So this bit of code has been working fine for me the last time I ran it… but has a recent update to torch caused the FID score to blow up ? I am a bit suspicious of this warning:
UserWarning: The torch.cuda.*DtypeTensor constructors are no longer recommended. It's best to use methods such as torch.tensor(data, dtype=*, device='cuda') to create tensors.
So whenever I am running the training process, I get:
<ipython-input-8-6bc3f7d49a34> in calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps)
32 if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
33 m = np.max(np.abs(covmean.imag))
---> 34 raise ValueError('Imaginary component {}'.format(m))
35 covmean = covmean.real
36
ValueError: Imaginary component 1.172166351981351e+117
This is just an example provided by my class (which I haven’t changed) but I am not sure how to go about debugging this especially since I have limited GPU usage on colab. Any ideas? Thanks!