这个定理基本上可以表述为:
F(f*g) = F(f)xF(g)
我知道这个定理,但是我无法使用PyTorch重现结果。
以下是可重现的代码:
import torch
import torch.nn.functional as F
# calculate f*g
f = torch.ones((1,1,5,5))
g = torch.tensor(list(range(9))).view(1,1,3,3).float()
conv = F.conv2d(f, g, bias=None, padding=2)
# calculate F(f*g)
F_fg = torch.rfft(conv, signal_ndim=2, onesided=False)
# calculate F x G
f = f.squeeze()
g = g.squeeze()
# need to pad into at least [w1+w2-1, h1+h2-1], which is 7 in our case.
size = f.size(0) + g.size(0) - 1
f_new = torch.zeros((7,7))
g_new = torch.zeros((7,7))
f_new[1:6,1:6] = f
g_new[2:5,2:5] = g
F_f = torch.rfft(f_new, signal_ndim=2, onesided=False)
F_g = torch.rfft(g_new, signal_ndim=2, onesided=False)
FxG = torch.mul(F_f, F_g)
print(FxG - F_fg)
这里是 print(FxG - F_fg) 的结果
tensor([[[[[ 0.0000e+00, 0.0000e+00],
[ 4.1426e+02, 1.7270e+02],
[-3.6546e+01, 4.7600e+01],
[-1.0216e+01, -4.1198e+01],
[-1.0216e+01, -2.0223e+00],
[-3.6546e+01, -6.2804e+01],
[ 4.1426e+02, -1.1427e+02]],
...
[[ 4.1063e+02, -2.2347e+02],
[-7.6294e-06, 2.2817e+01],
[-1.9024e+01, -9.0105e+00],
[ 7.1708e+00, -4.1027e+00],
[-2.6739e+00, -1.1121e+01],
[ 8.8471e+00, 7.1710e+00],
[ 4.2528e+01, 9.7559e+01]]]]])
您可以看到,差异并不总是为0。
有人能告诉我为什么,并且如何正确地做这件事吗?
谢谢。
F_g = torch.rfft(g_new.flip(0).flip(1), ...
,这应该可以让您更接近结果。由于DFT假定信号是周期性的(对于傅里叶变换是离散的必要条件),因此可能还存在一些填充差异。我稍后会验证这一点。 - jodag