def preprocess(pipeline, training_data, eval_data, predict_data, output_dir):
"""Read in input files, runs ml.Preprocess, and writes preprocessed output.
Args:
pipeline: beam pipeline
training_data, eval_data, predict_data: file paths to input csv files.
output_dir: file path to where to write all the output files.
Returns:
metadata and preprocessed features as pcollections.
"""
feature_set = iris.IrisFeatures()
coder_with_target = io.CsvCoder.from_feature_set(feature_set,
feature_set.csv_columns)
coder_without_target = io.CsvCoder.from_feature_set(feature_set,
feature_set.csv_columns,
has_target_columns=False)
train = (
pipeline
| 'ReadTrainingData'
>> beam.io.textio.ReadFromText(training_data, coder=coder_with_target))
evaluate = (
pipeline
| 'ReadEvalData'
>> beam.io.textio.ReadFromText(eval_data, coder=coder_with_target))
predict = (
pipeline
| 'ReadPredictData'
>> beam.io.textio.ReadFromText(predict_data, coder=coder_without_target))
# TODO(b/32726166) Update input_format and format_metadata to read from these
# values directly from the coder.
(metadata, train_features, eval_features, predict_features) = (
(train, evaluate, predict)
| 'Preprocess' >> ml.Preprocess(
feature_set,
input_format='csv',
format_metadata={
'headers': feature_set.csv_columns
}))
# Writes metadata.json - specified through METADATA_FILENAME- (text file),
# features_train, features_eval, and features_eval (TFRecord files)
(metadata | 'SaveMetadata'
>> io.SaveMetadata(os.path.join(output_dir, METADATA_FILE_NAME)))
# We turn off sharding of the feature files because the dataset is very small.
(train_features | 'SaveTrain'
>> io.SaveFeatures(os.path.join(output_dir, 'features_train')))
(eval_features | 'SaveEval'
>> io.SaveFeatures(os.path.join(output_dir, 'features_eval')))
(predict_features | 'SavePredict'
>> io.SaveFeatures(os.path.join(output_dir, 'features_predict')))
return metadata, train_features, eval_features, predict_features
def preprocess(pipeline, training_data, eval_data, output_dir,
frequency_threshold, metadata_file_name):
"""Run pre-processing step as a pipeline.
Args:
pipeline: beam pipeline
training_data: file paths to input csv files.
eval_data: file paths to input csv files.
output_dir: file path to where to write all the output files.
frequency_threshold: frequency threshold to use for categorical values.
metadata_file_name: one of metadata.{json|yaml}.
"""
feature_set, csv_columns = criteo.criteo_features(
frequency_threshold=frequency_threshold)
coder = io.CsvCoder.from_feature_set(feature_set, csv_columns, delimiter='\t')
train = (
pipeline
| 'ReadTrainingData'
>> beam.io.ReadFromText(
training_data, strip_trailing_newlines=True, coder=coder))
evaluate = (
pipeline
| 'ReadEvalData'
>> beam.io.ReadFromText(
eval_data, strip_trailing_newlines=True, coder=coder))
# TODO(b/32726166) Update input_format and format_metadata to read from these
# values directly from the coder.
(metadata, train_features, evaluate_features) = (
(train, evaluate)
| 'Preprocess' >> ml.Preprocess(
feature_set,
input_format='csv',
format_metadata={'headers': csv_columns,
'delimiter': '\t'}))
# Writes metadata.json, features_train, features_eval, and features_eval files
# pylint: disable=expression-not-assigned
(metadata
| 'SaveMetadata'
>> io.SaveMetadata(os.path.join(output_dir, metadata_file_name)))
(train_features
| 'WriteTraining'
>> io.SaveFeatures(os.path.join(output_dir, 'features_train')))
(evaluate_features
| 'WriteEval'
>> io.SaveFeatures(os.path.join(output_dir, 'features_eval')))
def preprocess(pipeline, training_data, eval_data, predict_data, output_dir,
metadata_file_name):
feature_set = iris.IrisFeatures()
read_training_data = beam.io.ReadFromText(
training_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(feature_set,
feature_set.csv_columns))
read_eval_data = beam.io.ReadFromText(eval_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(
feature_set,
feature_set.csv_columns))
read_predict_data = beam.io.ReadFromText(
predict_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(feature_set,
feature_set.csv_columns,
has_target_columns=False))
train = pipeline | 'ReadTrainingData' >> read_training_data
evaluate = pipeline | 'ReadEvalData' >> read_eval_data
predict = pipeline | 'ReadPredictData' >> read_predict_data
# TODO(b/32726166) Update input_format and format_metadata to read from these
# values directly from the coder.
(metadata, train_features, eval_features, predict_features) = (
(train, evaluate, predict)
| 'Preprocess' >> ml.Preprocess(
feature_set,
input_format='csv',
format_metadata={'headers': feature_set.csv_columns}))
# pylint: disable=expression-not-assigned
(metadata | 'SaveMetadata' >> io.SaveMetadata(
os.path.join(output_dir, metadata_file_name)))
# We turn off sharding of these feature files because the dataset very small.
(train_features | 'SaveTrain' >> io.SaveFeatures(
os.path.join(output_dir, 'features_train')))
(eval_features |
'SaveEval' >> io.SaveFeatures(os.path.join(output_dir, 'features_eval')))
(predict_features | 'SavePredict' >> io.SaveFeatures(
os.path.join(output_dir, 'features_predict')))
# pylint: enable=expression-not-assigned
return metadata, train_features, eval_features, predict_features
def preprocess(pipeline):
feature_set = iris.IrisFeatures()
training_data = beam.io.TextFileSource(args.training_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(
feature_set,
feature_set.csv_columns))
eval_data = beam.io.TextFileSource(args.eval_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(
feature_set,
feature_set.csv_columns))
predict_data = beam.io.TextFileSource(args.predict_data,
strip_trailing_newlines=True,
coder=io.CsvCoder.from_feature_set(
feature_set,
feature_set.csv_columns,
has_target_columns=False))
train = pipeline | beam.Read('ReadTrainingData', training_data)
evaluate = pipeline | beam.Read('ReadEvalData', eval_data)
predict = pipeline | beam.Read('ReadPredictData', predict_data)
(metadata, train_features, eval_features, predict_features) = (
(train, evaluate, predict)
| 'Preprocess' >> ml.Preprocess(
feature_set,
input_format='csv',
format_metadata={'headers': feature_set.csv_columns}))
# Writes metadata.yaml, features_train, features_eval, and features_eval files
# pylint: disable=expression-not-assigned
(metadata | 'SaveMetadata' >> io.SaveMetadata(
os.path.join(args.output_dir, 'metadata.yaml')))
# We turn off sharding of these feature files because the dataset very small.
(train_features | 'SaveTrain' >> io.SaveFeatures(
os.path.join(args.output_dir, 'features_train')))
(eval_features | 'SaveEval' >> io.SaveFeatures(
os.path.join(args.output_dir, 'features_eval')))
(predict_features | 'SavePredict' >> io.SaveFeatures(
os.path.join(args.output_dir, 'features_predict')))
# pylint: enable=expression-not-assigned
return metadata, train_features, eval_features, predict_features
