YOLOv4 Training with TAO vs Darknet: Differences in mAP Values

I’ve been training a YOLOv4 object detection model with TAO, but the results are not satisfactory. The highest mAP value I’ve achieved is around 0.455, which is quite low compared to the results I get when training with Darknet. In Darknet, my initial mAP value is around 0.64. My highest mAP is over 80. I’m trying to understand the problem causing this difference in performance.

Has anyone else experienced similar issues when training YOLOv4 with TAO compared to Darknet? Any suggestions on how to improve the mAP values when training with TAO?

TAO config:
yolov4_train_resnet101.txt (2.2 KB)

TAO training log:
training_log.csv (2.5 KB)

Darknet config:
yolov4-custom.cfg (12.0 KB)

Please refer to

Pretrained weights trained on the ImageNet dataset tend to provide good accuracy for object detection. Due to copyright issues, we can’t provide the ImageNet dataset or any ImageNet-pretrained models in TAO Toolkit.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

Hi,
Please refer to below link.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.