BCHW format about stn.pytorch HOT 13 OPEN

You can use transpose: img = img.transpose(1,2).transpose(2,3), this should change BCHW layout to BHWC

from stn.pytorch.

But, transpose(1,2).transpose(2,3) seems not to rearrange the internal array.
torch.FloatTensor(1,2,3,4).stride() and torch.FloatTensor(1,2,3,4).transpose(1,2).transpose(2,3).stride() are (24, 12, 4, 1) and (24, 4, 1, 12), respectively, while torch.FloatTensor(1,3,4,2).stride() is (24, 8, 2, 1).

So if I run the code, at line 44 and 45 in my_lib.c,

real yf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth];
real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1];

xf is not valid, because grids_strideWidth is still 1.

I guess it needs to be like

real xf = grids_data[b*grids_strideBatch + yOut*grids_strideHeight + xOut*grids_strideWidth + 1*grids_strideChannel];

although I have not tested it.

from stn.pytorch.

transpose(1,2).transpose(2,3) changes the internal array, you can use .size() to check, I have been using that all the time. test_conv_stn.ipynb actually uses that fyi.

On a separate note, I guess BCHW should be the standard because it follows pytorch conv layers convention. I probably will have a version for that later. Let me know what you think.

from stn.pytorch.

Oh sorry I misunderstood, you are talking about permutation for grid rather than image. Hmm, I always use the grid generator to generate grid in BHWC format directly so never run into the problem you mentioned.

from stn.pytorch.

Thank you. Meanwhile I'll use permute (or transpose) and then contiguous(). It seems to work properly so far :)

from stn.pytorch.

Thank Fei for the nice work. Do you have any update on BCHW support?

from stn.pytorch.

@junyanz Hi Junyan, thank you for your interest, it is likely to be added after the NIPS deadline. We do find the majority of users need BCHW instead of BHWC and will thus prioritize it :D .

from stn.pytorch.

@fxia22 Thanks for your prompt response. Good luck with your NIPS submission.

from stn.pytorch.

BCHW support added. example can be found in test.py

from stn.pytorch.

Thanks a lot!

from stn.pytorch.

@fxia22 to go from BHWC to BCHW just use img.permute(0, 3, 1, 2)

from stn.pytorch.

@edgarriba Thanks for your suggestion. As discussed above, the problem of BCHW for STN is that BCHW layout is not suitable for coalescing. Permutation itself doesn't change the memory layout, but .contiguous() after permute will work.

from stn.pytorch.

ah, right. Permute just recompute strides

from stn.pytorch.