model_lib.py 34 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
36
37
38
39
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
from object_detection.utils import shape_utils
from object_detection.utils import variables_helper
from object_detection.utils import visualization_utils as vis_utils

40
41
42
43
44
45
46
47
# 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,
48
49
50
51
52
53
    '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,
54
55
56
}


57
58
def _prepare_groundtruth_for_eval(detection_model, class_agnostic,
                                  max_number_of_boxes):
59
  """Extracts groundtruth data from detection_model and prepares it for eval.
60
61
62
63

  Args:
    detection_model: A `DetectionModel` object.
    class_agnostic: Whether the detections are class_agnostic.
64
    max_number_of_boxes: Max number of groundtruth boxes.
65
66
67
68

  Returns:
    A tuple of:
    groundtruth: Dictionary with the following fields:
69
70
71
72
73
      '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
74
        groundtruth)
75
76
77
78
      '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..
79
80
81
    class_agnostic: Boolean indicating whether detections are class agnostic.
  """
  input_data_fields = fields.InputDataFields()
82
83
84
  groundtruth_boxes = tf.stack(
      detection_model.groundtruth_lists(fields.BoxListFields.boxes))
  groundtruth_boxes_shape = tf.shape(groundtruth_boxes)
85
86
87
  # For class-agnostic models, groundtruth one-hot encodings collapse to all
  # ones.
  if class_agnostic:
88
89
    groundtruth_classes_one_hot = tf.ones(
        [groundtruth_boxes_shape[0], groundtruth_boxes_shape[1], 1])
90
  else:
91
92
    groundtruth_classes_one_hot = tf.stack(
        detection_model.groundtruth_lists(fields.BoxListFields.classes))
93
94
  label_id_offset = 1  # Applying label id offset (b/63711816)
  groundtruth_classes = (
95
      tf.argmax(groundtruth_classes_one_hot, axis=2) + label_id_offset)
96
97
98
99
100
  groundtruth = {
      input_data_fields.groundtruth_boxes: groundtruth_boxes,
      input_data_fields.groundtruth_classes: groundtruth_classes
  }
  if detection_model.groundtruth_has_field(fields.BoxListFields.masks):
101
102
103
    groundtruth[input_data_fields.groundtruth_instance_masks] = tf.stack(
        detection_model.groundtruth_lists(fields.BoxListFields.masks))

104
  if detection_model.groundtruth_has_field(fields.BoxListFields.is_crowd):
105
106
107
108
109
    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]]))
110
111
112
113
114
115
116
  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
117
  tensor_dict containing values that are lists of unstacked, unpadded tensors.
118
119
120
121
122
123

  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]

124
125
  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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
  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.
  """
144
145
146
  unbatched_tensor_dict = {
      key: tf.unstack(tensor) for key, tensor in tensor_dict.items()
  }
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
  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


def create_model_fn(detection_model_fn, configs, hparams, use_tpu=False):
  """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.

  Returns:
    `model_fn` for `Estimator`.
  """
  train_config = configs['train_config']
  eval_input_config = configs['eval_input_config']
202
  eval_config = configs['eval_config']
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220

  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
221
222
223
224

    # Make sure to set the Keras learning phase. True during training,
    # False for inference.
    tf.keras.backend.set_learning_phase(is_training)
225
226
    detection_model = detection_model_fn(
        is_training=is_training, add_summaries=(not use_tpu))
227
228
229
230
231
232
233
    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:
234
235
236
237
238
      # 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())
239
      unpad_groundtruth_tensors = boxes_shape[1] is not None and not use_tpu
240
241
      labels = unstack_batch(
          labels, unpad_groundtruth_tensors=unpad_groundtruth_tensors)
242
243
244
245
246
247
248
249
250
251
252

    if mode in (tf.estimator.ModeKeys.TRAIN, tf.estimator.ModeKeys.EVAL):
      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]
253
254
255
      gt_weights_list = None
      if fields.InputDataFields.groundtruth_weights in labels:
        gt_weights_list = labels[fields.InputDataFields.groundtruth_weights]
256
257
258
259
      gt_confidences_list = None
      if fields.InputDataFields.groundtruth_confidences in labels:
        gt_confidences_list = labels[
            fields.InputDataFields.groundtruth_confidences]
260
      gt_is_crowd_list = None
261
262
      if fields.InputDataFields.groundtruth_is_crowd in labels:
        gt_is_crowd_list = labels[fields.InputDataFields.groundtruth_is_crowd]
263
264
265
      detection_model.provide_groundtruth(
          groundtruth_boxes_list=gt_boxes_list,
          groundtruth_classes_list=gt_classes_list,
266
          groundtruth_confidences_list=gt_confidences_list,
267
          groundtruth_masks_list=gt_masks_list,
268
          groundtruth_keypoints_list=gt_keypoints_list,
269
          groundtruth_weights_list=gt_weights_list,
270
          groundtruth_is_crowd_list=gt_is_crowd_list)
271
272

