
505
506
507
508
509
510
511
      return tf.contrib.tpu.TPUEstimatorSpec(
          mode=mode,
          scaffold_fn=scaffold_fn,
          predictions=detections,
          loss=total_loss,
          train_op=train_op,
          eval_metrics=eval_metric_ops,
          export_outputs=export_outputs)
    else:
512
513
514
515
516
517
518
519
520
      if scaffold is None:
        keep_checkpoint_every_n_hours = (
            train_config.keep_checkpoint_every_n_hours)
        saver = tf.train.Saver(
            sharded=True,
            keep_checkpoint_every_n_hours=keep_checkpoint_every_n_hours,
            save_relative_paths=True)
        tf.add_to_collection(tf.GraphKeys.SAVERS, saver)
        scaffold = tf.train.Scaffold(saver=saver)
521
522
523
524
525
526
      return tf.estimator.EstimatorSpec(
          mode=mode,
          predictions=detections,
          loss=total_loss,
          train_op=train_op,
          eval_metric_ops=eval_metric_ops,
527
528
          export_outputs=export_outputs,
          scaffold=scaffold)
529
530
531
532

  return model_fn


533
534
535
def create_estimator_and_inputs(run_config,
                                hparams,
                                pipeline_config_path,
536
                                config_override=None,
537
                                train_steps=None,
pkulzc's avatar
pkulzc committed
538
                                sample_1_of_n_eval_examples=None,
539
                                sample_1_of_n_eval_on_train_examples=1,
540
541
542
543
544
                                model_fn_creator=create_model_fn,
                                use_tpu_estimator=False,
                                use_tpu=False,
                                num_shards=1,
                                params=None,
545
                                override_eval_num_epochs=True,
546
                                save_final_config=False,
547
548
                                postprocess_on_cpu=False,
                                export_to_tpu=None,
549
550
                                **kwargs):
  """Creates `Estimator`, input functions, and steps.
551
552
553
554
555

  Args:
    run_config: A `RunConfig`.
    hparams: A `HParams`.
    pipeline_config_path: A path to a pipeline config file.
556
557
    config_override: A pipeline_pb2.TrainEvalPipelineConfig text proto to
      override the config from `pipeline_config_path`.
558
559
    train_steps: Number of training steps. If None, the number of training steps
      is set from the `TrainConfig` proto.
560
561
562
563
564
    sample_1_of_n_eval_examples: Integer representing how often an eval example
      should be sampled. If 1, will sample all examples.
    sample_1_of_n_eval_on_train_examples: Similar to
      `sample_1_of_n_eval_examples`, except controls the sampling of training
      data for evaluation.
565
566
567
568
569
570
571
572
573
574
    model_fn_creator: A function that creates a `model_fn` for `Estimator`.
      Follows the signature:

      * Args:
        * `detection_model_fn`: Function that returns `DetectionModel` instance.
        * `configs`: Dictionary of pipeline config objects.
        * `hparams`: `HParams` object.
      * Returns:
        `model_fn` for `Estimator`.

575
576
577
578
579
580
581
582
    use_tpu_estimator: Whether a `TPUEstimator` should be returned. If False,
      an `Estimator` will be returned.
    use_tpu: Boolean, whether training and evaluation should run on TPU. Only
      used if `use_tpu_estimator` is True.
    num_shards: Number of shards (TPU cores). Only used if `use_tpu_estimator`
      is True.
    params: Parameter dictionary passed from the estimator. Only used if
      `use_tpu_estimator` is True.
583
584
    override_eval_num_epochs: Whether to overwrite the number of epochs to 1 for
      eval_input.
585
586
    save_final_config: Whether to save final config (obtained after applying
      overrides) to `estimator.model_dir`.
587
588
589
590
591
    postprocess_on_cpu: When use_tpu and postprocess_on_cpu are true,
      postprocess is scheduled on the host cpu.
    export_to_tpu: When use_tpu and export_to_tpu are true,
      `export_savedmodel()` exports a metagraph for serving on TPU besides the
      one on CPU.
592
593
594
    **kwargs: Additional keyword arguments for configuration override.

