Why EDVR use 'sum' as the loss reduction method?

[hcwang95]

I am thinking using sum will let the loss magnitude change if we change the batch size or input resolution. Is there any reason to use 'sum' instead of 'mean'? Thanks a lot!

Reference config: train_EDVR_M_x4_SR_REDS.yml:

  # losses
  pixel_opt:
    type: CharbonnierLoss
    loss_weight: 1.0
    reduction: sum