def check_expected_structure(self, sampler):
"""Checks the stream of episode descriptions is as expected."""
chunk_sizes = sampler.compute_chunk_sizes()
batch_size = sum(chunk_sizes)
placeholder_id = len(self.dataset_spec.get_classes(self.split))
# We need to go through TF and back because
# `reader.decompress_episode_representation` operates on TF tensors.
generator = functools.partial(reader.episode_representation_generator,
dataset_spec=self.dataset_spec,
split=self.split,
pool=None,
sampler=sampler)
tf_generator = tf.data.Dataset.from_generator(
generator, tf.int64,
tf.TensorShape([None,
2])).map(reader.decompress_episode_representation)
iterator = tf_generator.make_one_shot_iterator()
next_item = iterator.get_next()
for _ in range(3):
with self.cached_session() as sess:
batch = sess.run(next_item)
self.assertEqual(len(batch), batch_size)
flush_chunk, support_chunk, query_chunk = split_into_chunks(
batch, chunk_sizes)
# flush_chunk is slightly oversized: if we actually had support_chunk_size
# + query_chunk_size examples remaining, we could have used them.
# Therefore, the last element of flush_chunk should be padding.
self.assertEqual(flush_chunk[-1], placeholder_id)
# TODO(lamblinp): check more about the content of flush_chunk
# The padding should be at the end of each chunk.
for chunk in (flush_chunk, support_chunk, query_chunk):
num_actual_examples = sum(class_id != placeholder_id
for class_id in chunk)
self.assertNotIn(placeholder_id, chunk[:num_actual_examples])
self.assertTrue(
all(placeholder_id == class_id
for class_id in chunk[num_actual_examples:]))
def create_ta(s):
return tf.TensorArray(dtype=s.dtype,
size=self._train_interval + 1,
element_shape=tf.TensorShape(
[batch_size]).concatenate(s.shape))
def create_ta(s):
return tf.TensorArray(dtype=s.dtype,
size=maximum_iterations,
element_shape=tf.TensorShape(
[batch_size]).concatenate(s.shape))
