Skip to content

CheckpointConfiguration ​

ABC for configurations that include checkpoint settings.

Intended to be used as an interface so that the checkpoint code only sees the checkpoint-related properties.

Source

Values ​

checkpoint_frequency ​

python
checkpoint_frequency: Annotated[int, Field(gt=0)] = 10

Source

checkpoint_movie_command ​

python
checkpoint_movie_command: Annotated[str, Field(min_length=1)] = 'oakemu'

The OakEmu tool executable used to record checkpoint movies. May be an absolute path; a bare name is resolved against PATH and then ~/.local/bin (GUI-launched processes often lack the latter in PATH).

Source

checkpoint_movie_extra_end_frames ​

python
checkpoint_movie_extra_end_frames: Annotated[int, Field(ge=0)] = 50

Emulator frames to keep recording after the episode ends, so the movie shows what the final step led to. Stop conditions fire the frame the game decides an outcome, often before it is drawn -- Green Beret's death routine stop fires before any death animation ; with no extra frames a movie can end with the player apparently fine. 50 frames is one second.

Source

checkpoint_movie_filename_template ​

python
checkpoint_movie_filename_template: Annotated[str, Field(min_length=1)] = '{game}_checkpoint_{checkpoint}.mp4'

Relative movie output path inside the checkpoint directory. Supports the same fields as checkpoint_onnx_filename_template.

Source

checkpoint_movie_maximum_steps ​

python
checkpoint_movie_maximum_steps: Annotated[int, Field(gt=0)] = 5000

Step cap for checkpoint movies so games without a natural end (for example Deathchase) still terminate.

Source

checkpoint_onnx_filename_template ​

python
checkpoint_onnx_filename_template: Annotated[str, Field(min_length=1)] = '{game}_checkpoint_{checkpoint}.onnx'

Relative ONNX output path inside the checkpoint directory. Supports {game}, {game_name}, {experiment}, {experiment_name}, {checkpoint}, and {checkpoint_number}.

Source

compress_checkpoints ​

python
compress_checkpoints: bool = True

Whether to zstd-compress checkpoint files. Disabling can significantly speed up checkpoint saving at the cost of larger files.

Source

keep_last_n_checkpoints ​

python
keep_last_n_checkpoints: Annotated[int | None, Field(gt=0)] = 10

Source

save_movies_with_checkpoints ​

python
save_movies_with_checkpoints: bool = True

Whether to record a movie of each checkpoint's ONNX model playing greedily into the checkpoint directory. Runs the OakEmu tool as a detached background process so training is not blocked. Requires save_onnx_with_checkpoints. Movies are pruned along with their checkpoint directories.

Source

save_onnx_with_checkpoints ​

python
save_onnx_with_checkpoints: bool = True

Whether to export an ONNX model into each checkpoint directory after saving checkpoint files.

Source

save_trajectory_store_keys_csv ​

python
save_trajectory_store_keys_csv: bool = False

Whether to save trajectory-store debug CSVs (keys, per-entry rows, and frontier summaries) in each checkpoint.

Source