Some questions

[xiaopoche1221]

Hello, I tried to reproduce your code, but an error occurred. I don't know what version of the package you are using. This error occurred:
“ImportError: cannot import name 'TanhNormal' from partially initialized module 'garage.torch.distributions.tanh_normal' ”
Which version of garage is it? It seems that the file is missing.

[navigator8972]

Hi, thanks for your interests. I think garage has some updates that broke legacy code.
Just checked my local garage and the commit is

5aecf5a163fca64841e89662f8e9d4939171ed7e

Please try to compile this version and let me know if that fix the error.

[xiaopoche1221]

Hi, thanks for your interests. I think garage has some updates that broke legacy code. Just checked my local garage and the commit is

5aecf5a163fca64841e89662f8e9d4939171ed7e

Please try to compile this version and let me know if that fix the error.

Thank you for your reply
I tried to find your garage version, but failed.
I went to the garage community to find out, but someone said that you need the torch version<1.8. Is your torch version 1.12?
The problem connection is this: rlworkgroup/garage#2312
In addition, I am running test_ koopman_ lqr
The problem encountered in the code is: RuntimeError: svd_ cpu: the updating process of SBDSDC did not converge (error: 24)

[navigator8972]

Hi, torch version 1.12 should not be mandatory for the RL experiments. The most recent test in test_koopman_lqr with 1.12 is trying to use newer pytorch API to solve matrix inversion, which appeared to be unstable before with float32. The RL experiments were not using this.