FuseMedML 2

[mosheraboh]

Hi all!
We are planning significant refactoring and upgrades in the coming month.
We hold all the changes in a branch called "fuse2" - https://github.com/IBM/fuse-med-ml/tree/fuse2
"fuse2" will be merged into master when finalized.
Those modifications will break the backward compatibility. However, porting old code to fuse2 style should be easy.
To avoid conflicts, we won't merge pull requests to master until fuse2 branch will be finalized. New pull request should be merged directly to fuse2 branch instead.

Here are the most significant refactoring and upgrades:

1. The end-to-end examples *fuse_examples" moved to a separate package (still maintained in this repository though).
   To install FuseMedML core without the examples - pip install -e .
   To install FuseMedML core and the end-to-end examples -

   pip install -e .
   pip install -e examples
   

2. FuseMedML structure changed and is now built from 4 packages (fuse.utils, fuse.eval, fuse.data, fuse.dl).
   Components that weren't under fuse.utils, fuse.eval and fuse.data, moved to fuse.dl.
   For example: fuse.losses moved to fuse.dl.losses
3. "Fuse" prefix removed from components name. For example, FuseLossDefault changed to be LossDefault.
4. New data package (replacing the old one)! See details here: https://github.com/IBM/fuse-med-ml/blob/fuse2/fuse/data/README.md
5. FuseUtilsHierarchicalDict is now deprecated (but still supported for now).
   Instead we are using NDict, which allows to read values from nested dict directly.
   For example:
   batch_dict["data.input.image"]
   instead of FuseUtilsHierarchicalDict.get(batch_dict, "data.input.image")
   Currently in PR - FuseUtilsHierarchicalDict -> NDict #66

To finalize fuse2, we need to:

1. Port our examples to fuse2 style - currently just mnist example ported to fuse2 style (see here the modifications for reference - https://github.com/IBM/fuse-med-ml/pull/62/files)
2. Rewrite the main fuse-med-ml README file to better reflect fuse2 style.

You are welcome to send updates and ask questions in this discussion thread.

[mosheraboh]

Hi,
Another modification was pushed to fuse2 branch.
@YoelShoshan and I simplified OpBase interface to make it much easier to follow.
If you've implemented Ops and still didn't push it to fuse2 branch, simply remove the op_id argument from op.call method to make it work again.
CC: @taltlusty @itaijj @SagiPolaczek @egozi @alex-golts @liamhazan

[mosheraboh]

We will start to maintain a list of already implemented ops:
General ops should be added here: https://github.com/IBM/fuse-med-ml/tree/fuse2/fuse/data#operators-list
Imaging ops should be added here: https://github.com/IBM/fuse-med-ml/tree/fuse2/fuseimg#imaging-operators

@liamhazan , @SagiPolaczek , @itaijj , @michalozeryflato , if your PRs include new ops, please also add them to the lists above.