    preprocessed_images = features[fields.InputDataFields.image]
273
274
275
276
277
278
279
280
281
282
283
284
    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])
        for k, v in prediction_dict.items():
          if v.dtype == tf.bfloat16:
            prediction_dict[k] = tf.cast(v, tf.float32)
    else:
      prediction_dict = detection_model.predict(
          preprocessed_images,
          features[fields.InputDataFields.true_image_shape])
285
286
287
    if mode in (tf.estimator.ModeKeys.EVAL, tf.estimator.ModeKeys.PREDICT):
      detections = detection_model.postprocess(
          prediction_dict, features[fields.InputDataFields.true_image_shape])
288
289
290

    if mode == tf.estimator.ModeKeys.TRAIN:
      if train_config.fine_tune_checkpoint and hparams.load_pretrained:
291
292
293
294
295
296
297
298
        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'
299
        asg_map = detection_model.restore_map(
300
            fine_tune_checkpoint_type=train_config.fine_tune_checkpoint_type,
301
302
303
304
            load_all_detection_checkpoint_vars=(
                train_config.load_all_detection_checkpoint_vars))
        available_var_map = (
            variables_helper.get_variables_available_in_checkpoint(
305
306
                asg_map,
                train_config.fine_tune_checkpoint,
307
308
                include_global_step=False))
        if use_tpu:
309

310
311
312
313
          def tpu_scaffold():
            tf.train.init_from_checkpoint(train_config.fine_tune_checkpoint,
                                          available_var_map)
            return tf.train.Scaffold()
314

315
316
317
318
319
320
321
322
          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])
323
      losses = [loss_tensor for loss_tensor in losses_dict.values()]
324
      if train_config.add_regularization_loss:
325
        regularization_losses = detection_model.regularization_losses()
326
        if regularization_losses:
327
328
          regularization_loss = tf.add_n(
              regularization_losses, name='regularization_loss')
329
          losses.append(regularization_loss)
330
          losses_dict['Loss/regularization_loss'] = regularization_loss
331
      total_loss = tf.add_n(losses, name='total_loss')
332
      losses_dict['Loss/total_loss'] = total_loss
333

334
335
336
337
338
      if 'graph_rewriter_config' in configs:
        graph_rewriter_fn = graph_rewriter_builder.build(
            configs['graph_rewriter_config'], is_training=is_training)
        graph_rewriter_fn()

339
340
      # TODO(rathodv): Stop creating optimizer summary vars in EVAL mode once we
      # can write learning rate summaries on TPU without host calls.
341
342
343
344
      global_step = tf.train.get_or_create_global_step()
      training_optimizer, optimizer_summary_vars = optimizer_builder.build(
          train_config.optimizer)

345
    if mode == tf.estimator.ModeKeys.TRAIN:
346
      if use_tpu:
347
        training_optimizer = tf.contrib.tpu.CrossShardOptimizer(
348
349
350
351
            training_optimizer)

      # Optionally freeze some layers by setting their gradients to be zero.
      trainable_variables = None
352
353
354
355
356
357
358
359
360
361
      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)
362
363
364
365
366
367
368
369
370

      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
371
372
      if train_config.summarize_gradients:
        summaries = ['gradients', 'gradient_norm', 'global_gradient_norm']
373
374
375
376
377
378
      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,
379
          update_ops=detection_model.updates(),
380
381
382
383
384
          variables=trainable_variables,
          summaries=summaries,
          name='')  # Preventing scope prefix on all variables.

    if mode == tf.estimator.ModeKeys.PREDICT:
385
      exported_output = exporter_lib.add_output_tensor_nodes(detections)
386
387
      export_outputs = {
          tf.saved_model.signature_constants.PREDICT_METHOD_NAME:
388
              tf.estimator.export.PredictOutput(exported_output)
389
390
391
      }

    eval_metric_ops = None
392
    scaffold = None
393
    if mode == tf.estimator.ModeKeys.EVAL:
394
395
      class_agnostic = (
          fields.DetectionResultFields.detection_classes not in detections)
396
397
398
      groundtruth = _prepare_groundtruth_for_eval(
          detection_model, class_agnostic,
          eval_input_config.max_number_of_boxes)
399
      use_original_images = fields.InputDataFields.original_image in features
pkulzc's avatar
pkulzc committed
400
      if use_original_images:
401
402
403
404
405
        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
406
407
      else:
        eval_images = features[fields.InputDataFields.image]
408
409
        true_image_shapes = None
        original_image_spatial_shapes = None
pkulzc's avatar
pkulzc committed
410

