openai / baselines-results Goto Github PK

I was hoping to run through the acktr_ppo_acer_a2c_atari notebook. One of the first steps is to load results from ./atari50.

Would you be able to add that path to the repository, or provide a mechanism for downloading it like the dqn_results one does?

Hey,
Is there any chance getting the code to create the graphs in the param-noise directory (as done in the jupyter notebook)? I want to use the results for comparison, but I need the data itself and not the graphs.

Thanks in advance!

It can help a lot because my computational resources are not enough for reproducing all of the deepq results by myself.

It would be good to have Mujoco baseline results for ACKTR, A2C, TRPO , PPO and DDPG for Mujoco v2 environment update. It would be good to have benchmark results compare against.

Just wondering if the notebook files are meant to be reproducible.
If that's the case could you please provide any link to the tinkerbell dependency in https://github.com/openai/baselines-results/blob/master/acktr_ppo_acer_a2c_atari.ipynb as I'm not able to find any public package with that name.
Many thanks in advance.

The evaluation results provided in the "dqn_result.pkl" only contain 19 games. Does the rest of the game's training process has no record file? If someone have, please release it.THANK YOU VERY MUCH

experiment = "atari-prior-a"  
game_data = defaultdict(lambda: [])
for run_name, data in run_to_episode_data.items():
    if run_name.startswith(experiment):
        game = data['env_id'][:-len('NoFrameskip-v3')]
        t = np.cumsum(data['episode_data']["episode_lengths"])
        r = np.array(data['episode_data']["episode_rewards"])
        # Ensure all mesurements after the deadline of 2e8 are thrown away
        t_fltr, r_fltr = t[t < MAX_TSTEPS], r[t < MAX_TSTEPS]
        game_data[game].append((t_fltr, r_fltr))
print(len(game_data.keys()))
# 19