How to add rotation augmentation option to data transforms?

[ncullen93]

How to add rotation augmentation option to data transforms?

[ncullen93]

That looks almost right to me. You might need it after ToTensor though since it works on torch.Tensor. The issue with a lot of the actual pytorch transforms is that they only work on PIL images or numpy arrays instead of torch tensors.

Here's an example:

from torchsample.transforms import *
import torch as th
x = th.zeros(3,256,256)
x[:,50:150,50:150] = 1
xr = Rotate(10)(x)
xs = Shear(0.1)(x)
xz = Zoom((0.6,0.9))(x)

Then try to plot those to see how they change:

import matplotlib.pyplot as plt
plt.imshow(x[0].numpy())
plt.show()
plt.imshow(xr[0].numpy())
plt.show()
plt.imshow(xs[0].numpy())
plt.show()
plt.imshow(xz[0].numpy())
plt.show()

Original:

Rotated:

Sheared:

Zoomed:

And of course you can combine all of these into one transform using torchsample.transforms.Affine or torchsample.transforms.AffineCompose.. You'll want to use one of those if you're doing more than one affine transform so that it only does a single interpolation!

The cool and unique thing here is that If you just want to generate the transformation matrix, you can do so:

x = torch.rand(3,256,256)
rotation_tform = Rotate(30, lazy=True)(x)
print(rotation_tform)
# 0.8856 -0.4644  0.0000
# 0.4644  0.8856  0.0000
# 0.0000  0.0000  1.0000
#[torch.FloatTensor of size 3x3]

[wangg12]

@ncullen93 So I need to put Rotate(30) after ToTensor? BTW, the docstring of Rotate about lazy seems wrong. https://github.com/ncullen93/torchsample/blob/master/torchsample/transforms/affine_transforms.py#L184

[ncullen93]

correct. I believe all or nearly all of the torchsample transforms operate directly on torch tensors. Therefore, if you're reading images you should call ToTensor first before any of them. and correct, they're switched thanks. I'll fix it.

[wangg12]

@ncullen93 The fill_mode argument is not implemented?

[ncullen93]

No it just fills with 0's. Are you looking for something specific? I can implement it pretty quickly, but it will slow things down a fraction.

[wangg12]

Is filling 0 a common practice?

[ncullen93]

Yes if your data is normalized to 0-1 or some constant otherwise. It's what I use

[wangg12]

OK, Thanks.