def _init_model(self, multi_model, validation):
# The benchmark runner will instantiate this class twice - once to determine
# the benchmarks to run, and once to actually to run them. However, Keras
# freezes if we try to load the same model twice. As such, we have to pull
# the model loading out of the constructor into a separate method which we
# call before each benchmark.
if multi_model:
metric_specs = metric_specs_util.specs_from_metrics(
[tf.keras.metrics.AUC(name="auc", num_thresholds=10000)],
model_names=["candidate", "baseline"])
if validation:
# Only one metric, adding a threshold for all slices.
metric_specs[0].metrics[0].threshold.CopyFrom(
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={"value": 0.5},
upper_bound={"value": 0.5}),
change_threshold=tfma.GenericChangeThreshold(
absolute={"value": -0.001},
direction=tfma.MetricDirection.HIGHER_IS_BETTER)))
self._eval_config = tfma.EvalConfig(model_specs=[
tfma.ModelSpec(name="candidate", label_key="tips"),
tfma.ModelSpec(name="baseline",
label_key="tips",
is_baseline=True)
],
metrics_specs=metric_specs)
self._eval_shared_models = {
"candidate":
tfma.default_eval_shared_model(
self._dataset.trained_saved_model_path(),
eval_config=self._eval_config,
model_name="candidate"),
"baseline":
tfma.default_eval_shared_model(
self._dataset.trained_saved_model_path(),
eval_config=self._eval_config,
model_name="baseline")
}
else:
metric_specs = metric_specs_util.specs_from_metrics(
[tf.keras.metrics.AUC(name="auc", num_thresholds=10000)])
if validation:
# Only one metric, adding a threshold for all slices.
metric_specs[0].metrics[0].threshold.CopyFrom(
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={"value": 0.5},
upper_bound={"value": 0.5})))
self._eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key="tips")],
metrics_specs=metric_specs)
self._eval_shared_models = {
"":
tfma.default_eval_shared_model(
self._dataset.trained_saved_model_path(),
eval_config=self._eval_config)
}
def _get_eval_config() -> tfma.EvalConfig:
return tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key=LABEL_KEY)],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.01}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': 1e-10})))
])
])
def eval_metrics_spec(self, eval_accuracy_threshold):
class_name = (
"BinaryAccuracy"
if self.vocab_size + self.out_of_vocab_buckets == 2
else "SparseCategoricalAccuracy"
)
return tfma.MetricsSpec(
metrics=[
tfma.MetricConfig(
class_name=class_name,
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={"value": eval_accuracy_threshold}
),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={"value": -1e-10},
),
),
)
]
)
def get_accuracy_eval_config(accuracy_threshold):
accuracy_threshold = tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': accuracy_threshold},
upper_bound={'value': 0.99}),
change_threshold=tfma.GenericChangeThreshold(
absolute={'value': 0.0001},
direction=tfma.MetricDirection.HIGHER_IS_BETTER))
metrics_specs = tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='BinaryAccuracy',
threshold=accuracy_threshold),
tfma.MetricConfig(class_name='ExampleCount')
])
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='income_bracket')],
metrics_specs=[metrics_specs],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['occupation'])
])
return eval_config
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir: Text,
metadata_path: Text,
beam_pipeline_args: List[Text]) -> pipeline.Pipeline:
"""Implements the imdb sentiment analysis pipline with TFX."""
output = example_gen_pb2.Output(split_config=example_gen_pb2.SplitConfig(
splits=[
example_gen_pb2.SplitConfig.Split(name='train', hash_buckets=9),
example_gen_pb2.SplitConfig.Split(name='eval', hash_buckets=1)
]))
# Brings data in to the pipline
example_gen = CsvExampleGen(input_base=data_root, output_config=output)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that trains a model.
trainer = Trainer(module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=trainer_pb2.TrainArgs(num_steps=500),
eval_args=trainer_pb2.EvalArgs(num_steps=200))
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='label')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
# Increase this threshold when training on complete
# dataset.
lower_bound={'value': 0.01}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-2})))
])
])
evaluator = Evaluator(examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
components = [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
enable_cache=True,
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir: Text,
metadata_path: Text,
direct_num_workers: int) -> pipeline.Pipeline:
"""Implements the Iris flowers pipeline with TFX."""
examples = external_input(data_root)
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=examples)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that trains a model using TF-Learn.
trainer = Trainer(
module_file=module_file,
custom_executor_spec=executor_spec.ExecutorClassSpec(GenericExecutor),
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=trainer_pb2.TrainArgs(num_steps=2000),
eval_args=trainer_pb2.EvalArgs(num_steps=5))
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute an evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='variety')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
# Performs infra validation of a candidate model to prevent unservable model
# from being pushed. This config will launch a model server of the latest
# TensorFlow Serving image in a local docker engine.
infra_validator = InfraValidator(
model=trainer.outputs['model'],
examples=example_gen.outputs['examples'],
serving_spec=infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(
tags=['latest']),
local_docker=infra_validator_pb2.LocalDockerConfig()),
request_spec=infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.
TensorFlowServingRequestSpec()))
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
infra_blessing=infra_validator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
infra_validator,
pusher,
],
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
# TODO(b/142684737): The multi-processing API might change.
beam_pipeline_args=['--direct_num_workers=%d' % direct_num_workers],
)
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_root: Text,
module_file: Text,
ai_platform_training_args: Dict[Text, Text],
ai_platform_serving_args: Dict[Text, Text],
enable_tuning: bool,
beam_pipeline_args: List[Text],
) -> tfx.dsl.Pipeline:
"""Implements the penguin pipeline with TFX and Kubeflow Pipeline.
Args:
pipeline_name: name of the TFX pipeline being created.
pipeline_root: root directory of the pipeline. Should be a valid GCS path.
data_root: uri of the penguin data.
module_file: uri of the module files used in Trainer and Transform
components.
ai_platform_training_args: Args of CAIP training job. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.jobs#Job
for detailed description.
ai_platform_serving_args: Args of CAIP model deployment. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.models
for detailed description.
enable_tuning: If True, the hyperparameter tuning through CloudTuner is
enabled.
beam_pipeline_args: List of beam pipeline options. Please refer to
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options.
Returns:
A TFX pipeline object.
"""
# Number of epochs in training.
train_steps = tfx.dsl.RuntimeParameter(
name='train_steps',
default=100,
ptype=int,
)
# Number of epochs in evaluation.
eval_steps = tfx.dsl.RuntimeParameter(
name='eval_steps',
default=50,
ptype=int,
)
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = tfx.components.CsvExampleGen(
input_base=os.path.join(data_root, 'labelled'))
# Computes statistics over data for visualization and example validation.
statistics_gen = tfx.components.StatisticsGen(
examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = tfx.components.ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = tfx.components.Transform(
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
# The Tuner component launches 1 AIP Training job for flock management of
# parallel tuning. For example, 2 workers (defined by num_parallel_trials)
# in the flock management AIP Training job, each runs a search loop for
# trials as shown below.
# Tuner component -> CAIP job X -> CloudTunerA -> tuning trials
# -> CloudTunerB -> tuning trials
#
# Distributed training for each trial depends on the Tuner
# (kerastuner.BaseTuner) setup in tuner_fn. Currently CloudTuner is single
# worker training per trial. DistributingCloudTuner (a subclass of
# CloudTuner) launches remote distributed training job per trial.
#
# E.g., single worker training per trial
# ... -> CloudTunerA -> single worker training
# -> CloudTunerB -> single worker training
# vs distributed training per trial
# ... -> DistributingCloudTunerA -> CAIP job Y -> master,worker1,2,3
# -> DistributingCloudTunerB -> CAIP job Z -> master,worker1,2,3
tuner = tfx.extensions.google_cloud_ai_platform.Tuner(
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
tune_args=tfx.proto.TuneArgs(
# num_parallel_trials=3 means that 3 search loops are
# running in parallel.
num_parallel_trials=3),
custom_config={
# Note that this TUNING_ARGS_KEY will be used to start the CAIP job
# for parallel tuning (CAIP job X above).
#
# num_parallel_trials will be used to fill/overwrite the
# workerCount specified by TUNING_ARGS_KEY:
# num_parallel_trials = workerCount + 1 (for master)
tfx.extensions.google_cloud_ai_platform.experimental.TUNING_ARGS_KEY:
ai_platform_training_args,
# This working directory has to be a valid GCS path and will be used
# to launch remote training job per trial.
tfx.extensions.google_cloud_ai_platform.experimental.REMOTE_TRIALS_WORKING_DIR_KEY:
os.path.join(_pipeline_root, 'trials'),
})
# Uses user-provided Python function that trains a model.
trainer = tfx.extensions.google_cloud_ai_platform.Trainer(
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
# If Tuner is in the pipeline, Trainer can take Tuner's output
# best_hyperparameters artifact as input and utilize it in the user module
# code.
#
# If there isn't Tuner in the pipeline, either use ImporterNode to import
# a previous Tuner's output to feed to Trainer, or directly use the tuned
# hyperparameters in user module code and set hyperparameters to None
# here.
#
# Example of ImporterNode,
# hparams_importer = ImporterNode(
# source_uri='path/to/best_hyperparameters.txt',
# artifact_type=HyperParameters).with_id('import_hparams')
# ...
# hyperparameters = hparams_importer.outputs['result'],
hyperparameters=(tuner.outputs['best_hyperparameters']
if enable_tuning else None),
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
custom_config={
tfx.extensions.google_cloud_ai_platform.TRAINING_ARGS_KEY:
ai_platform_training_args
})
# Get the latest blessed model for model validation.
model_resolver = tfx.dsl.Resolver(
strategy_class=tfx.dsl.experimental.LatestBlessedModelStrategy,
model=tfx.dsl.Channel(type=tfx.types.standard_artifacts.Model),
model_blessing=tfx.dsl.Channel(
type=tfx.types.standard_artifacts.ModelBlessing)).with_id(
'latest_blessed_model_resolver')
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='species')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = tfx.components.Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
pusher = tfx.extensions.google_cloud_ai_platform.Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
custom_config={
tfx.extensions.google_cloud_ai_platform.experimental.PUSHER_SERVING_ARGS_KEY:
ai_platform_serving_args
})
components = [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
if enable_tuning:
components.append(tuner)
return tfx.dsl.Pipeline(pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=True,
beam_pipeline_args=beam_pipeline_args)
def create_pipeline(pipeline_name: str, pipeline_root: str, data_root: str,
module_file: str, serving_model_dir: str,
beam_pipeline_args: List[str]) -> pipeline.Pipeline:
"""Implements the chicago taxi pipeline with TFX."""
