def validate(pipeline_params, model, encoder, track, date_cutoff, max_date):
target_name = 'unit_sales'
validate_stream = (row for row in stream_data(pipeline_params)
if validate_filter(row, date_cutoff, max_date))
encoded_validate_stream = encoder.encode_data_stream(validate_stream)
validation_predictions = [float(model.predict([row]))
for row in encoded_validate_stream]
target = [row[target_name] for row in stream_data(pipeline_params)
if validate_filter(row, date_cutoff, max_date)]
print("Calculating metrics")
validation_metrics = {'r2_score': metrics.r2_score(
y_true=target, y_pred=validation_predictions)}
track.log_metrics(validation_metrics)
fluentd_logger.log('validation_metrics', validation_metrics)
write_predictions_and_score(validation_metrics)
print("Evaluation done with metrics {}.".format(
json.dumps(validation_metrics)))
write_model(model)
track.log_artifact(file_names['model'])
make_validation_plot(target, validation_predictions, track)
def run_ml_model(pipeline_params, encoder, track, date_cutoff, seed=None):
target_name = 'unit_sales'
train_stream = (row for row in stream_data(pipeline_params)
if train_filter(row, date_cutoff))
encoded_train_stream = encoder.encode_data_stream(train_stream)
print('Encoding data')
# batch step, read it all in
encoded_train_data = list(encoded_train_stream)
print('Getting target')
# read it all in
target = [
row[target_name] for row in stream_data(pipeline_params)
if train_filter(row, date_cutoff)
]
model_name = pipeline_params['model_name']
params = pipeline_params['model_params'][model_name]
track.log_ml_params(params)
track.log_pipeline_params(pipeline_params)
trained_model, params = train_model(encoded_train_data,
target,
model_name,
params,
seed=seed)
return trained_model, params
def test_get_encoder_from_stream():
stream = stream_data(pipeline_params)
stream_small = (next(stream) for _ in range(100))
encoder = get_encoder_from_stream(stream_small)
assert isinstance(encoder, OneHotEncoder)
stream = stream_data(pipeline_params)
stream_small = (next(stream) for _ in range(100))
encoded = list(encoder.encode_data_stream(stream_small))
assert len(encoded) == 100
row_in = {
'id': '88219279',
'date': '2016-08-16',
'item_nbr': '103520',
'unit_sales': '10.0',
'family': 'GROCERY I',
'class': '1028',
'perishable': '0',
'year': '2016',
'month': '8',
'day': '16',
'dayofweek': '1',
'days_til_end_of_data': '364',
'dayoff': 'False'
}
encoded = encoder.encode_data([row_in])
decoded = encoder.decode_data(encoded)
print(decoded[0].keys())
print(row_in.keys())
del row_in['date'], row_in['id'], row_in['unit_sales']
assert decoded[0].keys() == row_in.keys()
# make a level not seen
row_in['class'] = 'FOO'
encoded = encoder.encode_data([row_in])
decoded = encoder.decode_data(encoded)
print(decoded[0].keys())
print(row_in.keys())
assert decoded[0].keys() == row_in.keys()
print(decoded)
assert decoded[0]['class'] == 'UNKNOWN_CATEGORICAL_LEVEL'
def test_stream_data():
stream = stream_data(pipeline_params)
row = next(stream)
assert isinstance(row, dict)
assert 'perishable' in row
assert isinstance(row['perishable'], int)
def get_max_date(pipeline_params):
# batch step
# '2017-08-15'
print('Getting max date')
max_date = max(
date_string_to_date(row["date"])
for row in stream_data(pipeline_params))
print('Max date: %s' % max_date)
return max_date.strftime('%Y-%m-%d')