From a922f83cca2b67d1a947e6673c304f99ba9b0230 Mon Sep 17 00:00:00 2001 From: Jean-Marc Valin Date: Tue, 21 Aug 2018 13:02:26 -0400 Subject: [PATCH] Fix input noise --- dnn/test_wavenet_audio.py | 6 ++++-- dnn/train_wavenet_audio.py | 10 ++++++---- 2 files changed, 10 insertions(+), 6 deletions(-) diff --git a/dnn/test_wavenet_audio.py b/dnn/test_wavenet_audio.py index 420a48bc..782f8977 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('wavenet3h13_30.h5') +model.load_weights('wavenet3h21_30.h5') order = 16 @@ -90,11 +90,13 @@ for c in range(1, nb_frames): fexc[0, 0, 1] = np.minimum(127, lin2ulaw(pred/32768.)) + 128 p, state = dec.predict([fexc, 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] = 32768*ulaw2lin(iexc[0, 0, 0]*1.0) - print(iexc[0, 0, 0], out_data[f*frame_size + i, 0], pcm[f*frame_size + i, 0]) + 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 df8c39cb..4659afab 100755 --- a/dnn/train_wavenet_audio.py +++ b/dnn/train_wavenet_audio.py @@ -45,21 +45,23 @@ data = data[:nb_frames*pcm_chunk_size] features = features[:nb_frames*feature_chunk_size*nb_features] in_data = np.concatenate([data[0:1], data[:-1]]); -in_data = in_data + np.random.randint(-1, 1, len(data)) +noise = np.concatenate([np.zeros((len(data)//3)), np.random.randint(-2, 2, len(data)//3), np.random.randint(-1, 1, len(data)//3)]) +in_data = in_data + noise +in_data = np.maximum(-127, np.minimum(127, in_data)) features = np.reshape(features, (nb_frames*feature_chunk_size, nb_features)) pred = np.fromfile(pred_file, dtype='int16') pred = pred[:nb_frames*pcm_chunk_size] -pred_in = 32768.*ulaw2lin(data) +pred_in = 32768.*ulaw2lin(in_data) for i in range(2, nb_frames*feature_chunk_size): pred[i*frame_size:(i+1)*frame_size] = 0 if i % 100000 == 0: print(i) for k in range(16): pred[i*frame_size:(i+1)*frame_size] = pred[i*frame_size:(i+1)*frame_size] - \ - pred_in[i*frame_size-k-1:(i+1)*frame_size-k-1]*features[i, nb_features-16+k] + pred_in[i*frame_size-k:(i+1)*frame_size-k]*features[i, nb_features-16+k] pred = np.minimum(127, lin2ulaw(pred/32768.)) #pred = pred + np.random.randint(-1, 1, len(data)) @@ -90,7 +92,7 @@ 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('wavenet3h13_{epoch:02d}.h5') +checkpoint = ModelCheckpoint('wavenet3h21_{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'])