def train(pipeline,
output_dir,
train_args_dict,
train_features=None,
eval_features=None,
metadata=None):
if not train_features:
train_features = (pipeline
| 'ReadTrain' >> io.LoadFeatures(
os.path.join(output_dir, 'features_train*')))
if not eval_features:
eval_features = (pipeline
| 'ReadEval' >> io.LoadFeatures(
os.path.join(output_dir, 'features_eval*')))
if not metadata:
metadata = (pipeline
| 'ReadMetadata' >> io.LoadMetadata(
os.path.join(output_dir, METADATA_FILE_NAME)))
trained_model, results = ((train_features, eval_features)
| 'Train' >> ml.Train(**train_args_dict))
trained_model | 'SaveModel' >> io.SaveModel(
os.path.join(output_dir, 'saved_model'))
results | io.SaveTrainingJobResult(
os.path.join(output_dir, 'train_results'))
return trained_model, results
def model_analysis(pipeline, output_dir, evaluation_data=None, metadata=None):
if not metadata:
metadata = (pipeline
| 'LoadMetadataForAnalysis' >> io.LoadMetadata(
os.path.join(output_dir, METADATA_FILE_NAME)))
if not evaluation_data:
coder = io.CsvCoder(
column_names=[
'key', 'target', 'predicted', 'score', 'target_label',
'predicted_label', 'all_scores'
],
numeric_column_names=['target', 'predicted', 'score'])
evaluation_data = (
pipeline
| 'ReadEvaluation' >> beam.io.ReadFromText(
os.path.join(output_dir, 'model_evaluations*'), coder=coder))
confusion_matrix, precision_recall, logloss = (
evaluation_data
| 'AnalyzeModel' >> ml.AnalyzeModel(metadata))
confusion_matrix | io.SaveConfusionMatrixCsv(
os.path.join(output_dir, 'analyzer_cm.csv'))
precision_recall | io.SavePrecisionRecallCsv(
os.path.join(output_dir, 'analyzer_pr.csv'))
(logloss
| 'WriteLogLoss' >> beam.io.WriteToText(
os.path.join(output_dir, 'analyzer_logloss'), file_name_suffix='.csv')
)
return confusion_matrix, precision_recall, logloss
def model_analysis(pipeline, evaluation_data=None, metadata=None):
if not metadata:
metadata = pipeline | io.LoadMetadata(
os.path.join(args.output_dir, "metadata.yaml"))
if not evaluation_data:
coder = io.CsvCoder(['key', 'target', 'predicted', 'score'],
['target', 'predicted', 'score'])
evaluation_data = read_text_file(pipeline, 'ReadEvaluation',
'model_evaluations', coder=coder)
confusion_matrix, precision_recall, logloss = (
evaluation_data | 'AnalyzeModel' >> ml.AnalyzeModel(metadata))
confusion_matrix | io.SaveConfusionMatrixCsv(
os.path.join(args.output_dir, 'analyzer_cm.csv'))
precision_recall | io.SavePrecisionRecallCsv(
os.path.join(args.output_dir, 'analyzer_pr.csv'))
write_text_file(logloss, 'Write Log Loss', 'analyzer_logloss.csv')
return confusion_matrix, precision_recall, logloss