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input_base=data_root)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that implements a model.
trainer = Trainer(
module_file=module_file,
transformed_examples=transform.outputs['transformed_examples'],
schema=schema_gen.outputs['schema'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=10000),
eval_args=trainer_pb2.EvalArgs(num_steps=5000))
# Get the latest blessed model for model validation.
model_resolver = resolver.Resolver(
strategy_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(
type=ModelBlessing)).with_id('latest_blessed_model_resolver')
# Uses TFMA to compute a evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(signature_name='eval')],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
],
metrics_specs=[
tfma.MetricsSpec(
thresholds={
'accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
])
evaluator = Evaluator(examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
config = kubernetes_dag_runner.get_default_kubernetes_metadata_config()
return pipeline.Pipeline(pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
],
enable_cache=False,
metadata_connection_config=config,
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_root: Text,
module_file: Text,
accuracy_threshold: float,
serving_model_dir: Text,
metadata_path: Text,
user_provided_schema_path: Optional[Text],
enable_tuning: bool,
enable_bulk_inferrer: bool,
examplegen_input_config: Optional[tfx.proto.Input],
examplegen_range_config: Optional[tfx.proto.RangeConfig],
resolver_range_config: Optional[tfx.proto.RangeConfig],
beam_pipeline_args: List[Text],
) -> tfx.dsl.Pipeline:
"""Implements the penguin pipeline with TFX.
Args:
pipeline_name: name of the TFX pipeline being created.
pipeline_root: root directory of the pipeline.
data_root: directory containing the penguin data.
module_file: path to files used in Trainer and Transform components.
accuracy_threshold: minimum accuracy to push the model.
serving_model_dir: filepath to write pipeline SavedModel to.
metadata_path: path to local pipeline ML Metadata store.
user_provided_schema_path: path to user provided schema file.
enable_tuning: If True, the hyperparameter tuning through KerasTuner is
enabled.
enable_bulk_inferrer: If True, the generated model will be used for a
batch inference.
examplegen_input_config: ExampleGen's input_config.
examplegen_range_config: ExampleGen's range_config.
resolver_range_config: SpansResolver's range_config. Specify this will
enable SpansResolver to get a window of ExampleGen's output Spans for
transform and training.
beam_pipeline_args: list of beam pipeline options for LocalDAGRunner. Please
refer to https://beam.apache.org/documentation/runners/direct/.
Returns:
A TFX pipeline object.
"""
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = tfx.components.CsvExampleGen(
input_base=os.path.join(data_root, 'labelled'),
input_config=examplegen_input_config,
range_config=examplegen_range_config)
# Computes statistics over data for visualization and example validation.
statistics_gen = tfx.components.StatisticsGen(
examples=example_gen.outputs['examples'])
if user_provided_schema_path:
# Import user-provided schema.
schema_importer = tfx.dsl.Importer(
source_uri=user_provided_schema_path,
artifact_type=tfx.types.standard_artifacts.Schema).with_id(
'schema_importer')
schema = schema_importer.outputs['result']
else:
# Generates schema based on statistics files.
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
schema = schema_gen.outputs['schema']
# Performs anomaly detection based on statistics and data schema.
example_validator = tfx.components.ExampleValidator(
statistics=statistics_gen.outputs['statistics'], schema=schema)
# Gets multiple Spans for transform and training.
if resolver_range_config:
examples_resolver = tfx.dsl.Resolver(
strategy_class=tfx.dsl.experimental.SpanRangeStrategy,
config={
'range_config': resolver_range_config
},
examples=tfx.dsl.Channel(
type=tfx.types.standard_artifacts.Examples,
producer_component_id=example_gen.id)).with_id('span_resolver')
# Performs transformations and feature engineering in training and serving.
transform = tfx.components.Transform(
examples=(examples_resolver.outputs['examples'] if
resolver_range_config else example_gen.outputs['examples']),
schema=schema,
module_file=module_file)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
tuner = tfx.components.Tuner(
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
train_args=tfx.proto.TrainArgs(num_steps=20),
eval_args=tfx.proto.EvalArgs(num_steps=5))
# Uses user-provided Python function that trains a model.
trainer = tfx.components.Trainer(
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema,
# If Tuner is in the pipeline, Trainer can take Tuner's output
# best_hyperparameters artifact as input and utilize it in the user module
# code.
#
# If there isn't Tuner in the pipeline, either use ImporterNode to import
# a previous Tuner's output to feed to Trainer, or directly use the tuned
# hyperparameters in user module code and set hyperparameters to None
# here.
#
# Example of ImporterNode,
# hparams_importer = ImporterNode(
# source_uri='path/to/best_hyperparameters.txt',
# artifact_type=HyperParameters).with_id('import_hparams')
# ...
# hyperparameters = hparams_importer.outputs['result'],
hyperparameters=(tuner.outputs['best_hyperparameters']
if enable_tuning else None),
train_args=tfx.proto.TrainArgs(num_steps=100),
eval_args=tfx.proto.EvalArgs(num_steps=5))
# Get the latest blessed model for model validation.
model_resolver = tfx.dsl.Resolver(
strategy_class=tfx.dsl.experimental.LatestBlessedModelStrategy,
model=tfx.dsl.Channel(type=tfx.types.standard_artifacts.Model),
model_blessing=tfx.dsl.Channel(
type=tfx.types.standard_artifacts.ModelBlessing)).with_id(
'latest_blessed_model_resolver')
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='species')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': accuracy_threshold}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = tfx.components.Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = tfx.components.Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=tfx.proto.PushDestination(
filesystem=tfx.proto.PushDestination.Filesystem(
base_directory=serving_model_dir)))
# Showcase for BulkInferrer component.
if enable_bulk_inferrer:
# Generates unlabelled examples.
example_gen_unlabelled = tfx.components.CsvExampleGen(
input_base=os.path.join(data_root, 'unlabelled')).with_id(
'CsvExampleGen_Unlabelled')
# Performs offline batch inference.
bulk_inferrer = tfx.components.BulkInferrer(
examples=example_gen_unlabelled.outputs['examples'],
model=trainer.outputs['model'],
# Empty data_spec.example_splits will result in using all splits.
data_spec=tfx.proto.DataSpec(),
model_spec=tfx.proto.ModelSpec())
components_list = [
example_gen,
statistics_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
if user_provided_schema_path:
components_list.append(schema_importer)
else:
components_list.append(schema_gen)
if resolver_range_config:
components_list.append(examples_resolver)
if enable_tuning:
components_list.append(tuner)
if enable_bulk_inferrer:
components_list.append(example_gen_unlabelled)
components_list.append(bulk_inferrer)
return tfx.dsl.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components_list,
enable_cache=True,
metadata_connection_config=tfx.orchestration.metadata.
sqlite_metadata_connection_config(metadata_path),
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(pipeline_root: Text,
csv_input_location: data_types.RuntimeParameter,
taxi_module_file: data_types.RuntimeParameter,
enable_cache: bool):
"""Creates a simple Kubeflow-based Chicago Taxi TFX pipeline.
Args:
pipeline_root: The root of the pipeline output.
csv_input_location: The location of the input data directory.
taxi_module_file: The location of the module file for Transform/Trainer.
enable_cache: Whether to enable cache or not.
Returns:
A logical TFX pipeline.Pipeline object.
"""
examples = external_input(csv_input_location)
example_gen = CsvExampleGen(input=examples)
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
infer_schema = SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False,
)
validate_stats = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=infer_schema.outputs['schema'],
)
transform = Transform(
examples=example_gen.outputs['examples'],
schema=infer_schema.outputs['schema'],
module_file=taxi_module_file,
)
trainer = Trainer(
module_file=taxi_module_file,
transformed_examples=transform.outputs['transformed_examples'],
schema=infer_schema.outputs['schema'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=10),
eval_args=trainer_pb2.EvalArgs(num_steps=5),
)
# Set the TFMA config for Model Evaluation and Validation.
eval_config = tfma.EvalConfig(
model_specs=[
# Using signature 'eval' implies the use of an EvalSavedModel. To use
# a serving model remove the signature to defaults to 'serving_default'
# and add a label_key.
tfma.ModelSpec(signature_name='eval')
],
metrics_specs=[
tfma.MetricsSpec(
# The metrics added here are in addition to those saved with the
# model (assuming either a keras model or EvalSavedModel is used).
# Any metrics added into the saved model (for example using
# model.compile(..., metrics=[...]), etc) will be computed
# automatically.
metrics=[tfma.MetricConfig(class_name='ExampleCount')],
# To add validation thresholds for metrics saved with the model,
# add them keyed by metric name to the thresholds map.
thresholds={
'binary_accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
],
slicing_specs=[
# An empty slice spec means the overall slice, i.e. the whole dataset.
tfma.SlicingSpec(),
# Data can be sliced along a feature column. In this case, data is
# sliced along feature column trip_start_hour.
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
])
model_analyzer = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
eval_config=eval_config,
)
pusher = Pusher(
model=trainer.outputs['model'],
model_blessing=model_analyzer.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=os.path.join(str(pipeline.ROOT_PARAMETER),
'model_serving'))),
)
return pipeline.Pipeline(
pipeline_name='parameterized_tfx_oss',
pipeline_root=pipeline_root,
components=[
example_gen, statistics_gen, infer_schema, validate_stats,
transform, trainer, model_analyzer, pusher
],
enable_cache=enable_cache,
)
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_path: Text,
preprocessing_fn: Text,
run_fn: Text,
train_args: trainer_pb2.TrainArgs,
eval_args: trainer_pb2.EvalArgs,
eval_accuracy_threshold: float,
serving_model_dir: Text,
metadata_connection_config: Optional[
metadata_store_pb2.ConnectionConfig
] = None,
beam_pipeline_args: Optional[List[Text]] = None,
ai_platform_training_args: Optional[Dict[Text, Text]] = None,
ai_platform_serving_args: Optional[Dict[Text, Any]] = None,
) -> pipeline.Pipeline:
"""Implements the complaint prediction pipeline with TFX."""
