- 説明:
D4RL は、オフライン強化学習用のオープンソース ベンチマークです。トレーニングおよびベンチマーク アルゴリズム用の標準化された環境とデータセットを提供します。
データセットはRLDS 形式に従って、ステップとエピソードを表します。
構成の説明: タスクとそのバージョンの詳細については、 https://github.com/rail-berkeley/d4rl/wiki/Tasks#adroitを参照してください。
バージョン:
-
1.0.0: 初期リリース。 -
1.1.0(デフォルト): is_last を追加しました。
-
監視キー(
as_superviseddocを参照):None図( tfds.show_examples ): サポートされていません。
引用:
@misc{fu2020d4rl,
title={D4RL: Datasets for Deep Data-Driven Reinforcement Learning},
author={Justin Fu and Aviral Kumar and Ofir Nachum and George Tucker and Sergey Levine},
year={2020},
eprint={2004.07219},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
d4rl_adroit_door/v0-human (デフォルト設定)
ダウンロードサイズ:
2.97 MiBデータセットのサイズ:
3.36 MiB自動キャッシュ(ドキュメント): はい
分割:
| スプリット | 例 |
|---|---|
'train' | 50 |
- 機能の構造:
FeaturesDict({
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float32,
'infos': FeaturesDict({
'qpos': Tensor(shape=(30,), dtype=float32),
'qvel': Tensor(shape=(30,), dtype=float32),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float32),
'reward': float32,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float32 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/情報/qpos | テンソル | (30,) | float32 | |
| ステップ/情報/qvel | テンソル | (30,) | float32 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float32 | |
| 歩数/報酬 | テンソル | float32 |
- 例( tfds.as_dataframe ):
d4rl_adroit_door/v0-cloned
ダウンロードサイズ:
602.42 MiBデータセットのサイズ:
497.47 MiB自動キャッシュ(ドキュメント): いいえ
分割:
| スプリット | 例 |
|---|---|
'train' | 6,214 |
- 機能の構造:
FeaturesDict({
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float64,
'infos': FeaturesDict({
'qpos': Tensor(shape=(30,), dtype=float64),
'qvel': Tensor(shape=(30,), dtype=float64),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float64),
'reward': float64,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float64 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/情報/qpos | テンソル | (30,) | float64 | |
| ステップ/情報/qvel | テンソル | (30,) | float64 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float64 | |
| 歩数/報酬 | テンソル | float64 |
- 例( tfds.as_dataframe ):
d4rl_adroit_door/v0-expert
ダウンロードサイズ:
511.05 MiBデータセットのサイズ:
710.30 MiB自動キャッシュ(ドキュメント): いいえ
分割:
| スプリット | 例 |
|---|---|
'train' | 5,000 |
- 機能の構造:
FeaturesDict({
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float32,
'infos': FeaturesDict({
'action_logstd': Tensor(shape=(28,), dtype=float32),
'action_mean': Tensor(shape=(28,), dtype=float32),
'qpos': Tensor(shape=(30,), dtype=float32),
'qvel': Tensor(shape=(30,), dtype=float32),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float32),
'reward': float32,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float32 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/infos/action_logstd | テンソル | (28,) | float32 | |
| ステップ/情報/アクション_平均 | テンソル | (28,) | float32 | |
| ステップ/情報/qpos | テンソル | (30,) | float32 | |
| ステップ/情報/qvel | テンソル | (30,) | float32 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float32 | |
| 歩数/報酬 | テンソル | float32 |
- 例( tfds.as_dataframe ):
d4rl_adroit_door/v1-human
ダウンロードサイズ:
2.98 MiBデータセットのサイズ:
3.42 MiB自動キャッシュ(ドキュメント): はい
分割:
| スプリット | 例 |
|---|---|
'train' | 25 |
- 機能の構造:
FeaturesDict({
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float32,
'infos': FeaturesDict({
'door_body_pos': Tensor(shape=(3,), dtype=float32),
'qpos': Tensor(shape=(30,), dtype=float32),
'qvel': Tensor(shape=(30,), dtype=float32),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float32),
'reward': float32,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float32 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/情報/door_body_pos | テンソル | (3,) | float32 | |
| ステップ/情報/qpos | テンソル | (30,) | float32 | |
| ステップ/情報/qvel | テンソル | (30,) | float32 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float32 | |
| 歩数/報酬 | テンソル | float32 |
- 例( tfds.as_dataframe ):
d4rl_adroit_door/v1-cloned
ダウンロードサイズ:
280.72 MiBデータセットのサイズ:
1.85 GiB自動キャッシュ(ドキュメント): いいえ
分割:
| スプリット | 例 |
|---|---|
'train' | 4,358 |
- 機能の構造:
FeaturesDict({
'algorithm': string,
'policy': FeaturesDict({
'fc0': FeaturesDict({
