All success rates are zero while training

[ssam2s]

Thanks for your great work !

Recently, I reproducing the experimental results with given codes.
My reproduction is conducted with your example codes in README.
There were no problem in skill discovery and clustering.

However, while training the meta controller, all evaluation metrics are zero.
I saw that some issues mentioned about this problem and they said it will be solved after enough training epochs.
But it is still zero after 50 epochs with LIBERO-OBJECT.

I wonder if i missed something while training.

Also, I found that there is a minor issue in script "lotus/skill_learning/multisensory_repr/dinov2_repr.py".
Your example codes of README are handle LIBERO-OBJECT dataset, but dinov2_repr.py script leverages LIBERO-SPATIAL.

[pengzhi1998]

Hi, thank you for your post! I'm wondering whether you have figured out how to solve the problem of the unexpected performance?

[ssam2s]

@pengzhi1998

Unfortunately, I haven't been able to figure out the exact cause and couldn't reproduce the results. However, I noticed an odd problem during the clustering process in the Skill Discovery Stage, where all but one or two of the skill cluster distances or scores were zero. I believe the issue might be related to that.

I hope this helps, and good luck with resolving it!

[pengzhi1998]

Got it. Thank you for your reply and help!!

[BrightMoonStar]

知道了。谢谢你的回复和帮助！！

HiHave you solved this problem? Thanks for your reply @pengzhi1998

[yfpeng1234]

Thanks for your post, I also encounter this issue. In addition to that, I also have two other questions:
1.in the agglomoration_script.py

I guess the 1-6 tasks are selected for pre-training and 7-10 are selected for fine-tuning. Then after running this code, the result look like this:

However, in the https://github.com/UT-Austin-RPL/Lotus/blob/master/lotus/script/skill_policy_libero_object.sh, the dinov2_libero_goal_image_only_6 folder is also needed. So I change the agglomeration_script.py like this:

2.when running https://github.com/UT-Austin-RPL/Lotus/blob/master/lotus/script/skill_policy_libero_object.sh, the terminal return a very strange message:

seems that skill2 and skill3 are not trained

[SmallStriker]

Thanks for your great work !

Recently, I reproducing the experimental results with given codes. My reproduction is conducted with your example codes in README. There were no problem in skill discovery and clustering.

However, while training the meta controller, all evaluation metrics are zero. I saw that some issues mentioned about this problem and they said it will be solved after enough training epochs. But it is still zero after 50 epochs with LIBERO-OBJECT.

I wonder if i missed something while training.

Also, I found that there is a minor issue in script "lotus/skill_learning/multisensory_repr/dinov2_repr.py". Your example codes of README are handle LIBERO-OBJECT dataset, but dinov2_repr.py script leverages LIBERO-SPATIAL.

So did you finally solve this problem about zero metrics during training? Thank you very much and look forward to your reply