components = []
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=external_input(data_path))
components.append(example_gen)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs["examples"])
components.append(statistics_gen)
# Generates schema based on statistics files.
schema_gen = SchemaGen(
statistics=statistics_gen.outputs["statistics"],
infer_feature_shape=True,
)
components.append(schema_gen)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator( # pylint: disable=unused-variable
statistics=statistics_gen.outputs["statistics"],
schema=schema_gen.outputs["schema"],
)
components.append(example_validator)
# Performs transformations and feature engineering in training and serving.
transform = Transform(
examples=example_gen.outputs["examples"],
schema=schema_gen.outputs["schema"],
preprocessing_fn=preprocessing_fn,
)
components.append(transform)
# Uses user-provided Python function that implements a model using TF-Learn.
trainer_args = {
"run_fn": run_fn,
"transformed_examples": transform.outputs["transformed_examples"],
"schema": schema_gen.outputs["schema"],
"transform_graph": transform.outputs["transform_graph"],
"train_args": train_args,
"eval_args": eval_args,
"custom_executor_spec": executor_spec.ExecutorClassSpec(
trainer_executor.GenericExecutor
),
}
if ai_platform_training_args is not None:
trainer_args.update(
{
"custom_executor_spec": executor_spec.ExecutorClassSpec(
ai_platform_trainer_executor.GenericExecutor
),
"custom_config": {
ai_platform_trainer_executor.TRAINING_ARGS_KEY: ai_platform_training_args, # noqa
},
}
)
trainer = Trainer(**trainer_args)
components.append(trainer)
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name="latest_blessed_model_resolver",
resolver_class=latest_blessed_model_resolver.LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing),
)
components.append(model_resolver)
# Uses TFMA to compute a evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key="big_tipper")],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(
metrics=[
tfma.MetricConfig(
class_name="BinaryAccuracy",
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={"value": eval_accuracy_threshold}
),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={"value": -1e-10},
),
),
)
]
)
],
)
evaluator = Evaluator(
examples=example_gen.outputs["examples"],
model=trainer.outputs["model"],
baseline_model=model_resolver.outputs["model"],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config,
)
components.append(evaluator)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher_args = {
"model": trainer.outputs["model"],
"model_blessing": evaluator.outputs["blessing"],
"push_destination": pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir
)
),
}
if ai_platform_serving_args is not None:
pusher_args.update(
{
"custom_executor_spec": executor_spec.ExecutorClassSpec(
ai_platform_pusher_executor.Executor
),
"custom_config": {
ai_platform_pusher_executor.SERVING_ARGS_KEY: ai_platform_serving_args # noqa
},
}
)
pusher = Pusher(**pusher_args) # pylint: disable=unused-variable
components.append(pusher)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=True,
metadata_connection_config=metadata_connection_config,
beam_pipeline_args=beam_pipeline_args,
)
def _create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_root: Text,
transform_module_file: Text,
train_module_file: Text,
serving_model_dir: Text,
direct_num_workers: int,
) -> pipeline.Pipeline:
# Component 1: Data Ingestion
example_gen = CsvExampleGen(input_base=data_root)
# Component 2: Statistics
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'], )
# Component 3: Schema
schema_gen = SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False,
)
# Component 4: Data Validator
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'],
)
# Component 5: Transform (Feature Engineering)
transform = Transform(
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=transform_module_file,
)
# Component 6: Trainer
trainer = Trainer(
module_file=train_module_file,
custom_executor_spec=executor_spec.ExecutorClassSpec(GenericExecutor),
examples=transform.outputs['transformed_examples'],
schema=schema_gen.outputs['schema'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=10000),
eval_args=trainer_pb2.EvalArgs(num_steps=5000),
)
# Component 7: Evaluate
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing),
)
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='tips')],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='ExampleCount'),
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))),
], ),
],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['trip_start_hour']),
],
)
evaluator = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config,
)
# Component 8: Push
pusher = Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=_serving_model_dir, ), ),
)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen, statistics_gen, schema_gen, example_validator,
transform, trainer, model_resolver, evaluator, pusher
],
beam_pipeline_args=[f'--direct_num_workers={direct_num_workers}'],
enable_cache=True,
)
def create_pipeline(pipeline_name: Text,
pipeline_root: Text,
data_root_uri,
trainer_config: TrainerConfig,
tuner_config: TunerConfig,
pusher_config: PusherConfig,
runtime_parameters_config: RuntimeParametersConfig = None,
str_runtime_parameters_supported = False,
int_runtime_parameters_supported = False,
local_run: bool = True,
beam_pipeline_args: Optional[List[Text]] = None,
enable_cache: Optional[bool] = True,
code_folder = '',
metadata_connection_config: Optional[metadata_store_pb2.ConnectionConfig] = None
) -> pipeline.Pipeline:
"""Trains and deploys the Keras Titanic Classifier with TFX and Kubeflow Pipeline on Google Cloud.
Args:
pipeline_name: name of the TFX pipeline being created.
pipeline_root: root directory of the pipeline. Should be a valid GCS path.
data_root_uri: uri of the dataset.
train_steps: runtime parameter for number of model training steps for the Trainer component.
eval_steps: runtime parameter for number of model evaluation steps for the Trainer component.
enable_tuning: If True, the hyperparameter tuning through CloudTuner is
enabled.
ai_platform_training_args: Args of CAIP training job. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.jobs#Job
for detailed description.
ai_platform_serving_args: Args of CAIP model deployment. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.models
for detailed description.
beam_pipeline_args: Optional list of beam pipeline options. Please refer to
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options.
When this argument is not provided, the default is to use GCP
DataflowRunner with 50GB disk size as specified in this function. If an
empty list is passed in, default specified by Beam will be used, which can
be found at
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options
enable_cache: Optional boolean
Returns:
A TFX pipeline object.
"""
#pydevd_pycharm.settrace('localhost', port=9091, stdoutToServer=True, stderrToServer=True)
absl.logging.info('pipeline_name: %s' % pipeline_name)
absl.logging.info('pipeline root: %s' % pipeline_root)
absl.logging.info('data_root_uri for training: %s' % data_root_uri)
absl.logging.info('train_steps for training: %s' % trainer_config.train_steps)
absl.logging.info('tuner_steps for tuning: %s' % tuner_config.tuner_steps)
absl.logging.info('eval_steps for evaluating: %s' % trainer_config.eval_steps)
absl.logging.info('os default list dir: %s' % os.listdir('.'))
schema_proper_folder = os.path.join(os.sep, code_folder, SCHEMA_FOLDER)
absl.logging.info('schema_proper_folder: %s' % schema_proper_folder)
preprocessing_proper_file = os.path.join(os.sep, code_folder, TRANSFORM_MODULE_FILE)
absl.logging.info('preprocessing_proper_file: %s' % preprocessing_proper_file)
model_proper_file = os.path.join(os.sep, code_folder, TRAIN_MODULE_FILE)
absl.logging.info('model_proper_file: %s' % model_proper_file)
hyperparameters_proper_folder = os.path.join(os.sep, code_folder, HYPERPARAMETERS_FOLDER)
absl.logging.info('hyperparameters_proper_folder: %s' % hyperparameters_proper_folder)
# Brings data into the pipeline and splits the data into training and eval splits
output_config = example_gen_pb2.Output(
split_config=example_gen_pb2.SplitConfig(splits=[
example_gen_pb2.SplitConfig.Split(name='train', hash_buckets=4),
example_gen_pb2.SplitConfig.Split(name='eval', hash_buckets=1)
]))
# examples = external_input(data_root_uri)
if str_runtime_parameters_supported and runtime_parameters_config is not None:
data_root_uri = runtime_parameters_config.data_root_runtime
examplegen = CsvExampleGen(input_base=data_root_uri, output_config=output_config)
# examplegen = CsvExampleGen(input_base=data_root_uri)
# Computes statistics over data for visualization and example validation.
statisticsgen = StatisticsGen(examples=examplegen.outputs.examples)
# Generates schema based on statistics files. Even though, we use user-provided schema
# we still want to generate the schema of the newest data for tracking and comparison
schemagen = SchemaGen(statistics=statisticsgen.outputs.statistics)
# Import a user-provided schema
import_schema = Importer(
source_uri=schema_proper_folder,
artifact_type=Schema).with_id('import_user_schema')
# Performs anomaly detection based on statistics and data schema.
examplevalidator = ExampleValidator(
statistics=statisticsgen.outputs.statistics,
schema=import_schema.outputs.result)
# Performs transformations and feature engineering in training and serving.
transform = Transform(
examples=examplegen.outputs.examples,
schema=import_schema.outputs.result,
module_file=preprocessing_proper_file)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
hparams_importer = Importer(
source_uri=hyperparameters_proper_folder,
artifact_type=HyperParameters).with_id('import_hparams')
# apparently only str RuntimeParameters are supported in airflow :/
if int_runtime_parameters_supported and runtime_parameters_config is not None:
train_steps = runtime_parameters_config.train_steps_runtime
eval_steps = runtime_parameters_config.eval_steps_runtime
else:
train_steps = trainer_config.train_steps
eval_steps = trainer_config.eval_steps
absl.logging.info('train_steps: %s' % train_steps)
absl.logging.info('eval_steps: %s' % eval_steps)
if tuner_config.enable_tuning:
tuner_args = {
'module_file': model_proper_file,
'examples': transform.outputs.transformed_examples,
'transform_graph': transform.outputs.transform_graph,
'train_args': {'num_steps': tuner_config.tuner_steps},
'eval_args': {'num_steps': tuner_config.eval_tuner_steps},
'custom_config': {'max_trials': tuner_config.max_trials, 'is_local_run': local_run}
# 'tune_args': tuner_pb2.TuneArgs(num_parallel_trials=3),
}
if tuner_config.ai_platform_tuner_args is not None:
tuner_args.update({
'custom_config': {
ai_platform_trainer_executor.TRAINING_ARGS_KEY: tuner_config.ai_platform_tuner_args
},
'tune_args': tuner_pb2.TuneArgs(num_parallel_trials=3)
})
absl.logging.info("tuner_args: " + str(tuner_args))
tuner = Tuner(**tuner_args)
hyperparameters = tuner.outputs.best_hyperparameters if tuner_config.enable_tuning else hparams_importer.outputs['result']
# Trains the model using a user provided trainer function.
trainer_args = {
'module_file': model_proper_file,
'transformed_examples': transform.outputs.transformed_examples,
'schema': import_schema.outputs.result,
'transform_graph': transform.outputs.transform_graph,
# train_args={'num_steps': train_steps},
'train_args': {'num_steps': train_steps},
'eval_args': {'num_steps': eval_steps},
#'hyperparameters': tuner.outputs.best_hyperparameters if tunerConfig.enable_tuning else None,
'hyperparameters': hyperparameters,
'custom_config': {'epochs': trainer_config.epochs, 'train_batch_size': trainer_config.train_batch_size,
'eval_batch_size': trainer_config.eval_batch_size,
}
}
if trainer_config.ai_platform_training_args is not None:
trainer_args['custom_config'].update({
ai_platform_trainer_executor.TRAINING_ARGS_KEY:
trainer_config.ai_platform_training_args,
})
trainer_args.update({
'custom_executor_spec':
executor_spec.ExecutorClassSpec(ai_platform_trainer_executor.GenericExecutor),
# 'custom_config': {
# ai_platform_trainer_executor.TRAINING_ARGS_KEY:
# ai_platform_training_args,
# }
})
else:
trainer_args.update({
'custom_executor_spec':
executor_spec.ExecutorClassSpec(trainer_executor.GenericExecutor),
#executor_spec.ExecutorClassSpec(custom_trainer_executor.CustomGenericExecutor), # for debugging purposes
})
trainer = Trainer(**trainer_args)
# Get the latest blessed model for model validation.
model_resolver = resolver.Resolver(
#instance_name='latest_blessed_model_resolver', # instance_name is deprecated, use with_id()
strategy_class=latest_blessed_model_resolver.LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing)\
).with_id('latest_blessed_model_resolver')
# Uses TFMA to compute a evaluation statistics over features of a model.
accuracy_threshold = tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5},
upper_bound={'value': 0.995}),
)
metrics_specs = tfma.MetricsSpec(
metrics=[
tfma.MetricConfig(class_name='BinaryAccuracy',
threshold=accuracy_threshold),
tfma.MetricConfig(class_name='ExampleCount')])
eval_config = tfma.EvalConfig(
model_specs=[
tfma.ModelSpec(label_key='Survived')
],
metrics_specs=[metrics_specs],
slicing_specs=[
tfma.SlicingSpec()
,tfma.SlicingSpec(feature_keys=['Sex'])
,tfma.SlicingSpec(feature_keys=['Age'])
,tfma.SlicingSpec(feature_keys=['Parch'])
]
)
evaluator = Evaluator(
examples=examplegen.outputs.examples,
model=trainer.outputs.model,
baseline_model=model_resolver.outputs.model,
eval_config=eval_config
)
# Validate model can be loaded and queried in sand-boxed environment
# mirroring production.
serving_config = None
if local_run:
serving_config = infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(tags=['latest']),
local_docker=infra_validator_pb2.LocalDockerConfig() # Running on local docker.