411
412
413
      eval_dict = eval_util.result_dict_for_batched_example(
          eval_images,
          features[inputs.HASH_KEY],
414
415
416
          detections,
          groundtruth,
          class_agnostic=class_agnostic,
417
418
419
          scale_to_absolute=True,
          original_image_spatial_shapes=original_image_spatial_shapes,
          true_image_shapes=true_image_shapes)
420
421
422
423
424
425

      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)
426
      vis_metric_ops = None
427
      if not use_tpu and use_original_images:
428
429
430
431
432
433
434
435
        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)
436

437
438
      # Eval metrics on a single example.
      eval_metric_ops = eval_util.get_eval_metric_ops_for_evaluators(
DefineFC's avatar
DefineFC committed
439
          eval_config, list(category_index.values()), eval_dict)
440
441
442
443
      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())
444
445
      if vis_metric_ops is not None:
        eval_metric_ops.update(vis_metric_ops)
446
      eval_metric_ops = {str(k): v for k, v in eval_metric_ops.items()}
447

448
449
450
451
452
453
454
455
456
457
      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)

458
459
    # EVAL executes on CPU, so use regular non-TPU EstimatorSpec.
    if use_tpu and mode != tf.estimator.ModeKeys.EVAL:
460
461
462
463
464
465
466
467
468
      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:
469
470
471
472
473
474
475
476
477
      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)
478
479
480
481
482
483
      return tf.estimator.EstimatorSpec(
          mode=mode,
          predictions=detections,
          loss=total_loss,
          train_op=train_op,
          eval_metric_ops=eval_metric_ops,
484
485
          export_outputs=export_outputs,
          scaffold=scaffold)
486
487
488
489

  return model_fn


490
491
492
def create_estimator_and_inputs(run_config,
                                hparams,
                                pipeline_config_path,
493
                                config_override=None,
494
                                train_steps=None,
495
496
                                sample_1_of_n_eval_examples=1,
                                sample_1_of_n_eval_on_train_examples=1,
497
498
499
500
501
                                model_fn_creator=create_model_fn,
                                use_tpu_estimator=False,
                                use_tpu=False,
                                num_shards=1,
                                params=None,
502
                                override_eval_num_epochs=True,
503
                                save_final_config=False,
504
505
                                **kwargs):
  """Creates `Estimator`, input functions, and steps.
506
507
508
509
510

  Args:
    run_config: A `RunConfig`.
    hparams: A `HParams`.
    pipeline_config_path: A path to a pipeline config file.
511
512
    config_override: A pipeline_pb2.TrainEvalPipelineConfig text proto to
      override the config from `pipeline_config_path`.
513
514
    train_steps: Number of training steps. If None, the number of training steps
      is set from the `TrainConfig` proto.
515
516
517
518
519
    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.
520
521
522
523
524
525
526
527
528
529
    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`.

530
531
532
533
534
535
536
537
    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.
538
539
    override_eval_num_epochs: Whether to overwrite the number of epochs to
      1 for eval_input.
540
541
    save_final_config: Whether to save final config (obtained after applying
      overrides) to `estimator.model_dir`.
542
543
544
    **kwargs: Additional keyword arguments for configuration override.

  Returns:
545
546
547
    A dictionary with the following fields:
    'estimator': An `Estimator` or `TPUEstimator`.
    'train_input_fn': A training input function.
548
549
    'eval_input_fns': A list of all evaluation input functions.
    'eval_input_names': A list of names for each evaluation input.
550
    'eval_on_train_input_fn': An evaluation-on-train input function.
551
552
553
    'predict_input_fn': A prediction input function.
    'train_steps': Number of training steps. Either directly from input or from
      configuration.
554
  """
555
556
557
558
  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']
559
560
  create_pipeline_proto_from_configs = MODEL_BUILD_UTIL_MAP[
      'create_pipeline_proto_from_configs']
561
562
563
564
  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']

565
566
  configs = get_configs_from_pipeline_file(pipeline_config_path,
                                           config_override=config_override)
567
568
  kwargs.update({
      'train_steps': train_steps,
pkulzc's avatar
pkulzc committed
569
      'sample_1_of_n_eval_examples': sample_1_of_n_eval_examples
570
571
572
573
574
  })
  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.')
575
  configs = merge_external_params_with_configs(
576
      configs, hparams, kwargs_dict=kwargs)
577
578
579
580
  model_config = configs['model']
  train_config = configs['train_config']
  train_input_config = configs['train_input_config']
  eval_config = configs['eval_config']
581
582
583
584
585
586
587
588
589
590
591
  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
592

