def preprocess_data(en_tokenizer, fr_tokenizer, en_text, fr_text, en_timesteps,
fr_timesteps):
""" Preprocessing data and getting a sequence of word indices """
en_seq = sents2sequences(en_tokenizer,
en_text,
reverse=False,
padding_type='pre',
pad_length=en_timesteps)
fr_seq = sents2sequences(fr_tokenizer, fr_text, pad_length=fr_timesteps)
logger.info('Vocabulary size (English): {}'.format(np.max(en_seq) + 1))
logger.info('Vocabulary size (French): {}'.format(np.max(fr_seq) + 1))
logger.debug('En text shape: {}'.format(en_seq.shape))
logger.debug('Fr text shape: {}'.format(fr_seq.shape))
return en_seq, fr_seq
def infer_nmt(encoder_model, decoder_model, test_en_seq, en_vsize, fr_vsize):
"""
Infer logic
:param encoder_model: keras.Model
:param decoder_model: keras.Model
:param test_en_seq: sequence of word ids
:param en_vsize: int
:param fr_vsize: int
:return:
"""
test_fr_seq = sents2sequences(fr_tokenizer, ['sos'], fr_vsize)
test_en_onehot_seq = to_categorical(test_en_seq, num_classes=en_vsize)
test_fr_onehot_seq = np.expand_dims(
to_categorical(test_fr_seq, num_classes=fr_vsize), 1)
enc_outs, enc_fwd_state, enc_back_state = encoder_model.predict(
test_en_onehot_seq)
dec_state = np.concatenate([enc_fwd_state, enc_back_state], axis=-1)
attention_weights = []
fr_text = ''
for i in range(fr_timesteps):
dec_out, attention, dec_state = decoder_model.predict(
[enc_outs, dec_state, test_fr_onehot_seq])
dec_ind = np.argmax(dec_out, axis=-1)[0, 0]
if dec_ind == 0:
break
test_fr_seq = sents2sequences(fr_tokenizer, [fr_index2word[dec_ind]],
fr_vsize)
test_fr_onehot_seq = np.expand_dims(
to_categorical(test_fr_seq, num_classes=fr_vsize), 1)
attention_weights.append((dec_ind, attention))
fr_text += fr_index2word[dec_ind] + ' '
return fr_text, attention_weights
fr_vsize=fr_vsize)
n_epochs = 10 if not debug else 3
train(full_model, en_seq, fr_seq, batch_size, n_epochs)
""" Save model """
if not os.path.exists(os.path.join('..', 'h5.models')):
os.mkdir(os.path.join('..', 'h5.models'))
full_model.save(os.path.join('..', 'h5.models', 'nmt.h5'))
""" Index2word """
en_index2word = dict(
zip(en_tokenizer.word_index.values(), en_tokenizer.word_index.keys()))
fr_index2word = dict(
zip(fr_tokenizer.word_index.values(), fr_tokenizer.word_index.keys()))
""" Inferring with trained model """
test_en = ts_en_text[0]
logger.info('Translating: {}'.format(test_en))
test_en_seq = sents2sequences(en_tokenizer, [test_en],
pad_length=en_timesteps)
test_fr, attn_weights = infer_nmt(encoder_model=infer_enc_model,
decoder_model=infer_dec_model,
test_en_seq=test_en_seq,
en_vsize=en_vsize,
fr_vsize=fr_vsize)
logger.info('\tFrench: {}'.format(test_fr))
""" Attention plotting """
plot_attention_weights(test_en_seq,
attn_weights,
en_index2word,
fr_index2word,
base_dir=base_dir)