)
else:
serving_config = infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(tags=['latest']),
kubernetes=infra_validator_pb2.KubernetesConfig() # Running on K8s.
)
validation_config = infra_validator_pb2.ValidationSpec(
max_loading_time_seconds=60,
num_tries=3,
)
request_config = infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServingRequestSpec(),
num_examples=3,
)
infravalidator = InfraValidator(
model=trainer.outputs.model,
examples=examplegen.outputs.examples,
serving_spec=serving_config,
validation_spec=validation_config,
request_spec=request_config,
)
# Checks whether the model passed the validation steps and pushes the model
# to CAIP Prediction if checks are passed.
pusher_args = {
'model': trainer.outputs.model,
'model_blessing': evaluator.outputs.blessing,
'infra_blessing': infravalidator.outputs.blessing
}
if local_run:
pusher_args.update({'push_destination':
pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=pusher_config.serving_model_dir))})
if pusher_config.ai_platform_serving_args is not None:
pusher_args.update({
'custom_executor_spec':
executor_spec.ExecutorClassSpec(ai_platform_pusher_executor.Executor
),
'custom_config': {
ai_platform_pusher_executor.SERVING_ARGS_KEY:
pusher_config.ai_platform_serving_args
},
})
pusher = Pusher(**pusher_args) # pylint: disable=unused-variable
components = [
examplegen,
statisticsgen,
schemagen,
import_schema,
examplevalidator,
transform,
trainer,
model_resolver,
evaluator,
infravalidator,
pusher
]
if tuner_config.enable_tuning:
components.append(tuner)
else:
components.append(hparams_importer)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=enable_cache,
metadata_connection_config=metadata_connection_config,
beam_pipeline_args=beam_pipeline_args
)
def create_pipeline(pipeline_name: Text,
pipeline_root: Text,
data_root_uri: data_types.RuntimeParameter,
train_steps: data_types.RuntimeParameter,
eval_steps: data_types.RuntimeParameter,
ai_platform_training_args: Dict[Text, Text],
ai_platform_serving_args: Dict[Text, Text],
beam_pipeline_args: List[Text],
enable_cache: Optional[bool] = False) -> pipeline.Pipeline:
"""Trains and deploys the Covertype classifier."""
# Brings data into the pipeline and splits the data into training and eval splits
examples = external_input(data_root_uri)
output_config = example_gen_pb2.Output(
split_config=example_gen_pb2.SplitConfig(splits=[
example_gen_pb2.SplitConfig.Split(name='train', hash_buckets=4),
example_gen_pb2.SplitConfig.Split(name='eval', hash_buckets=1)
]))
generate_examples = CsvExampleGen(input=examples)
# Computes statistics over data for visualization and example validation.
generate_statistics = StatisticsGen(
examples=generate_examples.outputs.examples)
# Import a user-provided schema
import_schema = ImporterNode(instance_name='import_user_schema',
source_uri=SCHEMA_FOLDER,
artifact_type=Schema)
# Generates schema based on statistics files.Even though, we use user-provided schema
# we still want to generate the schema of the newest data for tracking and comparison
infer_schema = SchemaGen(statistics=generate_statistics.outputs.statistics)
# Performs anomaly detection based on statistics and data schema.
validate_stats = ExampleValidator(
statistics=generate_statistics.outputs.statistics,
schema=import_schema.outputs.result)
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=generate_examples.outputs.examples,
schema=import_schema.outputs.result,
module_file=TRANSFORM_MODULE_FILE)
# Trains the model using a user provided trainer function.
train = Trainer(
custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_trainer_executor.GenericExecutor),
# custom_executor_spec=executor_spec.ExecutorClassSpec(trainer_executor.GenericExecutor),
module_file=TRAIN_MODULE_FILE,
transformed_examples=transform.outputs.transformed_examples,
schema=import_schema.outputs.result,
transform_graph=transform.outputs.transform_graph,
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
custom_config={'ai_platform_training_args': ai_platform_training_args})
# Get the latest blessed model for model validation.
resolve = ResolverNode(instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute a evaluation statistics over features of a model.
accuracy_threshold = tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(lower_bound={'value': 0.5},
upper_bound={'value':
0.99}),
change_threshold=tfma.GenericChangeThreshold(
absolute={'value': 0.0001},
direction=tfma.MetricDirection.HIGHER_IS_BETTER),
)
metrics_specs = tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='SparseCategoricalAccuracy',
threshold=accuracy_threshold),
tfma.MetricConfig(class_name='ExampleCount')
])
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='Cover_Type')],
metrics_specs=[metrics_specs],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['Wilderness_Area'])
])
analyze = Evaluator(examples=generate_examples.outputs.examples,
model=train.outputs.model,
baseline_model=resolve.outputs.model,
eval_config=eval_config)
# Validate model can be loaded and queried in sand-boxed environment
# mirroring production.
serving_config = infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(
tags=['latest']),
kubernetes=infra_validator_pb2.KubernetesConfig(),
)
validation_config = infra_validator_pb2.ValidationSpec(
max_loading_time_seconds=60,
num_tries=3,
)
request_config = infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServingRequestSpec(),
num_examples=3,
)
infra_validate = InfraValidator(
model=train.outputs['model'],
examples=generate_examples.outputs['examples'],
serving_spec=serving_config,
validation_spec=validation_config,
request_spec=request_config,
)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
deploy = Pusher(custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_pusher_executor.Executor),
model=train.outputs['model'],
model_blessing=analyze.outputs['blessing'],
infra_blessing=infra_validate.outputs['blessing'],
custom_config={
ai_platform_pusher_executor.SERVING_ARGS_KEY:
ai_platform_serving_args
})
return pipeline.Pipeline(pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
generate_examples, generate_statistics,
import_schema, infer_schema, validate_stats,
transform, train, resolve, analyze,
infra_validate, deploy
],
enable_cache=enable_cache,
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, accuracy_threshold: float,
serving_model_dir: Text, metadata_path: Text,
beam_pipeline_args: List[Text],
make_warmup: bool) -> pipeline.Pipeline:
"""Implements the penguin pipeline with TFX."""
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input_base=data_root)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that trains a model using TF-Learn.
trainer = Trainer(module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=trainer_pb2.TrainArgs(num_steps=2000),
eval_args=trainer_pb2.EvalArgs(num_steps=5))
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='species')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': accuracy_threshold}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator(examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
# Performs infra validation of a candidate model to prevent unservable model
# from being pushed. This config will launch a model server of the latest
# TensorFlow Serving image in a local docker engine.
infra_validator = InfraValidator(
model=trainer.outputs['model'],
examples=example_gen.outputs['examples'],
serving_spec=infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(
tags=['latest']),
local_docker=infra_validator_pb2.LocalDockerConfig()),
request_spec=infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.
TensorFlowServingRequestSpec(),
# If this flag is set, InfraValidator will produce a model with
# warmup requests (in its outputs['blessing']).
make_warmup=make_warmup))
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
if make_warmup:
# If InfraValidator.request_spec.make_warmup = True, its output contains
# a model so that Pusher can push 'infra_blessing' input instead of
# 'model' input.
pusher = Pusher(model_blessing=evaluator.outputs['blessing'],
infra_blessing=infra_validator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
else:
# Otherwise, 'infra_blessing' does not contain a model and is used as a
# conditional checker just like 'model_blessing' does. This is the typical
# use case.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
infra_blessing=infra_validator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
infra_validator,
pusher,
],
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(
pipeline_root: str,
csv_input_location: str,
taxi_module_file: tfx.dsl.experimental.RuntimeParameter,
push_destination: tfx.dsl.experimental.RuntimeParameter,
enable_cache: bool,
):
"""Creates a simple Kubeflow-based Chicago Taxi TFX pipeline.
Args:
pipeline_root: The root of the pipeline output.
csv_input_location: The location of the input data directory.
taxi_module_file: The location of the module file for Transform/Trainer.
enable_cache: Whether to enable cache or not.
Returns:
A logical TFX pipeline.Pipeline object.
"""
example_gen = tfx.components.CsvExampleGen(input_base=csv_input_location)
statistics_gen = tfx.components.StatisticsGen(
examples=example_gen.outputs['examples'])
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False,
)
example_validator = tfx.components.ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'],
)
transform = tfx.components.Transform(
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=taxi_module_file,
)
trainer = tfx.components.Trainer(
module_file=taxi_module_file,
examples=transform.outputs['transformed_examples'],
schema=schema_gen.outputs['schema'],
transform_graph=transform.outputs['transform_graph'],
train_args=tfx.proto.TrainArgs(num_steps=10),
eval_args=tfx.proto.EvalArgs(num_steps=5),
)
# Set the TFMA config for Model Evaluation and Validation.
eval_config = tfma.EvalConfig(
model_specs=[
tfma.ModelSpec(
signature_name='serving_default',
label_key='tips_xf',
preprocessing_function_names=['transform_features'])
],
metrics_specs=[
tfma.MetricsSpec(
# The metrics added here are in addition to those saved with the
# model (assuming either a keras model or EvalSavedModel is used).
# Any metrics added into the saved model (for example using
# model.compile(..., metrics=[...]), etc) will be computed
# automatically.
metrics=[tfma.MetricConfig(class_name='ExampleCount')],
# To add validation thresholds for metrics saved with the model,
# add them keyed by metric name to the thresholds map.
thresholds={
'binary_accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
],
slicing_specs=[
# An empty slice spec means the overall slice, i.e. the whole dataset.
tfma.SlicingSpec(),
# Data can be sliced along a feature column. In this case, data is
# sliced along feature column trip_start_hour.
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
])
evaluator = tfx.components.Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
eval_config=eval_config,
)
pusher = tfx.components.Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=push_destination,
)
return tfx.dsl.Pipeline(
pipeline_name='parameterized_tfx_oss',
pipeline_root=pipeline_root,
components=[
example_gen, statistics_gen, schema_gen, example_validator,
transform, trainer, evaluator, pusher
],
enable_cache=enable_cache,
)
],
metrics_specs=[
tfma.MetricsSpec(
# The metrics added here are in addition to those saved with the
# model (assuming either a keras model or EvalSavedModel is used).
# Any metrics added into the saved model (for example using
# model.compile(..., metrics=[...]), etc) will be computed
# automatically.
metrics=[tfma.MetricConfig(class_name='ExampleCount')],
# To add validation thresholds for metrics saved with the model,
# add them keyed by metric name to the thresholds map.
thresholds={
'accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
],
slicing_specs=[
# An empty slice spec means the overall slice, i.e. the whole dataset.
tfma.SlicingSpec(),
# Data can be sliced along a feature column. In this case, data is
# sliced along feature column trip_start_hour.
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
])
# Use TFMA to compute a evaluation statistics over features of a model and
# validate them against a baseline.
def create_pipeline(pipeline_name: Text,
pipeline_root: Text,
data_root_uri: data_types.RuntimeParameter,
train_steps: data_types.RuntimeParameter,
eval_steps: data_types.RuntimeParameter,
enable_tuning: bool,
ai_platform_training_args: Dict[Text, Text],
ai_platform_serving_args: Dict[Text, Text],
beam_pipeline_args: List[Text],
enable_cache: Optional[bool] = False) -> pipeline.Pipeline:
"""Trains and deploys the Keras Covertype Classifier with TFX and Kubeflow Pipeline on Google Cloud.
