COVID-19 library in torchvision

[AlexTS1980]

🚀 Feature

Library of models and dataset interfaces for COVID-19 models

Motivation

There are quite a few models (feature extractors, mask segmentation, classifiers) for COVID-19, both in pytorch and tensorflow. They use different datasets, making it hard to scientists to compare results and extend their findings. It would be good to (re-implement) at least some models and dataset interfaces as a library in torchvision

Pitch

Similar to the models and datasets in torchvision: models for the published models (see below), at least those that come with pretrained weights (e.g. COVIDNet-CT), and datasets for open-source labelled dataset interfaces: eg. CNCB-CT, UCSD, MedSeg, Zenodo, especially mask extraction.

Alternatives

None that I know of

Additional context

Some candidates include COVIDNet (x-rays), COVIDNet-CT (ct-scans), COVNet (ct-scans), JCS (ct-scans).

cc @pmeier

[vfdev-5]

@AlexTS1980 thanks for proposing this Feature Request. For instance, project maintainers/owners are working on defining the rules of what can be added as model or dataset. This detailed information will be exposed in CONTRIBUTING guide (https://github.com/pytorch/vision/pull/2663). cc @fmassa

As alternatives, I can think of MONAI: https://github.com/Project-MONAI/MONAI nevertheless handling medical imagery is also a direction that torchvision would like to support...

[AlexTS1980]

@AlexTS1980 thanks for proposing this Feature Request. For instance, project maintainers/owners are working on defining the rules of what can be added as model or dataset. This detailed information will be exposed in CONTRIBUTING guide (#2663). cc @fmassa

As alternatives, I can think of MONAI: https://github.com/Project-MONAI/MONAI nevertheless handling medical imagery is also a direction that torchvision would like to support...

What I mean is for example take this or this model and implement them as a part of torchvision library, so that they could be instantiated like VGG or ResNet, e.g. through covid_models package, with pretrained weights, etc. Same with the datasets.

[oke-aditya]

I guess an easy way could be to publish weights to torch hub. People can simply instantiate alexnet or vgg from torchvision and load the weights using torch.hub.load().

[dvolgyes]

Alexnet, VGG are proven architectures, and e.g. VGG is very frequently used in transfer learning. There are hundreds of new networks, but as long as they don't provide a clear reusable architecture, i don't see any point to incorporate it into a core library. Torch hub is perfect for this. As these covid networks trained on very specific dataset, often not even too diverse dataset, i don't see that they could be generic enough for an inclusion.

What would be the advantage of having these networks in torchvision, instead of torch.hub? The only point i see is publicity, but no offense, i haven't read anything in these models which would raised them clearly above average.