def chat(self, question, chat_settings):
"""Chat with the chatbot model by predicting an answer to a question.
'question' and 'answer' in this context are generic terms for the interactions in a dialog exchange
and can be statements, remarks, queries, requests, or any other type of dialog speech.
For example:
Question: "How are you?" Answer: "Fine."
Question: "That's great." Answer: "Yeah."
Args:
question: The input question for which the model should predict an answer.
chat_settings: The ChatSettings instance containing the chat settings and inference hyperparameters
Returns:
q_with_hist: question with history if chat_settings.show_question_context = True otherwise None.
answers: array of answer beams if chat_settings.show_all_beams = True otherwise the single selected answer.
"""
#Process the question by cleaning it and converting it to an integer encoded vector
if chat_settings.enable_auto_punctuation:
question = Vocabulary.auto_punctuate(question)
question = Vocabulary.clean_text(question, normalize_words = chat_settings.inference_hparams.normalize_words)
question = self.input_vocabulary.words2ints(question)
#Prepend the currently tracked steps of the conversation history separated by EOS tokens.
#This allows for deeper dialog context to influence the answer prediction.
question_with_history = []
for i in range(len(self.conversation_history)):
question_with_history += self.conversation_history[i] + [self.input_vocabulary.eos_int()]
question_with_history += question
#Get the answer prediction
batch = np.zeros((1, len(question_with_history)))
batch[0] = question_with_history
max_output_sequence_length = chat_settings.inference_hparams.max_answer_words + 1 # + 1 since the EOS token is counted as a timestep
predicted_answer_info = self.predict_batch(inputs = batch,
input_sequence_length = np.array([len(question_with_history)]),
max_output_sequence_length = max_output_sequence_length,
beam_length_penalty_weight = chat_settings.inference_hparams.beam_length_penalty_weight,
sampling_temperature = chat_settings.inference_hparams.sampling_temperature,
log_summary = chat_settings.inference_hparams.log_summary)
#Read the answer prediction
answer_beams = []
if self.beam_width > 0:
#For beam search decoding: if show_all_beams is enabeled then output all beams (sequences), otherwise take the first beam.
# The beams (in the "predictions" matrix) are ordered with the highest ranked beams first.
beam_count = 1 if not chat_settings.show_all_beams else len(predicted_answer_info["predictions_seq_lengths"][0])
for i in range(beam_count):
predicted_answer_seq_length = predicted_answer_info["predictions_seq_lengths"][0][i] - 1 #-1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][0][:predicted_answer_seq_length, i].tolist()
answer_beams.append(predicted_answer)
else:
#For greedy / sampling decoding: only one beam (sequence) is returned, based on the argmax for greedy decoding
# or the sampling distribution for sampling decoding. Return this beam.
beam_count = 1
predicted_answer_seq_length = predicted_answer_info["predictions_seq_lengths"][0] - 1 #-1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][0][:predicted_answer_seq_length].tolist()
answer_beams.append(predicted_answer)
#Add new conversation steps to the end of the history and trim from the beginning if it is longer than conv_history_length
#Answers need to be converted from output_vocabulary ints to input_vocabulary ints (since they will be fed back in to the encoder)
self.conversation_history.append(question)
answer_for_history = self.output_vocabulary.ints2words(answer_beams[0], is_punct_discrete_word = True, capitalize_i = False)
answer_for_history = self.input_vocabulary.words2ints(answer_for_history)
self.conversation_history.append(answer_for_history)
self.trim_conversation_history(chat_settings.inference_hparams.conv_history_length)
#Convert the answer(s) to text and return
answers = []
for i in range(beam_count):
answer = self.output_vocabulary.ints2words(answer_beams[i])
answers.append(answer)
q_with_hist = None if not chat_settings.show_question_context else self.input_vocabulary.ints2words(question_with_history)
if chat_settings.show_all_beams:
return q_with_hist, answers
else:
return q_with_hist, answers[0]
def chat(self, question, chat_settings):
"""
chat with the seq2seq model
:param question: input question in which the model should predict an answer
:param chat_settings: chat settings
:return: answer
"""
# Process the question by cleaning it and converting it to an integer encoded vector
if chat_settings.enable_auto_punctuation:
question = Vocabulary.auto_punctuate(question)
question = Vocabulary.clean_text(
question,
normalize_words=chat_settings.inference_hparams.normalize_words)
question = self.vocabulary.words2ints(question)
# Get the answer prediction
batch = np.expand_dims(question, 0)
max_output_sequence_length = chat_settings.inference_hparams.max_answer_words + 1
predicted_answer_info = self.predict_batch(
inputs=batch,
input_batch_lengths=1,
max_output_sequence_length=max_output_sequence_length,
beam_length_penalty_weight=chat_settings.inference_hparams.
