Loading policy from persitence dir
Author: AnsgarKasselmannTUMCreated Apr 14, 2026Updated Apr 14, 2026
Hey, I'm encountering issues when loading a policy from the persistence dir. My current approach is
experiment = Experiment.from_directory(experiment_dir.__str__(), restore_policy=True)
world = experiment.create_experiment_world(raise_error_on_dirname_collision=False)However, this results in the following error, due to the changes in torchs default model loading:
_pickle.UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, do those steps only if you trust the source of the checkpoint.
(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
WeightsUnpickler error: Unsupported global: GLOBAL tianshou.algorithm.modelfree.sac.SAC was not an allowed global by default. Please use `torch.serialization.add_safe_globals([tianshou.algorithm.modelfree.sac.SAC])` or the `torch.serialization.safe_globals([tianshou.algorithm.modelfree.sac.SAC])` context manager to allowlist this global if you trust this class/function.Is this behavior intentional? I think whitelisting ever class, as suggested in the error message, is not really feasible. My current workaround it to set the env variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1. However, this doesn't feel like the proper way either.
Source: thu-ml/tianshou