search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.57k stars 7.95k forks source link

Diferent map scores for '-map' flag and ' detector map' post training. #7632

Open Duarte-Nunes opened 3 years ago

Duarte-Nunes commented 3 years ago

I trained both yolov4-tiny and yolov4-tiny-3l on a custom dataset.

Whilst training the mAP for tiny was around 84% and for tiny-3l was around 85%.

To double check i ran the detector map command and for the tiny-3l the result was the same 85% but for the tiny im gettin 67% instead of the expected 84%.

lsd1994 commented 3 years ago

There are xx_best.weights and xx_last.weights after training, which one do you use?

stephanecharette commented 3 years ago

Use xx_best.weights.

Duarte-Nunes commented 3 years ago

I used the last weights indeed and that could have an impact but the last mAP calculation done during training also returned the 84%, since then i trained to additional times with some different settings and the "detector map" results have always been consistent with the training map calculation.

It was only that one exception, but i will try with best_weights and see, however it shouldn't be a problem that i expect to run into the future.