def run(p, args, aggregator_dict, cloud_logger=None):
"""Run the pipeline with the args and dataflow pipeline option."""
# Create a PCollection for model directory.
model_dir = p | "Create Model Directory" >> beam.Create([args.model_dir])
input_file_format = args.input_file_format.lower()
# Create one pcollection per input file or file pattern. And then flatten
# them into one pcollection. The duplicated names need to be removed as the
# file name is used to create unique labels for the PTransform.
readers = []
for pattern in list(
set(args.input_file_patterns.split(FILE_LIST_SEPARATOR))):
# Setup reader.
#
# TODO(user): Perhaps simplify the batch prediction code by using
# CompressionTypes.AUTO.
if input_file_format.startswith("tfrecord"):
if input_file_format == "tfrecord_gzip":
compression_type = CompressionTypes.GZIP
else:
assert input_file_format == "tfrecord"
compression_type = CompressionTypes.UNCOMPRESSED
reader = "READ_TFRECORD_FILES_%s" % pattern >> ReadFromTFRecord(
pattern, compression_type=compression_type)
else:
assert input_file_format == "text"
reader = "READ_TEXT_FILES_%s" % pattern >> ReadFromText(pattern)
# Put the pcollections into a list and flatten later.
readers.append(p | reader)
# Setup the whole pipeline.
results, errors = (readers
| beam.Flatten()
| "BATCH_PREDICTION" >> batch_prediction.BatchPredict(
beam.pvalue.AsSingleton(model_dir),
batch_size=args.batch_size,
aggregator_dict=aggregator_dict,
cloud_logger=cloud_logger))
# Convert predictions to JSON and then write to output files.
_ = (results
| "TO_JSON" >> beam.Map(json.dumps)
| "WRITE_PREDICTION_RESULTS" >> WriteToText(
os.path.join(args.output_location,
OUTPUT_RESULTS_FILES_BASENAME_)))
# Write prediction errors counts to output files.
_ = (
errors
| "GROUP_BY_ERROR_TYPE" >> beam.combiners.Count.PerKey()
| "WRITE_ERRORS" >> WriteToText(
os.path.join(args.output_location, OUTPUT_ERRORS_FILES_BASENAME_)))
return p.run()
def run(p, args, aggregator_dict):
"""Run the pipeline with the args and dataflow pipeline option."""
# Create a PCollection for model directory.
model_dir = p | "Create Model Directory" >> beam.Create([args.model_dir])
input_file_format = args.input_file_format.lower()
input_file_patterns = args.input_file_patterns
# Setup reader.
if input_file_format == "text":
reader = p | "READ_TEXT_FILES" >> ReadFromMultiFilesText(
input_file_patterns)
elif input_file_format == "tfrecord":
reader = p | "READ_TF_FILES" >> ReadFromMultiFilesTFRecord(
input_file_patterns)
elif input_file_format == "tfrecord_gzip":
reader = p | "READ_TFGZIP_FILES" >> ReadFromMultiFilesTFRecordGZip(
input_file_patterns)
# Setup the whole pipeline.
results, errors = (reader
| "BATCH_PREDICTION" >> batch_prediction.BatchPredict(
beam.pvalue.AsSingleton(model_dir),
tags=args.tags,
signature_name=args.signature_name,
batch_size=args.batch_size,
aggregator_dict=aggregator_dict,
user_project_id=args.user_project_id,
user_job_id=args.user_job_id,
framework=args.framework))
# Convert predictions to JSON and then write to output files.
_ = (results
| "TO_JSON" >> beam.Map(json.dumps)
|
"WRITE_PREDICTION_RESULTS" >> WriteToText(args.output_result_prefix))
# Write prediction errors counts to output files.
_ = (errors
| "GROUP_BY_ERROR_TYPE" >> beam.combiners.Count.PerKey()
| "WRITE_ERRORS" >> WriteToText(args.output_error_prefix))
return p.run()
def run(p, args, aggregator_dict):
"""Run the pipeline with the args and dataflow pipeline option."""
# Create a PCollection for model directory.
model_dir = p | "Create Model Directory" >> beam.Create([args.model_dir])
input_file_format = args.input_file_format.lower()
input_file_patterns = args.input_file_patterns
# Setup reader.
if input_file_format == "json":
reader = p | "READ_TEXT_FILES" >> ReadFromMultiFilesText(
input_file_patterns)
elif input_file_format == "tfrecord":
reader = p | "READ_TF_FILES" >> ReadFromMultiFilesTFRecord(
input_file_patterns)
elif input_file_format == "tfrecord_gzip":
reader = p | "READ_TFGZIP_FILES" >> ReadFromMultiFilesTFRecordGZip(
input_file_patterns)
# Setup the whole pipeline.
results, errors = (reader
| "BATCH_PREDICTION" >> batch_prediction.BatchPredict(
beam.pvalue.AsSingleton(model_dir),
tags=args.tags,
signature_name=args.signature_name,
batch_size=args.batch_size,
aggregator_dict=aggregator_dict,
user_project_id=args.user_project_id,
user_job_id=args.user_job_id,
framework=args.framework))
output_file_format = args.output_file_format.lower()
# Convert predictions to target format and then write to output files.
if output_file_format == "json":
_ = (results
| "TO_JSON" >> beam.Map(json.dumps)
| "WRITE_PREDICTION_RESULTS" >> WriteToText(
args.output_result_prefix))
elif output_file_format == "csv":
fields = (
results
| "SAMPLE_SINGLE_ELEMENT" >> Sample.FixedSizeGlobally(1)
| "GET_KEYS" >> beam.Map(
# entry could be None if no inputs were valid
lambda entry: entry[0].keys() if entry else []))
_ = (fields
| "KEYS_TO_CSV" >> beam.Map(keys_to_csv)
| "WRITE_KEYS" >> WriteToText(args.output_result_prefix,
file_name_suffix="_header.csv",
shard_name_template=""))
_ = (results
| "VALUES_TO_CSV" >> beam.Map(values_to_csv,
beam.pvalue.AsSingleton(fields))
| "WRITE_PREDICTION_RESULTS" >> WriteToText(
args.output_result_prefix,
file_name_suffix=".csv",
append_trailing_newlines=False))
# Write prediction errors counts to output files.
_ = (errors
| "GROUP_BY_ERROR_TYPE" >> beam.combiners.Count.PerKey()
| "WRITE_ERRORS" >> WriteToText(args.output_error_prefix))
return p.run()
