[tune] experiment_analysis.py nan

[javadan]

I have a case where I can't get an Analysis object on some experiment directories.

It is looking for episode_reward_mean, but it is nan sometimes because it's an average.

When debugging, I get to the below, when looking for the best config or best log dir.
idx = df[metric].idxmax() makes idx 'nan'.

    def _retrieve_rows(self, metric=None, mode=None):
        assert mode is None or mode in ["max", "min"]
        rows = {}
        for path, df in self.trial_dataframes.items():
            if mode == "max":
                idx = df[metric].idxmax()
            elif mode == "min":
                idx = df[metric].idxmin()
            else:
                idx = -1
            rows[path] = df.iloc[idx].to_dict()

        return rows



    def get_best_config(self, metric, mode="max"):
        """Retrieve the best config corresponding to the trial.

        Args:
            metric (str): Key for trial info to order on.
            mode (str): One of [min, max].
        """
        rows = self._retrieve_rows(metric=metric, mode=mode)
        if not rows:
            # only nans encountered when retrieving rows
            logger.warning("Not able to retrieve the best config for {} "
                           "according to the specified metric "
                           "(only nans encountered).".format(
                               self._experiment_dir))
            return None
        all_configs = self.get_all_configs()
        compare_op = max if mode == "max" else min
        best_path = compare_op(rows, key=lambda k: rows[k][metric])
        return all_configs[best_path]

Could _retrieve_rows return a default value like None, rather than throwing an exception on the df.iloc[idx] ?

[amogkam]

@javadan what version of Ray is this? I believe this error has been fixed in this PR #9381. It will be included in the next 0.8.7 release. In the meantime you can install ray from the latest wheels (https://docs.ray.io/en/master/installation.html#latest-snapshots-nightlies)

[javadan]

Thanks, 0.8.6. I'll try out the wheels plan now.

[javadan]

Hmm so I updated to ray-0.9.0.dev0
and am getting some error in exploration.py and a few other files.
AttributeError: module 'gym.spaces' has no attribute 'Space'
Guess it's always a risk, running latest snapshots.
Do you think I need to roll back to a specific gym version maybe?
I think they must have changed their package names at some point, from gym to gym.spaces or vice versa.
How would I check what version of gym rllib is looking for?

[amogkam]

Good question, I think downgrading your gym version should work? @ericl @sven1977 any advice here?

[javadan]

I worked around for now, by globally replacing gym.spaces with gym
Good enough for now. Thanks.