Higher-level API development

[KickItLikeShika]

I have experimented a lot of different ideas for this Higher-level API, but at the end i went with this one. It's obviously not done yet so i got some questions:
1- What do you think so far? is there anything i should do? anything i should replace? better implementation? better idea? better way to do this?
2- I got an error provided in the github gist below, i need help, or a better way to implement it

note:

• The reason why i chose this way to implement the trainer.fit() just because i found it so cool! like the one provided in sklearn, define the trainer, then fit it, then predict. And all of the this work fine btw.

File on Github Gist:
Trainer Implementation

[vfdev-5]

@KickItLikeShika thanks for exploring this question! I think discussing that we could find something interesting and helpful.
Maybe, a generic though about that: please, keep in mind that adopted approach have to be "generic" = user could handle single/multiple models/optimizers, data sources etc.

I think, it would be very helpful to take a look and analyze existing solutions referenced in the issue and have a clear idea of what is already done and the pros/cons of those approaches.

What do you think ?

[sdesrozis]

Thank you for this work. Having an exhaustive comparison of apis is a good idea. We must be aware to limit subjective opinion even if it is not completely possible. Provide sample codes would be very interesting.

Maybe we could first list relevant apis ?

[KickItLikeShika]

Sure, tomorrow a bunch of sample codes will be ready and posted here. Can i join slack btw?

[vfdev-5]

We are using pytorch slack which we are not admin, so if you are already on it you can find us on #pytorch-ignite channel. Otherwise, I'm using discord to communicate with folks (vfdev-5#4458). On the other hand, it is also nice to use github discussions such that things could help the community...

[KickItLikeShika]

@vfdev-5 @sdesrozis
I did as @sdesrozis asked. I have provided some sample codes and made APIs similar to MONAI, Transfer-NLP, ARGUS, trainers. I really tried to make it much smaller and made some function raise NotImplementedError just to worry more about the idea. I don't provide a lot features just for the simplicity for now, and sure used PyTorch-Ignite engine. I hope it's clear enough for you to understand the idea. I see you updated GSoC 2021 Development of a Higher-level API and it's more clear now and i hope you keep updating it and provide more details. And what do you think is gonna be more interesting? generic API like Argus/MONAI or task specific API like Transfer-NLP? Both seem really interesting. I wait for you reply to start building the API.

Implementation Similar To MONAI

Implementation Similar to Transfer-NLP

Implementation Similar to Argus

[sdesrozis]

@KickItLikeShika Thank you ! I take look at your samples as soon as possible.