Feature proposal: TorchIO hub - A system to store and fetch transform objects for reproducibility about torchio HOT 6 OPEN

Hi, @tomvars. Thanks for the proposal. I think this is an excellent idea.

One potential solution would be leveraging the PyTorch Hub tools. I got this code working. What do you think?

import torchio as tio
fpg = tio.datasets.FPG()
fpg.plot(reorient=False)

import torch
repo = 'fepegar/resseg:add-preprocessing-hubconf'
function_name = 'get_preprocessing_transform'
input_path = fpg.t1.path
preprocess = torch.hub.load(repo, function_name, input_path, image_name='t1', force_reload=True)
preprocessed = preprocess(fpg)
preprocessed.plot(reorient=False)

from torchio.

I really like this API! You could maybe create a new repo like fepegar/torchiohub:main and have a single hubconf.py file as the access point to different preprocessing functions. In the repo users could append their transform functions to a large transforms.py file and the hubconf.py would have lines such as from transforms import ronneberger_unet_2015_transform

from torchio.

You mean something like this?

@classmethod
def from_hub(cls, *args, **kwargs):
    return torch.hub.load(*args, **kwargs)

from torchio.

I think it's more convenient to allow users to use their own hubconf in their repos because

1. This is what PyTorch does, so people are familiar with the syntax etc.
2. Sometimes, getting a transform needs some special code. The snippet I shared is an example in which additional libraries or files might be needed just to compute the transform, and we wouldn't want to put everyone's code in the same repo.

So the contribution to this library (which I'm happy to write) would be documentation on how to set up transforms for reproducibility on top of PyTorch Hub. Does that sound good?

from torchio.

That makes sense 👍 thoughts on introducing a class method for the Transform called from_hub which would wrap the torch.hub.load call and pass in the relevant arguments?

from torchio.

Hey, I forgot to share some experiments I conducted. The code below needs unet to be PIP-installed:

import torch
import torchio as tio

colin = tio.datasets.Colin27()
path = colin.t1.path
torch.hub.load('fepegar/resseg:add-preprocessing-hubconf', 'get_preprocessing_transform', path)
transform = torch.hub.load('fepegar/resseg:add-preprocessing-hubconf', 'get_preprocessing_transform', path, image_name='t1')
transform(colin).plot()

Here, HistogramStandardization makes it a bit awkward, but things work. We should write a tutorial about this. If you think the class method would be helpful, feel free to contribute with a PR!

from torchio.