Splitting image into small batches

vision
1

Hi,
I am new to Pytorch but would like to do the following.
Lets say I have 37 images with two channels and size 256x256.
In other words, my data is the shape (37,2,256,256).
I want to split each image into (256/stride)x(256/stride) images of size kernel_shape assuming circular/periodic padding. I am assuming that each small batch is independent from each other.

For example, assuming stride=1 and kernel_shape=(5,5), a data like (37,2,256,256) will become (37*256*256,2,5,5). I initially thought that Conv2D could be used for that, but after seeing examples I realized this is actually for creating convolutional layers.

2

I think unfold should work. This post gives you an example.

3

I haven’t tried yet, but I think I understand the code. Thanks!
That solves the part o breaking into patches. Do Pytorch have anything for adding the padding needed for the periodic/circular padding? I want that the every pixel has it’s own patch.
Edit: Just noticed F.pad has option for periodic boundary condition.

4

I was exploring the padding options and

x = torch.randn(2, 37, 256, 256)+10
S = x.shape

kh, kw = 101, 101  # kernel size
dh, dw = 1, 1  # stride

pad = (S[-1]%kw // 2, S[-1]%kw // 2, S[-2]%kh // 2, S[-2]%kh // 2, 0, 0, 0, 0)
# Pad to multiples of 32
x = F.pad(x, pad)

plt.imshow(x[0,0,:,:])

Returns
image

But selecting mode="circular" gives NotImplementedError.

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
/home/gridsan/groups/dedalus_group/gfd22/stochastic_closures_repo/src/01-model/ml/cnn.py in <cell line: 11>()
      40 pad = (S[-1]%kw // 2, S[-1]%kw // 2, S[-2]%kh // 2, S[-2]%kh // 2, 0, 0, 0, 0)
     41 # Pad to multiples of 32
---> 42 x = F.pad(x, pad, mode="circular")
     44 plt.imshow(x[0,0,:,:])

File /state/partition1/llgrid/pkg/anaconda/anaconda3-2022b/lib/python3.8/site-packages/torch/nn/functional.py:4397, in _pad(input, pad, mode, value)
   4395         raise NotImplementedError
   4396 else:
-> 4397     raise NotImplementedError("Only 2D, 3D, 4D, 5D padding with non-constant padding are supported for now")

NotImplementedError: Only 2D, 3D, 4D, 5D padding with non-constant padding are supported for now

torch.__version__ = 1.11.0+cu113
Not sure I understand this error, because for me, this is a 4D padding.

I tried to perform the padding for each channel and snapshot and got the same error.

Similar code works for numpy.pad.

5

Circular padding seems to have more limitations as given in the docs so you might need to pass 4 values only:

x = torch.randn(2, 37, 256, 256)+10
S = x.shape

kh, kw = 101, 101  # kernel size
dh, dw = 1, 1  # stride

pad = (S[-1]%kw // 2, S[-1]%kw // 2, S[-2]%kh // 2, S[-2]%kh // 2, 0, 0, 0, 0)
y1 = F.pad(x, pad)
print(y1.shape)

pad = (S[-1]%kw // 2, S[-1]%kw // 2, S[-2]%kh // 2, S[-2]%kh // 2)
y2 = F.pad(x, pad, mode="circular")
print(y2.shape)