  Returns:
595
596
597
    A dictionary with the following fields:
    'estimator': An `Estimator` or `TPUEstimator`.
    'train_input_fn': A training input function.
598
599
    'eval_input_fns': A list of all evaluation input functions.
    'eval_input_names': A list of names for each evaluation input.
600
    'eval_on_train_input_fn': An evaluation-on-train input function.
601
602
603
    'predict_input_fn': A prediction input function.
    'train_steps': Number of training steps. Either directly from input or from
      configuration.
604
  """
605
606
607
608
  get_configs_from_pipeline_file = MODEL_BUILD_UTIL_MAP[
      'get_configs_from_pipeline_file']
  merge_external_params_with_configs = MODEL_BUILD_UTIL_MAP[
      'merge_external_params_with_configs']
609
610
  create_pipeline_proto_from_configs = MODEL_BUILD_UTIL_MAP[
      'create_pipeline_proto_from_configs']
611
612
613
  create_train_input_fn = MODEL_BUILD_UTIL_MAP['create_train_input_fn']
  create_eval_input_fn = MODEL_BUILD_UTIL_MAP['create_eval_input_fn']
  create_predict_input_fn = MODEL_BUILD_UTIL_MAP['create_predict_input_fn']
614
  detection_model_fn_base = MODEL_BUILD_UTIL_MAP['detection_model_fn_base']
615

616
617
  configs = get_configs_from_pipeline_file(
      pipeline_config_path, config_override=config_override)
618
619
  kwargs.update({
      'train_steps': train_steps,
620
      'use_bfloat16': configs['train_config'].use_bfloat16 and use_tpu
621
  })
pkulzc's avatar
pkulzc committed
622
623
624
625
  if sample_1_of_n_eval_examples >= 1:
    kwargs.update({
        'sample_1_of_n_eval_examples': sample_1_of_n_eval_examples
    })
626
627
628
629
  if override_eval_num_epochs:
    kwargs.update({'eval_num_epochs': 1})
    tf.logging.warning(
        'Forced number of epochs for all eval validations to be 1.')
630
  configs = merge_external_params_with_configs(
631
      configs, hparams, kwargs_dict=kwargs)
632
633
634
635
  model_config = configs['model']
  train_config = configs['train_config']
  train_input_config = configs['train_input_config']
  eval_config = configs['eval_config']
636
637
638
639
640
641
642
643
644
645
646
  eval_input_configs = configs['eval_input_configs']
  eval_on_train_input_config = copy.deepcopy(train_input_config)
  eval_on_train_input_config.sample_1_of_n_examples = (
      sample_1_of_n_eval_on_train_examples)
  if override_eval_num_epochs and eval_on_train_input_config.num_epochs != 1:
    tf.logging.warning('Expected number of evaluation epochs is 1, but '
                       'instead encountered `eval_on_train_input_config'
                       '.num_epochs` = '
                       '{}. Overwriting `num_epochs` to 1.'.format(
                           eval_on_train_input_config.num_epochs))
    eval_on_train_input_config.num_epochs = 1
647

648
649
650
  # update train_steps from config but only when non-zero value is provided
  if train_steps is None and train_config.num_steps != 0:
    train_steps = train_config.num_steps
651
652

  detection_model_fn = functools.partial(
653
      detection_model_fn_base, model_config=model_config)
654

655
  # Create the input functions for TRAIN/EVAL/PREDICT.
656
  train_input_fn = create_train_input_fn(
657
658
659
      train_config=train_config,
      train_input_config=train_input_config,
      model_config=model_config)
660
661
662
663
664
665
666
667
668
  eval_input_fns = [
      create_eval_input_fn(
          eval_config=eval_config,
          eval_input_config=eval_input_config,
          model_config=model_config) for eval_input_config in eval_input_configs
  ]
  eval_input_names = [
      eval_input_config.name for eval_input_config in eval_input_configs
  ]
669
670
  eval_on_train_input_fn = create_eval_input_fn(
      eval_config=eval_config,
671
      eval_input_config=eval_on_train_input_config,
672
      model_config=model_config)
673
  predict_input_fn = create_predict_input_fn(
674
      model_config=model_config, predict_input_config=eval_input_configs[0])
675