Args:
pipeline_name: name of the TFX pipeline being created.
pipeline_root: root directory of the pipeline. Should be a valid GCS path.
data_root_uri: uri of the dataset.
train_steps: runtime parameter for number of model training steps for the Trainer component.
eval_steps: runtime parameter for number of model evaluation steps for the Trainer component.
enable_tuning: If True, the hyperparameter tuning through CloudTuner is
enabled.
ai_platform_training_args: Args of CAIP training job. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.jobs#Job
for detailed description.
ai_platform_serving_args: Args of CAIP model deployment. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.models
for detailed description.
beam_pipeline_args: Optional list of beam pipeline options. Please refer to
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options.
When this argument is not provided, the default is to use GCP
DataflowRunner with 50GB disk size as specified in this function. If an
empty list is passed in, default specified by Beam will be used, which can
be found at
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options
enable_cache: Optional boolean
Returns:
A TFX pipeline object.
"""
# Brings data into the pipeline and splits the data into training and eval splits
output = example_gen_pb2.Output(split_config=example_gen_pb2.SplitConfig(
splits=[
example_gen_pb2.SplitConfig.Split(name='train', hash_buckets=4),
example_gen_pb2.SplitConfig.Split(name='eval', hash_buckets=1)
]))
examplegen = CsvExampleGen(input_base=data_root_uri, output_config=output)
# Computes statistics over data for visualization and example validation.
statisticsgen = StatisticsGen(examples=examplegen.outputs.examples)
# Generates schema based on statistics files. Even though, we use user-provided schema
# we still want to generate the schema of the newest data for tracking and comparison
schemagen = SchemaGen(statistics=statisticsgen.outputs.statistics)
# Import a user-provided schema
import_schema = ImporterNode(instance_name='import_user_schema',
source_uri=SCHEMA_FOLDER,
artifact_type=Schema)
# Performs anomaly detection based on statistics and data schema.
examplevalidator = ExampleValidator(
statistics=statisticsgen.outputs.statistics,
schema=import_schema.outputs.result)
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=examplegen.outputs.examples,
schema=import_schema.outputs.result,
module_file=TRANSFORM_MODULE_FILE)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
# The Tuner component launches 1 AI Platform Training job for flock management.
# For example, 3 workers (defined by num_parallel_trials) in the flock
# management AI Platform Training job, each runs Tuner.Executor.
tuner = Tuner(
module_file=TRAIN_MODULE_FILE,
examples=transform.outputs.transformed_examples,
transform_graph=transform.outputs.transform_graph,
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
tune_args=tuner_pb2.TuneArgs(
# num_parallel_trials=3 means that 3 search loops are running in parallel.
num_parallel_trials=3),
custom_config={
# Configures Cloud AI Platform-specific configs. For details, see
# https://cloud.google.com/ai-platform/training/docs/reference/rest/v1/projects.jobs#traininginput.
ai_platform_trainer_executor.TRAINING_ARGS_KEY:
ai_platform_training_args
})
# Trains the model using a user provided trainer function.
trainer = Trainer(
custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_trainer_executor.GenericExecutor),
module_file=TRAIN_MODULE_FILE,
transformed_examples=transform.outputs.transformed_examples,
schema=import_schema.outputs.result,
transform_graph=transform.outputs.transform_graph,
hyperparameters=(tuner.outputs.best_hyperparameters
if enable_tuning else None),
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
custom_config={'ai_platform_training_args': ai_platform_training_args})
# Get the latest blessed model for model validation.
resolver = ResolverNode(instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute a evaluation statistics over features of a model.
accuracy_threshold = tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(lower_bound={'value': 0.5},
upper_bound={'value':
0.99}), )
metrics_specs = tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='SparseCategoricalAccuracy',
threshold=accuracy_threshold),
tfma.MetricConfig(class_name='ExampleCount')
])
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='Cover_Type')],
metrics_specs=[metrics_specs],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['Wilderness_Area'])
])
evaluator = Evaluator(examples=examplegen.outputs.examples,
model=trainer.outputs.model,
baseline_model=resolver.outputs.model,
eval_config=eval_config)
# Validate model can be loaded and queried in sand-boxed environment
# mirroring production.
serving_config = infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(
tags=['latest']),
kubernetes=infra_validator_pb2.KubernetesConfig(),
)
validation_config = infra_validator_pb2.ValidationSpec(
max_loading_time_seconds=60,
num_tries=3,
)
request_config = infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServingRequestSpec(),
num_examples=3,
)
infravalidator = InfraValidator(
model=trainer.outputs.model,
examples=examplegen.outputs.examples,
serving_spec=serving_config,
validation_spec=validation_config,
request_spec=request_config,
)
# Checks whether the model passed the validation steps and pushes the model
# to CAIP Prediction if checks are passed.
pusher = Pusher(custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_pusher_executor.Executor),
model=trainer.outputs.model,
model_blessing=evaluator.outputs.blessing,
infra_blessing=infravalidator.outputs.blessing,
custom_config={
ai_platform_pusher_executor.SERVING_ARGS_KEY:
ai_platform_serving_args
})
components = [
examplegen, statisticsgen, schemagen, import_schema, examplevalidator,
transform, trainer, resolver, evaluator, infravalidator, pusher
]
if enable_tuning:
components.append(tuner)
return pipeline.Pipeline(pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=enable_cache,
beam_pipeline_args=beam_pipeline_args)
def create_pipeline_components(
pipeline_root: Text,
transform_module: Text,
trainer_module: Text,
bigquery_query: Text = '',
csv_input_location: Text = '',
) -> List[base_node.BaseNode]:
"""Creates components for a simple Chicago Taxi TFX pipeline for testing.
Args:
pipeline_root: The root of the pipeline output.
transform_module: The location of the transform module file.
trainer_module: The location of the trainer module file.
bigquery_query: The query to get input data from BigQuery. If not empty,
BigQueryExampleGen will be used.
csv_input_location: The location of the input data directory.
Returns:
A list of TFX components that constitutes an end-to-end test pipeline.
"""
if bool(bigquery_query) == bool(csv_input_location):
raise ValueError(
'Exactly one example gen is expected. ',
'Please provide either bigquery_query or csv_input_location.')
if bigquery_query:
example_gen = big_query_example_gen_component.BigQueryExampleGen(
query=bigquery_query)
else:
example_gen = components.CsvExampleGen(input_base=csv_input_location)
statistics_gen = components.StatisticsGen(
examples=example_gen.outputs['examples'])
schema_gen = components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False)
example_validator = components.ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
transform = components.Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=transform_module)
latest_model_resolver = resolver.Resolver(
strategy_class=latest_artifacts_resolver.LatestArtifactsResolver,
model=channel.Channel(type=standard_artifacts.Model)).with_id(
'Resolver.latest_model_resolver')
trainer = components.Trainer(
custom_executor_spec=executor_spec.ExecutorClassSpec(Executor),
transformed_examples=transform.outputs['transformed_examples'],
schema=schema_gen.outputs['schema'],
base_model=latest_model_resolver.outputs['model'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=10),
eval_args=trainer_pb2.EvalArgs(num_steps=5),
module_file=trainer_module,
)
# Get the latest blessed model for model validation.
model_resolver = resolver.Resolver(
strategy_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=channel.Channel(type=standard_artifacts.Model),
model_blessing=channel.Channel(
type=standard_artifacts.ModelBlessing)).with_id(
'Resolver.latest_blessed_model_resolver')
# Set the TFMA config for Model Evaluation and Validation.
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(signature_name='eval')],
metrics_specs=[
tfma.MetricsSpec(
metrics=[tfma.MetricConfig(class_name='ExampleCount')],
thresholds={
'binary_accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
])
evaluator = components.Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
pusher = components.Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=os.path.join(pipeline_root, 'model_serving'))))
return [
example_gen, statistics_gen, schema_gen, example_validator, transform,
latest_model_resolver, trainer, model_resolver, evaluator, pusher
]
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
serving_model_uri: Text,
data_root_uri: Union[Text, data_types.RuntimeParameter],
schema_folder_uri: Union[Text, data_types.RuntimeParameter],
train_steps: Union[int, data_types.RuntimeParameter],
eval_steps: Union[int, data_types.RuntimeParameter],
beam_pipeline_args: List[Text],
trainer_custom_executor_spec: Optional[executor_spec.ExecutorSpec] = None,
trainer_custom_config: Optional[Dict[Text, Any]] = None,
enable_tuning: Optional[bool] = False,
enable_cache: Optional[bool] = False,
metadata_connection_config: Optional[
metadata_store_pb2.ConnectionConfig] = None
) -> pipeline.Pipeline:
"""Trains and deploys the Keras Covertype Classifier with TFX and AI Platform Pipelines."""
