Improving code readability of torchvision.models.detection

[Umong51]

🚀 The feature, motivation and pitch

Dear @pmeier,

As a student who is learning about the Faster R-CNN family and trying to understand the implementation provided in the torchvision.model.detection module, I find the code difficult to read and comprehend. Specifically, some blocks of code could be simplified further, such as the code block shown below:

#https://github.com/pytorch/vision/blob/2b25d67925df9741ba2a75a07bc3046302969e87/torchvision/models/detection/_utils.py#L162
box_sum = 0
for val in boxes_per_image:
    box_sum += val

Therefore, I would like to propose some improvements to enhance the code's readability and make it more accessible to beginners like myself. I believe these changes will improve the overall quality of the module. Here are some possible improvements that I could make:

• Add more comments and docstrings to explain the purpose and functionality of each function and module.
  Simplify and reorganize complex code blocks.
• I would be happy to collaborate with other contributors to implement these changes. Please let me know if you have any feedback or suggestions.

Thank you for your consideration.

Alternatives

No response

Additional context

No response

[pmeier]

Hey @Umong51 and thanks for your interest in contributing.

some blocks of code could be simplified further, such as the code block shown below

Not sure what you want to simplify in this example. Naively, I would write it like

box_sum = sum(boxes_per_image)

but I don't know if torchscript is ok with that. In any case, feel free to propose simplifications, but if any of that causes issues with torchscript, it is unlikely that we go forward with it. IMO the best way is to just send a PR and see if the CI is green.

Add more comments and docstrings to explain the purpose and functionality of each function and module.

Documentation improvements are always welcome, so go for it!

[NicolasHug]

Hi @Umong51 ,

Thanks for the proposal. The detection be a little hairy indeed. Feel free to submit PRs for the parts that deserve some improvements - ideally small PRs, so they're easy to review. I'd suggest focusing on docs rather than on code for now, because modifying the code involves risk of correctness (and our model testing scripts aren't exactly strong right now).

[Umong51]

Hi @pmeier & @NicolasHug,
Thanks for your response. I appreciate your suggestion to focus on documentation for now. I'll submit small PRs for improvements. Looking forward to contributing further.

[oke-aditya]

Detection is a tricky module to debug in torchvision indeed. But here are few places where you can start!

1. Quite a few functions are internal, not exposed to docs on pytorch.org/vision. These lack docstrings in the code, You can probably dive into them and add docs?
2. Although typing may improve readability slightly. It's not easy to get it compatible with JIT and go through code reviews. Takes lot of developer effort. Might not be worth the run, but maybe some small places you can add?
3. Maybe break down a model, generate anchors, then generate ROIs, visualize them. Load weights, then visualize these. Do a prediction without NMS, then with NMS, and visualize it. Visual guides are best way to understand least on high level.