search

pytorch / vision

Datasets, Transforms and Models specific to Computer Vision
https://pytorch.org/vision
BSD 3-Clause "New" or "Revised" License
16.18k stars 6.95k forks source link

Documentation and examples for prototype torchvision #4992

Open oke-aditya opened 2 years ago

oke-aditya commented 2 years ago

📚 The doc issue

First of all, hats off to all the maintainers @fmassa @datumbox @NicolasHug @pmeier It's been a great revamp, and I think all the new features being added are great! If software engineering is an art then torchvision is Picaso!

Any thoughts or ideas on how to effectively use the new prototype torchvision? The new models, datasets and transforms. As an end user perspective

It would great to read documentation on why these were created over legacy ones and what are the benefits.

Also examples pertaining to same would be nice, maybe before next release ?

Also porting the current examples to the new APIs?

Suggest a potential alternative/fix

P.S. I would be glad to contribute :smile: It will take some time for me to dive into the codebase and understand though. But I'm up for it.

cc @bjuncek

datumbox commented 2 years ago

@oke-aditya Thanks for bringing this up!

Concerning the multi-weight support and the models prototype, I plan to write a blogpost once the API is cleaned up. You can see at #4652 some of the issues we still want to discuss before introducing it properly to the community. Any feedback related to the open issues or the API would be awesome.

You can make contributions to any part of the code-base. Nothing is off limits; standard open-source practices apply.

Concerning updating documentation, tutorials, references etc, this is likely something that is going to happen once the API is more mature and closer to the release. See #4679 for pendings on the model prototype.

pmeier commented 2 years ago

It would great to read documentation on why these were created over legacy ones and what are the benefits.

The lack on documentation is somewhat intentional (at least from my side). We are currently moving quite fast and there are a lot of moving parts. Having a documentation will only slow us down, because it will become outdated often. From my side the most "stable" submodule is datasets. There is one more major issue I need to tackle before I can start to promote and document it. That should happen in the next two weeks.

Although this doesn't replace a good documentation, we will have some presence at the PyTorch developer day 2021. @fmassa will give a talk and I will be present at a poster doing live demos of the current state. Feel free to come by and talk to us.

oke-aditya commented 2 years ago

I understand your thoughts Philip, as documenting such modules having lot of friction is hard, and would slow down development.

That's great. It would be really nice to hear during the PyTorch developer day. It would be indeed a great chance to get an opportunity to interact with developers :pray:

Edit: I guess I won't be able to attend the networking session :( Since I haven't got access to gather.town (had applied for it) Will look forward towards the poster and presentation.

pmeier commented 2 years ago

@oke-aditya I just added #5133 that is the starting point for the prototype datasets documentation.

oke-aditya commented 2 years ago

Great 😃 will have a read.