676
677
678
  # Read export_to_tpu from hparams if not passed.
  if export_to_tpu is None:
    export_to_tpu = hparams.get('export_to_tpu', False)
679
680
  tf.logging.info('create_estimator_and_inputs: use_tpu %s, export_to_tpu %s',
                  use_tpu, export_to_tpu)
681
682
  model_fn = model_fn_creator(detection_model_fn, configs, hparams, use_tpu,
                              postprocess_on_cpu)
683
  if use_tpu_estimator:
684
    estimator = tf.contrib.tpu.TPUEstimator(
685
686
687
688
689
690
        model_fn=model_fn,
        train_batch_size=train_config.batch_size,
        # For each core, only batch size 1 is supported for eval.
        eval_batch_size=num_shards * 1 if use_tpu else 1,
        use_tpu=use_tpu,
        config=run_config,
691
692
        export_to_tpu=export_to_tpu,
        eval_on_tpu=False,  # Eval runs on CPU, so disable eval on TPU
pkulzc's avatar
pkulzc committed
693
        params=params if params else {})
694
695
  else:
    estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)
696

697
  # Write the as-run pipeline config to disk.
698
  if run_config.is_chief and save_final_config:
699
    pipeline_config_final = create_pipeline_proto_from_configs(configs)
700
    config_util.save_pipeline_config(pipeline_config_final, estimator.model_dir)
701

702
  return dict(
703
704
      estimator=estimator,
      train_input_fn=train_input_fn,
705
706
      eval_input_fns=eval_input_fns,
      eval_input_names=eval_input_names,
707
      eval_on_train_input_fn=eval_on_train_input_fn,
708
      predict_input_fn=predict_input_fn,
709
      train_steps=train_steps)
710
711
712


def create_train_and_eval_specs(train_input_fn,
713
                                eval_input_fns,
714
                                eval_on_train_input_fn,
715
716
717
718
                                predict_input_fn,
                                train_steps,
                                eval_on_train_data=False,
                                final_exporter_name='Servo',
719
                                eval_spec_names=None):
720
721
722
723
  """Creates a `TrainSpec` and `EvalSpec`s.

  Args:
    train_input_fn: Function that produces features and labels on train data.
724
725
    eval_input_fns: A list of functions that produce features and labels on eval
      data.
726
727
    eval_on_train_input_fn: Function that produces features and labels for
      evaluation on train data.
728
729
730
731
732
    predict_input_fn: Function that produces features for inference.
    train_steps: Number of training steps.
    eval_on_train_data: Whether to evaluate model on training data. Default is
      False.
    final_exporter_name: String name given to `FinalExporter`.
733
    eval_spec_names: A list of string names for each `EvalSpec`.
734
735

  Returns:
736
737
738
    Tuple of `TrainSpec` and list of `EvalSpecs`. If `eval_on_train_data` is
    True, the last `EvalSpec` in the list will correspond to training data. The
    rest EvalSpecs in the list are evaluation datas.
739
740
741
742
  """
  train_spec = tf.estimator.TrainSpec(
      input_fn=train_input_fn, max_steps=train_steps)

743
  if eval_spec_names is None:
744
    eval_spec_names = [str(i) for i in range(len(eval_input_fns))]
745
746

  eval_specs = []
747
748
749
750
751
752
753
754
  for index, (eval_spec_name, eval_input_fn) in enumerate(
      zip(eval_spec_names, eval_input_fns)):
    # Uses final_exporter_name as exporter_name for the first eval spec for
    # backward compatibility.
    if index == 0:
      exporter_name = final_exporter_name
    else:
      exporter_name = '{}_{}'.format(final_exporter_name, eval_spec_name)
755
756
757
758
759
760
761
762
    exporter = tf.estimator.FinalExporter(
        name=exporter_name, serving_input_receiver_fn=predict_input_fn)
    eval_specs.append(
        tf.estimator.EvalSpec(
            name=eval_spec_name,
            input_fn=eval_input_fn,
            steps=None,
            exporters=exporter))
763
764
765
766

  if eval_on_train_data:
    eval_specs.append(
        tf.estimator.EvalSpec(
767
            name='eval_on_train', input_fn=eval_on_train_input_fn, steps=None))
768
769

  return train_spec, eval_specs
770
771


772
def continuous_eval(estimator, model_dir, input_fn, train_steps, name):
773
774
775
776
777
778
779
780
781
782
  """Perform continuous evaluation on checkpoints written to a model directory.

