rlplot is an easy to use and highly encapsulated RL plot library (including basic error bar lineplot and a wrapper to rliable).

2023.10.19: Added some experimental data (see exp) to specify data format.

Installation

To install rlplot, please run:

git clone https://github.com/mantle2048/rlplot
cd rlplot
pip install -e .

Requirements

hydra-core >= 1.3.0
arch == 5.3.0
scipy >= 1.7.0
numpy >= 0.9.0
absl-py >= 1.16.4
seaborn >= 0.11.2

After installation, open your python console and type

import rlplot
print(rlplot.__version__)

If no error occurs, you have successfully installed rlplot.

Demo

Here, we provide an illustration on how to utilize this library for analysis and plot of your RL experiments.

0. Preliminary

Assum that you have four algorithms [D, C, B, A] and four tasks [HalfCheetah-v4, Ant-v4, Walker2d-v4, Humanoid-v4].

The experimental logs must be put in exps folder with the following file tree:

└── ${algo}_${task}(i.e. A_Ant-v4) # exp_name
    └── ${timestamp}_${algo}_${task}_${seed}(i.e. 2023-01-14_02-36-30_A_Ant-v4_101) # run_name, "timestamp" is optional
        └── progress.csv # experiment log, including metrics like "AverageEvalReward", "TotalEnvInteracts", "Epoch", etc

1. The create of diagnosis

Convert all progress.csv files in exps to yaml file and store them in diagnosis folder.

For instance, you can run the following command:

python main.py \
    type=create_diagnosis \
    n_epoch=200 \
    epoch_len=5000 \
    milestone_gap=100000 \
    metric='AverageEvalReward' \
    algos=['D','C','B','A'] \
    tasks=['HalfCheetah-v4','Ant-v4','Walker2d-v4','Humanoid-v4']

Alternatively, you can manually modify cfgs/config.yaml and cfgs/type/create_diagnosis.yaml and then just run make diag.

2. Plot figures

2.1. Metric curve

python main.py \
    type=metric_curve \
    n_epoch=200 \
    epoch_len=5000 \
    metric='AverageEvalReward' \
    aggregate_name='IQM' \
    smooth_size=1 \
    algos=['D','C','B','A'] \
    tasks=['HalfCheetah-v4','Ant-v4','Walker2d-v4','Humanoid-v4']

Across all tasks	HalfCheetah-v4	Ant-v4	Walker2d-v4	Humanoid-v4

2.2. Metric value

python main.py \
    type=metric_value \
    milestone='1m' \
    aggregate_names=['Mean','IQM'] \
    algos=['D','C','B','A'] \

2.3. Performance profiles

python main.py \
    type=performance_profiles \
    milestones=['300k','500k','1m'] \
    algos=['D','C','B','A']

2.4. Probability of improvement

python main.py \
    type=probability_of_improvement \
    milestone='1m' \
    algos=['D','C','B','A']

2.5. Sample efficiency curve

python main.py \
    type=sample_efficiency_curve \
    steps=[1,10,25,50,75,100,125,150,175,200] \
    algos=['D','C','B','A']

2.6. Overall ranks

python main.py \
    type=overall_ranks \
    milestones=['300k','500k','1m'] \
    algos=['D','C','B','A']

2.7. One command for all plot

For a "lazy" person :), you can modify the yaml config files in cfgs for your case , then just run make all.

Video

Trivia

The main code structure in rlplot is very simple, and most of it is inherited from rliable. You can completely modify your favorite painting style in plot_utils.py. All credit goes to the author of rliable. For more details about the specific meaning of above plots, please refer to the original library rliable.

mantle2048 / rlplot