593
594
595
  # 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
596
597
598
599

  detection_model_fn = functools.partial(
      model_builder.build, model_config=model_config)

600
  # Create the input functions for TRAIN/EVAL/PREDICT.
601
  train_input_fn = create_train_input_fn(
602
603
604
      train_config=train_config,
      train_input_config=train_input_config,
      model_config=model_config)
605
606
607
608
609
610
611
612
613
  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
  ]
614
615
  eval_on_train_input_fn = create_eval_input_fn(
      eval_config=eval_config,
616
      eval_input_config=eval_on_train_input_config,
617
      model_config=model_config)
618
  predict_input_fn = create_predict_input_fn(
619
      model_config=model_config, predict_input_config=eval_input_configs[0])
620

621
622
623
  export_to_tpu = hparams.get('export_to_tpu', False)
  tf.logging.info('create_estimator_and_inputs: use_tpu %s, export_to_tpu %s',
                  use_tpu, export_to_tpu)
624
625
  model_fn = model_fn_creator(detection_model_fn, configs, hparams, use_tpu)
  if use_tpu_estimator:
626
    estimator = tf.contrib.tpu.TPUEstimator(
627
628
629
630
631
632
        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,
633
        # TODO(lzc): Remove conditional after CMLE moves to TF 1.9
634
635
636
        params=params if params else {})
  else:
    estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)
637

638
  # Write the as-run pipeline config to disk.
639
  if run_config.is_chief and save_final_config:
640
    pipeline_config_final = create_pipeline_proto_from_configs(configs)
641
    config_util.save_pipeline_config(pipeline_config_final, estimator.model_dir)
642

643
  return dict(
644
645
      estimator=estimator,
      train_input_fn=train_input_fn,
646
647
      eval_input_fns=eval_input_fns,
      eval_input_names=eval_input_names,
648
      eval_on_train_input_fn=eval_on_train_input_fn,
649
      predict_input_fn=predict_input_fn,
650
      train_steps=train_steps)
651
652
653


def create_train_and_eval_specs(train_input_fn,
654
                                eval_input_fns,
655
                                eval_on_train_input_fn,
656
657
658
659
                                predict_input_fn,
                                train_steps,
                                eval_on_train_data=False,
                                final_exporter_name='Servo',
660
                                eval_spec_names=None):
661
662
663
664
  """Creates a `TrainSpec` and `EvalSpec`s.

  Args:
    train_input_fn: Function that produces features and labels on train data.
665
666
    eval_input_fns: A list of functions that produce features and labels on eval
      data.
667
668
    eval_on_train_input_fn: Function that produces features and labels for
      evaluation on train data.
669
670
671
672
673
    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`.
674
    eval_spec_names: A list of string names for each `EvalSpec`.
675
676

  Returns:
677
678
679
    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.
680
681
682
683
  """
  train_spec = tf.estimator.TrainSpec(
      input_fn=train_input_fn, max_steps=train_steps)

684
  if eval_spec_names is None:
685
    eval_spec_names = [str(i) for i in range(len(eval_input_fns))]
686
687

  eval_specs = []
688
689
690
691
692
693
694
695
  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)
696
697
698
699
700
701
702
703
    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))
704
705
706
707

  if eval_on_train_data:
    eval_specs.append(
        tf.estimator.EvalSpec(
708
            name='eval_on_train', input_fn=eval_on_train_input_fn, steps=None))
709
710

  return train_spec, eval_specs
711
712


713
def continuous_eval(estimator, model_dir, input_fn, train_steps, name):
714
715
716
717
718
719
720
721
722
723
  """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.
  """
724

725
726
727
728
729
730
731
732
733
734
735
  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(
736
          input_fn=input_fn, steps=None, checkpoint_path=ckpt, name=name)
737
738
739
740
741
742
743
744
745
746
747
748
749
750
      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)


751
752
753
754
755
756
757
758
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.
759

760
761
  EXPERIMENT CLASS IS DEPRECATED. Please switch to
  tf.estimator.train_and_evaluate. As an example, see model_main.py.
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
  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,
797
      save_final_config=True,
798
799
800
      **kwargs)
  estimator = train_and_eval_dict['estimator']
  train_input_fn = train_and_eval_dict['train_input_fn']
801
  eval_input_fns = train_and_eval_dict['eval_input_fns']
802
803
804
805
806
807
808
809
810
811
812
  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,
813
      eval_input_fn=eval_input_fns[0],
814
      train_steps=train_steps,
815
      eval_steps=None,
816
      export_strategies=export_strategies,
817
818
      eval_delay_secs=120,
  )