search

xueleichen / PyTorch-Underwater-Image-Enhancement

The repo for "Underwater Image Enhancement based on Deep Learning and Image Formation Model"
MIT License
87 stars 20 forks source link

why the batch size is 1 #4

Closed fengxiaohu closed 1 year ago

fengxiaohu commented 1 year ago

Usually I set the batch size 64 or 128 during the training process. I assume that the parameter is depend on your GPU memory,So can I set it to 64 when I own 3080ti to accelerate the training

xueleichen commented 1 year ago

Hi @fengxiaohu , I think it is a convention that people use small batch sizes in low-level computer vision tasks (like enhancement or superresolution). I just followed this rule.

Some examples are: https://github.com/Li-Chongyi/Water-Net_Code https://github.com/yulunzhang/RCAN https://github.com/xahidbuffon/SRDRM These three github repos use the batch size of 1 or 2.

But you can try large batch sizes and see what the results look like.

fengxiaohu commented 1 year ago

thanks for reply,I will try to increase the batch_size and see what will happen