[Question] Training & Validating Models with Subset of COCO Dataset

[hashshura]

Hello,

I want to evaluate YOLOv4 models (yolov4, yolov4-csp, yolov4-tiny, and yolov4x-mish) for vehicle subset from COCO dataset (which includes only cars, trucks, buses, and motorcycles at ~19k images). I have already filtered the dataset and successfully trained some models.

My questions are:

• What's the difference between .conv and .weights files in YOLOv4 pre-release?
• Should I use them (the convs) to train (and then evaluate) the models with COCO dataset (knowing AFAIK, those were pre-trained using COCO too)?
• If I shouldn't, what conv I can use to train & evaluate with COCO subdataset?
• Is it OK for the YOLOv4-CSP and YOLOv4x-mish to have high loss (> 100, but increasing mAP) when training?

I asked this because it seems like my YOLOv4 resulted model validation (train:val is 95:5) has too high mAP (even higher than YOLOv4-CSP and YOLOv4x-mish, all of them 416x416):

I saw similar issues, but my questions didn't really get answered there.

Thank you very much for this awesome work!

[KKDDD]

I have the same problem. I want to evaluate a custom data set with coco metrics, but I can’t find a suitable method. At the same time, I i found that all“image_id” of the JSON file is 0.