Dropout2d doesn't drop channels for (C, H, W)

[OverLordGoldDragon]

🐛 Describe the bug

Docs read that C is dropped, which does not occur for (C, H, W)

import torch
torch.manual_seed(0)
ipt = torch.ones((2, 3, 4))
print(torch.nn.Dropout2d(p=0.5)(ipt))

tensor([[[2., 2., 2., 2.],
         [0., 0., 0., 0.],
         [0., 0., 0., 0.]],
        [[2., 2., 2., 2.],
         [2., 2., 2., 2.],
         [0., 0., 0., 0.]]])

but does for (N, C, H, W)

torch.manual_seed(0)
print(torch.nn.Dropout2d(p=0.5)(ipt[None]))

tensor([[[[2., 2., 2., 2.],
          [2., 2., 2., 2.],
          [2., 2., 2., 2.]],
         [[0., 0., 0., 0.],
          [0., 0., 0., 0.],
          [0., 0., 0., 0.]]]])

It appears Dropout2d is instead implementing Dropout1d for (N, C, T).

Versions

1.10.0 via Anaconda; Windows 10

cc @ezyang @gchanan @zou3519 @bdhirsh @albanD @mruberry @jbschlosser @walterddr @kshitij12345

[kshitij12345]

Even Dropout3d is affected with the same issue. Seems like the underlying implementation always assumes that the input will be batched.

pytorch/aten/src/ATen/native/Dropout.cpp

Lines 19 to 20 in fc37e5b

	sizes.push_back(input_sizes[0]);
	sizes.push_back(input_sizes[1]);

[albanD]

cc @jbschlosser should this be fixed when no batch dim support is added?

[kshitij12345]

FYI, as per the tracker issue #60585 it is already supported.

[albanD]

Ok, so most likely a bug then!

Tentative high pri

[jbschlosser]

Reopening this thanks to #79549

TL;DR: The fix for this issue was silently BC-breaking for those who depended on Dropout2d to do 1D channel-wise dropout for 3D inputs. For some reason, Dropout1d didn't exist before, so Dropout2d was the only option. We'll revert to the old behavior for 1.12 with a warning to move to the newly-added Dropout1d. In a release, we'll switch back to a no-batch-dim interpretation for 3D inputs.