def warmup_predictor(nn_model, prediction_mode):
if prediction_mode in {
PREDICTION_MODES.beamsearch_reranking,
PREDICTION_MODES.sampling_reranking
}:
prediction_config = PredictionConfig(prediction_mode)
predictor_factory(nn_model, prediction_mode,
prediction_config.get_options_dict())
def get_nn_response_ids(context_token_ids,
nn_model,
mode,
output_candidates_num=1,
output_seq_len=MAX_PREDICTIONS_LENGTH,
condition_ids=None,
**kwargs):
"""
Predicts several responses for every context.
:param context_token_ids: np.array; shape (batch_size, context_size, context_len); dtype=int
Represents all tokens ids to use for predicting
:param nn_model: CakeChatModel
:param mode: one of PREDICTION_MODES mode
:param output_candidates_num: Number of candidates to generate.
When mode is either 'beamsearch', 'beamsearch-reranking' or 'sampling-reranking', the candidates with the
highest score are returned. When mode is 'sampling', the candidates_num of samples are generated independently.
:param condition_ids: List with ids of conditions responding for each context.
:param output_seq_len: Number of tokens to generate.
:param kwargs: Other prediction parameters, passed into predictor constructor.
Might be different depending on mode. See PredictionConfig for the details.
:return: np.array; shape (batch_size, output_candidates_num, output_seq_len); dtype=int
Generated predictions.
"""
if mode == PREDICTION_MODES.sampling:
kwargs['samples_num'] = output_candidates_num
prediction_config = PredictionConfig(mode, **kwargs)
_logger.debug('Generating predicted response for the following params: %s' % prediction_config)
predictor = predictor_factory(nn_model, mode, prediction_config.get_options_dict())
responses = predictor.predict_responses(context_token_ids, output_seq_len, condition_ids, output_candidates_num)
return responses
def get_nn_response_ids(context_token_ids,
nn_model,
mode,
output_candidates_num=1,
output_seq_len=MAX_PREDICTIONS_LENGTH,
condition_ids=None,
**kwargs):
"""
Predicts several responses for every context.
:param context_token_ids: np.array; shape=(batch_size x context_size x context_len); dtype=int
Represents all tokens ids to use for predicting
:param nn_model: CakeChatModel
:param mode: one of PREDICTION_MODES mode
:param output_candidates_num: Number of candidates to generate.
When mode is either 'beamsearch', 'beamsearch-reranking' or 'sampling-reranking', the candidates with the
highest score are returned. When mode is 'sampling', the candidates_num of samples are generated independently.
:param condition_ids: List with ids of conditions responding for each context.
:param output_seq_len: Number of tokens to generate.
:param kwargs: Other prediction parameters, passed into predictor constructor.
Might be different depending on mode. See PredictionConfig for the details.
:return: np.array; shape=(responses_num x output_candidates_num x output_seq_len); dtype=int
Generated predictions.
"""
if mode == PREDICTION_MODES.sampling:
kwargs['samples_num'] = output_candidates_num
prediction_config = PredictionConfig(mode, **kwargs)
_logger.debug('Generating predicted response for the following params: %s' % prediction_config)
predictor = predictor_factory(nn_model, mode, prediction_config.get_options_dict())
return np.array(
predictor.predict_responses(context_token_ids, output_seq_len, condition_ids, output_candidates_num))
def warmup_predictor(nn_model, prediction_mode):
if prediction_mode in {PREDICTION_MODES.beamsearch_reranking, PREDICTION_MODES.sampling_reranking}:
prediction_config = PredictionConfig(prediction_mode)
predictor_factory(nn_model, prediction_mode, prediction_config.get_options_dict())