def preprocessor(self):
"""See base class."""
base_preprocessor = vrgripper_env_models.DefaultVRGripperPreprocessor(
model_feature_specification_fn=self._episode_feature_specification,
model_label_specification_fn=self._episode_label_specification)
self._preprocessor = preprocessors.FixedLenMetaExamplePreprocessor(
base_preprocessor=base_preprocessor)
return self._preprocessor
def test_meta_example_preprocess(
self,
num_condition_samples_per_task,
num_inference_samples_per_task,
outer_batch_size):
base_preprocessor = MockBasePreprocessor()
meta_example_preprocessor = preprocessors.FixedLenMetaExamplePreprocessor(
base_preprocessor=base_preprocessor,
num_condition_samples_per_task=num_condition_samples_per_task,
num_inference_samples_per_task=num_inference_samples_per_task)
mock_tensors = self._create_mock_tensors(
meta_example_preprocessor, outer_batch_size)
with self.session() as sess:
dataset = tf.data.Dataset.from_tensor_slices(mock_tensors)
dataset = dataset.batch(outer_batch_size, drop_remainder=True)
preprocess_fn = functools.partial(
meta_example_preprocessor.preprocess,
mode=tf.estimator.ModeKeys.TRAIN)
dataset = dataset.map(map_func=preprocess_fn, num_parallel_calls=1)
raw_meta_features, raw_meta_labels = (
dataset.make_one_shot_iterator().get_next())
np_raw_meta_features, np_raw_meta_labels = sess.run(
[raw_meta_features, raw_meta_labels])
ref_features, ref_labels = mock_tensors
self.assertEqual(
list(np_raw_meta_features.condition.features.keys()),
list(np_raw_meta_features.inference.features.keys()))
# The labels and the condition labels have to have the same keys.
self.assertEqual(
list(np_raw_meta_features.condition.labels.keys()),
list(np_raw_meta_labels.keys()))
# The image has been resized. Therefore, we ensure that its shape is
# correct. Note, we have to strip the outer and inner batch dimensions.
self.assertEqual(
np_raw_meta_features.condition.features.image.shape[2:],
_DEFAULT_OUT_IMAGE_SHAPE)
self.assertEqual(
np_raw_meta_features.inference.features.image.shape[2:],
_DEFAULT_OUT_IMAGE_SHAPE)
for i in range(num_condition_samples_per_task):
np.testing.assert_array_almost_equal(
np_raw_meta_features.condition.features['action'][:, i, Ellipsis],
ref_features['condition/features/action/{:d}'.format(i)])
for label_name in np_raw_meta_features.condition.labels.keys():
np.testing.assert_array_almost_equal(
np_raw_meta_features.condition.labels[label_name][:, i, Ellipsis],
ref_features['condition/labels/{:s}/{:d}'.format(
label_name, i)])
for i in range(num_inference_samples_per_task):
np.testing.assert_array_almost_equal(
np_raw_meta_features.inference.features['action'][:, i, Ellipsis],
ref_features['inference/features/action/{:d}'.format(i)])
for label_name in np_raw_meta_features.condition.labels.keys():
np.testing.assert_array_almost_equal(
np_raw_meta_labels[label_name][:, i, Ellipsis],
ref_labels[six.ensure_str(label_name) + '/{:d}'.format(i)])
