shithub: opus

Info  •  Files  •  Log  •  Branches

Remove useless (and possibly hurtful) residual connection

I guess it's a bad idea to forward inputs directly

--- a/dnn/lpcnet.c

+++ b/dnn/lpcnet.c

@@ -81,7 +81,6 @@

     compute_conv1d(&feature_conv2, conv2_out, net->feature_conv2_state, conv1_out);

     celt_assert(FRAME_INPUT_SIZE == FEATURE_CONV2_OUT_SIZE);

     if (lpcnet->frame_count < FEATURES_DELAY) RNN_CLEAR(conv2_out, FEATURE_CONV2_OUT_SIZE);

-    for (i=0;i<FEATURE_CONV2_OUT_SIZE;i++) conv2_out[i] += lpcnet->old_input[FEATURES_DELAY-1][i];

     memmove(lpcnet->old_input[1], lpcnet->old_input[0], (FEATURES_DELAY-1)*FRAME_INPUT_SIZE*sizeof(in[0]));

     memcpy(lpcnet->old_input[0], in, FRAME_INPUT_SIZE*sizeof(in[0]));

     compute_dense(&feature_dense1, dense1_out, conv2_out);

--- a/dnn/lpcnet.py

+++ b/dnn/lpcnet.py

@@ -122,7 +122,7 @@

     dec_state2 = Input(shape=(rnn_units2,))

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

-    fconv2 = Conv1D(102, 3, padding='same', activation='tanh', name='feature_conv2')

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

     embed = Embedding(256, embed_size, embeddings_initializer=PCMInit(), name='embed_sig')

     cpcm = Reshape((-1, embed_size*2))(embed(pcm))

@@ -137,7 +137,6 @@

     fdense1 = Dense(128, activation='tanh', name='feature_dense1')

     fdense2 = Dense(128, activation='tanh', name='feature_dense2')

-    cfeat = Add()([cfeat, cat_feat])

     cfeat = fdense2(fdense1(cfeat))

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

--- a/dnn/test_lpcnet.py

+++ b/dnn/test_lpcnet.py

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

-model.load_weights('lpcnet9_384_10_G16_120.h5')

+model.load_weights('lpcnet20c_384_10_G16_80.h5')

 order = 16

--- a/dnn/train_lpcnet.py

+++ b/dnn/train_lpcnet.py

@@ -97,7 +97,7 @@

 del pred

 # dump models to disk as we go

-checkpoint = ModelCheckpoint('lpcnet20_384_10_G16_{epoch:02d}.h5')

+checkpoint = ModelCheckpoint('lpcnet20c_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