austinmoehle / wavernn Goto Github PK

@austinmoehle Just as your explanation in the notebook, cudnn version gru is very fast than the tf implementation. Do you know if it's possible to use cudnn version gru when inference?

In your pb model, it seems that you use while loop rather than cudnn version gru.

Nice work, gave me a lot of inspiration, thanks!
I encountered a problem，as described in the title. The following is the details:

InvalidArgumentError (see above for traceback): Cannot colocate nodes 'model/optimizer/global_norm/L2Loss_62' and 'model/optimizer/gradients/model/inference/CudnnRNN_1_grad/CudnnRNNBackprop' because no device type supports both of those nodes and the other nodes colocated with them.
Colocation Debug Info:
Colocation group had the following types and devices:
CudnnRNNBackprop: GPU
Identity:
L2Loss: CPU

what`s that means, and how to solve.

Your work is very impressive on me. But it seems that the traning code is not provided in the repo. Could you release the training code or send me a copy? Thanks in advance.

AttributeError Traceback (most recent call last)
in
11
12 # Outputs from explicit-TF GRU.
---> 13 outputs_explicit = run_cudnn_gru_explicit(inp, INPUT_CHANNELS, RECURRENT_SIZE)
14
15 init = tf.global_variables_initializer()

in run_cudnn_gru_explicit(inputs, input_channels, recurrent_size)
25 saveable = cudnn_rnn_ops.CudnnGRUSaveable(
26 kernel, 1, recurrent_size, input_channels)
---> 27 weights, biases = saveable._OpaqueParamsToCanonical()
28
29 # Build the individual weights and biases.

AttributeError: 'CudnnGRUSaveable' object has no attribute '_OpaqueParamsToCanonical'

I used tensorboard to inspect your model structure and found that the pb model you provided just uses one softmax with 256 outputs (8 bits).

However, the paper uses two separated DNNs to predict the coarse and fine part of a sample. Is that because your model reuse the matrix of O1 and O3 (O2 and O4) or you just support 8 bits with mu-law compression?