def encode(texts, USE_QA_model, ConvRT_model, contexts=None):
texts = [utils.simple_preprocess(text) for text in texts]
contexts = [utils.simple_preprocess(text) for text in contexts]
ConvRT_encodings = ConvRT_model.encode_responses(texts)
USE_embeddings = USE_QA_model.signatures['response_encoder'](
input=tf.constant(texts), context=tf.constant(contexts))["outputs"]
return np.concatenate(
[np.asarray(USE_embeddings),
np.asarray(ConvRT_encodings)], axis=-1)
def encode(texts, USE_QA_model, ConvRT_model, contexts=None):
texts = [utils.simple_preprocess(text) for text in texts]
if contexts is not None:
new_contexts = []
for context in contexts:
new_context = [utils.simple_preprocess(text) for text in context]
new_contexts.append(new_context)
contexts = new_contexts
context_encodings = ConvRT_model.encode_contexts(texts,
extra_contexts=contexts)
question_embeddings = USE_QA_model.signatures['question_encoder'](
tf.constant(texts))["outputs"]
return np.concatenate(
[np.asarray(question_embeddings),
np.asarray(context_encodings)],
axis=-1)
def para_encode(text):
global ParaNMT_embd
text, tokenized_text = utils.simple_preprocess(text)
paraembds = []
for token in tokenized_text:
if token in ParaNMT_embd:
paraembds.append(ParaNMT_embd[token])
for word in ParaNMT_embd:
paradim = ParaNMT_embd[word].shape[0]
break
paravec = np.zeros((paradim), np.float32)
if paraembds:
paravec = np.sum(np.asarray(paraembds), axis=0) / math.sqrt(
len(paraembds))
return paravec
model_reverse,
utterance,
query_encoding,
candidates,
response_context,
conversation_history,
bias=biases)
print("\n")
if len(str(response).split(" ")) <= 100:
if flags.voice:
entry = utils.simple_preprocess(str(response).lower(),
for_speech=True,
return_tokenized=True)
entry = " ".join(entry)
wavefiles = text2speech.process(entry)
def f1():
utils.delay_print("Bot: " + response)
def f2():
text2speech.play(wavefiles)
p1 = Process(target=f1)
p2 = Process(target=f2)
p1.start()
p2.start()