ref: b6af21f31c9309e1ee031bb4c0cf51b213cf7f3c
parent: 211435f5d3ad2afad1aca3dc5ef8d6dfee564ff0
author: Jean-Marc Valin <jmvalin@jmvalin.ca>
date: Mon Jul 23 13:05:21 EDT 2018
wip...
--- a/dnn/gatedconv.py
+++ b/dnn/gatedconv.py
@@ -1,6 +1,6 @@
from keras import backend as K
from keras.engine.topology import Layer
-from keras.layers import activations, initializers, regularizers, constraints, InputSpec, Conv1D
+from keras.layers import activations, initializers, regularizers, constraints, InputSpec, Conv1D, Dense
import numpy as np
class GatedConv(Conv1D):
@@ -42,7 +42,7 @@
self.out_dims = filters
self.nongate_activation = activations.get(activation)
- def call(self, inputs, memory=None):
+ def call(self, inputs, cond=None, memory=None):
if memory is None:
mem = K.zeros((K.shape(inputs)[0], self.mem_size, K.shape(inputs)[-1]))
else:
@@ -49,6 +49,9 @@
mem = K.variable(K.cast_to_floatx(memory))
inputs = K.concatenate([mem, inputs], axis=1)
ret = super(GatedConv, self).call(inputs)
+ if cond is not None:
+ d = Dense(2*self.out_dims, use_bias=False, activation='linear')
+ ret = ret + d(cond)
ret = self.nongate_activation(ret[:, :, :self.out_dims]) * activations.sigmoid(ret[:, :, self.out_dims:])
if self.return_memory:
ret = ret, inputs[:, :self.mem_size, :]
--- a/dnn/wavenet.py
+++ b/dnn/wavenet.py
@@ -4,6 +4,7 @@
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 VarianceScaling
from mdense import MDense
import numpy as np
import h5py
@@ -34,12 +35,20 @@
rfeat = rep(cfeat)
#tmp = Concatenate()([pcm, rfeat])
tmp = pcm
+ init = VarianceScaling(scale=1.5,mode='fan_avg',distribution='uniform')
for k in range(10):
res = tmp
- tmp = Concatenate()([tmp, rfeat])
dilation = 9-k if fftnet else k
- c = GatedConv(units, 2, dilation_rate=2**dilation, activation='tanh')
- tmp = Dense(units, activation='relu')(c(tmp))
+ '''#tmp = Concatenate()([tmp, rfeat])
+ c = GatedConv(units, 2, dilation_rate=2**dilation, activation='tanh', kernel_initializer=init)
+ tmp = Dense(units, activation='relu')(c(tmp, cond=rfeat))'''
+
+ tmp = Concatenate()([tmp, rfeat])
+ c1 = CausalConv(units, 2, dilation_rate=2**dilation, activation='tanh')
+ c2 = CausalConv(units, 2, dilation_rate=2**dilation, activation='sigmoid')
+ tmp = Multiply()([c1(tmp), c2(tmp)])
+ tmp = Dense(units, activation='relu')(tmp)
+
if k != 0:
tmp = Add()([tmp, res])
--
⑨