beam_length_penalty_weight)
# Read the answer prediction
answer_beams = []
if self.beam_width > 0:
# For beam search decoding: if show_all_beams is enabeled then output all beams (sequences),
# otherwise take the first beam.
# The beams (in the "predictions" matrix) are ordered with the highest ranked beams first.
beam_count = 1 if not chat_settings.show_all_beams else len(
predicted_answer_info["predictions_seq_lengths"][0])
for i in range(beam_count):
predicted_answer_seq_length = predicted_answer_info[
"predictions_seq_lengths"][0][
i] - 1 # -1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][
0][:predicted_answer_seq_length, i].tolist()
answer_beams.append(predicted_answer)
else:
# For greedy / sampling decoding: only one beam (sequence) is returned,
# based on the argmax for greedy decoding
# or the sampling distribution for sampling decoding. Return this beam.
beam_count = 1
predicted_answer_seq_length = predicted_answer_info[
"predictions_seq_lengths"][0] - 1 # -1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][
0][:predicted_answer_seq_length].tolist()
answer_beams.append(predicted_answer)
# Convert the answer(s) to text and return
answers = []
for i in range(beam_count):
answer = self.vocabulary.ints2words(answer_beams[i])
answers.append(answer)
if chat_settings.show_all_beams:
return answers
else:
return answers[0]
def chat(self, question, chat_settings):
if chat_settings.enable_auto_punctuation:
question = Vocabulary.auto_punctuate(question)
question = Vocabulary.clean_text(
question,
normalize_words=chat_settings.inference_hparams.normalize_words)
question = self.input_vocabulary.words2ints(question)
question_with_history = []
for i in range(len(self.conversation_history)):
question_with_history += self.conversation_history[i] + [
self.input_vocabulary.eos_int()
]
question_with_history += question
#Get the answer prediction
batch = np.zeros((1, len(question_with_history)))
batch[0] = question_with_history
max_output_sequence_length = chat_settings.inference_hparams.max_answer_words + 1 # + 1 since the EOS token is counted as a timestep
predicted_answer_info = self.predict_batch(
inputs=batch,
input_sequence_length=np.array([len(question_with_history)]),
max_output_sequence_length=max_output_sequence_length,
beam_length_penalty_weight=chat_settings.inference_hparams.
beam_length_penalty_weight,
sampling_temperature=chat_settings.inference_hparams.
sampling_temperature,
log_summary=chat_settings.inference_hparams.log_summary)
#Read the answer prediction
answer_beams = []
if self.beam_width > 0:
#For beam search decoding: if show_all_beams is enabeled then output all beams (sequences), otherwise take the first beam.
# The beams (in the "predictions" matrix) are ordered with the highest ranked beams first.
beam_count = 1 if not chat_settings.show_all_beams else len(
predicted_answer_info["predictions_seq_lengths"][0])
for i in range(beam_count):
predicted_answer_seq_length = predicted_answer_info[
"predictions_seq_lengths"][0][
i] - 1 #-1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][
0][:predicted_answer_seq_length, i].tolist()
answer_beams.append(predicted_answer)
else:
#For greedy / sampling decoding: only one beam (sequence) is returned, based on the argmax for greedy decoding
# or the sampling distribution for sampling decoding. Return this beam.
beam_count = 1
predicted_answer_seq_length = predicted_answer_info[
"predictions_seq_lengths"][0] - 1 #-1 to exclude the EOS token
predicted_answer = predicted_answer_info["predictions"][
0][:predicted_answer_seq_length].tolist()
answer_beams.append(predicted_answer)
#Add new conversation steps to the end of the history and trim from the beginning if it is longer than conv_history_length
#Answers need to be converted from output_vocabulary ints to input_vocabulary ints (since they will be fed back in to the encoder)
self.conversation_history.append(question)
answer_for_history = self.output_vocabulary.ints2words(
answer_beams[0], is_punct_discrete_word=True, capitalize_i=False)
answer_for_history = self.input_vocabulary.words2ints(
answer_for_history)
self.conversation_history.append(answer_for_history)
self.trim_conversation_history(
chat_settings.inference_hparams.conv_history_length)
#Convert the answer(s) to text and return
answers = []
for i in range(beam_count):
answer = self.output_vocabulary.ints2words(answer_beams[i])
answers.append(answer)
q_with_hist = None if not chat_settings.show_question_context else self.input_vocabulary.ints2words(
question_with_history)
if chat_settings.show_all_beams:
return q_with_hist, answers
else:
return q_with_hist, answers[0]
