Evaluation on Windows fails due to missing ninja (use standard COCOeval)

[cmaranes]

On Windows, the evaluation step tries to import the optimized Cython version (COCOeval_opt), which in turn requires a working ninja install for JIT compilation. If ninja isn’t available, the import fails and the entire eval pipeline breaks.

🔧 Temporary Workaround

Force use of the pure-Python evaluator by swapping out the import. In yolox/evaluators/coco_evaluator.py (around line 294): https://github.com/Megvii-BaseDetection/YOLOX/blob/d872c71bf63e1906ef7b7bb5a9d7a529c7a59e6a/yolox/evaluators/coco_evaluator.py#L294

, replace:

- from yolox.evaluators.coco_opt import COCOeval_opt
+ from pycocotools.cocoeval import COCOeval

This change bypasses the ninja dependency and lets evaluation complete using the standard COCO API.

Feel free to let me know if you’d like a PR implementing the fallback logic!