MAELoss may be not averaged along batch dimension

[MiZhenxing]

Hi, thank you for your excellent work. I have a question about MAELoss.

PointMVSNet/pointmvsnet/networks.py

Lines 170 to 181 in cacb2d7

	class MAELoss(nn.Module):
	def forward(self, pred_depth_image, gt_depth_image, depth_interval):
	"""non zero mean absolute loss for one batch"""
	# shape = list(pred_depth_image)
	depth_interval = depth_interval.view(-1)
	mask_valid = (~torch.eq(gt_depth_image, 0.0)).type(torch.float)
	denom = torch.sum(mask_valid, dim=(1, 2, 3)) + 1e-7
	masked_abs_error = mask_valid * torch.abs(pred_depth_image - gt_depth_image)
	masked_mae = torch.sum(masked_abs_error, dim=(1, 2, 3))
	masked_mae = torch.sum((masked_mae / depth_interval) / denom)
	return masked_mae

As above codes show, the Line 179 takes the sum of the average loss of each depthmap, so the resulting loss is not the average loss but the sum loss of each depthmap in one batch. I would like to ask if this would affect the training of the network, especially when the batch_size changes. Or this code is written for a purpose I don't know? Thank you!