I am trying to replace my existing BN layers with GroupNorm. However, after doing so, my model needs nearly twice the GPU memory compared to before with BN layers. Is this supposed to be the case or might there be something wrong? Thanks!
I am trying to replace my existing BN layers with GroupNorm. However, after doing so, my model needs nearly twice the GPU memory compared to before with BN layers. Is this supposed to be the case or might there be something wrong? Thanks!