paddle.nn.functional. unfold ( x, kernel_sizes, strides=1, paddings=0, dilations=1, name=None ) [source]

Return a col buffer of sliding local blocks of input x, also known as im2col for batched 2D image tensors. For each block under the convolution filter, all element will be rearranged as a column. While the convolution filter sliding over the input feature map, a series of such columns will be formed.

For each input \(x\) with shape [N, C, H, W], the output shape [N, Cout, Lout] can be calculated as following.

\[ \begin{align}\begin{aligned}dkernel[0] &= dilations[0] \times (kernel\_sizes[0] - 1) + 1\\dkernel[1] &= dilations[1] \times (kernel\_sizes[1] - 1) + 1\\hout &= \frac{H + paddings[0] + paddings[2] - dkernel[0]}{strides[0]} + 1\\wout &= \frac{W + paddings[1] + paddings[3] - dkernel[1]}{strides[1]} + 1\\Cout &= C \times kernel\_sizes[0] \times kernel\_sizes[1]\\Lout &= hout \times wout\end{aligned}\end{align} \]
  • x (Tensor) – 4-D Tensor, input tensor of format [N, C, H, W], data type can be float32 or float64

  • kernel_sizes (int|list|tuple) – The size of convolution kernel, should be [k_h, k_w] or an integer k treated as [k, k].

  • strides (int|list|tuple, optional) – The strides, should be [stride_h, stride_w] or an integer stride treated as [sride, stride]. For default, strides will be [1, 1].

  • paddings (int|list|tuple, optional) – The paddings of each dimension, should be [padding_top, padding_left, padding_bottom, padding_right] or [padding_h, padding_w] or an integer padding. If [padding_h, padding_w] was given, it will expanded to [padding_h, padding_w, padding_h, padding_w]. If an integer padding was given, [padding, padding, padding, padding] will be used. For default, paddings will be [0, 0, 0, 0]

  • dilations (int|list|tuple, optional) – the dilations of convolution kernel, should be [dilation_h, dilation_w], or an integer dilation treated as [dilation, dilation]. For default, it will be [1, 1].

  • name (str, optional) – The default value is None. Normally there is no need for user to set this property. For more information, please refer to Name


Tensor, The tensor corresponding to the sliding local blocks. The output shape is [N, Cout, Lout] as decriabled above. Cout is the total number of values within each block, and Lout is the total number of such blocks. The data type of output is the same as the input \(x\)


>>> import paddle
>>> import paddle.nn.functional as F

>>> x = paddle.randn((100,3,224,224))
>>> y = F.unfold(x, [3, 3], 1, 1, 1)