def run(self, input_dataset, eval_datasets=[], return_pipeline=False):
"""
Converts internal pipeline architecture dict into pipeline and runs it.
Args:
- input_dataset: Input dataset to train
- eval_dataset: Dataset to evaluate
- return_pipeline: Whether to return the pipeline which fitted and produced the preds
Returns:
- If return_pipeline is False, returns just the predictions, otherwise returns a tuple
(preds, pipeline)
"""
pipeline = self.load_pipeline_architecture(self.pipeline_architecture_dict)
pipeline.check()
runtime = Runtime(pipeline, context=Context.TESTING)
runtime.fit(inputs=[input_dataset], return_values=['outputs.0'])
all_preds = []
for dataset in eval_datasets:
all_preds.append(runtime.produce(inputs=[dataset], return_values=['outputs.0']))
results = all_preds
if return_pipeline:
results = (all_preds, pipeline)
return results
def score_pipeline(dataset_root, problem, pipeline_path):
train_dataset = load_dataset(dataset_root, 'TRAIN')
test_dataset = load_dataset(dataset_root, 'SCORE', 'TEST')
pipeline = load_pipeline(pipeline_path)
# Creating an instance on runtime with pipeline description and problem description.
runtime = Runtime(
pipeline=pipeline,
problem_description=problem,
context=Context.TESTING
)
print("Fitting the pipeline")
fit_results = runtime.fit(inputs=[train_dataset])
fit_results.check_success()
# Producing results using the fitted pipeline.
print("Producing predictions")
produce_results = runtime.produce(inputs=[test_dataset])
produce_results.check_success()
predictions = produce_results.values['outputs.0']
metrics = problem['problem']['performance_metrics']
print("Computing the score")
scoring_pipeline = load_pipeline('ta2/pipelines/scoring_pipeline.yml')
scores, scoring_pipeline_run = score(
scoring_pipeline, problem, predictions, [test_dataset], metrics,
context=Context.TESTING, random_seed=0,
)
return scores.iloc[0].value
def run_pipeline(pipeline, dataset_name, datasets_path):
ensure_downloaded(dataset_name, datasets_path)
root_path = os.path.join(os.path.abspath(datasets_path), dataset_name)
train_dataset = load_dataset(root_path, 'TRAIN')
train_problem = load_problem(root_path, 'TRAIN')
# Creating an instance on runtime with pipeline description and problem description.
runtime = Runtime(pipeline=pipeline,
problem_description=train_problem,
context=Context.TESTING)
# Fitting pipeline on input dataset.
fit_results = runtime.fit(inputs=[train_dataset])
fit_results.check_success()
# Producing results using the fitted pipeline.
test_dataset = load_dataset(root_path, 'TEST')
produce_results = runtime.produce(inputs=[test_dataset])
produce_results.check_success()
print('Pipeline run successfully')
output = list(produce_results.values.values())[0]
print(output.shape)
print(output.head())
def __run_pipeline(self,
pipeline_description,
data,
volume_dir='/volumes'):
runtime = Runtime(pipeline=pipeline_description,
context=metadata_base.Context.TESTING,
volumes_dir=volume_dir)
fit_result = runtime.fit([data])
return fit_result
def fitproduce(self, input_item):
problem_doc, pipeline_json, dataset_train, dataset_test = input_item[
1:]
# Run pipeline
pipeline = Pipeline.from_json(pipeline_json)
pipeline_runtime = Runtime(pipeline, context=Context.TESTING)
pipeline_runtime.fit(inputs=[dataset_train],
return_values=['outputs.0'])
score_predictions = pipeline_runtime.produce(
inputs=[dataset_test], return_values=['outputs.0'])
score_predictions = score_predictions.values['outputs.0']
# Write predictions to output path
path = self.get_predictions_save_path()
utils.utils.write_predictions_to_file(score_predictions, path,
problem_doc)
path_uri = "file://%s" % path
return path_uri
def score(self, input_item):
problem_doc, metric, pipeline_json, dataset_train, dataset_test = input_item[
1:]
# Run pipeline
pipeline = Pipeline.from_json(pipeline_json)
pipeline_runtime = Runtime(pipeline, context=Context.TESTING)
pipeline_runtime.fit(inputs=[dataset_train],
return_values=['outputs.0'])
score_predictions = pipeline_runtime.produce(
inputs=[dataset_test], return_values=['outputs.0'])
score_predictions = score_predictions.values['outputs.0']
# Evaluate scores on score dir
achieved_score = utils.train_utils.score(score_predictions,
dataset_test,
problem_doc,
override_metric_key=metric)
return achieved_score
def score_pipeline(dataset,
problem,
pipeline_path,
static=None,
output_path=None):
pipeline = load_pipeline(pipeline_path)
# Creating an instance on runtime with pipeline description and problem description.
runtime = Runtime(
pipeline=pipeline,
problem_description=problem,
context=Context.EVALUATION,
volumes_dir=static,
)
LOGGER.info("Fitting pipeline %s", pipeline_path)
fit_results = runtime.fit(inputs=[dataset])
fit_results.check_success()
dataset_doc_path = dataset.metadata.query(())['location_uris'][0]
dataset_root = dataset_doc_path[:-len(
'/TRAIN/dataset_TRAIN/datasetDoc.json')]
test_dataset = load_dataset(dataset_root, 'SCORE', 'TEST')
