FID calculation resaves real images everytime about lightweight-gan

[Dok11]

oh, what do you mean by this?

PNG files have too large size, up to 1 MB per 1024×1024px. But this size dont justified to use in FID. Jpeg will give same metric on FID but with 10 times less size, about 100 KB per same image.

[lucidrains]

@cr458 Hi Christopher! Someone else actually wrote that code originally, so this is news to me.

Reflecting on how it is written, I think it would be more efficient to copy N images over from the data directory. I'll see what I can do tomorrow

[Dok11]

Why we need to copy files if we know where they places? Just use some count of known dataset, easy :D

[lucidrains]

@Dok11 ah got it! thanks for clarifying :)

[Dok11]

And one more real_path + '{}.png' png uses more space on HDD but to calc FID this format does not have any usable info

[lucidrains]

Why we need to copy files if we know where they places? Just use some count of known dataset, easy :D

oh yes, of course 🤦 lol i'll do just that

[lucidrains]

And one more real_path + '{}.png' png uses more space on HDD but to calc FID this format does not have any usable info

oh, what do you mean by this?

[lucidrains]

@cr458 Hi Christopher! Let me know if https://github.com/lucidrains/lightweight-gan/releases/tag/0.16.4 works for you!

[cr458]

@lucidrains looks like it's working fine!

[cr458]

@lucidrains looking at the pytorch-fid code all we need to do is save the mean and covariance matrix of the inception embeddings of the real images.

so here's what I'd suggest:

1. Calculate Inception activations once for the real images, using the whole dataset and the calculate_activation_statistics from here store the them as npz files in the fid_real directory

2. In subsequent FID calculations, use these stored values, and only save the generated images

This should speed up the FID calculation even more, what do you think?