shithub: opus

Info  •  Files  •  Log  •  Branches

using features (except pitch gain which has NaNs for now)

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -12,12 +12,19 @@

 rnn_units=64

 pcm_bits = 8

 pcm_levels = 2**pcm_bits

+nb_used_features = 37

+

 def new_wavernn_model():

     pcm = Input(shape=(None, 1))

+    feat = Input(shape=(None, nb_used_features))

+

+    rep = Lambda(lambda x: K.repeat_elements(x, 160, 1))

+

     rnn = CuDNNGRU(rnn_units, return_sequences=True)

+    rnn_in = Concatenate()([pcm, rep(feat)])

     md = MDense(pcm_levels, activation='softmax')

-    ulaw_prob = md(rnn(pcm))

+    ulaw_prob = md(rnn(rnn_in))

-    model = Model(pcm, ulaw_prob)

+    model = Model([pcm, feat], ulaw_prob)

     return model

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -24,7 +24,7 @@

 pcmfile = sys.argv[1]

 feature_file = sys.argv[2]

 nb_features = 54

-nb_used_features = 38

+nb_used_features = lpcnet.nb_used_features

 feature_chunk_size = 15

 pcm_chunk_size = 160*feature_chunk_size

@@ -44,8 +44,8 @@

 features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features))

 features = features[:, :, :nb_used_features]

-checkpoint = ModelCheckpoint('lpcnet1b_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet1c_{epoch:02d}.h5')

 #model.load_weights('wavernn1c_01.h5')

 model.compile(optimizer=Adam(0.002, amsgrad=True, decay=2e-4), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])

-model.fit(in_data, out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])

+model.fit([in_data, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])

-- 

⑨