evaluator.py 9.75 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
# 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.
# ==============================================================================
"""Detection model evaluator.

This file provides a generic evaluation method that can be used to evaluate a
DetectionModel.
"""
20

21
22
23
24
25
26
import logging
import tensorflow as tf

from object_detection import eval_util
from object_detection.core import prefetcher
from object_detection.core import standard_fields as fields
27
from object_detection.metrics import coco_evaluation
28
29
30
31
32
33
from object_detection.utils import object_detection_evaluation

# A dictionary of metric names to classes that implement the metric. The classes
# in the dictionary must implement
# utils.object_detection_evaluation.DetectionEvaluator interface.
EVAL_METRICS_CLASS_DICT = {
34
    'pascal_voc_detection_metrics':
35
        object_detection_evaluation.PascalDetectionEvaluator,
36
    'weighted_pascal_voc_detection_metrics':
37
        object_detection_evaluation.WeightedPascalDetectionEvaluator,
38
39
40
41
42
    'pascal_voc_instance_segmentation_metrics':
        object_detection_evaluation.PascalInstanceSegmentationEvaluator,
    'weighted_pascal_voc_instance_segmentation_metrics':
        object_detection_evaluation.WeightedPascalInstanceSegmentationEvaluator,
    'open_images_detection_metrics':
43
44
45
46
47
        object_detection_evaluation.OpenImagesDetectionEvaluator,
    'coco_detection_metrics':
        coco_evaluation.CocoDetectionEvaluator,
    'coco_mask_metrics':
        coco_evaluation.CocoMaskEvaluator,
48
49
}

50
51
EVAL_DEFAULT_METRIC = 'pascal_voc_detection_metrics'

52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69

def _extract_prediction_tensors(model,
                                create_input_dict_fn,
                                ignore_groundtruth=False):
  """Restores the model in a tensorflow session.

  Args:
    model: model to perform predictions with.
    create_input_dict_fn: function to create input tensor dictionaries.
    ignore_groundtruth: whether groundtruth should be ignored.

  Returns:
    tensor_dict: A tensor dictionary with evaluations.
  """
  input_dict = create_input_dict_fn()
  prefetch_queue = prefetcher.prefetch(input_dict, capacity=500)
  input_dict = prefetch_queue.dequeue()
  original_image = tf.expand_dims(input_dict[fields.InputDataFields.image], 0)
70
71
72
73
  preprocessed_image, true_image_shapes = model.preprocess(
      tf.to_float(original_image))
  prediction_dict = model.predict(preprocessed_image, true_image_shapes)
  detections = model.postprocess(prediction_dict, true_image_shapes)
74

75
  groundtruth = None
76
  if not ignore_groundtruth:
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
112
113
114
115
116
117
    groundtruth = {
        fields.InputDataFields.groundtruth_boxes:
            input_dict[fields.InputDataFields.groundtruth_boxes],
        fields.InputDataFields.groundtruth_classes:
            input_dict[fields.InputDataFields.groundtruth_classes],
        fields.InputDataFields.groundtruth_area:
            input_dict[fields.InputDataFields.groundtruth_area],
        fields.InputDataFields.groundtruth_is_crowd:
            input_dict[fields.InputDataFields.groundtruth_is_crowd],
        fields.InputDataFields.groundtruth_difficult:
            input_dict[fields.InputDataFields.groundtruth_difficult]
    }
    if fields.InputDataFields.groundtruth_group_of in input_dict:
      groundtruth[fields.InputDataFields.groundtruth_group_of] = (
          input_dict[fields.InputDataFields.groundtruth_group_of])
    if fields.DetectionResultFields.detection_masks in detections:
      groundtruth[fields.InputDataFields.groundtruth_instance_masks] = (
          input_dict[fields.InputDataFields.groundtruth_instance_masks])

  return eval_util.result_dict_for_single_example(
      original_image,
      input_dict[fields.InputDataFields.source_id],
      detections,
      groundtruth,
      class_agnostic=(
          fields.DetectionResultFields.detection_classes not in detections),
      scale_to_absolute=True)


def get_evaluators(eval_config, categories):
  """Returns the evaluator class according to eval_config, valid for categories.

  Args:
    eval_config: evaluation configurations.
    categories: a list of categories to evaluate.
  Returns:
    An list of instances of DetectionEvaluator.

  Raises:
    ValueError: if metric is not in the metric class dictionary.
  """
118
119
120
121
122
123
124
  eval_metric_fn_keys = eval_config.metrics_set
  if not eval_metric_fn_keys:
    eval_metric_fn_keys = [EVAL_DEFAULT_METRIC]
  evaluators_list = []
  for eval_metric_fn_key in eval_metric_fn_keys:
    if eval_metric_fn_key not in EVAL_METRICS_CLASS_DICT:
      raise ValueError('Metric not found: {}'.format(eval_metric_fn_key))
Zhichao Lu's avatar
Zhichao Lu committed
125
126
    evaluators_list.append(
        EVAL_METRICS_CLASS_DICT[eval_metric_fn_key](categories=categories))
127
  return evaluators_list
128
129
130


def evaluate(create_input_dict_fn, create_model_fn, eval_config, categories,
131
             checkpoint_dir, eval_dir, graph_hook_fn=None):
132
133
134
135
136
137
138
139
140
141
  """Evaluation function for detection models.

