[Question] Can I train or finetune a model in observation term?

[HerocatUED]

Question

I noticed that there is an observation term named image_features, but it keeps the model frozen. I wonder can we train or finetune it despite of some hard-code parameters such as mean or std. Can gradient successfully backwards if I just set the model at train mode? I'm trying but I might meet bugs written by myself. I need to ensure that set the model to train model is okay or not.