@foralliance Hi！I think we should speak English because your questions may help the people in other countries.

[defei-coder]

@foralliance Hi！I think we should speak English because your questions may help the people in other countries.

1. No. When you train the model with large input image size (e.g., 800x1300), the batch-size will reduce to 1-2 due to the limited GPU memory. Then the effect of BN will be constrained. If so, please replace BN with GN. GN does not care batch-size.

2. This question is very complex. I can just answer your question that they are not relevant.

Originally posted by @KimSoybean in #3 (comment)

[defei-coder]

@KimSoybean Hi
I read the answer above,you mean one GPU can not reach the large batch_size, I think 128 means accum_batch_size, we can use one GPU read 4 batch_size by 32 iter_size. So the batch_size of one GPU(such as 4) will Influence the effect of BN? Or why you recommend GN.

[defei-coder]

Or B of BN means batch_size that one gpu can reach.Looking forward to your answer。

[KimSoybean]

BN is evaluated by the statistics in a batch, so batchsize will impact the perfoemance. I have tried GN on mmdetection on ScratchDet, and get 1mAP lower result. I haven't found the reason.