unilab.base.curriculum
Curriculum learning for adaptive difficulty adjustment.
Classes
-
class unilab.base.curriculum.EpisodeLengthTracker[source]
Bases: object
Track a weighted moving average of episode lengths on one tensor.
- Parameters:
num_envs (int)
window_size (int)
-
__init__(num_envs, window_size=1000, *, device=None)[source]
- Parameters:
num_envs (int)
window_size (int)
-
update(episode_lengths)[source]
Recursively update the average without synchronizing each batch.
- Parameters:
episode_lengths (Tensor)
- Return type:
None
-
update_masked(episode_lengths, mask)[source]
Update from a fixed-width episode-length carrier and bool mask.
This boundary avoids materializing reset
rows; false values are excluded with fixed-shape where operations.
- Parameters:
-
- Return type:
None
-
property average: Tensor
Recursive average as a device-resident scalar tensor.