  Args:
    create_input_dict_fn: a function to create a tensor input dictionary.
    create_model_fn: a function that creates a DetectionModel.
    eval_config: a eval_pb2.EvalConfig protobuf.
    categories: a list of category dictionaries. Each dict in the list should
                have an integer 'id' field and string 'name' field.
    checkpoint_dir: directory to load the checkpoints to evaluate from.
    eval_dir: directory to write evaluation metrics summary to.
142
143
144
145
    graph_hook_fn: Optional function that is called after the training graph is
      completely built. This is helpful to perform additional changes to the
      training graph such as optimizing batchnorm. The function should modify
      the default graph.
146
147
148
149

  Returns:
    metrics: A dictionary containing metric names and values from the latest
      run.
150
151
152
153
154
155
156
157
158
159
160
161
162
  """

  model = create_model_fn()

  if eval_config.ignore_groundtruth and not eval_config.export_path:
    logging.fatal('If ignore_groundtruth=True then an export_path is '
                  'required. Aborting!!!')

  tensor_dict = _extract_prediction_tensors(
      model=model,
      create_input_dict_fn=create_input_dict_fn,
      ignore_groundtruth=eval_config.ignore_groundtruth)

163
  def _process_batch(tensor_dict, sess, batch_index, counters):
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
    """Evaluates tensors in tensor_dict, visualizing the first K examples.

    This function calls sess.run on tensor_dict, evaluating the original_image
    tensor only on the first K examples and visualizing detections overlaid
    on this original_image.

    Args:
      tensor_dict: a dictionary of tensors
      sess: tensorflow session
      batch_index: the index of the batch amongst all batches in the run.
      counters: a dictionary holding 'success' and 'skipped' fields which can
        be updated to keep track of number of successful and failed runs,
        respectively.  If these fields are not updated, then the success/skipped
        counter values shown at the end of evaluation will be incorrect.

    Returns:
      result_dict: a dictionary of numpy arrays
    """
    try:
183
      result_dict = sess.run(tensor_dict)
184
185
186
187
188
      counters['success'] += 1
    except tf.errors.InvalidArgumentError:
      logging.info('Skipping image')
      counters['skipped'] += 1
      return {}
189
    global_step = tf.train.global_step(sess, tf.train.get_global_step())
190
191
192
    if batch_index < eval_config.num_visualizations:
      tag = 'image-{}'.format(batch_index)
      eval_util.visualize_detection_results(
193
194
195
196
          result_dict,
          tag,
          global_step,
          categories=categories,
197
198
          summary_dir=eval_dir,
          export_dir=eval_config.visualization_export_dir,
199
200
201
202
203
204
205
206
207
          show_groundtruth=eval_config.visualize_groundtruth_boxes,
          groundtruth_box_visualization_color=eval_config.
          groundtruth_box_visualization_color,
          min_score_thresh=eval_config.min_score_threshold,
          max_num_predictions=eval_config.max_num_boxes_to_visualize,
          skip_scores=eval_config.skip_scores,
          skip_labels=eval_config.skip_labels,
          keep_image_id_for_visualization_export=eval_config.
          keep_image_id_for_visualization_export)
208
209
210
    return result_dict

  variables_to_restore = tf.global_variables()
211
  global_step = tf.train.get_or_create_global_step()
212
  variables_to_restore.append(global_step)
213
214
215

  if graph_hook_fn: graph_hook_fn()

216
217
218
219
  if eval_config.use_moving_averages:
    variable_averages = tf.train.ExponentialMovingAverage(0.0)
    variables_to_restore = variable_averages.variables_to_restore()
  saver = tf.train.Saver(variables_to_restore)
220

221
222
223
224
  def _restore_latest_checkpoint(sess):
    latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
    saver.restore(sess, latest_checkpoint)

225
  metrics = eval_util.repeated_checkpoint_run(
226
227
      tensor_dict=tensor_dict,
      summary_dir=eval_dir,
228
      evaluators=get_evaluators(eval_config, categories),
229
230
231
232
233
234
      batch_processor=_process_batch,
      checkpoint_dirs=[checkpoint_dir],
      variables_to_restore=None,
      restore_fn=_restore_latest_checkpoint,
      num_batches=eval_config.num_examples,
      eval_interval_secs=eval_config.eval_interval_secs,
235
236
237
      max_number_of_evaluations=(1 if eval_config.ignore_groundtruth else
                                 eval_config.max_evals
                                 if eval_config.max_evals else None),
238
239
240
      master=eval_config.eval_master,
      save_graph=eval_config.save_graph,
      save_graph_dir=(eval_dir if eval_config.save_graph else ''))
241
242

  return metrics