'bias': Tensor(shape=(256,), dtype=float32),
'weight': Tensor(shape=(39, 256), dtype=float32),
}),
'fc1': FeaturesDict({
'bias': Tensor(shape=(256,), dtype=float32),
'weight': Tensor(shape=(256, 256), dtype=float32),
}),
'last_fc': FeaturesDict({
'bias': Tensor(shape=(28,), dtype=float32),
'weight': Tensor(shape=(256, 28), dtype=float32),
}),
'nonlinearity': string,
'output_distribution': string,
}),
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float32,
'infos': FeaturesDict({
'door_body_pos': Tensor(shape=(3,), dtype=float32),
'qpos': Tensor(shape=(30,), dtype=float32),
'qvel': Tensor(shape=(30,), dtype=float32),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float32),
'reward': float32,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| アルゴリズム | テンソル | 弦 | ||
| ポリシー | 特徴辞書 | |||
| ポリシー/fc0 | 特徴辞書 | |||
| ポリシー/fc0/バイアス | テンソル | (256,) | float32 | |
| ポリシー/fc0/重み | テンソル | (39,256) | float32 | |
| ポリシー/fc1 | 特徴辞書 | |||
| ポリシー/FC1/バイアス | テンソル | (256,) | float32 | |
| ポリシー/fc1/重み | テンソル | (256、256) | float32 | |
| ポリシー/last_fc | 特徴辞書 | |||
| ポリシー/last_fc/バイアス | テンソル | (28,) | float32 | |
| ポリシー/last_fc/weight | テンソル | (256、28) | float32 | |
| ポリシー/非線形性 | テンソル | 弦 | ||
| ポリシー/出力_配布 | テンソル | 弦 | ||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float32 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/情報/door_body_pos | テンソル | (3,) | float32 | |
| ステップ/情報/qpos | テンソル | (30,) | float32 | |
| ステップ/情報/qvel | テンソル | (30,) | float32 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float32 | |
| 歩数/報酬 | テンソル | float32 |
- 例( tfds.as_dataframe ):
d4rl_adroit_door/v1-expert
ダウンロードサイズ:
511.22 MiBデータセットのサイズ:
803.48 MiB自動キャッシュ(ドキュメント): いいえ
分割:
| スプリット | 例 |
|---|---|
'train' | 5,000 |
- 機能の構造:
FeaturesDict({
'algorithm': string,
'policy': FeaturesDict({
'fc0': FeaturesDict({
'bias': Tensor(shape=(32,), dtype=float32),
'weight': Tensor(shape=(32, 39), dtype=float32),
}),
'fc1': FeaturesDict({
'bias': Tensor(shape=(32,), dtype=float32),
'weight': Tensor(shape=(32, 32), dtype=float32),
}),
'last_fc': FeaturesDict({
'bias': Tensor(shape=(28,), dtype=float32),
'weight': Tensor(shape=(28, 32), dtype=float32),
}),
'last_fc_log_std': FeaturesDict({
'bias': Tensor(shape=(28,), dtype=float32),
'weight': Tensor(shape=(28, 32), dtype=float32),
}),
'nonlinearity': string,
'output_distribution': string,
}),
'steps': Dataset({
'action': Tensor(shape=(28,), dtype=float32),
'discount': float32,
'infos': FeaturesDict({
'action_log_std': Tensor(shape=(28,), dtype=float32),
'action_mean': Tensor(shape=(28,), dtype=float32),
'door_body_pos': Tensor(shape=(3,), dtype=float32),
'qpos': Tensor(shape=(30,), dtype=float32),
'qvel': Tensor(shape=(30,), dtype=float32),
}),
'is_first': bool,
'is_last': bool,
'is_terminal': bool,
'observation': Tensor(shape=(39,), dtype=float32),
'reward': float32,
}),
})
- 機能ドキュメント:
| 特徴 | クラス | 形 | Dタイプ | 説明 |
|---|---|---|---|---|
| 特徴辞書 | ||||
| アルゴリズム | テンソル | 弦 | ||
| ポリシー | 特徴辞書 | |||
| ポリシー/fc0 | 特徴辞書 | |||
| ポリシー/fc0/バイアス | テンソル | (32,) | float32 | |
| ポリシー/fc0/重み | テンソル | (32、39) | float32 | |
| ポリシー/fc1 | 特徴辞書 | |||
| ポリシー/FC1/バイアス | テンソル | (32,) | float32 | |
| ポリシー/fc1/重み | テンソル | (32, 32) | float32 | |
| ポリシー/last_fc | 特徴辞書 | |||
| ポリシー/last_fc/バイアス | テンソル | (28,) | float32 | |
| ポリシー/last_fc/weight | テンソル | (28、32) | float32 | |
| ポリシー/last_fc_log_std | 特徴辞書 | |||
| ポリシー/last_fc_log_std/bias | テンソル | (28,) | float32 | |
| ポリシー/last_fc_log_std/weight | テンソル | (28、32) | float32 | |
| ポリシー/非線形性 | テンソル | 弦 | ||
| ポリシー/出力_配布 | テンソル | 弦 | ||
| ステップ | データセット | |||
| ステップ/アクション | テンソル | (28,) | float32 | |
| 歩数/割引 | テンソル | float32 | ||
| 手順/情報 | 特徴辞書 | |||
| ステップ/infos/action_log_std | テンソル | (28,) | float32 | |
| ステップ/情報/アクション_平均 | テンソル | (28,) | float32 | |
| ステップ/情報/door_body_pos | テンソル | (3,) | float32 | |
| ステップ/情報/qpos | テンソル | (30,) | float32 | |
| ステップ/情報/qvel | テンソル | (30,) | float32 | |
| ステップ/is_first | テンソル | ブール | ||
| ステップ/is_last | テンソル | ブール | ||
| ステップ/is_terminal | テンソル | ブール | ||
| ステップ/観察 | テンソル | (39,) | float32 | |
| 歩数/報酬 | テンソル | float32 |
- 例( tfds.as_dataframe ):