parameters are missing in the pretrained weights

[dingquanyu]

Hi,

I've downloaded the pretrained weights and trying to run inference using the multimer.unifold.pt but it crashed because aux_heads.experimentally_resolved.linear.weight and aux_heads.experimentally_resolved.linear.bias are not in the state_dict. I tried both multimer_af2 and multimer_ft but both failed.

[dingquanyu]

Hi,

I've also tried the sourced codes from release 2.0.0 and 2.1.0 but still got the same error. In all cases I only run the run_unifold.sh using multimer.unifold.pt

[ZiyaoLi]

I don't meet your error, but according to your error msg, you may manually close the experimentally resolved head in config.py. seems that your ckp is not trained with that head.

[dingquanyu]

Hi,

Thanks but I didn't got this error from training. It occurred when I freshly cloned this repo and run run_unifold.sh without any modifications.

[ZiyaoLi]

this seems unlikely as you reported the error using model name multimer_ft, but in multimer_ft we do not enable the experimentally resolved head. see

Uni-Fold/unifold/config.py

Line 588 in 1a30171

elif name == "multimer_ft":