SlidingWindow¶

class SlidingWindow(kernel_size, padding=0, stride=1, dilation=1, **kwargs)[source]¶

Apply a sliding window to input tensor and copy content in the window to corresponding output location. Assume input shape is \((N, C, IH, IW)\), then output shape would be \((N, C, OH, OW, window_h, window_w)\) where \((OH, OW)\) would be computed from padding, stride, window and \((IH, IW)\), as in convolution. For each output location, we have;

\[\begin{split}out_{n, c, oh, ow, wh, ww} &= src_{n, c, ih+wh, iw+ww} \\ \text{where } & ih=-pad_h+oh \times stride_h + (wh-1) \times (dilation_h-1) \\ & iw=-pad_w+ow \times stride_w + (ww-1) \times (dilation_w-1)\end{split}\]

Parameters

kernel_size (Union[int, Tuple[int, int]]) – the size of the window to take a max over.
padding (Union[int, Tuple[int, int]]) – implicit zero padding to be added on both sides. Default: 0
stride (Union[int, Tuple[int, int]]) – the stride of the window. Default: 1
dilation (Union[int, Tuple[int, int]]) – the dilation of the window. Default: 1

Example

>>> import numpy as np
>>> inp = Tensor(np.arange(30).reshape(1,1,5,6))
>>> op = M.SlidingWindow(kernel_size=3, padding=1, stride=2, dilation=2)
>>> out = op(inp)
>>> print(out.numpy())
[[[[[[ 0  0  0]
     [ 0  7  9]
     [ 0 19 21]]

    [[ 0  0  0]
     [ 7  9 11]
     [19 21 23]]]


   [[[ 0  7  9]
     [ 0 19 21]
     [ 0  0  0]]

    [[ 7  9 11]
     [19 21 23]
     [ 0  0  0]]]]]]

DeformableConv2d

SlidingWindowTranspose