# Brings data into the pipeline and splits the data into training and eval splits
output_config = example_gen_pb2.Output(
split_config=example_gen_pb2.SplitConfig(splits=[
example_gen_pb2.SplitConfig.Split(name='train', hash_buckets=4),
example_gen_pb2.SplitConfig.Split(name='eval', hash_buckets=1)
]))
examplegen = CsvExampleGen(input_base=data_root_uri)
# Computes statistics over data for visualization and example validation.
statisticsgen = StatisticsGen(examples=examplegen.outputs.examples)
# Generates schema based on statistics files. Even though, we use user-provided schema
# we still want to generate the schema of the newest data for tracking and comparison
schemagen = SchemaGen(statistics=statisticsgen.outputs.statistics)
# Import a user-provided schema
import_schema = ImporterNode(
instance_name='import_user_schema',
#source_uri=SCHEMA_FOLDER,
source_uri=schema_folder_uri,
artifact_type=Schema)
# Performs anomaly detection based on statistics and data schema.
examplevalidator = ExampleValidator(
statistics=statisticsgen.outputs.statistics,
schema=import_schema.outputs.result)
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=examplegen.outputs.examples,
schema=import_schema.outputs.result,
module_file=TRANSFORM_MODULE_FILE)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
# The Tuner component launches 1 AI Platform Training job for flock management.
# For example, 3 workers (defined by num_parallel_trials) in the flock
# management AI Platform Training job, each runs Tuner.Executor.
tuner = Tuner(
module_file=TRAIN_MODULE_FILE,
examples=transform.outputs.transformed_examples,
transform_graph=transform.outputs.transform_graph,
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
tune_args=tuner_pb2.TuneArgs(
# num_parallel_trials=3 means that 3 search loops are running in parallel.
num_parallel_trials=3),
custom_config=custom_config)
# Trains the model using a user provided trainer function.
trainer = Trainer(
custom_executor_spec=trainer_custom_executor_spec,
module_file=TRAIN_MODULE_FILE,
transformed_examples=transform.outputs.transformed_examples,
schema=import_schema.outputs.result,
transform_graph=transform.outputs.transform_graph,
hyperparameters=(tuner.outputs.best_hyperparameters
if enable_tuning else None),
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
custom_config=trainer_custom_config)
# Get the latest blessed model for model validation.
resolver = ResolverNode(instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute a evaluation statistics over features of a model.
accuracy_threshold = tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(lower_bound={'value': 0.5},
upper_bound={'value':
0.99}), )
metrics_specs = tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='SparseCategoricalAccuracy',
threshold=accuracy_threshold),
tfma.MetricConfig(class_name='ExampleCount')
])
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='Cover_Type')],
metrics_specs=[metrics_specs],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['Wilderness_Area'])
])
evaluator = Evaluator(examples=examplegen.outputs.examples,
model=trainer.outputs.model,
baseline_model=resolver.outputs.model,
eval_config=eval_config)
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_uri)))
components = [
examplegen, statisticsgen, schemagen, import_schema, examplevalidator,
transform, trainer, resolver, evaluator, pusher
]
if enable_tuning:
components.append(tuner)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=enable_cache,
beam_pipeline_args=beam_pipeline_args,
metadata_connection_config=metadata_connection_config)
def init_components(data_dir,
module_file,
serving_model_dir=None,
ai_platform_training_args=None,
ai_platform_serving_args=None,
training_steps=1000,
eval_steps=200):
"""
This function is to initialize tfx components
"""
if serving_model_dir and ai_platform_serving_args:
raise NotImplementedError(
"Can't set ai_platform_serving_args and serving_model_dir at "
"the same time. Choose one deployment option.")
output = example_gen_pb2.Output(split_config=example_gen_pb2.SplitConfig(
splits=[
example_gen_pb2.SplitConfig.Split(name="train", hash_buckets=99),
example_gen_pb2.SplitConfig.Split(name="eval", hash_buckets=1),
]))
example_gen = CsvExampleGen(input_base=data_dir, output_config=output)
statistics_gen = StatisticsGen(examples=example_gen.outputs["examples"])
schema_gen = SchemaGen(
statistics=statistics_gen.outputs["statistics"],
infer_feature_shape=False,
)
example_validator = ExampleValidator(
statistics=statistics_gen.outputs["statistics"],
schema=schema_gen.outputs["schema"],
)
transform = Transform(
examples=example_gen.outputs["examples"],
schema=schema_gen.outputs["schema"],
module_file=module_file,
)
training_kwargs = {
"module_file": module_file,
"examples": transform.outputs["transformed_examples"],
"schema": schema_gen.outputs["schema"],
"transform_graph": transform.outputs['transform_graph'],
"train_args": trainer_pb2.TrainArgs(num_steps=training_steps),
"eval_args": trainer_pb2.EvalArgs(num_steps=eval_steps),
}
if ai_platform_training_args:
training_kwargs.update({
"custom_executor_spec":
executor_spec.ExecutorClassSpec(
aip_trainer_executor.GenericExecutor),
"custom_config": {
aip_trainer_executor.TRAINING_ARGS_KEY:
ai_platform_training_args # noqa
},
})
else:
training_kwargs.update({
"custom_executor_spec":
executor_spec.ExecutorClassSpec(GenericExecutor)
})
trainer = Trainer(**training_kwargs)
model_resolver = ResolverNode(
instance_name="latest_blessed_model_resolver",
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing),
)
#model_resolver for tfx==0.30.0
# model_resolver = tfx.dsl.Resolver(
# strategy_class=tfx.dsl.experimental.LatestBlessedModelStrategy,
# model=tfx.dsl.Channel(type=tfx.types.standard_artifacts.Model),
# model_blessing=tfx.dsl.Channel(
# type=tfx.types.standard_artifacts.ModelBlessing)).with_id(
# 'latest_blessed_model_resolver')
#the book's eval_config might be wrong,
#threshold has to be set within the tfma.MetricConfig() with each metric
#this seems to have caused the models not be blessed
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key="consumer_disputed")],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=["product"]),
],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(class_name='ExampleCount'),
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={"value": 0.01},
),
)),
# tfma.MetricConfig(
# class_name='AUC',
# threshold=tfma.MetricThreshold(
# value_threshold=tfma.GenericValueThreshold(
# lower_bound={'value': 0.5}
# ),
# change_threshold=tfma.GenericChangeThreshold(
# direction=tfma.MetricDirection.HIGHER_IS_BETTER,
# absolute={"value": 0.01},
# ),
# )
# ),
])
],
)
evaluator = Evaluator(
examples=example_gen.outputs["examples"],
model=trainer.outputs["model"],
# baseline_model=model_resolver.outputs["model"],
eval_config=eval_config,
)
pusher_kwargs = {
"model": trainer.outputs["model"],
"model_blessing": evaluator.outputs["blessing"],
}
if ai_platform_serving_args:
pusher_kwargs.update({
"custom_executor_spec":
executor_spec.ExecutorClassSpec(aip_pusher_executor.Executor),
"custom_config": {
aip_pusher_executor.SERVING_ARGS_KEY:
ai_platform_serving_args # noqa
},
})
elif serving_model_dir:
pusher_kwargs.update({
"push_destination":
pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir))
})
else:
raise NotImplementedError(
"Provide ai_platform_serving_args or serving_model_dir.")
pusher = Pusher(**pusher_kwargs)
#compile all components in a list
components = [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
return components
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir: Text,
metadata_path: Text, enable_tuning: bool,
beam_pipeline_args: List[Text]) -> pipeline.Pipeline:
"""Implements the Iris flowers pipeline with TFX."""
examples = external_input(data_root)
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=examples)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
tuner = Tuner(module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=20),
eval_args=trainer_pb2.EvalArgs(num_steps=5))
# Uses user-provided Python function that trains a model.
trainer = Trainer(
module_file=module_file,
custom_executor_spec=executor_spec.ExecutorClassSpec(GenericExecutor),
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
# If Tuner is in the pipeline, Trainer can take Tuner's output
# best_hyperparameters artifact as input and utilize it in the user module
# code.
#
# If there isn't Tuner in the pipeline, either use ImporterNode to import
# a previous Tuner's output to feed to Trainer, or directly use the tuned
# hyperparameters in user module code and set hyperparameters to None
# here.
#
# Example of ImporterNode,
# hparams_importer = ImporterNode(
# instance_name='import_hparams',
# source_uri='path/to/best_hyperparameters.txt',
# artifact_type=HyperParameters)
# ...
# hyperparameters = hparams_importer.outputs['result'],
hyperparameters=(tuner.outputs['best_hyperparameters']
if enable_tuning else None),
train_args=trainer_pb2.TrainArgs(num_steps=100),
eval_args=trainer_pb2.EvalArgs(num_steps=5))
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute an evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='variety')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
components = [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
if enable_tuning:
components.append(tuner)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
beam_pipeline_args=beam_pipeline_args)
def create_e2e_components(
pipeline_root: Text,
csv_input_location: Text,
transform_module: Text,
trainer_module: Text,
) -> List[BaseComponent]:
"""Creates components for a simple Chicago Taxi TFX pipeline for testing.
Args:
pipeline_root: The root of the pipeline output.
csv_input_location: The location of the input data directory.
transform_module: The location of the transform module file.
trainer_module: The location of the trainer module file.
Returns:
A list of TFX components that constitutes an end-to-end test pipeline.
"""
example_gen = CsvExampleGen(input_base=csv_input_location)
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'])
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=transform_module)
latest_model_resolver = resolver.Resolver(
strategy_class=latest_artifact_strategy.LatestArtifactStrategy,
latest_model=Channel(type=Model)).with_id('latest_model_resolver')
trainer = Trainer(
transformed_examples=transform.outputs['transformed_examples'],
schema=schema_gen.outputs['schema'],
base_model=latest_model_resolver.outputs['latest_model'],
transform_graph=transform.outputs['transform_graph'],
train_args=trainer_pb2.TrainArgs(num_steps=10),
eval_args=trainer_pb2.EvalArgs(num_steps=5),
module_file=trainer_module,
)
# Set the TFMA config for Model Evaluation and Validation.
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(signature_name='eval')],
metrics_specs=[
tfma.MetricsSpec(
metrics=[tfma.MetricConfig(class_name='ExampleCount')],
thresholds={
'accuracy':
tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.5}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10}))
})
],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['trip_start_hour'])
])
evaluator = Evaluator(examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
eval_config=eval_config)
infra_validator = InfraValidator(
model=trainer.outputs['model'],
examples=example_gen.outputs['examples'],
serving_spec=infra_validator_pb2.ServingSpec(
tensorflow_serving=infra_validator_pb2.TensorFlowServing(
tags=['latest']),
kubernetes=infra_validator_pb2.KubernetesConfig()),
request_spec=infra_validator_pb2.RequestSpec(
tensorflow_serving=infra_validator_pb2.
