Average EPE score with PyTorch implementation

[zohrehazizi]

Hi. I loaded the pwc_net.pth.tar with the pytorch model and ran the inference on all the MPI Single dataset. I used the following way to calculate the average EPE loss:

for each image:
epe = torch.norm(flow_gt-flo, p=2, dim=0).mean()
epe_list.append(epe)

(after the loop:)
epe_list.mean()

where dim=0 is the axis corresponding to flow channels and has size 2.
And I get a score of 3.20 on Clean and 3.61 on Final . This is different from the numbers you reported for MPI Sintel. Could you please help me realize what I might be doing wrong? Thank you.