Dynamically freezing parts of net during training

[Swami]

I am wondering if it is possible (using pycaffe) to dynamically set/modify lr_mult parameters for different layers during training. Any help would be really appreciated.