def save_model(self, save_path):
ensure_dir(os.path.dirname(save_path))
all_params = get_all_params(self._net['dist'])
with open(save_path, 'wb') as f:
params = {v.name: v.get_value() for v in all_params}
np.savez(f, **params)
_logger.info('\nSaved model:\n{}\n'.format(save_path))
def save_model(self, save_model_path):
ensure_dir(os.path.dirname(save_model_path))
ordered_params = get_all_param_values(self._net['dist'])
with open(save_model_path, 'wb') as f:
np.savez(f, *ordered_params)
_logger.info('\nSaved model:\n{}\n'.format(save_model_path))
def download(self, remote_file_name, local_file_name):
"""
Download file from AWS S3 to the local one
"""
# create dir if not exists for storing file from s3
ensure_dir(os.path.dirname(local_file_name))
self._logger.info('Getting file %s from AWS S3 and saving it as %s' % (remote_file_name, local_file_name))
self._bucket_client.download_file(remote_file_name, local_file_name)
self._logger.info('Got file %s from S3' % remote_file_name)
def download(self, remote_file_name, local_file_name):
"""
Download file from AWS S3 to the local one
"""
remote_file_name = os.path.normpath(remote_file_name)
# create dir if not exists for storing file from s3
ensure_dir(os.path.dirname(local_file_name))
self._logger.info('Getting file %s from AWS S3 and saving it as %s' % (remote_file_name, local_file_name))
self._bucket_client.download_file(remote_file_name, local_file_name)
self._logger.info('Got file %s from S3' % remote_file_name)
def log_predictions(predictions_path,
x_test,
nn_model,
mode,
candidates_num=None,
stats_info=None,
cur_perplexity=None,
output_seq_len=MAX_PREDICTIONS_LENGTH,
**kwargs):
_logger.info('Logging responses to test lines')
# Create all the directories for the prediction path in case they don't exist
prediction_dir = os.path.dirname(predictions_path)
if prediction_dir:
ensure_dir(prediction_dir)
with open(predictions_path, 'w') as test_res_fh:
csv_writer = init_csv_writer(test_res_fh, mode, output_seq_len)
if cur_perplexity:
csv_writer.writerow(['Current perplexity: %.2f' % cur_perplexity])
if stats_info:
csv_writer.writerow([_get_iteration_stats(stats_info)])
csv_writer.writerow(
['input sentence'] +
['candidate #{}'.format(v + 1) for v in xrange(candidates_num)])
questions = transform_context_token_ids_to_sentences(
x_test, nn_model.index_to_token)
_logger.info(
'Start predicting responses of length {out_len} for {n_samples} samples with mode {mode}'
.format(out_len=output_seq_len,
n_samples=x_test.shape[0],
mode=mode))
nn_responses = get_nn_responses(x_test, nn_model, mode, candidates_num,
output_seq_len, **kwargs)
_logger.info('Logging generated predictions...')
for idx, (question,
responses) in enumerate(zip(questions, nn_responses)):
csv_writer.writerow([question] + responses)
_logger.info(
'Succesfully dumped {n_resp} of {n_resp} responses'.format(
n_resp=len(questions)))
_logger.info('Here they are: {}'.format(predictions_path))
def log_predictions(predictions_path,
contexts_token_ids,
nn_model,
prediction_modes,
output_seq_len=MAX_PREDICTIONS_LENGTH,
**kwargs):
"""
Generate responses for provided contexts and save the results on the disk. For a given context
several responses will be generated - one for each mode from the prediction_modes list.
:param predictions_path: Generated responses will be saved to this file
:param contexts_token_ids: contexts token ids, numpy array of shape (batch_size, context_len, INPUT_SEQUENCE_LENGTH)
:param nn_model: instance of CakeChatModel class
:param prediction_modes: See PREDICTION_MODES for available options
:param output_seq_len: Max number of tokens in generated responses
"""
_logger.info('Logging responses for test set')
# Create all the directories for the prediction path in case they don't exist
ensure_dir(os.path.dirname(predictions_path))
_init_csv_writer(predictions_path, output_seq_len, nn_model.model_name)
contexts = transform_context_token_ids_to_sentences(
contexts_token_ids, nn_model.index_to_token)
predictions_data = pd.DataFrame()
predictions_data['contexts'] = contexts
for prediction_mode in prediction_modes:
predicted_responses = get_nn_responses(contexts_token_ids, nn_model,
prediction_mode, **kwargs)
# list of lists of strings, shape (contexts_num, 1)
predicted_responses = [response[0] for response in predicted_responses]
# list of strings, shape (contexts_num)
predictions_data[prediction_mode] = predicted_responses
predictions_data.to_csv(predictions_path,
sep='\t',
index=False,
encoding='utf-8',
mode='a',
float_format='%.2f')
_logger.info('Dumped {} predicted responses to {}'.format(
len(contexts), predictions_path))
def log_predictions(predictions_path,
x_test,
nn_model,
mode,
candidates_num=None,
stats_info=None,
cur_perplexity=None,
output_seq_len=MAX_PREDICTIONS_LENGTH,
**kwargs):
_logger.info('Logging responses to test lines')
# Create all the directories for the prediction path in case they don't exist
prediction_dir = os.path.dirname(predictions_path)
if prediction_dir:
ensure_dir(prediction_dir)
with open(predictions_path, 'w') as test_res_fh:
csv_writer = init_csv_writer(test_res_fh, mode, output_seq_len)
if cur_perplexity:
csv_writer.writerow(['Current perplexity: %.2f' % cur_perplexity])
if stats_info:
csv_writer.writerow([_get_iteration_stats(stats_info)])
csv_writer.writerow(['input sentence'] + ['candidate #{}'.format(v + 1) for v in xrange(candidates_num)])
questions = transform_context_token_ids_to_sentences(x_test, nn_model.index_to_token)
_logger.info('Start predicting responses of length {out_len} for {n_samples} samples with mode {mode}'.format(
out_len=output_seq_len, n_samples=x_test.shape[0], mode=mode))
nn_responses = get_nn_responses(x_test, nn_model, mode, candidates_num, output_seq_len, **kwargs)
_logger.info('Logging generated predictions...')