TensorFlowServingRequestSpec()))
pusher = Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=os.path.join(pipeline_root, 'model_serving'))))
return [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
latest_model_resolver,
trainer,
evaluator,
infra_validator,
pusher,
]
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir: Text,
metadata_path: Text, beam_pipeline_args: List[Text]):
"""Creates pipeline."""
pipeline_root = os.path.join(pipeline_root, 'pipelines', pipeline_name)
example_gen = ImportExampleGen(
input_base=data_root,
# IMPORTANT: must set FORMAT_PROTO
payload_format=example_gen_pb2.FORMAT_PROTO)
data_view_provider = provider_component.TfGraphDataViewProvider(
module_file=module_file,
create_decoder_func='make_decoder')
data_view_binder = binder_component.DataViewBinder(
example_gen.outputs['examples'],
data_view_provider.outputs['data_view'])
statistics_gen = StatisticsGen(
examples=data_view_binder.outputs['output_examples'])
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'])
transform = Transform(
examples=data_view_binder.outputs['output_examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file,
# important: must disable Transform materialization.
materialize=False)
trainer = Trainer(
examples=data_view_binder.outputs['output_examples'],
transform_graph=transform.outputs['transform_graph'],
module_file=module_file,
train_args=trainer_pb2.TrainArgs(num_steps=1000),
schema=schema_gen.outputs['schema'],
eval_args=trainer_pb2.EvalArgs(num_steps=10))
eval_config = tfma.EvalConfig(
model_specs=[
tfma.ModelSpec(
signature_name='',
label_key='relevance',
padding_options=tfma.config.PaddingOptions(
label_float_padding=-1.0, prediction_float_padding=-1.0))
],
slicing_specs=[
tfma.SlicingSpec(),
tfma.SlicingSpec(feature_keys=['query_tokens']),
],
metrics_specs=[
tfma.MetricsSpec(
per_slice_thresholds={
'metric/ndcg_10':
tfma.config.PerSliceMetricThresholds(thresholds=[
tfma.PerSliceMetricThreshold(
# The overall slice.
slicing_specs=[tfma.SlicingSpec()],
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6})))
])
})
])
evaluator = Evaluator(
examples=data_view_binder.outputs['output_examples'],
model=trainer.outputs['model'],
eval_config=eval_config,
schema=schema_gen.outputs['schema'])
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen, data_view_provider, data_view_binder,
statistics_gen,
schema_gen,
transform,
trainer,
evaluator,
pusher,
],
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
beam_pipeline_args=beam_pipeline_args)
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_path: Text,
preprocessing_fn: Text,
run_fn: Text,
train_args: trainer_pb2.TrainArgs,
eval_args: trainer_pb2.EvalArgs,
eval_accuracy_threshold: float,
serving_model_dir: Text,
metadata_connection_config: Optional[
metadata_store_pb2.ConnectionConfig] = None,
beam_pipeline_args: Optional[List[Text]] = None,
) -> pipeline.Pipeline:
"""Implements the penguin pipeline with TFX."""
components = []
# Brings data into the pipeline or otherwise joins/converts training data.
# TODO(step 2): Might use another ExampleGen class for your data.
example_gen = CsvExampleGen(input_base=data_path)
components.append(example_gen)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
components.append(statistics_gen)
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
components.append(schema_gen)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator( # pylint: disable=unused-variable
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
components.append(example_validator)
# Performs transformations and feature engineering in training and serving.
transform = Transform( # pylint: disable=unused-variable
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
preprocessing_fn=preprocessing_fn)
# TODO(step 3): Uncomment here to add Transform to the pipeline.
# components.append(transform)
# Uses user-provided Python function that implements a model using Tensorflow.
trainer = Trainer(
run_fn=run_fn,
examples=example_gen.outputs['examples'],
# Use outputs of Transform as training inputs if Transform is used.
# examples=transform.outputs['transformed_examples'],
# transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=train_args,
eval_args=eval_args)
# TODO(step 4): Uncomment here to add Trainer to the pipeline.
# components.append(trainer)
# Get the latest blessed model for model validation.
model_resolver = resolver.Resolver(
strategy_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(
type=ModelBlessing)).with_id('latest_blessed_model_resolver')
# TODO(step 5): Uncomment here to add Resolver to the pipeline.
# components.append(model_resolver)
# Uses TFMA to compute a evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key=features.LABEL_KEY)],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': eval_accuracy_threshold}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator( # pylint: disable=unused-variable
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
# TODO(step 5): Uncomment here to add Evaluator to the pipeline.
# components.append(evaluator)
# Pushes the model to a file destination if check passed.
pusher = Pusher( # pylint: disable=unused-variable
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
# TODO(step 5): Uncomment here to add Pusher to the pipeline.
# components.append(pusher)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
# Change this value to control caching of execution results. Default value
# is `False`.
# enable_cache=True,
metadata_connection_config=metadata_connection_config,
beam_pipeline_args=beam_pipeline_args,
)
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir_lite: Text,
metadata_path: Text, labels_path: Text,
beam_pipeline_args: List[Text]) -> pipeline.Pipeline:
"""Implements the CIFAR10 image classification pipeline using TFX."""
# This is needed for datasets with pre-defined splits
# Change the pattern argument to train_whole/* and test_whole/* to train
# on the whole CIFAR-10 dataset
input_config = example_gen_pb2.Input(splits=[
example_gen_pb2.Input.Split(name='train', pattern='train/*'),
example_gen_pb2.Input.Split(name='eval', pattern='test/*')
])
# Brings data into the pipeline.
example_gen = ImportExampleGen(input_base=data_root,
input_config=input_config)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that trains a model.
# When traning on the whole dataset, use 18744 for train steps, 156 for eval
# steps. 18744 train steps correspond to 24 epochs on the whole train set, and
# 156 eval steps correspond to 1 epoch on the whole test set. The
# configuration below is for training on the dataset we provided in the data
# folder, which has 128 train and 128 test samples. The 160 train steps
# correspond to 40 epochs on this tiny train set, and 4 eval steps correspond
# to 1 epoch on this tiny test set.
trainer = Trainer(module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=trainer_pb2.TrainArgs(num_steps=160),
eval_args=trainer_pb2.EvalArgs(num_steps=4),
custom_config={'labels_path': labels_path})
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(
type=ModelBlessing)).with_id('latest_blessed_model_resolver')
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compare to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[
tfma.ModelSpec(label_key='label_xf', model_type='tf_lite')
],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.55}),
# Change threshold will be ignored if there is no
# baseline model resolved from MLMD (first run).
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-3})))
])
])
# Uses TFMA to compute the evaluation statistics over features of a model.
# We evaluate using the materialized examples that are output by Transform
# because
# 1. the decoding_png function currently performed within Transform are not
# compatible with TFLite.
# 2. MLKit requires deserialized (float32) tensor image inputs
# Note that for deployment, the same logic that is performed within Transform
# must be reproduced client-side.
evaluator = Evaluator(examples=transform.outputs['transformed_examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir_lite)))
components = [
example_gen, statistics_gen, schema_gen, example_validator, transform,
trainer, model_resolver, evaluator, pusher
]
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
beam_pipeline_args=beam_pipeline_args)
def _create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_root: Text,
trainer_module_file: Text,
evaluator_module_file: Text,
serving_model_dir: Text,
metadata_path: Text,
beam_pipeline_args: List[Text],
) -> tfx.dsl.Pipeline:
"""Implements the Penguin pipeline with TFX."""
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = tfx.components.CsvExampleGen(
input_base=os.path.join(data_root, 'labelled'))
# Computes statistics over data for visualization and example validation.
statistics_gen = tfx.components.StatisticsGen(
examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = tfx.components.ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# TODO(humichael): Handle applying transformation component in Milestone 3.
# Uses user-provided Python function that trains a model using TF-Learn.
# Num_steps is not provided during evaluation because the scikit-learn model
# loads and evaluates the entire test set at once.
trainer = tfx.components.Trainer(
module_file=trainer_module_file,
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
train_args=tfx.proto.TrainArgs(num_steps=2000),
eval_args=tfx.proto.EvalArgs())
# Get the latest blessed model for model validation.
model_resolver = tfx.dsl.Resolver(
strategy_class=tfx.dsl.experimental.LatestBlessedModelStrategy,
model=tfx.dsl.Channel(type=tfx.types.standard_artifacts.Model),
model_blessing=tfx.dsl.Channel(
type=tfx.types.standard_artifacts.ModelBlessing)).with_id(
'latest_blessed_model_resolver')
# Uses TFMA to compute evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='species')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='Accuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = tfx.components.Evaluator(
module_file=evaluator_module_file,
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
eval_config=eval_config)
pusher = tfx.components.Pusher(
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
push_destination=tfx.proto.PushDestination(
filesystem=tfx.proto.PushDestination.Filesystem(
base_directory=serving_model_dir)))
return tfx.dsl.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen,
statistics_gen,
schema_gen,
example_validator,
trainer,
model_resolver,
evaluator,
pusher,
],
enable_cache=True,
metadata_connection_config=tfx.orchestration.metadata.
sqlite_metadata_connection_config(metadata_path),
beam_pipeline_args=beam_pipeline_args,
)
def _create_pipeline(pipeline_name: Text, pipeline_root: Text, data_root: Text,
module_file: Text, serving_model_dir: Text,
metadata_path: Text,
beam_pipeline_args: List[Text]) -> pipeline.Pipeline:
"""Implements the chicago taxi pipeline with TFX."""
examples = external_input(data_root)
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=examples)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
infer_schema = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False)
# Performs anomaly detection based on statistics and data schema.
validate_stats = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=infer_schema.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=infer_schema.outputs['schema'],
module_file=module_file)
# Uses user-provided Python function that implements a model using TF-Learn.
trainer = Trainer(
module_file=module_file,
custom_executor_spec=executor_spec.ExecutorClassSpec(GenericExecutor),
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=infer_schema.outputs['schema'],
train_args=trainer_pb2.TrainArgs(num_steps=10000),
eval_args=trainer_pb2.EvalArgs(num_steps=5000))
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute a evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='tips')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
model_analyzer = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher = Pusher(model=trainer.outputs['model'],
model_blessing=model_analyzer.outputs['blessing'],
push_destination=pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)))
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=[
example_gen,
statistics_gen,
infer_schema,
validate_stats,
transform,
trainer,
model_resolver,
model_analyzer,
pusher,
],
enable_cache=True,
metadata_connection_config=metadata.sqlite_metadata_connection_config(
metadata_path),
beam_pipeline_args=beam_pipeline_args)
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_path: Text,
# TODO(step 7): (Optional) Uncomment here to use BigQuery as a data source.
# query: Text,
preprocessing_fn: Text,
run_fn: Text,
train_args: trainer_pb2.TrainArgs,
eval_args: trainer_pb2.EvalArgs,
eval_accuracy_threshold: float,
serving_model_dir: Text,
metadata_connection_config: Optional[
metadata_store_pb2.ConnectionConfig] = None,
beam_pipeline_args: Optional[List[Text]] = None,
ai_platform_training_args: Optional[Dict[Text, Text]] = None,
ai_platform_serving_args: Optional[Dict[Text, Any]] = None,
) -> pipeline.Pipeline:
"""Implements the chicago taxi pipeline with TFX."""
components = []
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=external_input(data_path))
# TODO(step 7): (Optional) Uncomment here to use BigQuery as a data source.
# example_gen = BigQueryExampleGen(query=query)
components.append(example_gen)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# TODO(step 5): Uncomment here to add StatisticsGen to the pipeline.
