Reproduction of precision and recall metrics

[see08-ai]

Hello, we are using adm_in256_stats.npz to reproduce the precision and recall indicators of MAR-L, and the results are 0.51 precision and 0.60 recall, which are different from 0.81 and 0.60 reported in the paper. Could you describe how the precision and recall results reported in the paper are calculated?

The following is our calculation code, setting prc to true：
metrics_dict = torch_fidelity.calculate_metrics( input1=save_folder, input2=input2, fid_statistics_file=fid_statistics_file, cuda=True, isc=True, fid=True, kid=False, prc=True, verbose=False, )

[LTH14]

The precision and recall metrics should be calculated using the 10000 reference images, instead of the fid statistics file. The reference images could be found here https://github.com/openai/guided-diffusion/tree/main/evaluations. You should store the images into one folder, and set input_1 to that folder, input_2 to the folder where you generate your 50000 images.