shithub: opus

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -11,12 +11,9 @@

 import h5py

 import sys

-rnn_units1=384

-rnn_units2=16

 pcm_bits = 8

 embed_size = 128

 pcm_levels = 2**pcm_bits

-nb_used_features = 38

 class Sparsify(Callback):

     def __init__(self, t_start, t_end, interval, density):

@@ -88,7 +85,7 @@

             'seed': self.seed

-def new_wavernn_model():

+def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 38):

     pcm = Input(shape=(None, 2))

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

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

@@ -127,6 +124,10 @@

     ulaw_prob = md(gru_out2)

     model = Model([pcm, exc, feat, pitch], ulaw_prob)

+    model.rnn_units1 = rnn_units1

+    model.rnn_units2 = rnn_units2

+    model.nb_used_features = nb_used_features

     encoder = Model([feat, pitch], cfeat)

     dec_rnn_in = Concatenate()([cpcm, cexc, dec_feat])

--- a/dnn/test_lpcnet.py

+++ b/dnn/test_lpcnet.py

@@ -15,12 +15,8 @@

 config.gpu_options.per_process_gpu_memory_fraction = 0.2

 set_session(tf.Session(config=config))

-nb_epochs = 40

-batch_size = 64

+model, enc, dec = lpcnet.new_lpcnet_model()

-#model = wavenet.new_wavenet_model(fftnet=True)

-model, enc, dec = lpcnet.new_wavernn_model()

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

 #model.summary()

@@ -27,7 +23,7 @@

 feature_file = sys.argv[1]

 frame_size = 160

 nb_features = 55

-nb_used_features = lpcnet.nb_used_features

+nb_used_features = model.nb_used_features

 features = np.fromfile(feature_file, dtype='float32')

 features = np.resize(features, (-1, nb_features))

@@ -48,8 +44,8 @@

 pcm = np.zeros((nb_frames*pcm_chunk_size, ))

 fexc = np.zeros((1, 1, 2), dtype='float32')

 iexc = np.zeros((1, 1, 1), dtype='int16')

-state1 = np.zeros((1, lpcnet.rnn_units1), dtype='float32')

-state2 = np.zeros((1, lpcnet.rnn_units2), dtype='float32')

+state1 = np.zeros((1, model.rnn_units1), dtype='float32')

+state2 = np.zeros((1, model.rnn_units2), dtype='float32')

 mem = 0

 coef = 0.85

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -28,8 +28,7 @@

 # Try reducing batch_size if you run out of memory on your GPU

 batch_size = 64

-# Note we are creating a LPCNet model

-model, _, _ = lpcnet.new_wavernn_model()

+model, _, _ = lpcnet.new_lpcnet_model()

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

 model.summary()

@@ -40,7 +39,7 @@

 pcm_file = sys.argv[4]     # 16 bit unsigned short PCM samples

 frame_size = 160

 nb_features = 55

-nb_used_features = lpcnet.nb_used_features

+nb_used_features = model.nb_used_features

 feature_chunk_size = 15

 pcm_chunk_size = frame_size*feature_chunk_size

--

⑨