[BUG] ProbabilisticTensorDictModule._dist_sample hasattr error

[olliepro]

Describe the bug

The following code section is logically flawed. Bug was introduced here

        elif interaction_type is InteractionType.MEAN:
            if hasattr(dist, "mean"):
                try:
                    return dist.mean
                except NotImplementedError:
                    pass
            if dist.has_rsample:
                return dist.rsample((self.n_empirical_estimate,)).mean(0)
            else:
                return dist.sample((self.n_empirical_estimate,)).mean(0)

The hasattr attempts to access the dist.mean attribute to assess its existence. If the dist.mean raises a NotImplementedError, then hasattr(dist,'mean') will raise a NotImplemented error. The subsequent try block is placed in the wrong spot to catch such an error.

To Reproduce

from tensordict.nn import ProbabilisticTensorDictModule, InteractionType
from torchrl.modules import TanhNormal

prob_module = ProbabilisticTensorDictModule(
     in_keys=["loc", "scale"],
     out_keys=["action"],
     distribution_class=TanhNormal,
     distribution_kwargs={
         "low": -1,
         "high": 1,
     },
     return_log_prob=True,
)

prob_module._dist_sample(
    dist=TanhNormal(low=-1, high=1, loc=0, scale=1),
    interaction_type=InteractionType.MEAN
)

Expected behavior

The code should correctly catch the NotImplementedError exception and then estimate the mean

Suggestion

        elif interaction_type is InteractionType.MEAN:
            try:
                return dist.mean
            except AttributeError, NotImplementedError:
                if dist.has_rsample:
                    return dist.rsample((self.n_empirical_estimate,)).mean(0)
                else:
                    return dist.sample((self.n_empirical_estimate,)).mean(0)

System info

Python 3.10
torch==2.5.1+cu121
torchrl==0.6.0
tensordict==0.6.2

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[vmoens]

Solved by #1212