Revamping our classification models docs

[NicolasHug]

We're re-writing our models docs to make them clearer, simpler, and to properly document the upcoming multi-weight API. This issue is about adding docs for the classification models.

Perhaps this is something @oke-aditya @YosuaMichael @lezwon @zhiqwang would be interested in :) ?

Our latest new docs are currently here (this link is likely outdated by the time you look at it, but it doesn't matter; the skeleton is there). We created a separate section that will eventually be merged into the main one. We have documented a few models, but most of them are still missing. The list of models that still need docs is listed below. If you'd like to participate, please comment below with a message saying "I'm working on XYZ" where XYZ is a model, so that others don't pick the same as you do. To keep things simple, please submit one PR per model, but feel free to contribute more than one model.

How to write docs for a model

Note: below are detailed instructions. This makes it look more complicated than it actually is. Don't be scared!

A great place to start is to look a the changes in this PR that documents SqueezeNet. You'll need to do exactly the same for your model:

• Create a new .rst file in https://github.com/pytorch/vision/tree/main/docs/source/models. The file should look like this, with a link to the original paper, and a list of the corresponding model builders. It should also mention the base model class and link to the .py file where it is defined.
• Update the list in https://github.com/pytorch/vision/blame/main/docs/source/models_new.rst to link to this new file (without the .rst suffix). Please keep the list alphabetically sorted
• Update the docstring of each new model builder, similarly to this one.
  • there is a 1:1 mapping between a model builder and a Weight enum. For example, the docstring of squeezenet1_0 makes direct references to SqueezeNet1_0_Weights.
  • don't forget the autoclass directive in the docstring. This will auto-generate documentation for the Weight enums. You don't need to understand how this is done, but if you're curious, it's done here.

To build the docs locally, please look at our contributing guide. You won't need to worry about the gallery example, so always use make html-noplot instead of make html to save time.

Please don't hesitate to ping us if you need any help / guidance or if you have any question!

---

Classification models that need docs are:

• Alexnet - AlexNet New Documentation. #5886
• ConvNeXt - Add docs for ConvNext #5869
• Inception v3 - Add inception docs #5924
• DenseNet - Added docs for DenseNet #5878
• EfficientNet - EfficientNet New Documentation #5854
• EfficientNet v2 - Efficientnetv2 #5864
• GoogLeNet - GoogLeNet New Documentation #5885
• MNASNet - updating mnasnet docs #5920
• MobileNet v2 - docs: Added MobileNetV2 to the new doc #5899
• MobileNet v3 - Add docs for Mobilenetv3 #5908
• RegNet - Adding revamp docs for vision_transformers and regnet #5856
• ResNext - Add docs for ResNeXt #5871
• ShuffleNet v2 - Add revamped docs for ShuffleNet V2 #5921
• VisionTransformer - Adding revamp docs for vision_transformers and regnet #5856
• Wide ResNet - revamp docs for Wide ResNet #5907

[YosuaMichael]

This is great @NicolasHug !
I would like to try adding doc for RegNet and VisionTransformer.

[datumbox]

Great initiative to split into bite-sized chunks. :)

Tagging a few more regulars in case they are interested in helping: @xiaohu2015 @abhi-glitchhg @frgfm @yassineAlouini

[lezwon]

Hey @NicolasHug, I'd like to start off with densenet.

[oke-aditya]

I will take MobileNets :) v2 v3

[abhi-glitchhg]

I would like to take ResNext.

[zhiqwang]

I would like to take ShuffleNet v2

[xiaohu2015]

I would like to take ConvNeXt.

[xiaohu2015]

@xiaohu2015 I think @YosuaMichael has already picked VisionTransformer.

Ok, I choose another model ConvNeXt

[yassineAlouini]

EfficientNet and EfficientNet v2 if it is still for grabs. 😸

[datumbox]

@yassineAlouini It's yours! :)

Note that EfficientNet V1 and V2 are implemented on the same class (EfficientNet). On the documentation we want to document them separately because they are 2 different papers.

[frgfm]

Hi @NicolasHug @datumbox, thanks for the tag 🙂
I'm on holidays until next Monday, I'll take a look at what's left then! Great idea to improve the doc by the way 🙌

[yassineAlouini]

@yassineAlouini It's yours! :)

Note that EfficientNet V1 and V2 are implemented on the same class (EfficientNet). On the documentation we want to document them separately because they are 2 different papers.

Thanks. Should I work on both documentations then? Or at least start with EfficientNet? 🤔

[NicolasHug]

@yassineAlouini sure, feel free to do both!

[yassineAlouini]

One last question: how should we link the PR that we create? Is there a list of issues or do we just reference this?

[NicolasHug]

You can just reference this one!

[yassineAlouini]

Since my previous MRs have been merged, I can work on GoogLeNet doc if it is available. 👌

[NicolasHug]

Please pick as many as you'd like @yassineAlouini :)

[yassineAlouini]

Next: Alexnet.

[abhi-glitchhg]

I would take Wide Resnet

[frgfm]

Alright I'm back 😁
Anyone wants to lend me one of those models? @oke-aditya mind if I take MobileNetV3 while you handle MobileNetV2 for instance? 🙏

[oke-aditya]

Sure Take v2. I'm back too. Was sick :(

[NicolasHug]

Thank you so much everyone who participated so far, your help is greatly appreciated!

For those who would like to pick more models, I just opened this new issue to document the rest of the models (semantic segmantation and object detection) #5897

[frgfm]

Sad to hear this @oke-aditya :/ Happy to take care of V3 as well while I'm at it if you want!

[oke-aditya]

I Will take care of v3, slowly getting back to work. :)

[zhiqwang]

Hi @NicolasHug , Sorry for delay, I will finish ShuffleNetV2 by tomorrow.

[NicolasHug]

No rush @zhiqwang :)

[jdsgomes]

I will work on the MNASNet - seems to be the last one

[yassineAlouini]

@NicolasHug Shouldn't the GoogLeNet be checked in the list? 🤔 Asking so that no one does the job again. 😄

[NicolasHug]

Looks like we're all done! Thank you so much everyone for your help <3 !

[yassineAlouini]

Next is object detection/segmentation I guess? 🤔 @NicolasHug

[NicolasHug]

@yassineAlouini I think we took care of that in #5897

The only remaining docs are the ones for the classification Quantized models. https://github.com/pytorch/vision/tree/main/torchvision/models/quantization. I think I'll create another issue for these ones soon, let me know if you'd be interested!

[yassineAlouini]

@NicolasHug I see, that's great. I am interested in helping. 👌