Detection on Custom Dataset

[pranjulparnami]

Hi,

Firstly, thanks for this great piece of work!

I would like to train YOLOv4-Tiny on my custom dataset and the use-case is to run the model on Jetson TX2. So, could you please tell, how can I train the model so that it remains compatible with this code (as you've mentioned in another repo that this doesn't work with AlexeyAB's model).

I have checked this WongKinYiu/PyTorch_YOLOv4 repo. But there isn't any support to train Tiny-YOLOv4.

I can train my model using Darknet/AlexeyAB's repo, so if you can share the necessary changes required to make it compatible, that will be great.

Either of the above will work.

Regards!

[tjuskyzhang]

Hi,

Firstly, thanks for this great piece of work!

I would like to train YOLOv4-Tiny on my custom dataset and the use-case is to run the model on Jetson TX2. So, could you please tell, how can I train the model so that it remains compatible with this code (as you've mentioned in another repo that this doesn't work with AlexeyAB's model).

I have checked this WongKinYiu/PyTorch_YOLOv4 repo. But there isn't any support to train Tiny-YOLOv4.

I can train my model using Darknet/AlexeyAB's repo, so if you can share the necessary changes required to make it compatible, that will be great.

Either of the above will work.

Regards!

This may help you: WongKinYiu/PyTorch_YOLOv4#3 (comment)

[pranjulparnami]

The training part is working fine. But detection results are different from the one I get using WongKinYiu/PyTorch_YOLOv4's detect.py for same weights with same confidence and nms threshold.

[tjuskyzhang]

Try to set the MAX_OUTPUT_BBOX_COUNT to a large value like 99999999.
And nms method maybe different from the pytorch.