def _build_coco_metrics(self):
"""Build COCO metrics evaluator."""
if (not self._task_config.model.include_mask
) or self._task_config.annotation_file:
self.coco_metric = coco_evaluator.COCOEvaluator(
annotation_file=self._task_config.annotation_file,
include_mask=self._task_config.model.include_mask,
per_category_metrics=self._task_config.per_category_metrics)
else:
# Builds COCO-style annotation file if include_mask is True, and
# annotation_file isn't provided.
annotation_path = os.path.join(self._logging_dir, 'annotation.json')
if tf.io.gfile.exists(annotation_path):
logging.info(
'annotation.json file exists, skipping creating the annotation'
' file.')
else:
if self._task_config.validation_data.num_examples <= 0:
logging.info('validation_data.num_examples needs to be > 0')
if not self._task_config.validation_data.input_path:
logging.info('Can not create annotation file for tfds.')
logging.info(
'Creating coco-style annotation file: %s', annotation_path)
coco_utils.scan_and_generator_annotation_file(
self._task_config.validation_data.input_path,
self._task_config.validation_data.file_type,
self._task_config.validation_data.num_examples,
self.task_config.model.include_mask, annotation_path,
regenerate_source_id=self._task_config.validation_data.decoder
.simple_decoder.regenerate_source_id)
self.coco_metric = coco_evaluator.COCOEvaluator(
annotation_file=annotation_path,
include_mask=self._task_config.model.include_mask,
per_category_metrics=self._task_config.per_category_metrics)
def build_metrics(self, training: bool = True):
"""Build detection metrics."""
metrics = []
if training:
metric_names = [
'total_loss',
'rpn_score_loss',
'rpn_box_loss',
'frcnn_cls_loss',
'frcnn_box_loss',
'mask_loss',
'model_loss'
]
for name in metric_names:
metrics.append(tf.keras.metrics.Mean(name, dtype=tf.float32))
else:
if (not self._task_config.model.include_mask
) or self._task_config.annotation_file:
self.coco_metric = coco_evaluator.COCOEvaluator(
annotation_file=self._task_config.annotation_file,
include_mask=self._task_config.model.include_mask,
per_category_metrics=self._task_config.per_category_metrics)
else:
# Builds COCO-style annotation file if include_mask is True, and
# annotation_file isn't provided.
annotation_path = os.path.join(self._logging_dir, 'annotation.json')
if tf.io.gfile.exists(annotation_path):
logging.info(
'annotation.json file exists, skipping creating the annotation'
' file.')
else:
if self._task_config.validation_data.num_examples <= 0:
logging.info('validation_data.num_examples needs to be > 0')
if not self._task_config.validation_data.input_path:
logging.info('Can not create annotation file for tfds.')
logging.info(
'Creating coco-style annotation file: %s', annotation_path)
coco_utils.scan_and_generator_annotation_file(
self._task_config.validation_data.input_path,
self._task_config.validation_data.file_type,
self._task_config.validation_data.num_examples,
self.task_config.model.include_mask, annotation_path)
self.coco_metric = coco_evaluator.COCOEvaluator(
annotation_file=annotation_path,
include_mask=self._task_config.model.include_mask,
per_category_metrics=self._task_config.per_category_metrics)
return metrics
def test_scan_and_generator_annotation_file(self):
num_samples = 10
example = tfexample_utils.create_detection_test_example(
image_height=512,
image_width=512,
image_channel=3,
num_instances=10)
tf_examples = [example] * num_samples
data_file = os.path.join(self.create_tempdir(), 'test.tfrecord')
tfexample_utils.dump_to_tfrecord(record_file=data_file,
tf_examples=tf_examples)
annotation_file = os.path.join(self.create_tempdir(),
'annotation.json')
coco_utils.scan_and_generator_annotation_file(
file_pattern=data_file,
file_type='tfrecord',
num_samples=num_samples,
include_mask=True,
annotation_file=annotation_file)
self.assertTrue(
tf.io.gfile.exists(annotation_file),
msg='Annotation file {annotation_file} does not exists.')
