d3rlpy-benchmarks

This repository contains the benchmark results of d3rlpy.

Library repository: https://github.com/takuseno/d3rlpy

baselines

The baselines directory contains the training results of the official implementations. You can see the summary of the results at baseline_table.csv.

reproductions

D4RL

The d4rl directory contains the training results of the d3rlpy's benchmark scripts with D4RL. You can see the summary of the results at d4rl_table.csv.

Atari 2600

The atari directory contains the training results of the d3rlpy's benchmark scripts with Atari 2600 dataset. You can see the summary of the results at atari_table.csv.

Finetuning

The finetuning directory contains the finetuning results of the d3rlpy's benchmark scripts with AntMaze datasets. You can see the summary of the results at finetuning_table.csv.

analysis tools

installation

$ git clone --depth 1 https://github.com/takuseno/d3rlpy-benchmarks
$ cd d3rlpy-benchmarks
$ pip install -e .

library

This repository provides lightweight analysis tools for researchers to conduct the further analysis. Here is the example snippet:

import matplotlib.pyplot as plt
from d3rlpy_benchmarks.data_loader import load_d4rl_score
from d3rlpy_benchmarks.plot_utils import plot_score_curve

score = load_d4rl_score("CQL", "hopper", "medium-v0")
plot_score_curve(score, window_size=100)
plt.show()

There are many ready-to-go analysis scripts in scripts/analysis directory.

$ python scripts/analysis/d4rl/plot_curve.py --env hopper --dataset medium-v0 --window 100

coding style

$ pip install black isort mypy
$ python scripts/utils/static_check.py