def _BatchElements(pcoll): # pylint: disable=invalid-name
"""Batches elements either automatically or to the given batch_size."""
desired_batch_size = Context.get_desired_batch_size()
kwargs = dict(
min_batch_size=desired_batch_size, max_batch_size=desired_batch_size
) if desired_batch_size is not None else {}
return pcoll | 'BatchElements' >> util.BatchElements(**kwargs)
def run(args):
"""Runs the embedding generation Beam pipeline."""
if tf.io.gfile.exists(args.embed_output_dir):
print('Removing embedding output directory...')
tf.io.gfile.rmtree(args.embed_output_dir)
print('Creating empty output directory...')
tf.io.gfile.makedirs(args.embed_output_dir)
options = beam.options.pipeline_options.PipelineOptions(**vars(args))
original_dim = hub.load(args.module_url)(['']).shape[1]
random_projection_matrix = generate_random_projection_weights(
original_dim, args.projected_dim, args.embed_output_dir)
print('Starting the Beam pipeline...')
with beam.Pipeline(runner=_RUNNER, options=options) as pipeline:
_ = (
pipeline
| 'Read sentences from files' >>
beam.io.ReadFromText(file_pattern=args.data_file_pattern)
| 'Batch elements' >> util.BatchElements(
min_batch_size=_BATCH_SIZE / 2, max_batch_size=_BATCH_SIZE)
| 'Generate embeddings' >> beam.Map(
generate_embeddings, args.module_url, random_projection_matrix)
| 'Encode to tf example' >> beam.FlatMap(to_tf_example)
| 'Write to TFRecords files' >> beam.io.WriteToTFRecord(
file_path_prefix='{}/emb'.format(args.embed_output_dir),
file_name_suffix='.tfrecords')
)
print('Beam pipeline completed.')
def test_constant_batch(self):
# Assumes a single bundle...
with TestPipeline() as p:
res = (p
| beam.Create(range(35))
| util.BatchElements(min_batch_size=10, max_batch_size=10)
| beam.Map(len))
assert_that(res, equal_to([10, 10, 10, 5]))
def test_grows_to_max_batch(self):
# Assumes a single bundle...
with TestPipeline() as p:
res = (p
| beam.Create(range(164))
| util.BatchElements(
min_batch_size=1, max_batch_size=50, clock=FakeClock())
| beam.Map(len))
assert_that(res, equal_to([1, 1, 2, 4, 8, 16, 32, 50, 50]))
def expand(self, pvalue):
return (pvalue
| FlatMap(self._create_image_annotation_pairs)
| util.BatchElements(min_batch_size=self.min_batch_size,
max_batch_size=self.max_batch_size)
| ParDo(
_ImageAnnotateFn(features=self.features,
retry=self.retry,
timeout=self.timeout,
client_options=self.client_options,
metadata=self.metadata)))
def test_windowed_batches(self):
# Assumes a single bundle, in order...
with TestPipeline() as p:
res = (
p
| beam.Create(range(47))
| beam.Map(lambda t: window.TimestampedValue(t, t))
| beam.WindowInto(window.FixedWindows(30))
| util.BatchElements(
min_batch_size=5, max_batch_size=10, clock=FakeClock())
| beam.Map(len))
assert_that(res, equal_to([
5, 5, 10, 10, # elements in [0, 30)
10, 7, # elements in [30, 47)
]))