Training problem

[BoSeal]

When I train the lane dataset, I use the hardware configuration： one Tesla K20 GPU with 5G memory. This causes OOM！To solve this problem, I reduce the batch size to 4 ，but still happen“OOM”. So I don't know if I'm going to keep reducing the batch size? If I continue to reduce the batch size, I worry that it will affect the accuracy of the results and not reach the accuracy obtained in your paper. Can you give me some advice?

[XingangPan]

@BoSeal Reducing batchsize would harm the performance of BN. One possible solution to this is to use frozen BN and smaller batchsize (like 2) as done in some semantic segmentation works.

[BoSeal]

Thanks very much！

[OrkunYilmaz]

Why does reducing batchsize harm the performance of BN? What is BN?

Thanks in advance