# components.append(statistics_gen)
# Generates schema based on statistics files.
schema_gen = SchemaGen(statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=False)
# TODO(step 5): Uncomment here to add SchemaGen to the pipeline.
# components.append(schema_gen)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator( # pylint: disable=unused-variable
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# TODO(step 5): Uncomment here to add ExampleValidator to the pipeline.
# components.append(example_validator)
# Performs transformations and feature engineering in training and serving.
transform = Transform(examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
preprocessing_fn=preprocessing_fn)
# TODO(step 6): Uncomment here to add Transform to the pipeline.
# components.append(transform)
# Uses user-provided Python function that implements a model using TF-Learn.
trainer_args = {
'run_fn':
run_fn,
'transformed_examples':
transform.outputs['transformed_examples'],
'schema':
schema_gen.outputs['schema'],
'transform_graph':
transform.outputs['transform_graph'],
'train_args':
train_args,
'eval_args':
eval_args,
'custom_executor_spec':
executor_spec.ExecutorClassSpec(trainer_executor.GenericExecutor),
}
if ai_platform_training_args is not None:
trainer_args.update({
'custom_executor_spec':
executor_spec.ExecutorClassSpec(
ai_platform_trainer_executor.GenericExecutor),
'custom_config': {
ai_platform_trainer_executor.TRAINING_ARGS_KEY:
ai_platform_training_args,
}
})
trainer = Trainer(**trainer_args)
# TODO(step 6): Uncomment here to add Trainer to the pipeline.
# components.append(trainer)
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.
LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# TODO(step 6): Uncomment here to add ResolverNode to the pipeline.
# components.append(model_resolver)
# Uses TFMA to compute a evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='tips')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='BinaryAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': eval_accuracy_threshold}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
# TODO(step 6): Uncomment here to add Evaluator to the pipeline.
# components.append(evaluator)
# Checks whether the model passed the validation steps and pushes the model
# to a file destination if check passed.
pusher_args = {
'model':
trainer.outputs['model'],
'model_blessing':
evaluator.outputs['blessing'],
'push_destination':
pusher_pb2.PushDestination(
filesystem=pusher_pb2.PushDestination.Filesystem(
base_directory=serving_model_dir)),
}
if ai_platform_serving_args is not None:
pusher_args.update({
'custom_executor_spec':
executor_spec.ExecutorClassSpec(
ai_platform_pusher_executor.Executor),
'custom_config': {
ai_platform_pusher_executor.SERVING_ARGS_KEY:
ai_platform_serving_args
},
})
pusher = Pusher(**pusher_args) # pylint: disable=unused-variable
# TODO(step 6): Uncomment here to add Pusher to the pipeline.
# components.append(pusher)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
# TODO(step 8): Change this value to control caching of execution results.
enable_cache=True,
metadata_connection_config=metadata_connection_config,
beam_pipeline_args=beam_pipeline_args,
)
def create_pipeline(
pipeline_name: Text,
pipeline_root: Text,
data_root: Text,
module_file: Text,
ai_platform_training_args: Dict[Text, Text],
ai_platform_serving_args: Dict[Text, Text],
enable_tuning: bool,
beam_pipeline_args: Optional[List[Text]] = None) -> pipeline.Pipeline:
"""Implements the Iris flowers pipeline with TFX and Kubeflow Pipeline.
Args:
pipeline_name: name of the TFX pipeline being created.
pipeline_root: root directory of the pipeline. Should be a valid GCS path.
data_root: uri of the Iris flowers data.
module_file: uri of the module files used in Trainer and Transform
components.
ai_platform_training_args: Args of CAIP training job. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.jobs#Job
for detailed description.
ai_platform_serving_args: Args of CAIP model deployment. Please refer to
https://cloud.google.com/ml-engine/reference/rest/v1/projects.models
for detailed description.
enable_tuning: If True, the hyperparameter tuning through CloudTuner is
enabled.
beam_pipeline_args: Optional list of beam pipeline options. Please refer to
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options.
When this argument is not provided, the default is to use GCP
DataflowRunner with 50GB disk size as specified in this function. If an
empty list is passed in, default specified by Beam will be used, which can
be found at
https://cloud.google.com/dataflow/docs/guides/specifying-exec-params#setting-other-cloud-dataflow-pipeline-options
Returns:
A TFX pipeline object.
"""
examples = external_input(data_root)
# Beam args to run data processing on DataflowRunner.
#
# TODO(b/151114974): Remove `disk_size_gb` flag after default is increased.
# TODO(b/151116587): Remove `shuffle_mode` flag after default is changed.
# TODO(b/156874687): Remove `machine_type` after IP addresses are no longer a
# scaling bottleneck.
if beam_pipeline_args is None:
beam_pipeline_args = [
'--runner=DataflowRunner',
'--project=' + _project_id,
'--temp_location=' + os.path.join(_output_bucket, 'tmp'),
'--region=' + _gcp_region,
# Temporary overrides of defaults.
'--disk_size_gb=50',
'--experiments=shuffle_mode=auto',
'--machine_type=n1-standard-8',
]
# Number of epochs in training.
train_steps = data_types.RuntimeParameter(
name='train_steps',
default=100,
ptype=int,
)
# Number of epochs in evaluation.
eval_steps = data_types.RuntimeParameter(
name='eval_steps',
default=50,
ptype=int,
)
# Brings data into the pipeline or otherwise joins/converts training data.
example_gen = CsvExampleGen(input=examples)
# Computes statistics over data for visualization and example validation.
statistics_gen = StatisticsGen(examples=example_gen.outputs['examples'])
# Generates schema based on statistics files.
schema_gen = SchemaGen(
statistics=statistics_gen.outputs['statistics'], infer_feature_shape=True)
# Performs anomaly detection based on statistics and data schema.
example_validator = ExampleValidator(
statistics=statistics_gen.outputs['statistics'],
schema=schema_gen.outputs['schema'])
# Performs transformations and feature engineering in training and serving.
transform = Transform(
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=module_file)
# Update ai_platform_training_args if distributed training was enabled.
# Number of worker machines used in distributed training.
worker_count = data_types.RuntimeParameter(
name='worker_count',
default=2,
ptype=int,
)
# Type of worker machines used in distributed training.
worker_type = data_types.RuntimeParameter(
name='worker_type',
default='standard',
ptype=str,
)
local_training_args = copy.deepcopy(ai_platform_training_args)
if FLAGS.distributed_training:
local_training_args.update({
# You can specify the machine types, the number of replicas for workers
# and parameter servers.
# https://cloud.google.com/ml-engine/reference/rest/v1/projects.jobs#ScaleTier
'scaleTier': 'CUSTOM',
'masterType': 'large_model',
'workerType': worker_type,
'parameterServerType': 'standard',
'workerCount': worker_count,
'parameterServerCount': 1,
})
# Tunes the hyperparameters for model training based on user-provided Python
# function. Note that once the hyperparameters are tuned, you can drop the
# Tuner component from pipeline and feed Trainer with tuned hyperparameters.
if enable_tuning:
# The Tuner component launches 1 AIP Training job for flock management.
# For example, 3 workers (defined by num_parallel_trials) in the flock
# management AIP Training job, each runs Tuner.Executor.
# Then, 3 AIP Training Jobs (defined by local_training_args) are invoked
# from each worker in the flock management Job for Trial execution.
tuner = Tuner(
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
tune_args=tuner_pb2.TuneArgs(
# num_parallel_trials=3 means that 3 search loops are
# running in parallel.
# Each tuner may include a distributed training job which can be
# specified in local_training_args above (e.g. 1 PS + 2 workers).
num_parallel_trials=3),
custom_config={
# Configures Cloud AI Platform-specific configs . For details, see
# https://cloud.google.com/ai-platform/training/docs/reference/rest/v1/projects.jobs#traininginput.
ai_platform_trainer_executor.TRAINING_ARGS_KEY:
local_training_args
})
# Uses user-provided Python function that trains a model.
trainer = Trainer(
custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_trainer_executor.GenericExecutor),
module_file=module_file,
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
# If Tuner is in the pipeline, Trainer can take Tuner's output
# best_hyperparameters artifact as input and utilize it in the user module
# code.
#
# If there isn't Tuner in the pipeline, either use ImporterNode to import
# a previous Tuner's output to feed to Trainer, or directly use the tuned
# hyperparameters in user module code and set hyperparameters to None
# here.
#
# Example of ImporterNode,
# hparams_importer = ImporterNode(
# instance_name='import_hparams',
# source_uri='path/to/best_hyperparameters.txt',
# artifact_type=HyperParameters)
# ...
# hyperparameters = hparams_importer.outputs['result'],
hyperparameters=(tuner.outputs['best_hyperparameters']
if enable_tuning else None),
train_args={'num_steps': train_steps},
eval_args={'num_steps': eval_steps},
custom_config={
ai_platform_trainer_executor.TRAINING_ARGS_KEY:
local_training_args
})
# Get the latest blessed model for model validation.
model_resolver = ResolverNode(
instance_name='latest_blessed_model_resolver',
resolver_class=latest_blessed_model_resolver.LatestBlessedModelResolver,
model=Channel(type=Model),
model_blessing=Channel(type=ModelBlessing))
# Uses TFMA to compute an evaluation statistics over features of a model and
# perform quality validation of a candidate model (compared to a baseline).
eval_config = tfma.EvalConfig(
model_specs=[tfma.ModelSpec(label_key='variety')],
slicing_specs=[tfma.SlicingSpec()],
metrics_specs=[
tfma.MetricsSpec(metrics=[
tfma.MetricConfig(
class_name='SparseCategoricalAccuracy',
threshold=tfma.MetricThreshold(
value_threshold=tfma.GenericValueThreshold(
lower_bound={'value': 0.6}),
change_threshold=tfma.GenericChangeThreshold(
direction=tfma.MetricDirection.HIGHER_IS_BETTER,
absolute={'value': -1e-10})))
])
])
evaluator = Evaluator(
examples=example_gen.outputs['examples'],
model=trainer.outputs['model'],
baseline_model=model_resolver.outputs['model'],
# Change threshold will be ignored if there is no baseline (first run).
eval_config=eval_config)
pusher = Pusher(
custom_executor_spec=executor_spec.ExecutorClassSpec(
ai_platform_pusher_executor.Executor),
model=trainer.outputs['model'],
model_blessing=evaluator.outputs['blessing'],
custom_config={
ai_platform_pusher_executor.SERVING_ARGS_KEY: ai_platform_serving_args
})
components = [
example_gen,
statistics_gen,
schema_gen,
example_validator,
transform,
trainer,
model_resolver,
evaluator,
pusher,
]
if enable_tuning:
components.append(tuner)
return pipeline.Pipeline(
pipeline_name=pipeline_name,
pipeline_root=pipeline_root,
components=components,
enable_cache=True,
beam_pipeline_args=beam_pipeline_args)