Support for Multiple Channels in th_affine2d

[ncullen93]

Support for Multiple Channels in th_affine2d

[ncullen93]

hey it's because the current version only supports a single channel. Let me add support for that right now and get back to you

EDIT: ok, it should be fixed. Try that code again. The thing is that multi-channel support makes it a littttle slower. I don't plan to add multi-channel support for 3D affines right now.

[fizzxed]

Now there is another issue.
In th_affine2d we assume matrix has dimensions (2, 3) seen:

if matrix.dim() == 2:
    matrix = matrix.view(-1,2,3)

But in the affine_transforms.py the matrix has size (3, 3)

[ncullen93]

should be fixed... lol i desperately need to write some tests, sorry.

I fix it by just indexing to the second row so if there is a third row it'll get cut off. This is what Keras does.

[bfortuner]

Haven't looked closely, but having trouble with both Pad() and Rotate() passing in a tensor (3,256,256) which works for the other Tensor transform types..

Here's the error I see when I call AffineTransforms like Rotate(), Shear(), Zoom(), etc

rotated_img = torchsample.transforms.Rotate(10)(img_tns)

Error:
`/torchsample/utils.py in th_affine2d(x, matrix, mode, center)
---> 86 if matrix.dim() == 2:
87 matrix = matrix[:2,:]
88 matrix = matrix.view(-1,2,3)

RuntimeError: size '[-1 x 2 x 3]' is invalid for input of with 9 elements at /py/conda-bld/pytorch_1490980628440/work/torch/lib/TH/THStorage.c:55`

[ncullen93]

Sorry I'm not sure what the issue is without seeing more code. These all work for me:

from torchsample.transforms import *
import torch as th
x = th.ones(3,256,256)
xr = Rotate(10)(x)
xs = Shear(0.1)(x)
xz = Zoom((0.8,1.2))(x)
xp = Pad((3,300,300))(x)
print(xp.size()) # torch.Size([3, 300, 300])

Maybe has to do with the new pytorch version? Are you using 0.1.11 or 0.1.12? Let me upgrade. EDIT: Still works on 0.1.12.

[bfortuner]

This works! I was "slicing" my tensor with narrow() to remove an extra dimension, but I wasn't doing it right. Thanks

[ncullen93]

Great!