Could anyone reproduce the rFID scores from the small-large enc-dec version?

[shashankg7]

Hi,

I am running the model on ImageNet data from scratch, using the best config from the paper (small encoder and large decoder), training on a cluster of 56 A100 GPUs for more than a week's time, but the rFID (reconstruction FID on validation) I am getting so far is around 21, which is far from the number reported in the paper.

Could anyone get close to the rFID reported in the paper (around 1.6)?