def test_experiment_sequence_combiner(csv_filename):
# Machine translation with attention
input_features_template = Template(
'[{name: english, type: sequence, vocab_size: 10,'
' max_len: 10, min_len: 10, encoder: rnn, cell_type: lstm,'
' reduce_output: null}, {name: spanish, type: sequence, vocab_size: 10,'
' max_len: 10, min_len: 10, encoder: rnn, cell_type: lstm,'
' reduce_output: null}, {name: category,'
' type: category, vocab_size: 10} ]')
output_features_string = "[{type: category, name: intent, reduce_input:" \
" sum, vocab_size: 10}]"
model_definition_template2 = Template(
'{input_features: ${input_name}, output_features: ${output_name}, '
'training: {epochs: 2}, combiner: {type: sequence_concat, encoder: rnn,'
'main_sequence_feature: random_sequence}}')
# Generate test data
rel_path = generate_data(
input_features_template.substitute(encoder1='rnn', encoder2='rnn'),
output_features_string, csv_filename)
for encoder1, encoder2 in zip(encoders, encoders):
input_features = input_features_template.substitute(encoder1=encoder1,
encoder2=encoder2)
model_definition = model_definition_template2.substitute(
input_name=input_features, output_name=output_features_string)
experiment(yaml.load(model_definition),
skip_save_processed_input=True,
skip_save_progress=True,
skip_save_unprocessed_output=True,
data_csv=rel_path)
def test_experiment_model_resume(csv_filename):
# Single sequence input, single category output
# Tests saving a model file, loading it to rerun training and predict
input_features = [sequence_feature(encoder='rnn', reduce_output='sum')]
output_features = [categorical_feature(vocab_size=2, reduce_input='sum')]
# Generate test data
rel_path = generate_data(input_features, output_features, csv_filename)
model_definition = {
'input_features': input_features,
'output_features': output_features,
'combiner': {
'type': 'concat',
'fc_size': 14
},
'training': {
'epochs': 2
}
}
exp_dir_name = experiment(model_definition, data_csv=rel_path)
logging.info('Experiment Directory: {0}'.format(exp_dir_name))
experiment(model_definition,
data_csv=rel_path,
model_resume_path=exp_dir_name)
full_predict(os.path.join(exp_dir_name, 'model'), data_csv=rel_path)
shutil.rmtree(exp_dir_name, ignore_errors=True)
def run_experiment(input_features, output_features, **kwargs):
"""
Helper method to avoid code repetition in running an experiment. Deletes
the data saved to disk after running the experiment
:param input_features: list of input feature dictionaries
:param output_features: list of output feature dictionaries
**kwargs you may also pass extra parameters to the experiment as keyword
arguments
:return: None
"""
model_definition = None
if input_features is not None and output_features is not None:
# This if is necessary so that the caller can call with
# model_definition_file (and not model_definition)
model_definition = {
'input_features': input_features,
'output_features': output_features,
'combiner': {
'type': 'concat',
'fc_size': 14
},
'training': {'epochs': 2}
}
args = {
'model_definition': model_definition,
'skip_save_processed_input': True,
'skip_save_progress': True,
'skip_save_unprocessed_output': True,
}
args.update(kwargs)
exp_dir_name = experiment(**args)
shutil.rmtree(exp_dir_name, ignore_errors=True)
def run_experiment(input_features, output_features, data_csv):
"""
Helper method to avoid code repetition in running an experiment
:param input_features: input schema
:param output_features: output schema
:param data_csv: path to data
:return: None
"""
model_definition = model_definition_template.substitute(
input_name=input_features, output_name=output_features)
experiment(yaml.load(model_definition),
skip_save_processed_input=True,
skip_save_progress=True,
skip_save_unprocessed_output=True,
data_csv=data_csv)
def test_experiment_sequence_combiner(csv_filename):
# Sequence combiner
input_features = [
sequence_feature(
name='english',
min_len=5,
max_len=5,
encoder='rnn',
cell_type='lstm',
reduce_output=None
),
sequence_feature(
name='spanish',
min_len=5,
max_len=5,
encoder='rnn',
cell_type='lstm',
reduce_output=None
),
category_feature(vocab_size=5)
]
output_features = [
category_feature(reduce_input='sum', vocab_size=5)
]
model_definition = {
'input_features': input_features,
'output_features': output_features,
'training': {
'epochs': 2
},
'combiner': {
'type': 'sequence_concat',
'encoder': 'rnn',
'main_sequence_feature': 'random_sequence'
}
}
# Generate test data
rel_path = generate_data(input_features, output_features, csv_filename)
for encoder in ENCODERS[:-2]:
logger.error('sequence combiner. encoders: {0}, {1}'.format(
encoder,
encoder
))
input_features[0]['encoder'] = encoder
input_features[1]['encoder'] = encoder
model_definition['input_features'] = input_features
exp_dir_name = experiment(
model_definition,
skip_save_processed_input=False,
skip_save_progress=True,
skip_save_unprocessed_output=True,
data_csv=rel_path
)
shutil.rmtree(exp_dir_name, ignore_errors=True)
def test_experiment_model_resume(csv_filename):
# Single sequence input, single category output
# Tests saving a model file, loading it to rerun training and predict
input_features = '[{name: utterance, type: sequence, vocab_size: 10,' \
' max_len: 10, encoder: rnn, reduce_output: sum}]'
output_features = "[{name: intent, type: category, vocab_size: 2," \
" reduce_input: sum}] "
# Generate test data
rel_path = generate_data(input_features, output_features, csv_filename)
model_definition = model_definition_template.substitute(
input_name=input_features, output_name=output_features)
exp_dir_name = experiment(yaml.load(model_definition), data_csv=rel_path)
logging.info('Experiment Directory: {0}'.format(exp_dir_name))
experiment(yaml.load(model_definition),
data_csv=rel_path,
model_resume_path=exp_dir_name)
full_predict(os.path.join(exp_dir_name, 'model'), data_csv=rel_path)
def test_experiment_seq_seq1_model_def_file(csv_filename, yaml_filename):
# seq-to-seq test to use model definition file instead of dictionary
input_features = ('[{name: utt, type: text, reduce_output: null, '
'vocab_size: 10, min_len: 10, max_len: 10,'
' encoder: embed}]')
output_features = ('[{name: iob, type: text, reduce_input: null, '
'vocab_size: 3, min_len: 10, max_len: 10,'
' decoder: tagger}]')
# Save the model definition to a yaml file
model_definition = yaml.load(
model_definition_template.substitute(input_name=input_features,
output_name=output_features))
with open(yaml_filename, 'w') as yaml_out:
yaml.dump(model_definition, yaml_out)
rel_path = generate_data(input_features, output_features, csv_filename)
experiment(model_definition=None,
model_definition_file=yaml_filename,
skip_save_processed_input=True,
skip_save_progress=True,
skip_save_unprocessed_output=True,
data_csv=rel_path)