I am trying to convert a numpy implementation of DFT using numpy.fft.fft2 to all PyTorch. When compairing torch.fft, the results are not the same. Why is this?
Numpy
>>> np.fft.fft2(test_cp, axes=(1,2)).shape
(100, 32, 32, 16)
>>> np.fft.fft2(test_cp, axes=(1,2))[0][0][0]
array([ 121.36078 +0.j, 862.5427 +0.j, -478.19095 +0.j, -3.937337+0.j,
124.33441 +0.j, 42.13216 +0.j, 401.703 +0.j, 43.894173+0.j,
-155.7095 +0.j, 513.4581 +0.j, -119.220665+0.j, -166.82913 +0.j,
-954.76184 +0.j, -401.38803 +0.j, -189.80426 +0.j, -231.5777 +0.j],
dtype=complex64)
>>> abs(np.fft.fft2(test_cp, axes=(1,2))[0][0][0]) # to compare directly to real numbers in torch
array([121.36078 , 862.5427 , 478.19095 , 3.937337, 124.33441 ,
42.13216 , 401.703 , 43.894173, 155.7095 , 513.4581 ,
119.220665, 166.82913 , 954.76184 , 401.38803 , 189.80426 ,
231.5777 ], dtype=float32)
Pytorch
Looking at the real numbers from the output of (*, 2)
>>> torch.fft(layer_acts, signal_ndim=2).shape
torch.Size([100, 32, 32, 16, 2])
>>> torch.fft(layer_acts, signal_ndim=2)[:, :,:,:,0][0][0][0]
tensor([ 6.9696, -41.5982, -144.1784, -54.1054, -2.1185, -78.1295,
72.2305, 107.3909, 172.7782, 107.3909, 72.2305, -78.1295,
-2.1185, -54.1054, -144.1784, -41.5982], grad_fn=<SelectBackward>)
What makes the results different in the way numpy computes compared to torch?
Thanks!