chrisdonahue/wavegan

Tensorflow2.4 support

luisarandas opened this issue · 5 comments

Hello.

Is this repository working with TF>2? I'm having a hard time with virtual environments and these older versions on linux.

Thanks

Unfortunately the repo has not yet been updated to work with TF>2. It might be fairly straightforward to patch it using tf.compat.v1, though I don't have much experience with TF2 :(

I think I managed a workaround on Linux with tf-gpu 2.2.0.0rc1 if you want I can fork and upload (with tf.compat.v1, which is not super nice but it is training right now). Tho I can't access the tensorboard with the newest cuda drivers.

ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

Had problems with the device search and added thecommon workaround

sess = tf.compat.v1.Session(config=tf.ConfigProto( allow_soft_placement=True, log_device_placement=True))

Will see how this goes. How much time/iterations do you recommend? Kind of blind here lol
Thank you

Hi @luisarandas,

Did you manage to get this working? It was working fine for me with TF 1.13 middle of last year, but now I'm revisiting it and it no longer works properly. The model is clearly in the GPU memory and it claims to be training, but for some reason it's progressing at CPU-like speeds. The same behaviour is happening locally and in Colab. I don't have such problems with other repos that use TF v2 hence I'm wondering if the tf.compat.v1 workaround works with this codebase. Obviously I will try it but these things take time and it would be great to hear if you have something working already!

Thanks,

Mark

I have created an updated fork of this repository migrated to run on Tensorflow 2. You may have more success getting it to run in modern environments, especially if you are aiming to train on newer GPU hardware. https://github.com/mattjwarren/wavegan

Thanks for sorting this out, appreciated. It was getting annoying to have to restart Colab every time it gave me an A100