Fix Windows-incompatible mkdir and add COCO auto-download to tutorial notebook

[ashishsoni-ai]

What

Fixes tutorial_image_detection.ipynb so it runs end-to-end on Windows without manual intervention.

Why

• !mkdir -p fails on Windows (cmd.exe doesn't support the -p flag), throwing a SyntaxError
• The COCO dataset cell had placeholder paths (<PATH_TO_COCO_VAL2017_IMAGES>) with no download guidance, blocking new contributors from running the notebook at all

Changes

• Replaced the Unix-only !mkdir -p calls with cross-platform os.makedirs(..., exist_ok=True)
• Added a cell that automatically downloads and extracts COCO annotations, with an option to either:
  • use a small subset (default, ~20 images, fast) for quick testing, or
  • download the full val2017 dataset for complete evaluation
• Updated the dataset-loading cell to use the downloaded paths instead of placeholders

Testing

Ran the full notebook end-to-end on Windows 11 (Python 3.11, PyTorch 2.4.1 CPU) — all cells execute successfully, including model inference, evaluation metrics, and the precision-recall curve output.

Fixes #612

[dpascualhe]

Hi @ashishsoni-ai , nice contribution! The only thing that needs fixing before merging is removing duplicated imports and moving those that are actually new to the first cell for consistency

[ashishsoni-ai]

Thanks for the review @dpascualhe . Cleaned up the duplicate imports and moved the new ones into the first cell. Let me know if anything else needs adjusting — happy to help with more issues if there are any you'd point me toward.

[ashishsoni-ai]

Hi @dpascualhe, just following up on this PR. I addressed the requested changes by removing the duplicate imports and moving the new imports into the first cell for consistency.

I also noticed the PR now has merge conflicts with the base branch. If you'd prefer, I can rebase and resolve the conflicts as well.

I'd appreciate another review whenever you have time. Thanks!

[dpascualhe]

Hi @ashishsoni-ai ! Thanks for addressing the changes, the new downloading logic is working perfectly fine. As you said, there are some conflicts now, can you resolve those? Besides, the cell that Run evaluation on a subset of the dataset can be simplified now, as you are already defining the size of the subset in the first cell and could use directly the dataset object instead of defining a new small_dataset. After that, we are ready to merge 😄

[ashishsoni-ai]

Hi @dpascualhe, both points are addressed and pushed:

• Conflicts: merged master into the branch — the only conflict was the notebook metadata (kernelspec / language_info), where I kept upstream's values. Interestingly, upstream had already landed the same os.makedirs fix, so that part merged cleanly.
• Evaluation cell: simplified as suggested — it now runs detection_model.eval(dataset=dataset, ...) directly, since the dataset is already limited to SUBSET_SIZE in the first cell. The small_dataset slice, the temporary dataset.dataset swap and its finally restore are gone.

The PR is conflict-free now. I'll refresh the stored notebook outputs (they still show the old 5-image evaluation) in a moment — after that we should be ready to merge. Thanks for the review!