Installing FastAI for NVidia RTX 3060

Got a new PC for Christmas, with a NVidia RTX-3060. Seems this is supported by CUDA 11 and PyTorch, but this isn’t working with FastAI. Any idea how to get this working? I’m on Ubuntu 20.

It does work for me. What error message did you get?

Okay, so I’ve installed torch using the following command:
ip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

And pip list gives me:
torch 1.10.0+cu113
torchaudio 0.10.0+cu113
torchvision 0.11.1+cu113

When I try to open a jupiter-notebook file used previously with FastAI it worked, but now it comes up with:

/usr/local/lib/python3.8/dist-packages/fastai/vision/models/xresnet.py in
11 # Cell
12 from …torch_basics import *
—> 13 from torchvision.models.utils import load_state_dict_from_url
14
15 # Cell

ModuleNotFoundError: No module named ‘torchvision.models.utils’

So I can load/import torch, but not fastai.

Apologies for the delay getting back.

The solution to this problem was to install the latest FastAI 2.5 by downloading the source from git and running the install as per the FastAI documentation. Using pip to install fast AI would cause torch to be moved back to 1.7, which would break things.

Also note that this broke opencv, and I needed to uninstall opencv-python-headless to get things working again.

pip uninstall opencv-python-headless

I suspect that installing FastAI via conda might also work, but I was running pip.

Also, I note that using the git repo causes a bunch of Microsoft Azure packages to be brought down. Not really sure what they are for, and appear unrelated to the distribution itself.

1 Like