How to use CmBN/self-adversarial mentioned in yolov4?

[yinruiying]

No description provided.

[ly0303521]

#5117
self-adversarial

[vb123er951]

Does there have answer for "how to use CmBN?"
I try to use CmBN on cspdarknet53
I change all the batch_normalize=1 in cfg file to batch_normalize=2, but the IOU and GIOU always 0, like this:

[yinruiying]

Does there have answer for "how to use CmBN?"
I try to use CmBN on cspdarknet53
I change all the batch_normalize=1 in cfg file to batch_normalize=2, but the IOU and GIOU always 0, like this:

I try batch_normalize=2 in my cfg, and the training process is normal.

[vb123er951]

Does there have answer for "how to use CmBN?"
I try to use CmBN on cspdarknet53
I change all the batch_normalize=1 in cfg file to batch_normalize=2, but the IOU and GIOU always 0, like this:

I try batch_normalize=2 in my cfg, and the training process is normal.

Your log shows you are using ciou? Do you change any other settings?
Can you tell me what pretrained weight you use and share your cfg with me?
Thank you very much.

[yinruiying]

I use the pretrained weight trained on my own dataset many times. Just changing batchnorm=2, the best_map increase .8%.
darknet53-yolov3_20200604.cfg.txt

[yinruiying]

#5117
self-adversarial

Thanks！I'll try it.

[kairavpatel]

I use the pretrained weight trained on my own dataset many times. Just changing batchnorm=2, the best_map increase .8%.
darknet53-yolov3_20200604.cfg.txt

Awesome result!!
Anyone has an idea, how many previous iterations are considered to calculate current statistics, while using CmBN? @AlexeyAB

[AlexeyAB]

Anyone has an idea, how many previous iterations are considered to calculate current statistics, while using CmBN?

This number is subdivisions in cfg file.

previous mini_batches inside one batch.

[kairavpatel]

Anyone has an idea, how many previous iterations are considered to calculate current statistics, while using CmBN?

This number is subdivisions in cfg file.

previous mini_batches inside one batch.
Ok..Thanks..