usernaamee / keras-wavenet Goto Github PK

the line in No.95

seed_audio[-1] = ampl_val_16
seed_audio[:-1] = seed_audio[1:]

should be

seed_audio[:-1] = seed_audio[1:]
seed_audio[-1] = ampl_val_16

The order can't be reverse

In frame_generator, shouldn't temp = audio[i + frame_size + 1] be temp = audio[i + frame_size]?
Otherwise you skip a sample.

The code assumes that validate.wav and train.wav are in the same folder, however, they are not shared as part of the repo. It would be great if a sample dataset or steps acquire one can be added.

I’m getting an assertion error for AtrousConvolution2D setting the atrous rate. Are you getting this?

Traceback (most recent call last):
File "simple-generative-model-regressor.py", line 83, in
model = get_basic_generative_model(frame_size)
File "simple-generative-model-regressor.py", line 30, in get_basic_generative_model
l1a, l1b = wavenetBlock(10, 5, 2, 1, 3)(input)
File "simple-generative-model-regressor.py", line 16, in f
border_mode='same')(input)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 515, in call
self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 573, in add_inbound_node
Node.create_node(self, inbound_layers, node_indices, tensor_indices)
File "/usr/local/lib/python2.7/dist-packages/keras/engine/topology.py", line 150, in create_node
output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0]))
File "/usr/local/lib/python2.7/dist-packages/keras/layers/convolutional.py", line 663, in call
filter_dilation=self.atrous_rate)
File "/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.py", line 1520, in conv2d

I noticed that you are using 'AtrousConvolution1D' in the WaveNet block, but is this function also implement causal convolution? Thx.

Hi,

I tried to run this today with Tensorflow as the backend.

Using TF, the atrous rate must be a symmetric tuple (ie, (2, 2)):
https://github.com/usernaamee/keras-wavenet/blob/master/simple-generative-model-regressor.py#L15

Hi, the causal conv descripted in deepmind paper is fully convolution, and it is left-right unbalance. Then they can get high parallel. But your implementation is to predict the 4097th sample by i:i+4096, this seems infeasible for speech data (16k per second).

Because I use keras model.summary and I noticed, that number of parameters stays the same if I change dilation_rate.