diff --git a/dnn/lpcnet.py b/dnn/lpcnet.py index 7d40c1a5..e92c48b0 100644 --- a/dnn/lpcnet.py +++ b/dnn/lpcnet.py @@ -12,6 +12,7 @@ import sys rnn_units=512 pcm_bits = 8 +embed_size = 128 pcm_levels = 2**pcm_bits nb_used_features = 38 @@ -42,6 +43,7 @@ class PCMInit(Initializer): def new_wavernn_model(): pcm = Input(shape=(None, 2)) + exc = Input(shape=(None, 1)) pitch = Input(shape=(None, 1)) feat = Input(shape=(None, nb_used_features)) dec_feat = Input(shape=(None, 32)) @@ -60,26 +62,27 @@ def new_wavernn_model(): cpcm = pcm cpitch = pitch - embed = Embedding(256, 128, embeddings_initializer=PCMInit()) - cpcm = Reshape((-1, 128*2))(embed(pcm)) - + embed = Embedding(256, embed_size, embeddings_initializer=PCMInit()) + cpcm = Reshape((-1, embed_size*2))(embed(pcm)) + embed2 = Embedding(256, embed_size, embeddings_initializer=PCMInit()) + cexc = Reshape((-1, embed_size))(embed2(exc)) cfeat = fconv2(fconv1(feat)) rep = Lambda(lambda x: K.repeat_elements(x, 160, 1)) rnn = CuDNNGRU(rnn_units, return_sequences=True, return_state=True) - rnn_in = Concatenate()([cpcm, rep(cfeat)]) + rnn_in = Concatenate()([cpcm, cexc, rep(cfeat)]) md = MDense(pcm_levels, activation='softmax') gru_out, state = rnn(rnn_in) ulaw_prob = md(gru_out) - model = Model([pcm, feat], ulaw_prob) + model = Model([pcm, exc, feat], ulaw_prob) encoder = Model(feat, cfeat) - dec_rnn_in = Concatenate()([cpcm, dec_feat]) + dec_rnn_in = Concatenate()([cpcm, cexc, dec_feat]) dec_gru_out, state = rnn(dec_rnn_in, initial_state=dec_state) dec_ulaw_prob = md(dec_gru_out) - decoder = Model([pcm, dec_feat, dec_state], [dec_ulaw_prob, state]) + decoder = Model([pcm, exc, dec_feat, dec_state], [dec_ulaw_prob, state]) return model, encoder, decoder diff --git a/dnn/test_wavenet_audio.py b/dnn/test_wavenet_audio.py index 828a4e5d..7ad41232 100755 --- a/dnn/test_wavenet_audio.py +++ b/dnn/test_wavenet_audio.py @@ -66,7 +66,7 @@ in_data = np.reshape(in_data, (nb_frames*pcm_chunk_size, 1)) out_data = np.reshape(data, (nb_frames*pcm_chunk_size, 1)) -model.load_weights('wavenet4a3_30.h5') +model.load_weights('wavenet4b_30.h5') order = 16 @@ -85,19 +85,19 @@ for c in range(1, nb_frames): period = int(50*features[c, fr, 36]+100) period = period - 4 for i in range(frame_size): - fexc[0, 0, 0] = iexc + 128 + #fexc[0, 0, 0] = iexc + 128 pred = -sum(a*pcm[f*frame_size + i - 1:f*frame_size + i - order-1:-1, 0]) fexc[0, 0, 1] = np.minimum(127, lin2ulaw(pred/32768.)) + 128 - p, state = dec.predict([fexc, cfeat[:, fr:fr+1, :], state]) + p, state = dec.predict([fexc, iexc, cfeat[:, fr:fr+1, :], state]) #p = p*p #p = p/(1e-18 + np.sum(p)) p = np.maximum(p-0.001, 0) p = p/(1e-5 + np.sum(p)) - iexc[0, 0, 0] = np.argmax(np.random.multinomial(1, p[0,0,:], 1))-128 - pcm[f*frame_size + i, 0] = pred + 32768*ulaw2lin(iexc[0, 0, 0]*1.0) - iexc[0, 0, 0] = lin2ulaw(pcm[f*frame_size + i, 0]/32768) + iexc[0, 0, 0] = np.argmax(np.random.multinomial(1, p[0,0,:], 1)) + pcm[f*frame_size + i, 0] = pred + 32768*ulaw2lin(iexc[0, 0, 0]-128) + fexc[0, 0, 0] = lin2ulaw(pcm[f*frame_size + i, 0]/32768) + 128 print(iexc[0, 0, 0], 32768*ulaw2lin(out_data[f*frame_size + i, 0]), pcm[f*frame_size + i, 0], pred) diff --git a/dnn/train_wavenet_audio.py b/dnn/train_wavenet_audio.py index cf39a778..7139ae2c 100755 --- a/dnn/train_wavenet_audio.py +++ b/dnn/train_wavenet_audio.py @@ -46,7 +46,7 @@ udata = udata[:nb_frames*pcm_chunk_size] features = features[:nb_frames*feature_chunk_size*nb_features] in_data = np.concatenate([data[0:1], data[:-1]]); -noise = np.concatenate([np.zeros((len(data)*2//5)), np.random.randint(-2, 2, len(data)//5), np.random.randint(-1, 1, len(data)*2//5)]) +noise = np.concatenate([np.zeros((len(data)*2//5)), np.random.randint(-1, 1, len(data)*3//5)]) in_data = in_data + noise in_data = np.maximum(-127, np.minimum(127, in_data)) @@ -78,9 +78,18 @@ in_pitch = np.reshape(pitch/16., (nb_frames, pcm_chunk_size, 1)) in_data = np.reshape(in_data, (nb_frames, pcm_chunk_size, 1)) in_data = (in_data.astype('int16')+128).astype('uint8') -out_data = np.reshape(lin2ulaw((udata-upred)/32768), (nb_frames, pcm_chunk_size, 1)) +out_data = lin2ulaw((udata-upred)/32768) +in_exc = np.concatenate([out_data[0:1], out_data[:-1]]); + +out_data = np.reshape(out_data, (nb_frames, pcm_chunk_size, 1)) out_data = np.maximum(-127, np.minimum(127, out_data)) out_data = (out_data.astype('int16')+128).astype('uint8') + +in_exc = np.reshape(in_exc, (nb_frames, pcm_chunk_size, 1)) +in_exc = np.maximum(-127, np.minimum(127, in_exc)) +in_exc = (in_exc.astype('int16')+128).astype('uint8') + + features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features)) features = features[:, :, :nb_used_features] pred = np.reshape(pred, (nb_frames, pcm_chunk_size, 1)) @@ -94,8 +103,8 @@ in_data = np.concatenate([in_data, pred], axis=-1) # f.create_dataset('data', data=in_data[:50000, :, :]) # f.create_dataset('feat', data=features[:50000, :, :]) -checkpoint = ModelCheckpoint('wavenet4a3_{epoch:02d}.h5') +checkpoint = ModelCheckpoint('wavenet4b_{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']) -model.fit([in_data, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint]) +model.fit([in_data, in_exc, features], out_data, batch_size=batch_size, epochs=30, validation_split=0.2, callbacks=[checkpoint])