How to generate the npz files for the custom reference dataset?

[ChongjianGE]

I notice that the the npz file for the reference images mainly include some information of
mu
sigma
mu_s
sigma_s
mu_clip
sigma_clip
arr_0
Since I am currently considering on generating new reference npz files on my custom dataset, could anyone explain how to generate the above information on a new set of reference images?

[Fushennn]

I am thinking about it, too. Has it been solved yet?

[TaylorJocelyn]

I’m experiencing a similar issue. May I ask if you were able to resolve it?

[ZachL1]

If the FID statistics haven't been calculated for the provided reference batch, the evaluation script will automatically compute them.

see line 172:

guided-diffusion/evaluations/evaluator.py

Lines 167 to 172 in 22e0df8

	obj = np.load(npz_path)
	if "mu" in list(obj.keys()):
	return FIDStatistics(obj["mu"], obj["sigma"]), FIDStatistics(
	obj["mu_s"], obj["sigma_s"]
	)
	return tuple(self.compute_statistics(x) for x in activations)

[escorciav]

when dumping your own data use np.savez or np.savez_compressed

for details, refer to:

NpzArrayReader, BatchIterator, StreamingNpzArrayReader, open_npz_array, _open_npy_file