Does BatchNorm make performance uncertain even worse? about cosnet

[carrierlxk]

Hi, you can fix the batch norm layer. In addition, to reduce the gpu memory, you can reduce the input image size from 473 to 378.

[lscelory]

Hi, you can fix the batch norm layer. In addition, to reduce the gpu memory, you can reduce the input image size from 473 to 378.

Thank you for your advice! I will try these settings.

[NoaHimesaka]

Hi, you can fix the batch norm layer. In addition, to reduce the gpu memory, you can reduce the input image size from 473 to 378.

Thank you for your advice! I will try these settings.

I'm sorry to bother you
But I'm curious, do you get a better result?
I've met the same question with you, I can only set batch_size=4 and change the input size to 378 using 2 GPU 1080ti

[lscelory]

According to my exps, it seems batch size is not the determinant factor as long as you have a strong pre-trained model.

[NoaHimesaka]

According to my exps, it seems batch size is not the determinant factor as long as you have a strong pre-trained model.

Thank you!