search

BradSegal / CXR_PGGAN

Source code and model weights for the PGGAN model utilised for the paper: Evaluating the Clinical Realism of Synthetic Chest X-Rays Generated Using Progressively Growing GANs
12 stars 9 forks source link

About FID calculation #3

Open lzh1-hub opened 2 years ago

lzh1-hub commented 2 years ago

Sorry to bother you, I am a college student, I am very interested in your work, your work is great. I would like to cite your paper《Evaluating the Clinical Realism of Synthetic Chest X‑Rays Generated Using Progressively Growing GANs》. I would like to ask how the FID is calculated in your paper? The FID I get for the medical image is very high, 130

BradSegal commented 2 years ago

Hi there,

Sure, I'll find and send the evaluation code I used for the FID calculations. I remember there's some variation dependent on the exact model weights used for the calculation and it is sample size dependent. In order to ensure adequate comparison between FIDs you need to ensure you use the original weights that the previous GAN works employed. Too small image sets also give wildly fluctuating FIDs which is to be expected as you need a large enough cohort to represent the average set of features.

lzh1-hub commented 2 years ago

Thank you very much, my email is 13657996227@163.com