add nn.CenterCrop1d and nn.CenterCrop2d

[bodokaiser]

There are some models which use "crop layers" (e.g. U-Net, ParseNet) so I think it wouldn't be bad to have a CenterCropNd layer.

As F.pad supports negative padding we just need to calculate padding offsets on top.

import torch

from torch.nn import functional as F
from torch.autograd import Variable

def center_crop(x, height, width):
    crop_h = torch.FloatTensor([x.size()[2]]).sub(height).div(-2)
    crop_w = torch.FloatTensor([x.size()[3]]).sub(width).div(-2)

    return F.pad(x, [
        crop_w.ceil().int()[0], crop_w.floor().int()[0],
        crop_h.ceil().int()[0], crop_h.floor().int()[0],
    ])

variable = Variable(torch.randn(1, 3, 60, 40))

print(center_crop(variable, 20, 20).size())
print(center_crop(variable, 20, 40).size())

I can send a PR on interest.

[fmassa]

Hum, I think the type of cropping that one might want to apply might be problem-specific (not necessarily only center crop).
Plus, cropping an image can be performed as efficiently as F.pad by indexing the tensor with slices like x[:, :, 3:7, 5:10], so I'm not so sure about the need of a special layer for that.

[bodokaiser]

What about caffe's crop layer?

[fmassa]

Caffe doesn't give you the ability to easily manipulate the tensors, so there is a need of such layer. The same applies for lua torch, as you need to define a module that will compute the operation that can be differentiated, but in pytorch you can just directly manipulate the tensors and the backprop will be performed.
if you check caffe's crop layer, it's almost the same as the snippet I sent in the previous message.

[apaszke]

Even though it's easy to do it with indexing, I think it'd be better to provide some standard crops in the library. It's much more convenient than messing with indices, and can save making a few errors along the way each time you need to crop.

[andrewgiessel]

+1 to this. I think standard cropping layers would be great!

[bodokaiser]

I could make a PR this week if there is consensus on API.

[apaszke]

I think F.center_crop(x, *args) and nn.CenterCrop(*args) (where len(args) == x.dim() - 2) would be the best API for it.

[andrewgiessel]

Is it reasonable to include non-symmetric cropping in this API as well?

[bodokaiser]

@andrewgiessel See the current proposition at #1349. I think it would be also reasonable to specify center crop in more detail (e.g. you want to crop 10 -> 3 you have one offset 3 and one 4 but you cants specify the order at the moment).
Is this what you thought of?

[fmassa]

This can be achieved by passing a negative pad in torch.nn.functional.pad.