search

med-air / DEX

[ICRA'23] Demonstration-Guided Reinforcement Learning with Efficient Exploration for Task Automation of Surgical Robot
MIT License
32 stars 4 forks source link

wandb problem #10

Closed kidd12138 closed 2 months ago

kidd12138 commented 2 months ago

I have created a wandb account and changed the wandb entity in train.yaml before running the command,but it appears the error below.Is there anything else should be modified?

屏幕截图 2024-07-21 104512

TaoHuang13 commented 2 months ago

Hi. I guess this is because the configuration parser views your wandb account as an integer (384684286) rather than a string. You should force a type transfer (str(x)) before wandb.init().

kidd12138 commented 2 months ago

Probably.I tried to create another wandb account with no integer and it didn't happen again.Thanks for the advice.And there is another question,I see there are quite a few parameters uploaded on the wandb,what parameters should be compared to evaluate the effectiveness of RL methods?I don't know if it's too simple,I'm still in the process of learning it.

TaoHuang13 commented 2 months ago

Glad to hear that. Regarding evaluating RL, the most direct metric is the evaluated episode reward (task success rate in our tasks). Other metrics, such as critic loss or policy loss, are somehow meaningless. It also depends on the RL algorithm you choose.