weighted sum in masked loss

[nehSgnaiL]

Hi,

Thanks for the remarkable work.

I would like to know more about the operations in defined loss. Since the mask has been normalized by mask /= torch.mean(mask), should we use the sum operation torch.sum(loss) rather than the mean operation torch.mean(loss) in returning loss?

Bigscity-LibCity/libcity/model/loss.py

Lines 75 to 87 in 38ff383

	def masked_mse_torch(preds, labels, null_val=np.nan):
	labels[torch.abs(labels) < 1e-4] = 0
	if np.isnan(null_val):
	mask = ~torch.isnan(labels)
	else:
	mask = labels.ne(null_val)
	mask = mask.float()
	mask /= torch.mean(mask)
	mask = torch.where(torch.isnan(mask), torch.zeros_like(mask), mask)
	loss = torch.square(torch.sub(preds, labels))
	loss = loss * mask
	loss = torch.where(torch.isnan(loss), torch.zeros_like(loss), loss)
	return torch.mean(loss)

I am not sure what I understand is right due to my limited knowledge. If you could respond, that would be greatly appreciated.

[nehSgnaiL]

The codes mask /= torch.mean(mask) and return torch.mean(loss) work properly. The outcome appears to be correct, but I find it difficult to read.