Post Paper Architectural Optimization ReadMe update

[artest08]

📚 The doc issue

The announcement about the V2 improvements of RetinaNet, FasterRCNN and MaskRCNN has been made (Post Paper Architectural Optimization). Thanks for the amazing work.

However, if the references/detection/Readme is updated conveniently with the new configuration, it would be amazing to reproduce the results and utilize the new training scheme in other datasets.

Thanks in advance

Suggest a potential alternative/fix

No response

[datumbox]

@artest08 Thanks for the issue.

We've added all this information on the docs. Every model comes with a recipe link that points to how it was trained so that it's fully reproducible.

The fact that you were obliged to open an issue might indicate that our docs are not discoverable. That's good feedback to know. Could you let me know if you looked on the documentation but didn't find the section or if you didn't consider looking there because previously we didn't have the section? Or perhaps you would prefer to have the commands also available on the references readmes?

[artest08]

Thanks for your fast response @datumbox.

Actually, I did not know that there is recipe information in the documentation of the models. So, I did not consider looking there. That is exactly what I was looking for.

Maybe, some recipe links of the crucial models (not all of them) can be added also in references/detection/ReadMe so that people would become aware about where to look when they need them.

[datumbox]

@artest08 Sounds reasonable. Once issue with this is that we now have a bunch of versions, so adding the commands in the readme in an unstructured manner can be quite a lot of text. On the other hand, I agree that having some of the models in the reference readmes is also very misleading...

@NicolasHug What you think? Shall we add all commands back to the readmes? Or perhaps remove the readmes and move them somewhere else?

[NicolasHug]

I guess we should start thinking about documenting the "recipes" in a more centralized place, to make all this more obvious. In our docs or in the READMEs, I don't really have an opinion yet.

But as an easy patch, perhaps we could start by clearly stating at the top of all the README something like:

NOTE: More details about how each model was trained can be found in their corresponding models documentation. See for example the recipe link of FasterRCNN_ResNet50_FPN_V2_Weights.

[datumbox]

@NicolasHug agreed that adding a quick patch can clarify things.

@artest08 would you be interested in sending a PR that adds it?