Bug in plot_writer_data

The following code:

from rlberry.agents import UCBVIAgent
from rlberry.envs import GridWorld
from rlberry.manager import AgentManager, plot_writer_data

manager = AgentManager(
    UCBVIAgent,
    (GridWorld, {}),
    fit_budget=10,
    output_dir='rlberry_data/ucbvi_experiment',
    outdir_id_style=None
)
manager.fit()
plot_writer_data(manager, tag='episode_rewards')

raises the error:

ValueError: input dir not found, verify that the agent are trained and that AgentManager.outdir_id_style="timestamp"

Since the agent manager is given to plot_writer_data, we should not care about outdir_id_style, all we need is already in the manager.

Also, this leads to a completely unexpected behavior if the code above is previously called with outdir_id_style="timestamp", since we’ll plot the data from the previous run, instead of data in the current manager.

[TimotheeMathieu]

Ok for me. I can do it.

[omardrwch]

I think we could distinguish two cases when we have a path:

• If input_obj is a path to a .pickle file, we load the file as an AgentManager (which should allow us to load older experiments);
• If input_obj is a folder, we take the .pickle file in the most recent subfolder (either by looking at the timestamp in the folder name, or by looking at its creation time).

What do you think?