just-d4rl

Easy-to-use D4RL offline dataset loader, focused solely on downloading and providing D4RL datasets without dependencies on gym or gymnasium.

Key Features

• Downloads and provides D4RL offline datasets
• No dependencies on gym or gymnasium
• Lightweight and focused functionality
• Easy integration with PyTorch

Installation

Install from PyPI:

pip install just-d4rl

Usage

from just_d4rl import D4RLDataset, d4rl_offline_dataset, D4RLScoreNormalizer

# Download and load a D4RL dataset
d4rl_dataset = d4rl_offline_dataset("hopper-medium-v2")

# Example datasets
d4rl_dataset = d4rl_offline_dataset("walker2d-random-v2")
d4rl_dataset = d4rl_offline_dataset("halfcheetah-medium-expert-v2")
d4rl_dataset = d4rl_offline_dataset("antmaze-umaze-v2")

dataset = d4rl_dataset
dataset['observations'].shape, dataset['actions'].shape, dataset['rewards'].shape, dataset['next_observations'].shape, dataset['terminals'].shape
# ((1000000, 11), (1000000, 3), (1000000,), (1000000, 11), (1000000,))

# Create a PyTorch Dataset
d4rl_dataset = d4rl_offline_dataset("hopper-medium-v2")
dataset = D4RLDataset(d4rl_dataset)

# Get a batch of data
batch = dataset[-16:]
batch["observation"].shape, batch["action"].shape, batch["reward"].shape, batch["next_observation"].shape, batch["terminal"].shape
# (torch.Size([16, 11]), torch.Size([16, 3]), torch.Size([16]), torch.Size([16, 11]), torch.Size([16]))


returns = np.random.rand(16, 1)
get_normalized_score = D4RLScoreNormalizer("hopper-medium-v2")
get_normalized_score(returns)