d3rlpy.datasets.get_atari_transitions

d3rlpy.datasets.get_atari_transitions(game_name, fraction=0.01, index=0, num_stack=None, sticky_action=True, pre_stack=False, render_mode=None)[source]

Returns atari dataset as a list of Transition objects and envrironment.

The dataset is provided through d4rl-atari. The difference from get_atari function is that this function will sample transitions from all epochs. This function is necessary for reproducing Atari experiments.

from d3rlpy.datasets import get_atari_transitions

# get 1% of transitions from all epochs (1M x 50 epoch x 1% = 0.5M)
dataset, env = get_atari_transitions('breakout', fraction=0.01)

References

Parameters:

  • game_name (str) – Atari 2600 game name in lower_snake_case.

  • fraction (float) – fraction of sampled transitions.

  • index (int) – index to specify which trial to load.

  • num_stack (Optional[int]) – the number of frames to stack (only applied to env).

  • sticky_action (bool) – Flag to enable sticky action.

  • pre_stack (bool) – Flag to pre-stack observations. If this is False, FrameStackTransitionPicker and FrameStackTrajectorySlicer will be used to stack observations at sampling-time.

  • render_mode (Optional[str]) – Mode of rendering (human, rgb_array).

Returns:

tuple of a list of d3rlpy.dataset.Transition and gym environment.

Return type:

Tuple[ReplayBuffer, Env[ndarray[Any, dtype[Any]], int]]