shithub: opus

Info  •  Files  •  Log  •  Branches

Implement FFTNet too

--- a/dnn/train_wavenet.py

+++ b/dnn/train_wavenet.py

@@ -16,9 +16,9 @@

 #set_session(tf.Session(config=config))

 nb_epochs = 40

-batch_size = 32

+batch_size = 64

-model = wavenet.new_wavenet_model()

+model = wavenet.new_wavenet_model(fftnet=True)

 model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])

 model.summary()

@@ -62,7 +62,7 @@

 # f.create_dataset('data', data=in_data[:50000, :, :])

 # f.create_dataset('feat', data=features[:50000, :, :])

-checkpoint = ModelCheckpoint('wavenet3a_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('wavenet3c_{epoch:02d}.h5')

 #model.load_weights('wavernn1c_01.h5')

 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=2e-4), loss='sparse_categorical_crossentropy', metrics=['sparse_categorical_accuracy'])

--- a/dnn/wavenet.py

+++ b/dnn/wavenet.py

@@ -10,13 +10,13 @@

 import sys

 from causalconv import CausalConv

-units=128

+units=256

 pcm_bits = 8

 pcm_levels = 2**pcm_bits

 nb_used_features = 38

-def new_wavenet_model():

+def new_wavenet_model(fftnet=False):

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

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

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

@@ -36,8 +36,9 @@

     for k in range(10):

         res = tmp

         tmp = Concatenate()([tmp, rfeat])

-        c1 = CausalConv(units, 2, dilation_rate=2**k, activation='tanh')

-        c2 = CausalConv(units, 2, dilation_rate=2**k, activation='sigmoid')

+        dilation = 9-k if fftnet else k

+        c1 = CausalConv(units, 2, dilation_rate=2**dilation, activation='tanh')

+        c2 = CausalConv(units, 2, dilation_rate=2**dilation, activation='sigmoid')

         tmp = Multiply()([c1(tmp), c2(tmp)])

         tmp = Dense(units, activation='relu')(tmp)

         if k != 0:

-- 

⑨