Spleeter: [Discussion] Proper instructions for spleeter-gpu

[aidv]

So the past two days I've been trying many attempts to properly use Spleeter both for inference and training.

Both failing to some degree.

When I do inference, I have to install using the old 202003 label in my Anaconda environment, and it works fine for inference, but during training it crashes with

TypeError: x and y must have the same dtype, got tf.string != tf.int32

That's on a Win10 machine with CUDA 10.1. CuDNN version is unknown to me, but I'd assume that spleeter separate ... would've complained if if the correct CuDNN version would've been installed.

So I'm just going to assume that the right version

I then tried to install Spleeter from the GitHub repo on Ubuntu 18.04 and 20.04 eventually after many, MANY, trial and errors finding out that I needed to manually install CUDA 10.0, but why? I thought that CUDA 10.1 was supposed to be installed?!

Anyways.

When it comes to CuDNN, I've tried
the following:

7.6.5
7.6.4
7.6.3
7.6.2
7.6.1
7.6.0
7.5.1
7.5.0
7.4.2
7.4.1

only to either have the CPU utilized or for it all to fail and end up with the error

Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.

I have tried both the Anaconda env and pip. Docker failed completely.

I have also tried tensorflow-gpu 2.2, 1.15.3, 1.15.2, 1.15.0 and 1.14

What is going on here?

Actually let me ask the following:

What is YOUR setup and configuration to reliably run Spleeter both for inference and training?

I'm eager to get this to run reliably as I have a multi GPU ML rig and I really look forward to start training models as I have assembled a huge library of audio files and I have been successful

[romi1502]

Hi @aidv,
I've just checked and both separation and training work fine on GPU with the spleeter:3.7-gpu docker image using nvidia-docker (I think unfortunately not available on Windows) : this is the setup we use on Debian.
The image uses CUDA 10.1.243 and CuDNN 7.6.3.30 and run the latest spleeter version (1.5.3). I tested on a RTX 2080 with driver v430.14.
I don't have access to a Windows GPU machine so I can't reproduce your configuration, but the CuDNN issue you have seems to be more a CUDA configuration issue than a spleeter one.

[furfury]

Which Debian version?

[aidv]

@romi1502 aah I see. Interesting.

What version of Debian are you using?

RTX 2080 graphics driver 430.14
Debian 9.9
CUDA 10.1.243
CuDNN 7.6.3.30
Spleeter 1.5.3

I could try this exact setup today.

[romi1502]

Debian 9.9

[aidv]

@romi1502 You mention that you use Docker and CUDA 10.1.243, but docker states ENV CUDA_VERSION 10.0.130.

Why? What am I missing?

[Thanhcaro]

I need to use GPU for edit faster too. Hope that will have solution.

[theblissprogrammer]

I have a gist to set up the latest spleeter-gpu with Ubuntu 18.04 LTS from scratch. Check it out https://gist.github.com/theblissprogrammer/a2e529e9e2ef54261844ff12af99f5cf