golbin/wavenet

Yet another WaveNet implementation in PyTorch.

wavenet's Issues

sigmoid and tanh on the same tensor?

Hi, thanks for the codes!

Although Figure 4 in the original paper is described as in your code, Eq. 2 (which I assume to be more correct) says there are two different convolutions, i.e., I think it should be

(ResidualBlock)

    def forward(self, x, skip_size)
        input_tanh = self.dilated(x)
        input_sigmoid = self.dilated(x)

        # pixelCNN
        gated_tanh = self.gated_tanh(input_tanh)  # [-1, 1]
        gated_sigmoid = self.gate_sigmoid(input_sigmoid)  # [0, 1]
        gated = gated_tanh * gated_sigmoid  # [0, 1]

instead of

    def forward(self, x, skip_size)
        output = self.dilated(x)

        # pixelCNN
        gated_tanh = self.gated_tanh(output)  # [-1, 1]
        gated_sigmoid = self.gate_sigmoid(output)  # [0, 1]
        gated = gated_tanh * gated_sigmoid  # [0, 1]

But I'm also guessing, have you possibly looked into it before?

What is the purpose of the transpose?

Thank you for your code! It is documented very well!

What is the purpose of swapping the channels and timestep in the input tensor of WaveNet.forward()?

WaveNet/wavenet/networks.py

Line 243 in 0554533

output = x.transpose(1, 2)

The program is "killed: 9" after running for about 3 minutes

The information about my computer is as below:
MacBook Pro (15-inch, 2017)
Processor: 2.9 GHz Intel Core i7
GPU: 16 GB 2133 MHz LPDDR3

I check the list of GPUs that supports CUDA and my GPU does appear on the list...
I also ran the test.py in this repository and there was no problem.
However, when I ran train.py, the program was killed after running for 3 minutes. I wonder if this is because my GPU does not support CUDA...

About conditioning on spectrogram

How to use this implementation for conditioned synthesis on spectrogram?

Non-causal dilated convolution?

Hello,

I think your dilated convolution class, as implemented, is not causal because of the 0 padding.

The last comment here proposes a solution that might work better:
pytorch/pytorch#1333

The program is "killed: 9" after running for about 3 minutes

The information about my computer is as below:
MacBook Pro (15-inch, 2017)
Processor: 2.9 GHz Intel Core i7
GPU: 16 GB 2133 MHz LPDDR3

I check the list of GPUs that supports CUDA and my GPU does appear on the list...
I also ran the test.py in this repository and there was no problem.
However, when I ran train.py, the program was killed after running for 3 minutes. I wonder if this is because my GPU does not support CUDA...

out of memory

We tried to run train.py, but there were errors as below:

python3 train.py --data_dir="./test/data" --output_dir="./outputs"
THCudaCheck FAIL file=/pytorch/torch/lib/THC/generic/THCStorage.cu line=66 error=2 : out of memory
Traceback (most recent call last):
File "train.py", line 59, in
trainer = Trainer(args)
File "train.py", line 17, in init
lr=args.lr)
...
...
return super(_CudaBase, cls).new(cls, *args, **kwargs)
RuntimeError: cuda runtime error (2) : out of memory at /pytorch/torch/lib/THC/generic/THCStorage.cu:66

I don't know why.

RuntimeError: cuda runtime error (2) : out of memory

Hi! I'm running the following command to train the model:
$ python train.py --data_dir=./test/data --output_dir=./outputs

The GPU I'm using has 16276MiB. However, I get an out of memory error immediately:

/wavenet/networks.py", line 88, in forward
gated = gated_tanh * gated_sigmoid
RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1524584710464/work/aten/src/THC/generic/THCStorage.cu:58

Any thoughts as to why this might be happening? Based on my calculations, the input size is 1x100,000x256 which should easily fit in the 16276 MiB of memory that the GPU has.

pytorch version

what does that mean of saying 'PyTorch >= 0.3.0'?
i cannot see the 0.3.0 version.

i can not understand this two line code

There may be a few bugs in this code.

wavenet/utils/data.py
there should be a time shift between the input audio and the target audio , But not so in the code

When generating: RuntimeError: CUDA out of memory

Hi, I am running on a GeForce RTX3070 with 8GB.
I can train the model if I adjust the sample_size, which I interpret as batch_size.

However when generating, this parameter obviously does not help.
duration = generator.generate()
4885/14662 samples are generated. Traceback (most recent call last): File "<stdin>", line 1, in <module> File "<stdin>", line 39, in generate wavenet\model.py", line 71, in generate outputs = self.net(inputs) RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 8.00 GiB total capacity; 6.95 GiB already allocated; 0 bytes free; 7.31 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
Any Tips on how to avoid this? Thanks!