def _do_inference(model_handle, examples_file, num_examples, schema, model_name):
"""Sends requests to the model and prints the results.
Args:
model_handle: handle to the model. This can be either
"mlengine:model:version" or "host:port"
examples_file: path to csv file containing examples, with the first line
assumed to have the column headers
num_examples: number of requests to send to the server
schema: a Schema describing the input data
Returns:
Response from model server
"""
filtered_features = [
feature for feature in schema.feature if feature.name != taxi.LABEL_KEY
]
del schema.feature[:]
schema.feature.extend(filtered_features)
csv_coder = taxi.make_csv_coder(schema)
proto_coder = taxi.make_proto_coder(schema)
input_file = open(examples_file, 'r')
input_file.readline() # skip header line
serialized_examples = []
for _ in range(num_examples):
one_line = input_file.readline()
if not one_line:
print('End of example file reached')
break
one_example = csv_coder.decode(one_line)
serialized_example = proto_coder.encode(one_example)
serialized_examples.append(serialized_example)
parsed_model_handle = model_handle.split(':')
if parsed_model_handle[0] == 'mlengine':
_do_mlengine_inference(
model=parsed_model_handle[1],
version=model_name,
serialized_examples=serialized_examples,
)
else:
_do_local_inference(
host=parsed_model_handle[0],
port=parsed_model_handle[1],
serialized_examples=serialized_examples,
model_name=model_name)
def _do_inference(model_handle, examples_file, num_examples, schema):
"""Sends requests to the model and prints the results.
Args:
model_handle: handle to the model. This can be either
"mlengine:model:version" or "host:port"
examples_file: path to csv file containing examples, with the first line
assumed to have the column headers
num_examples: number of requests to send to the server
schema: a Schema describing the input data
Returns:
Response from model server
"""
filtered_features = [
feature for feature in schema.feature if feature.name != taxi.LABEL_KEY
]
del schema.feature[:]
schema.feature.extend(filtered_features)
csv_coder = taxi.make_csv_coder(schema)
proto_coder = taxi.make_proto_coder(schema)
input_file = open(examples_file, 'r')
input_file.readline() # skip header line
serialized_examples = []
for _ in range(num_examples):
one_line = input_file.readline()
if not one_line:
print('End of example file reached')
break
one_example = csv_coder.decode(one_line)
serialized_example = proto_coder.encode(one_example)
serialized_examples.append(serialized_example)
parsed_model_handle = model_handle.split(':')
if parsed_model_handle[0] == 'mlengine':
_do_mlengine_inference(
model=parsed_model_handle[1],
version=parsed_model_handle[2],
serialized_examples=serialized_examples)
else:
_do_local_inference(
host=parsed_model_handle[0],
port=parsed_model_handle[1],
serialized_examples=serialized_examples)
def process_tfma(schema_file,
big_query_table=None,
eval_model_dir=None,
max_eval_rows=None,
pipeline_args=None,
publish_to_bq=False,
project=None,
metrics_table=None,
metrics_dataset=None):
"""Runs a batch job to evaluate the eval_model against the given input.
Args:
schema_file: A file containing a text-serialized Schema that describes the
eval data.
big_query_table: A BigQuery table name specified as DATASET.TABLE which
should be the input for evaluation. This can only be set if input_csv is
None.
eval_model_dir: A directory where the eval model is located.
max_eval_rows: Number of rows to query from BigQuery.
pipeline_args: additional DataflowRunner or DirectRunner args passed to
the beam pipeline.
publish_to_bq:
project:
metrics_dataset:
metrics_table:
Raises:
ValueError: if input_csv and big_query_table are not specified correctly.
"""
if big_query_table is None:
raise ValueError(
'--big_query_table should be provided.')
slice_spec = [
tfma.slicer.SingleSliceSpec(),
tfma.slicer.SingleSliceSpec(columns=['trip_start_hour'])
]
metrics_namespace = metrics_table
schema = taxi.read_schema(schema_file)
eval_shared_model = tfma.default_eval_shared_model(
eval_saved_model_path=eval_model_dir,
add_metrics_callbacks=[
tfma.post_export_metrics.calibration_plot_and_prediction_histogram(),
tfma.post_export_metrics.auc_plots()
])
metrics_monitor = None
if publish_to_bq:
metrics_monitor = MetricsReader(
publish_to_bq=publish_to_bq,
project_name=project,
bq_table=metrics_table,
bq_dataset=metrics_dataset,
filters=MetricsFilter().with_namespace(metrics_namespace)
)
pipeline = beam.Pipeline(argv=pipeline_args)
query = taxi.make_sql(big_query_table, max_eval_rows, for_eval=True)
raw_feature_spec = taxi.get_raw_feature_spec(schema)
raw_data = (
pipeline
| 'ReadBigQuery' >> ReadFromBigQuery(query=query, project=project,
use_standard_sql=True)
| 'Measure time: Start' >> beam.ParDo(MeasureTime(metrics_namespace))
| 'CleanData' >> beam.Map(lambda x: (
taxi.clean_raw_data_dict(x, raw_feature_spec))))
# Examples must be in clean tf-example format.
coder = taxi.make_proto_coder(schema)
# Prepare arguments for Extract, Evaluate and Write steps
extractors = tfma.default_extractors(
eval_shared_model=eval_shared_model,
slice_spec=slice_spec,
desired_batch_size=None,
materialize=False)
evaluators = tfma.default_evaluators(
eval_shared_model=eval_shared_model,
desired_batch_size=None,
num_bootstrap_samples=1)
_ = (
raw_data
| 'ToSerializedTFExample' >> beam.Map(coder.encode)
| 'Extract Results' >> tfma.InputsToExtracts()
| 'Extract and evaluate' >> tfma.ExtractAndEvaluate(
extractors=extractors,
evaluators=evaluators)
| 'Map Evaluations to PCollection' >> MapEvalToPCollection()
| 'Measure time: End' >> beam.ParDo(
MeasureTime(metrics_namespace))
)
result = pipeline.run()
result.wait_until_finish()
if metrics_monitor:
metrics_monitor.publish_metrics(result)
def process_tfma(eval_result_dir,
schema_file,
input_csv=None,
big_query_table=None,
eval_model_dir=None,
max_eval_rows=None,
pipeline_args=None):
"""Runs a batch job to evaluate the eval_model against the given input.
