Fidder for older PyTorch versions

[cyanching]

Hey Alister!

Since the highest CUDA version for my workstation GPU is 10.1, the highest stable PyTorch version I could use is 1.8.1. I would appreciate it a lot if you could release an additional pip install fidder version that can run with older PyTorch. E.g., you can maybe change torch.cuda.amp.autocast() into torch.autocast()? Though this is the only error I am aware of at the moment.

Thank you so much for your work!

Cheers,

Cyan

[alisterburt]

Hey @cyanching - thanks for reaching out! I'm sorry you're running into issues

I don't think I call autocast directly in my code, I'm depending on PyTorch-lightning which abstracts over torch. I just had a look at their support policy and it seems like they support the latest four pytorch versions.

Could you try installing an older version of pytorch-lightning and see if things work? If you run into compatibility issues with the weights do let me know and I'll look into training with an older version to make things possible for you 🙂

[cyanching]

Hey Alister!

Thank you for your prompt reply! I just did a compatibility check, the highest stable PyTorch-lightning version I could install is 1.6.5, which supports PyTorch 1.8.1 and its corresponding torchvision 0.9.1. Whereas fidder requires the get_dimensions attribute of the torchvision.transforms.functional module which is only available from torchvision 0.13 onwards (PyTorch has to be 1.12.0 and above). It seems there is no way around getting newer versions of PyTorch and torchvision to run fidder on my GPU, which is not possible. :(

If it won't be too much trouble for you, I would be very grateful if you could look into training with an older version of PyTorch-lightning! :)

Thank you so much for your help!

Wish you a great day!

Cheers,

Cyan

[alisterburt]

@cyanching I removed the calls to TF.get_dimensions in #19 - having some trouble with deployment but you can install from source and see if that works for you with the older pytorch lightning?

[alisterburt]

https://github.com/teamtomo/fidder/actions/runs/4242336177 - deployment of v0.0.6 with #19 should now work

[cyanching]

Hey Alister!

Thank you so much for your speedy update! I just did some tests, fidder worked like a charm on my CUDA 10.1 GPU (Quadro P6000)! :)

You are a star! I wish I could double, triple star fidder!

Have an awesome day! :)

Cheers,

Cyan

[alisterburt]

Awesome, glad we could get you up and running so easily! I'll add a note to the README for others who may have the same issue, thanks for reporting and have a great day yourself!