What is the agent_load_dir for l2o_train.py?

Open YiruS opened this issue 3 years ago • 2 comments

Hi there,

Nice work and thanks for sharing the code! One question: in both l2o_train.py and l2o_train_seg.py it requires the agent_load_dir. What is this pretrained policy parameters coming from? Is it a necessary or good to have?

Thanks!

Oct 20 '22 21:10 YiruS

Hi @YiruS! Thanks for your interest in our work!

I just updated our README. You may follow here to download the pretrained L2O policy and config the dir.

The L2O policy is good to have to not a must.

Oct 24 '22 00:10 chenwydj

Hi @YiruS

No, it should be model_old. We want the features_new['layer4'] can still be used by the old FC layer to make predictions similar to the old predictions, i.e. to minimize the KL_div.

You may check our Figure 2 for a better understanding. Thanks!

Oct 24 '22 05:10 chenwydj