def val_step(input_ids, target_ids_, step, write_output_seq):
dec_padding_mask = create_padding_mask(input_ids)
(draft_predictions, _, refine_predictions,
_) = Model.predict(input_ids, dec_padding_mask, False)
if config.use_refine_decoder:
predictions = refine_predictions
else:
predictions = draft_predictions
rouge, bert = tf_write_output_sequence(target_ids_[:, 1:], predictions[:,
1:],
step, write_output_seq)
return (rouge, bert)
def translate():
en_input = input('Enter the english sentence-> ')
en_input = preprocess(en_input)
input_CLS_ID = source_tokenizer.vocab_size
input_SEP_ID = source_tokenizer.vocab_size + 1
target_CLS_ID = target_tokenizer.vocab_size
target_SEP_ID = target_tokenizer.vocab_size + 1
input_ids = tf.convert_to_tensor(
[[input_CLS_ID] + source_tokenizer.encode(en_input) + [input_SEP_ID]])
dec_padding_mask = create_padding_mask(input_ids)
start = time.time()
preds_draft_summary, _, _, _ = Model.predict(input_ids, dec_padding_mask)
translated_sequence = target_tokenizer.decode([
i for i in tf.squeeze(preds_draft_summary)
if i not in [target_CLS_ID, target_SEP_ID, config.PAD_ID]
])
print(
f'Translated output --> {translated_sequence if translated_sequence else "EMPTY"}'
)
print(f'Time taken --> {round(time.time()-start)} seconds')
def generate():
en_input = input('Enter the sentence-> ')
en_input = preprocess(en_input)
input_ids = tf.constant(source_tokenizer.encode(en_input))[None, :]
dec_padding_mask = create_padding_mask(input_ids)
start = time.time()
(preds_draft_summary, _, preds_refine_summary, _,
_) = Model.predict(input_ids,
batch_size=1,
draft_decoder_type='topktopp',
beam_size=10,
length_penalty=0.6,
temperature=1,
top_p=0.9,
top_k=25)
generated_sequence = target_tokenizer.decode(
tf.squeeze(preds_refine_summary), skip_special_tokens=True)
print(
f'Translated output--> {generated_sequence if generated_sequence else "EMPTY"}'
)
print(f'Time to process --> {round(time.time()-start)} seconds')