Bad training performance
CharlesPeng1998 opened this issue · 3 comments
Hi. I ran your training script for a few days and I got following results.
The agent seemed to learn something at about 1k step. Then the training curve degraded and converged at bad performance. I am wondering what may cause this to happen? Could you tell me how your pre-trained model was trained?
Sorry for the late reply. Did you load pretrained weight from here https://github.com/carla-simulator/imitation-learning before training? I just used this pretrained weight to extract feature and launch ddpg_main.py.
Yes. I think the pretrained model was loaded as the printed log showed:
Restoring from /media/autolab/disk_2T/jiancongpeng/repos/rl_CARLA/environment/imitation/model/model.ckpt-450000
INFO:tensorflow:Restoring parameters from /media/autolab/disk_2T/jiancongpeng/repos/rl_CARLA/environment/imitation/model/model.ckpt-450000
INFO:tensorflow:Restoring parameters from /media/autolab/disk_2T/jiancongpeng/repos/rl_CARLA/environment/imitation/model/model.ckpt-450000
mmm, not sure. Have you tried to train it again, could be caused by random parameters? I don't have a specific suggestion..