Antonin RAFFIN

Thanks for the pointers @duburcqa =)! I see. In that case, it is already supported as SB3 exposes its PyTorch policy via the `.policy` attribute (this is a `nn.Module`, see...

I would actually welcome a PR that shows how to use `th.compile` in our doc =)

Hello, thanks for the proposal, I will try to take a look at it soon (my stack was quite last week). Could you add the missing imports so I could...

Thanks I could give it a quick try and it does work =) There seems to be a big overhead compared to `SubprocVecEnv` or `DummyVecEnv` (both at startup and during...

> Please note that the current version of Ray (1.5 and above) will automatically call ray. init() on the first use of a Ray remote API, which can cause the...

Hello, those gains are for orthogonal initialization only (https://pytorch.org/docs/stable/_modules/torch/nn/init.html#orthogonal_), when they are not used, the default pytorch initialization is used. The gains are from OpenAI Baselines, to keep results consistent,...

Hello, thanks for the proposal, I didn't know that this syntax worked. ```python # In the code, KEY_TO_GOAL_STRATEGY[my_strategy] # can be replaced by GoalSelectionStrategy[my_strategy.upper()] ``` I let @qgallouedec decide and...

Duplicate of https://github.com/DLR-RM/stable-baselines3/pull/1267 (although this is a different implementation)

Hello, what have you tried so far? Did you try adapting what is already in the documentation? https://stable-baselines3.readthedocs.io/en/master/guide/export.html#export-to-onnx You should also take a look at SB3 contrib code, to have...

> I didn't figure out some variables definitions, so the conversion is not successful so far！ could you provide what you have currently and the error messages? Please use markdown...