Ampere incompatibility?

[avermilov]

Hello and thank you so much for the amazing project.

My problem is that the setup process described in the repo seems to not work for Ampere GPUs (in my case RTX 3080 Ti).

First I use the e4e/environment/e4e_env.yaml to create the Conda env. Then I follow the commands in the first cell of stylize.ipynb. However, I get ValueError: Unknown CUDA arch (8.6) or GPU not supported.

I think this may be because the default CUDA installed is 10.x, but my attempts to fix this by setting up the env differently have so far been unsuccessful. Would it be possible to add a fix for Ampere GPUs? Thanks in advance!

[mchong6]

The code is written purely in pytorch so this issue seems like incompatibility with pytorch/cuda? Instead of installing using the yaml file, I suggest just installing the pytorch/cuda version that works for you and installing the rest of the required packages.

[avermilov]

@mchong6 Thank you! Installing a working Pytorch + CUDA and then everything in stylize.ipynb worked perfectly!

Although while the code is working, I seem to not be able to train a 4-image sketch model from scratch because of an immediate CUDA out of memory. I saw you mention that 4-image models are very memory hungry, but do you perhaps know exactly how much memory is needed to train a multi-image model for 2, 4, 8, etc images?

[mchong6]

Im not sure how much memory is exactly needed but I could train with 4 images with 20gb memory. If you dont have enough gpu memory, instead of training with a batch size of 4 (1 for each image), you can have a dataloader such that each iteration loads a different image, maintaining batch size of 1.