the drop_size cannot be higher than 5?

[PanXiebit]

dropblock/dropblock/dropblock.py

Line 78 in 04b759f

padding=int(np.ceil(self.block_size // 2) + 1))

the implement of computer block mask using conv1d is to handle with the overlapped? and the output of _computer_block_mask must have the same size with x, right?

but when I try block_size=7, the output of block_block_mask is [N, 1, 1]. and when I try block_size=9, I will get an error.

In the function of computer block mask, after conv2d, the size of height and weight is
mask_size + 2*(block_size//2+1) - block_size + 1,
the height and weight of input x is mask_size + block_size//2.
and the former must be larger than the latter, so the (block_size//2) connot higher than 3?

[miguelvr]

The block size can be bigger for sure. It is being tested in the unit tests.

What is the shape of your input?

[PanXiebit]

@miguelvr I test like this

x = torch.rand(100, 10, 16, 16)
drop_block = DropBlock2D(block_size=9, drop_prob=0.3)

and get the error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/panxie/.conda/envs/NLP/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/panxie/.conda/envs/NLP/lib/python3.6/site-packages/dropblock/dropblock.py", line 67, in forward
    out = x * block_mask[:, None, :, :]
RuntimeError: The size of tensor a (16) must match the size of tensor b (2) at non-singleton dimension 3

[miguelvr]

That input is too small for a 9x9 block. The Bernoulli mask is going to be computed on a subset of your input of shape (W - bsize // 2, H - bsize // 2)

[PanXiebit]

>>> x = torch.rand(100, 10, 64, 64)
>>> drop_block = DropBlock2D(block_size=9, drop_prob=0.3)
>>> regularized_x = drop_block(x)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/panxie/.conda/envs/NLP/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/panxie/.conda/envs/NLP/lib/python3.6/site-packages/dropblock/dropblock.py", line 67, in forward
    out = x * block_mask[:, None, :, :]
RuntimeError: The size of tensor a (64) must match the size of tensor b (2) at non-singleton dimension 3

I try bigger size input, still get the error.
I think the problem is the padding is to small, if padding=int(2 * np.ceil(self.block_size//2)) no matter how bigger block_size, the height_to_crop can be positive.

[miguelvr]

Are you using the last version? I'll check that tomorrow

[miguelvr]

Please use the bleeding edge version from github now, it should be fixed in #21

I'll do a version bump soon and upload to pypi, in order to install from pip

[miguelvr]

should be now in pypi