ImageNet1k validation results for mixer_b16_224.miil_in21k_ft_in1k not matching the reported results.

[OCM-7898]

Hi, for a project of mine I was benchmarking some models from timm, I wanted to ask if the mixer_b16_224.miil_in21k_ft_in1k model reported in the results is correct. I am getting around 28% on the imagenet validation set. I ran some other models too those seem quite accurate but wondering if this is a weight issue with this certain model

[rwightman]

@OCM-7898 fuzzy memory, I think you need to validate those in fp32, mixed precision or low precision hurts them...