for idx, (question, responses) in enumerate(zip(questions, nn_responses)):
csv_writer.writerow([question] + responses)
_logger.info('Succesfully dumped {n_resp} of {n_resp} responses'.format(n_resp=len(questions)))
_logger.info('Here they are: {}'.format(predictions_path))
def dump_index_to_item(index_to_item, path):
ensure_dir(os.path.dirname(path))
with codecs.open(path, 'w', 'utf-8') as fh:
json.dump(index_to_item, fh, ensure_ascii=False)
def save_weights(self, save_path):
ensure_dir(os.path.dirname(save_path))
ordered_params = get_all_param_values(self._net['dist'])
with open(save_path, 'wb') as f:
np.savez(f, *ordered_params)
def _save_model(model, model_path):
_logger.info('Saving model to {}'.format(model_path))
ensure_dir(os.path.dirname(model_path))
model.save(model_path, separately=[])
_logger.info('Model has been saved')
def dump_index_to_item(index_to_item, path):
ensure_dir(os.path.dirname(path))
with codecs.open(path, 'w', 'utf-8') as fh:
json.dump(index_to_item, fh, ensure_ascii=False)
def log_predictions(predictions_path,
x_test,
nn_model,
prediction_modes=(PREDICTION_MODE_FOR_TESTS, ),
stats_info=None,
cur_perplexity=None,
output_seq_len=MAX_PREDICTIONS_LENGTH,
**kwargs):
"""
Generate responses for provided contexts and save the results on the disk. For a given context
several responses will be generated - one for each mode from the prediction_modes list.
:param predictions_path: Generated responses will be saved to this file
:param x_test: context token ids, numpy array of shape (batch_size, context_len, INPUT_SEQUENCE_LENGTH)
:param nn_model: instance of CakeChatModel class
:param prediction_modes: Iterable of modes to be used for responses generation. See PREDICTION_MODES
for available options
:param stats_info: Info about current training status: total time passed, time spent on training,
processed batches number and estimated time for one epoch
:param cur_perplexity: Addition to stats_info - current perplexity metric calculated on validation dataset
:param output_seq_len: Max number of tokens in generated responses
"""
_logger.info('Logging responses to test lines')
# Create all the directories for the prediction path in case they don't exist
prediction_dir = os.path.dirname(predictions_path)
if prediction_dir:
ensure_dir(prediction_dir)
with open(predictions_path, 'w') as test_res_fh:
csv_writer = init_csv_writer(test_res_fh, output_seq_len,
nn_model.model_name)
if cur_perplexity:
csv_writer.writerow(['Current perplexity: %.2f' % cur_perplexity])
if stats_info:
csv_writer.writerow([_get_iteration_stats(stats_info)])
contexts = transform_context_token_ids_to_sentences(
x_test, nn_model.index_to_token)
predictions_data = pd.DataFrame()
predictions_data['contexts'] = contexts
for pred_mode in prediction_modes:
responses_batch = get_nn_responses(x_test, nn_model, pred_mode,
**kwargs)
# list of lists of strings, shape (contexts_num, 1)
first_responses_batch = [response[0] for response in responses_batch]
# list of strings, shape (contexts_num)
predictions_data[pred_mode] = first_responses_batch
predictions_data.to_csv(predictions_path,
sep='\t',
index=False,
encoding='utf-8',
mode='a',
float_format='%.2f')
message = '\nSuccesfully dumped {} responses.'.format(len(contexts))
message += '\nHere they are:\n{}\n'.format(predictions_path)
_logger.info(message)
def _save_model(model, model_path):
_logger.info('Saving model to %s' % model_path)
ensure_dir(os.path.dirname(model_path))
model.save(model_path, separately=[])
_logger.info('Model has been saved')
