Add input embedding

dnn/train_wavenet_audio.py

@@ -31,7 +31,7 @@ feature_chunk_size = 15
 pcm_chunk_size = frame_size*feature_chunk_size
 
 data = np.fromfile(pcmfile, dtype='int16')
-data = np.minimum(127, lin2ulaw(data[160:]/32768.))
+data = np.minimum(127, lin2ulaw(data[80:]/32768.))
 nb_frames = len(data)//pcm_chunk_size
 
 features = np.fromfile(feature_file, dtype='float32')
@@ -39,7 +39,7 @@ features = np.fromfile(feature_file, dtype='float32')
 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]])/16.;
+in_data = np.concatenate([data[0:1], data[:-1]]);
 
 features = np.reshape(features, (nb_frames*feature_chunk_size, nb_features))
 pitch = 1.*data
@@ -51,6 +51,7 @@ for i in range(2, nb_frames*feature_chunk_size):
 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(data, (nb_frames, pcm_chunk_size, 1))
 out_data = (out_data.astype('int16')+128).astype('uint8')
 features = np.reshape(features, (nb_frames, feature_chunk_size, nb_features))

dnn/wavenet.py

@@ -4,6 +4,7 @@ import math
 from keras.models import Model
 from keras.layers import Input, LSTM, CuDNNGRU, Dense, Embedding, Reshape, Concatenate, Lambda, Conv1D, Add, Multiply, Bidirectional, MaxPooling1D, Activation
 from keras import backend as K
+from keras.initializers import Initializer
 from keras.initializers import VarianceScaling
 from mdense import MDense
 import numpy as np
@@ -17,6 +18,30 @@ pcm_bits = 8
 pcm_levels = 2**pcm_bits
 nb_used_features = 38
 
+class PCMInit(Initializer):
+    def __init__(self, gain=.1, seed=None):
+        self.gain = gain
+        self.seed = seed
+
+    def __call__(self, shape, dtype=None):
+        num_rows = 1
+        for dim in shape[:-1]:
+            num_rows *= dim
+        num_cols = shape[-1]
+        flat_shape = (num_rows, num_cols)
+        if self.seed is not None:
+            np.random.seed(self.seed)
+        a = np.random.uniform(-1.7321, 1.7321, flat_shape)
+        #a[:,0] = math.sqrt(12)*np.arange(-.5*num_rows+.5,.5*num_rows-.4)/num_rows
+        #a[:,1] = .5*a[:,0]*a[:,0]*a[:,0]
+        a = a + np.reshape(math.sqrt(12)*np.arange(-.5*num_rows+.5,.5*num_rows-.4)/num_rows, (num_rows, 1))
+        return self.gain * a
+
+    def get_config(self):
+        return {
+            'gain': self.gain,
+            'seed': self.seed
+        }
 
 def new_wavenet_model(fftnet=False):
     pcm = Input(shape=(None, 1))
@@ -34,7 +59,8 @@ def new_wavenet_model(fftnet=False):
     activation='tanh'
     rfeat = rep(cfeat)
     #tmp = Concatenate()([pcm, rfeat])
-    tmp = pcm
+    embed = Embedding(256, units, embeddings_initializer=PCMInit())
+    tmp = Reshape((-1, units))(embed(pcm))
     init = VarianceScaling(scale=1.5,mode='fan_avg',distribution='uniform')
     for k in range(10):
         res = tmp