retPerLayer=True not working as expected

[paudom]

While trying different parameters for the LPIPS class, I encountered that when using retPerLayer=True in the forward method the returned values are not as I would expect.
I think the problem is due to the following lines:

107.   val = res[0]
108.   for l in range(1,self.L):
109.       val += res[l]

After debugging I notice the following:
Before doing the for on line 108

res = [tensor([[[[0.1195]]]...ackward1>), tensor([[[[0.0963]]]...ackward1>), tensor([[[[0.1021]]]...ackward1>), tensor([[[[0.1396]]]...ackward1>), tensor([[[[0.1210]]]...ackward1>)]

so val is 0.1195 at first, but after iterating through the for res[0] keeps updating at the same value of val so at the end I get the following:

res = [tensor([[[[0.5785]]]...ackward1>), tensor([[[[0.0963]]]...ackward1>), tensor([[[[0.1021]]]...ackward1>), tensor([[[[0.1396]]]...ackward1>), tensor([[[[0.1210]]]...ackward1>)]
val = 0.5785

Could maybe be solved by assigning val to 0 at line 107?

107.   val = 0
108.   for l in range(self.L):
109.       val += res[l]

I'm trying to use only the output of some layers instead of the accumulation of all the layers.
Thanks in advance!!!

[richzhang]

retPerLayer simply returns the layer-wise results, in addition to the accumulated results https://github.com/richzhang/PerceptualSimilarity/blob/master/lpips/lpips.py#L119. Try looking at the 2nd output

[paudom]

What I was saying is that when the layer-wise results are activated the first position res[0] is the same as the accumulated value instead of being the value of the first layer.

[richzhang]

Could you look at the 2nd output, which has a list containing every layer?

[paudom]

That's exactly what I'm referring to. The second output is a list, yes. But the first element of that list is exactly the same as the accumulated value! res[0] == val and that is not correct, as the first element should be the output of the first layer not the accumulated value.

[richzhang]

Yikes, you're totally right. I fixed it and uploaded a new version to pypi. Thanks!