Args:
eval_result_dir: A directory where the evaluation result should be written
to.
schema_file: A file containing a text-serialized Schema that describes the
eval data.
input_csv: A path to a csv file which should be the input for evaluation.
This can only be set if big_query_table is None.
big_query_table: A BigQuery table name specified as DATASET.TABLE which
should be the input for evaluation. This can only be set if input_csv is
None.
eval_model_dir: A directory where the eval model is located.
max_eval_rows: Number of rows to query from BigQuery.
pipeline_args: additional DataflowRunner or DirectRunner args passed to the
beam pipeline.
Raises:
ValueError: if input_csv and big_query_table are not specified correctly.
"""
if input_csv == big_query_table and input_csv is None:
raise ValueError(
'one of --input_csv or --big_query_table should be provided.')
slice_spec = [
slicer.SingleSliceSpec(),
slicer.SingleSliceSpec(columns=['trip_start_hour'])
]
schema = taxi.read_schema(schema_file)
with beam.Pipeline(argv=pipeline_args) as pipeline:
if input_csv:
csv_coder = taxi.make_csv_coder(schema)
raw_data = (
pipeline
| 'ReadFromText' >> beam.io.ReadFromText(
input_csv, skip_header_lines=1)
| 'ParseCSV' >> beam.Map(csv_coder.decode))
else:
assert big_query_table
query = taxi.make_sql(big_query_table, max_eval_rows, for_eval=True)
raw_feature_spec = taxi.get_raw_feature_spec(schema)
raw_data = (
pipeline
| 'ReadBigQuery' >> beam.io.Read(
beam.io.BigQuerySource(query=query, use_standard_sql=True))
| 'CleanData' >>
beam.Map(lambda x: (taxi.clean_raw_data_dict(x, raw_feature_spec))))
# Examples must be in clean tf-example format.
coder = taxi.make_proto_coder(schema)
_ = (
raw_data
| 'ToSerializedTFExample' >> beam.Map(coder.encode)
| 'EvaluateAndWriteResults' >> tfma.EvaluateAndWriteResults(
eval_saved_model_path=eval_model_dir,
slice_spec=slice_spec,
add_metrics_callbacks=[
post_export_metrics.calibration_plot_and_prediction_histogram(),
post_export_metrics.auc_plots()
],
output_path=eval_result_dir))
def process_tfma(eval_result_dir,
schema_file,
input_csv=None,
big_query_table=None,
eval_model_dir=None,
max_eval_rows=None,
pipeline_args=None):
"""Runs a batch job to evaluate the eval_model against the given input.
Args:
eval_result_dir: A directory where the evaluation result should be written
to.
schema_file: A file containing a text-serialized Schema that describes the
eval data.
input_csv: A path to a csv file which should be the input for evaluation.
This can only be set if big_query_table is None.
big_query_table: A BigQuery table name specified as DATASET.TABLE which
should be the input for evaluation. This can only be set if input_csv is
None.
eval_model_dir: A directory where the eval model is located.
max_eval_rows: Number of rows to query from BigQuery.
pipeline_args: additional DataflowRunner or DirectRunner args passed to the
beam pipeline.
Raises:
ValueError: if input_csv and big_query_table are not specified correctly.
"""
if input_csv == big_query_table and input_csv is None:
raise ValueError(
'one of --input_csv or --big_query_table should be provided.')
slice_spec = [
tfma.slicer.SingleSliceSpec(),
tfma.slicer.SingleSliceSpec(columns=['trip_start_hour'])
]
schema = taxi.read_schema(schema_file)
eval_shared_model = tfma.default_eval_shared_model(
eval_saved_model_path=eval_model_dir,
add_metrics_callbacks=[
tfma.post_export_metrics.calibration_plot_and_prediction_histogram(),
tfma.post_export_metrics.auc_plots()
])
with beam.Pipeline(argv=pipeline_args) as pipeline:
if input_csv:
csv_coder = taxi.make_csv_coder(schema)
raw_data = (
pipeline
| 'ReadFromText' >> beam.io.ReadFromText(
input_csv, skip_header_lines=1)
| 'ParseCSV' >> beam.Map(csv_coder.decode))
else:
assert big_query_table
query = taxi.make_sql(big_query_table, max_eval_rows, for_eval=True)
raw_feature_spec = taxi.get_raw_feature_spec(schema)
raw_data = (
pipeline
| 'ReadBigQuery' >> beam.io.Read(
beam.io.BigQuerySource(query=query, use_standard_sql=True))
| 'CleanData' >>
beam.Map(lambda x: (taxi.clean_raw_data_dict(x, raw_feature_spec))))
# Examples must be in clean tf-example format.
coder = taxi.make_proto_coder(schema)
_ = (
raw_data
| 'ToSerializedTFExample' >> beam.Map(coder.encode)
|
'ExtractEvaluateAndWriteResults' >> tfma.ExtractEvaluateAndWriteResults(
eval_shared_model=eval_shared_model,
slice_spec=slice_spec,
output_path=eval_result_dir))
