MaxPool2d is not fully invertible, since the non-maximal values are lost.
MaxUnpool2d takes in as input the output of MaxPool2d
including the indices of the maximal values and computes a partial inverse
in which all non-maximal values are set to zero.
Arguments
- kernel_size
(int or tuple): Size of the max pooling window.
- stride
(int or tuple): Stride of the max pooling window. It is set to
kernel_sizeby default.- padding
(int or tuple): Padding that was added to the input
Note
MaxPool2d can map several input sizes to the same output
sizes. Hence, the inversion process can get ambiguous.
To accommodate this, you can provide the needed output size
as an additional argument output_size in the forward call.
See the Inputs and Example below.
Inputs
input: the input Tensor to invertindices: the indices given out bynn_max_pool2d()output_size(optional): the targeted output size
Shape
Input: \((N, C, H_{in}, W_{in})\)
Output: \((N, C, H_{out}, W_{out})\), where $$ H_{out} = (H_{in} - 1) \times \mbox{stride[0]} - 2 \times \mbox{padding[0]} + \mbox{kernel\_size[0]} $$ $$ W_{out} = (W_{in} - 1) \times \mbox{stride[1]} - 2 \times \mbox{padding[1]} + \mbox{kernel\_size[1]} $$ or as given by
output_sizein the call operator
Examples
if (torch_is_installed()) {
pool <- nn_max_pool2d(2, stride = 2, return_indices = TRUE)
unpool <- nn_max_unpool2d(2, stride = 2)
input <- torch_randn(1, 1, 4, 4)
out <- pool(input)
unpool(out[[1]], out[[2]])
# specify a different output size than input size
unpool(out[[1]], out[[2]], output_size = c(1, 1, 5, 5))
}
#> torch_tensor
#> (1,1,.,.) =
#> 0.0000 0.0000 0.0000 0.0000 0.0000
#> 1.6753 0.6431 0.0000 0.0000 0.0000
#> 0.0000 0.0000 1.0992 0.0000 1.7359
#> 0.0000 0.0000 0.0000 0.0000 0.0000
#> 0.0000 0.0000 0.0000 0.0000 0.0000
#> [ CPUFloatType{1,1,5,5} ]