Observations still being compressed with LZ4 despite compress_observations: False in config

Hello everyone,

I am currently using RLlib for training a PPO model on the CartPole-v1 environment. I have set the compress_observations flag in the configuration file to False as shown below:

cartpole-ppo:
    env: CartPole-v1
    run: PPO
    stop:
        episode_reward_mean: 150
        timesteps_total: 100000
    config:
        framework: torch
        gamma: 0.99
        lr: 0.0003
        num_workers: 1
        observation_filter: MeanStdFilter
        num_sgd_iter: 6
        vf_loss_coeff: 0.01
        model:
            fcnet_hiddens: [32]
            fcnet_activation: linear
            vf_share_layers: true
        enable_connectors: true
        output: "/path/to/folder" 
        compress_observations: False

However, I have noticed that the observations are still being compressed using LZ4 compression. I was expecting the observations to be stored uncompressed when setting compress_observations to False . Can anyone help me understand why the observations are still being compressed and how to disable the LZ4 compression for observations? Any help would be greatly appreciated. Thank you!

If you want to disable LZ4 compression for your observations in the output data when using RLlib, you can set the compress_observations option to False and add the output_compress_columns option with an empty set in your configuration file. Here’s the updated configuration file:

cartpole-ppo:
    env: CartPole-v1
    run: PPO
    stop:
        episode_reward_mean: 150
        timesteps_total: 100000
    config:
        framework: torch
        gamma: 0.99
        lr: 0.0003
        num_workers: 1
        observation_filter: MeanStdFilter
        num_sgd_iter: 6
        vf_loss_coeff: 0.01
        model:
            fcnet_hiddens: [32]
            fcnet_activation: linear
            vf_share_layers: true
        enable_connectors: true
        output: "/path/to/folder" 
        compress_observations: False
        output_compress_columns: set()

By setting compress_observations to False and output_compress_columns to an empty set (set() ), you ensure that the observations in the output data are not compressed with LZ4. This should result in uncompressed observations in the output data.