conv2d_fft

deepinv.physics.functional.conv2d_fft(x, filter, real_fft=True, padding='valid')

A helper function performing the 2d convolution of images x and filter using FFT.

The adjoint of this operation is deepinv.physics.functional.conv_transpose2d_fft()

Note

The convolution here is a convolution, not a correlation as in torch.nn.functional.conv2d(). This function gives the same result as deepinv.physics.functional.conv2d() and is faster for large kernels.

Parameters:

• x (torch.Tensor) – Image of size (B, C, W, H).

• filter (torch.Tensor) – Filter of size (b, c, w, h) where b can be either 1 or B and c can be either 1 or C. If b = 1 or c = 1, then this function supports broadcasting as the same as numpy. Otherwise, each channel of each image is convolved with the corresponding kernel.

• real_fft (bool) – for real filters and images choose True (default) to accelerate computation.

• padding (str) – can be 'valid', 'circular', 'replicate', 'reflect', 'constant' or 'zeros'. If padding = 'valid' the output is smaller than the image (no padding), otherwise the output has the same size as the image. Note that 'constant' and 'zeros' are equivalent. Default is 'valid'.

Returns:

torch.Tensor: the output of the convolution of the shape size as x.

Return type:

Tensor