Assertion Error (using images of a different size)

[Wazaki-Ou]

It seems like using images with a different size throws this error. I would like to double check that modifying that part of the code to make it accept images with other size will not create an issue in calculating the flow. Also. if there is any specific part of the code I need to be careful with when modifying.

Thank you

[sniklaus]

I have tried in on different sizes as well without any issues. The important part is that the flow is being rescaled, just like in the official implementation.

pytorch-pwc/run.py

Line 328 in 9cf1e29

tensorFlow[:, 0, :, :] *= float(intWidth) / float(intPreprocessedWidth)

[Wazaki-Ou]

This is the error I get:

my picture is a jpg of size 1920*1080

[sniklaus]

I am afraid that I a little bit confused. It says remember that there is no guarantee for correctness, comment this line out if you acknowledge this and want to continue, have you tried commenting the line out?

[Wazaki-Ou]

For anyone who faces the same issue, commenting the lines restricting the size can solve the issue. Thanks !!

[sniklaus]

I wouldn't necessarily call it an issue though, it is more like an acknowledgement.

[SystemErrorWang]

in fact, I used inputs with same size and got size mismatch problem in the middle of the network.

File "/home/wangxinrui/bbox_net/flow_net.py", line 104, in forward
tensorInput = torch.cat([ tensorInput, self.tensorPartial ], 1)
RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 8 and 7 in dimension 2 at /opt/conda/conda-bld/pytorch_1535491974311/work/aten/src/THC/generic/THCTensorMath.cu:87

[sniklaus]

Would you mind sharing the images that cause this error? If not, what is their resolution?

[SystemErrorWang]

I tested again and found 224224 resolution caused this problem while 256256 worked well. I guess the image resolution should be divided by 64.

[sniklaus]

I am not exactly sure what happened in your case, see the following excerpt.

pytorch-pwc/run.py

Line 311 in e76ece2

intPreprocessedWidth = int(math.floor(math.ceil(intWidth / 64.0) * 64.0))

Which ensures that the resolution is divisible by 64 and resizes 224x224 to 256x256 accordingly.