shithub: opus

Info  •  Files  •  Log  •  Branches

Use real features at the chunk edges rather than zeros

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -113,7 +113,7 @@

             'seed': self.seed

         }

-def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 38, use_gpu=True):

+def new_lpcnet_model(rnn_units1=384, rnn_units2=16, nb_used_features = 38, training=False, use_gpu=True):

     pcm = Input(shape=(None, 3))

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

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

@@ -121,8 +121,9 @@

     dec_state1 = Input(shape=(rnn_units1,))

     dec_state2 = Input(shape=(rnn_units2,))

-    fconv1 = Conv1D(128, 3, padding='same', activation='tanh', name='feature_conv1')

-    fconv2 = Conv1D(128, 3, padding='same', activation='tanh', name='feature_conv2')

+    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))

--- a/dnn/test_lpcnet.py

+++ b/dnn/test_lpcnet.py

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

 config.gpu_options.per_process_gpu_memory_fraction = 0.2

 set_session(tf.Session(config=config))

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

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

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

 #model.summary()

@@ -63,7 +63,7 @@

-model.load_weights('lpcnet20g_384_10_G16_02.h5')

+model.load_weights('lpcnet20h_384_10_G16_80.h5')

 order = 16

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -51,7 +51,7 @@

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

 batch_size = 64

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

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

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

 model.summary()

@@ -89,6 +89,11 @@

 features = features[:, :, :nb_used_features]

 features[:,:,18:36] = 0

+fpad1 = np.concatenate([features[0:1, 0:2, :], features[:-1, -2:, :]], axis=0)

+fpad2 = np.concatenate([features[1:, :2, :], features[0:1, -2:, :]], axis=0)

+features = np.concatenate([fpad1, features, fpad2], axis=1)

+

+

 periods = (.1 + 50*features[:,:,36:37]+100).astype('int16')

 in_data = np.concatenate([sig, pred, in_exc], axis=-1)

@@ -98,7 +103,7 @@

 del in_exc

 # dump models to disk as we go

-checkpoint = ModelCheckpoint('lpcnet20g_384_10_G16_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet20h_384_10_G16_{epoch:02d}.h5')

 #model.load_weights('lpcnet9b_384_10_G16_01.h5')

 model.compile(optimizer=Adam(0.001, amsgrad=True, decay=5e-5), loss='sparse_categorical_crossentropy')

-- 

⑨