Provide pre-trained model option for tutorials

[marksgraham]

We have increasing numbers of DDPM tutorials that all start with the same model training code before we get into the 'meat' of the tutorial - e.g comparing different schedulers, doing inpainting, using likelihoods for anomaly detection. It means users probably will need an hour to run a tutorial and it also slows down our tutorial development.

I'm thinking it could be helpful to train a good model and give users the option to grab the pre-trained model rather than train their own. What do you guys think? We could store the weights on google drive and download them the way @Warvito dis for the RadImagenet models, here.

[Warvito]

I think it is a good suggestion! It will make the tutorials more concise and possibly return better values compared to the current ones.

As you mentioned, we could use a repository with a hubconf.py file (from torch hub) as a placeholder while the models are not available at the model zoo (as discussed in #30). It is not too hard to prepare it. It would be great to also make available the training code in case the users want to know how the model was created. Should we make something similar to this example to train a 2D DDPM model on the MedNISTDataset (at least for a start) on our servers and make it available?

I guess this way, we could also try to use more complex datasets (like the ones discussed on #93) in some tutorials. The ones from MONAI do not have too much extra information that could be useful in tutorials with conditioned models.

For the model zoo, I was planning to retrain the 3D LDM (and make the training code available) or convert the weights for the current classes. @mjorgecardoso recommended having at least two big models available in the modelzoo for the release. Depending on the model, we could also use them in some future tutorials.

[marksgraham]

Ok, I can make the temporary repository, looks straightforward!

Shall I start by making a version of the distributed script for just the hand class of MedNIST? This will allow us to replace the training steps for all the current 2d ddpm tutorials ('d propose leaving the training code in these tutorials to let the users see it, and then allowing the users set a flag to choose if they want to train themselves or just grab the pretrained model).

I could also extend the script to train on the full MedNIST if we want a class-conditioned DDPM

[Warvito]

Shall I start by making a version of the distributed script for just the hand class of MedNIST?

Yes, I think it is a good start. If it is not too complicated to make it work on the servers, I guess the distributed version would be cool. If it is too much work, following the format of a single .py file like the ddp example or this one would be good. In the end, I guess the MedNIST will train very fast on one of our new GPUs.

[marksgraham]

We agreed in the last meeting that we should implement the pretrained models in the MONAI Bundle format. We should be able to just pull the weights files from these bundles when they are needed in the tutorials.

However, doing that right now would mean committing public training code to the model zoo that requires this private repo to run. So I was thinking we should go with the solution in my PR for now (torch hub) and look to transition to the Bundle format once the generative code is public. What do you think @Warvito ?

[Warvito]

I agree, I think it is the best solution @marksgraham