def tranferModel(nmt_base_model,
en_len,
fr_len,
MAX_VOCAB,
ENC_EMB_DIM,
DEC_EMB_DIM,
HID_DIM,
EMB_TRAIN=True,
GRU_LAYER=True,
is_trainable=False):
fr_vocab = en_vocab = MAX_VOCAB
en_inputs = Input(shape=(en_len, ), name='en_inputs')
tr_en_emb_layer = Embedding(en_vocab + 1,
ENC_EMB_DIM,
input_length=en_len,
trainable=EMB_TRAIN,
name='en_emb')
tr_en_emb = tr_en_emb_layer(en_inputs)
tr_en_emb_masked = Masking(name='en_emb_mask')(tr_en_emb) ###
if GRU_LAYER:
tr_en_gru = GRU(HID_DIM, return_state=True, dropout=0.5, name='en_gru')
en_out, en_state = tr_en_gru(tr_en_emb_masked)
else:
tr_en_gru = LSTM(HID_DIM,
return_state=True,
dropout=0.5,
name='en_lstm')
en_out, en_state, en_carry = tr_en_gru(tr_en_emb_masked)
if not is_trainable:
tr_en_emb_layer.trainable = False
tr_en_gru.trainable = False
de_inputs = Input(shape=(fr_len - 1, ), name='de_inputs')
tr_de_emb_layer = Embedding(fr_vocab + 1,
DEC_EMB_DIM,
input_length=fr_len - 1,
trainable=EMB_TRAIN,
name='de_emb')
tr_de_emb = tr_de_emb_layer(de_inputs)
tr_de_emb_masked = Masking(name='de_emb_mask')(tr_de_emb) ###
if GRU_LAYER:
tr_de_gru = GRU(HID_DIM,
return_sequences=True,
return_state=True,
dropout=0.5,
name='de_gru')
de_out, _ = tr_de_gru(tr_de_emb_masked, initial_state=en_state)
else:
tr_de_gru = LSTM(HID_DIM,
return_sequences=True,
return_state=True,
dropout=0.5,
name='de_lstm')
de_out, _, _ = tr_de_gru(tr_de_emb_masked,
initial_state=[en_state, en_carry])
tr_de_dense_0 = TimeDistributed(Dense(HID_DIM,
activation='relu',
name='de_dense_0'),
name='de_timedense_0') # fr_vocab
tr_de_dense_0_output = tr_de_dense_0(de_out) ###
tr_de_dense_1 = Dense(fr_vocab, activation='softmax',
name='de_dense_1') # trainable
de_pred = TimeDistributed(tr_de_dense_1,
name='de_timedense_1')(tr_de_dense_0_output)
if not is_trainable:
tr_de_emb_layer.trainable = False
tr_de_gru.trainable = False
tr_de_dense_0.trainable = False
# Define the new transfer learning model which accepts encoder/decoder inputs and outputs predictions
nmt_emb = Model(inputs=[en_inputs, de_inputs],
outputs=de_pred,
name='transfer_learning_model')
return nmt_emb, en_inputs, tr_en_emb_layer, tr_en_gru, en_state, \
de_inputs, tr_de_emb_layer, tr_de_gru, tr_de_dense_0, tr_de_dense_1
def new_lpcnet_model(rnn_units1=384,
rnn_units2=16,
nb_used_features=38,
training=False,
adaptation=False):
pcm = Input(shape=(None, 3))
feat = Input(shape=(None, nb_used_features))
pitch = Input(shape=(None, 1))
dec_feat = Input(shape=(None, 128))
dec_state1 = Input(shape=(rnn_units1, ))
dec_state2 = Input(shape=(rnn_units2, ))
padding = 'valid' if training else 'same'
fconv1 = Conv1D(128,
3,
padding=padding,
activation='tanh',
name='feature_conv1')
fconv2 = Conv1D(128,
3,
padding=padding,
activation='tanh',
name='feature_conv2')
embed = Embedding(256,
embed_size,
embeddings_initializer=PCMInit(),
name='embed_sig')
cpcm = Reshape((-1, embed_size * 3))(embed(pcm))
pembed = Embedding(256, 64, name='embed_pitch')
cat_feat = Concatenate()([feat, Reshape((-1, 64))(pembed(pitch))])
cfeat = fconv2(fconv1(cat_feat))
fdense1 = Dense(128, activation='tanh', name='feature_dense1')
fdense2 = Dense(128, activation='tanh', name='feature_dense2')
cfeat = fdense2(fdense1(cfeat))
rep = Lambda(lambda x: K.repeat_elements(x, frame_size, 1))
rnn = GRU(rnn_units1,
return_sequences=True,
return_state=True,
recurrent_activation="sigmoid",
reset_after='true',
name='gru_a')
rnn2 = GRU(rnn_units2,
return_sequences=True,
return_state=True,
recurrent_activation="sigmoid",
reset_after='true',
name='gru_b')
rnn_in = Concatenate()([cpcm, rep(cfeat)])
md = MDense(pcm_levels, activation='softmax', name='dual_fc')
gru_out1, _ = rnn(rnn_in)
gru_out2, _ = rnn2(Concatenate()([gru_out1, rep(cfeat)]))
ulaw_prob = md(gru_out2)
if adaptation:
rnn.trainable = False
rnn2.trainable = False
md.trainable = False
embed.Trainable = False
model = Model([pcm, feat, pitch], ulaw_prob)
model.rnn_units1 = rnn_units1
model.rnn_units2 = rnn_units2
model.nb_used_features = nb_used_features
model.frame_size = frame_size
encoder = Model([feat, pitch], cfeat)
dec_rnn_in = Concatenate()([cpcm, dec_feat])
dec_gru_out1, state1 = rnn(dec_rnn_in, initial_state=dec_state1)
dec_gru_out2, state2 = rnn2(Concatenate()([dec_gru_out1, dec_feat]),
initial_state=dec_state2)
dec_ulaw_prob = md(dec_gru_out2)
decoder = Model([pcm, dec_feat, dec_state1, dec_state2],
[dec_ulaw_prob, state1, state2])
return model, encoder, decoder