  Args:
    estimator: Estimator object to use for evaluation.
    model_dir: Model directory to read checkpoints for continuous evaluation.
    input_fn: Input function to use for evaluation.
    train_steps: Number of training steps. This is used to infer the last
      checkpoint and stop evaluation loop.
    name: Namescope for eval summary.
  """
783

784
785
786
787
788
789
790
791
792
793
794
  def terminate_eval():
    tf.logging.info('Terminating eval after 180 seconds of no checkpoints')
    return True

  for ckpt in tf.contrib.training.checkpoints_iterator(
      model_dir, min_interval_secs=180, timeout=None,
      timeout_fn=terminate_eval):

    tf.logging.info('Starting Evaluation.')
    try:
      eval_results = estimator.evaluate(
795
          input_fn=input_fn, steps=None, checkpoint_path=ckpt, name=name)
796
797
798
799
800
801
802
803
804
805
806
807
808
809
      tf.logging.info('Eval results: %s' % eval_results)

      # Terminate eval job when final checkpoint is reached
      current_step = int(os.path.basename(ckpt).split('-')[1])
      if current_step >= train_steps:
        tf.logging.info(
            'Evaluation finished after training step %d' % current_step)
        break

    except tf.errors.NotFoundError:
      tf.logging.info(
          'Checkpoint %s no longer exists, skipping checkpoint' % ckpt)


810
811
812
813
814
815
816
817
def populate_experiment(run_config,
                        hparams,
                        pipeline_config_path,
                        train_steps=None,
                        eval_steps=None,
                        model_fn_creator=create_model_fn,
                        **kwargs):
  """Populates an `Experiment` object.
818

819
820
  EXPERIMENT CLASS IS DEPRECATED. Please switch to
  tf.estimator.train_and_evaluate. As an example, see model_main.py.
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
851
852
853
854
855
  Args:
    run_config: A `RunConfig`.
    hparams: A `HParams`.
    pipeline_config_path: A path to a pipeline config file.
    train_steps: Number of training steps. If None, the number of training steps
      is set from the `TrainConfig` proto.
    eval_steps: Number of evaluation steps per evaluation cycle. If None, the
      number of evaluation steps is set from the `EvalConfig` proto.
    model_fn_creator: A function that creates a `model_fn` for `Estimator`.
      Follows the signature:

      * Args:
        * `detection_model_fn`: Function that returns `DetectionModel` instance.
        * `configs`: Dictionary of pipeline config objects.
        * `hparams`: `HParams` object.
      * Returns:
        `model_fn` for `Estimator`.

    **kwargs: Additional keyword arguments for configuration override.

  Returns:
    An `Experiment` that defines all aspects of training, evaluation, and
    export.
  """
  tf.logging.warning('Experiment is being deprecated. Please use '
                     'tf.estimator.train_and_evaluate(). See model_main.py for '
                     'an example.')
  train_and_eval_dict = create_estimator_and_inputs(
      run_config,
      hparams,
      pipeline_config_path,
      train_steps=train_steps,
      eval_steps=eval_steps,
      model_fn_creator=model_fn_creator,
856
      save_final_config=True,
857
858
859
      **kwargs)
  estimator = train_and_eval_dict['estimator']
  train_input_fn = train_and_eval_dict['train_input_fn']
860
  eval_input_fns = train_and_eval_dict['eval_input_fns']
861
862
863
864
865
866
867
868
869
870
871
  predict_input_fn = train_and_eval_dict['predict_input_fn']
  train_steps = train_and_eval_dict['train_steps']

  export_strategies = [
      tf.contrib.learn.utils.saved_model_export_utils.make_export_strategy(
          serving_input_fn=predict_input_fn)
  ]

  return tf.contrib.learn.Experiment(
      estimator=estimator,
      train_input_fn=train_input_fn,
872
      eval_input_fn=eval_input_fns[0],
873
      train_steps=train_steps,
874
      eval_steps=None,
875
      export_strategies=export_strategies,
876
877
      eval_delay_secs=120,
  )