Hi,
In the original paper, batch norm is a good tool to get better performance. However, in my case, the DnCNN model without BN is superior to the one with BN. It always happens in other network architecture as well. Is there any idea about what happen with BN in denosing task.
(I use the original code which of building model and generating noise and there are some other informs:
batch size : 32
patch size 128
colored images.
)
Hi, In the original paper, batch norm is a good tool to get better performance. However, in my case, the DnCNN model without BN is superior to the one with BN. It always happens in other network architecture as well. Is there any idea about what happen with BN in denosing task.
(I use the original code which of building model and generating noise and there are some other informs: batch size : 32 patch size 128 colored images. )
Thanks.