# Producing results using the fitted pipeline.
LOGGER.info("Producing predictions for pipeline %s", pipeline_path)
produce_results = runtime.produce(inputs=[test_dataset])
produce_results.check_success()
predictions = produce_results.values['outputs.0']
metrics = problem['problem']['performance_metrics']
LOGGER.info("Computing the score for pipeline %s", pipeline_path)
scoring_pipeline = load_pipeline(DEFAULT_SCORING_PIPELINE_PATH)
scores, scoring_pipeline_run = score(
scoring_pipeline=scoring_pipeline,
problem_description=problem,
predictions=predictions,
score_inputs=[test_dataset],
metrics=metrics,
context=Context.EVALUATION,
random_seed=0,
)
evaluated_pipeline_run = produce_results.pipeline_run
evaluated_pipeline_run.is_standard_pipeline = True
evaluated_pipeline_run.set_scores(scores, metrics)
evaluated_pipeline_run.set_scoring_pipeline_run(
scoring_pipeline_run.pipeline_run, [dataset])
_to_yaml_run(evaluated_pipeline_run, output_path)
return scores.iloc[0].value
import utils.train_utils
if __name__ == "__main__":
# Get args
try:
path_to_pipeline_json = sys.argv[1]
inputdir = sys.argv[2]
# Load datasets
problem_doc, dataset = utils.utils.load_data_from_dir(inputdir)
# Create pipeline
with open(path_to_pipeline_json, "r") as f:
pipeline = Pipeline.from_json(f.read())
pipeline_runtime = Runtime(pipeline, context=Context.TESTING)
pipeline_runtime.fit(inputs=[dataset], return_values=['outputs.0'])
problem_doc_score, dataset_score = utils.utils.load_data_from_dir(inputdir, mode="score")
score_predictions = pipeline_runtime.produce(inputs=[dataset_score], return_values=['outputs.0'])
score_predictions = score_predictions.values['outputs.0']
# Evaluate scores on score dir
achieved_score = utils.train_utils.score(score_predictions, dataset_score, problem_doc_score)
print(achieved_score)
except:
print("N/A")
def write_pipeline_run(self, problem_description, dataset, filename_yaml):
runtime = Runtime(pipeline=self.pipeline_description, problem_description=problem_description, context=Context.TESTING, is_standard_pipeline=True)
output = runtime.fit(inputs=dataset)
pipeline_run = output.pipeline_run
with open(filename_yaml, "w") as out:
pipeline_run.to_yaml(file=filename_yaml)
outputdir = sys.argv[3]
################
# Load dataset #
################
problem_doc_train, dataset_train = (
utils.utils.load_data_from_dir(inputdir, mode="train"))
###################
# Create pipeline #
###################
with open(path_to_pipeline_json, "r") as f:
pipeline = utils.primitive_pipeline_utils.load_pipeline(f.read())
pipeline_runtime = Runtime(pipeline)
pipeline_runtime.fit(inputs=[dataset_train])
############################
# Try scoring on SCORE set #
############################
try:
problem_doc_score, dataset_score = utils.utils.load_data_from_dir(inputdir, mode="score")
score_predictions = pipeline_runtime.produce(inputs=[dataset_score]).values['outputs.0']
validation_score = utils.train_utils.get_score_on_score_set(inputdir, problem_doc_score, score_predictions)
print("PipelineId: %s, Score: %s" % (pipeline.id, validation_score))
except:
print("-------------------------------")
print("Failed to evaluate on SCORE set")
print("-------------------------------")
traceback.print_exc()
# Loading problem description.
problem_description = problem.parse_problem_description(problem_path)
# Loading dataset.
path = 'file://{uri}'.format(uri=os.path.abspath(dataset_train_path))
dataset = D3MDatasetLoader().load(dataset_uri=path)
path2 = 'file://{uri}'.format(uri=os.path.abspath(dataset_predict_path))
dataset_predict = D3MDatasetLoader().load(dataset_uri=path2)
# Loading pipeline description file.
with open(pipeline_path, 'r') as file:
pipeline_description = pipeline_module.Pipeline.from_json(
string_or_file=file)
# Creating an instance on runtime with pipeline description and problem description.
runtime = Runtime(pipeline=pipeline_description,
problem_description=problem_description,
context=metadata_base.Context.TESTING)
# Fitting pipeline on input dataset.
fit_results = runtime.fit(inputs=[dataset])
fit_results.check_success()
# Producing results using the fitted pipeline.
produce_results = runtime.produce(inputs=[dataset_predict])
produce_results.check_success()
print(produce_results.values)