stats#
Source code: tianshou/data/stats.py
- class SequenceSummaryStats(*, mean: float, std: float, max: float, min: float)[source]#
Bases:
DataclassPPrintMixinA data structure for storing the statistics of a sequence.
- mean: float#
- std: float#
- max: float#
- min: float#
- classmethod from_sequence(sequence: Sequence[float | int] | ndarray) SequenceSummaryStats[source]#
- compute_dim_to_summary_stats(arr: Sequence[Sequence[float]] | ndarray) dict[int, SequenceSummaryStats][source]#
Compute summary statistics for each dimension of a sequence.
- Parameters:
arr – a 2-dim arr (or sequence of sequences) from which to compute the statistics.
- Returns:
A dictionary of summary statistics for each dimension.
- class TimingStats(*, total_time: float = 0.0, train_time: float = 0.0, train_time_collect: float = 0.0, train_time_update: float = 0.0, test_time: float = 0.0, update_speed: float = 0.0)[source]#
Bases:
DataclassPPrintMixinA data structure for storing timing statistics.
- total_time: float = 0.0#
The total time elapsed.
- train_time: float = 0.0#
The total time elapsed for training (collecting samples plus model update).
- train_time_collect: float = 0.0#
The total time elapsed for collecting training transitions.
- train_time_update: float = 0.0#
The total time elapsed for updating models.
- test_time: float = 0.0#
The total time elapsed for testing models.
- update_speed: float = 0.0#
The speed of updating (env_step per second).
- class InfoStats(*, gradient_step: int, best_score: float, best_reward: float, best_reward_std: float, train_step: int, train_episode: int, test_step: int, test_episode: int, timing: TimingStats)[source]#
Bases:
DataclassPPrintMixinA data structure for storing information about the learning process.
- gradient_step: int#
The total gradient step.
- best_score: float#
The best score over the test results. The one with the highest score will be considered the best model.
- best_reward: float#
The best reward over the test results.
- best_reward_std: float#
Standard deviation of the best reward over the test results.
- train_step: int#
The total collected step of training collector.
- train_episode: int#
The total collected episode of training collector.
- test_step: int#
The total collected step of test collector.
- test_episode: int#
The total collected episode of test collector.
- timing: TimingStats#
The timing statistics.
- class EpochStats(*, epoch: int, train_collect_stat: CollectStatsBase, test_collect_stat: CollectStats | None, training_stat: TrainingStats | None, info_stat: InfoStats)[source]#
Bases:
DataclassPPrintMixinA data structure for storing epoch statistics.
- epoch: int#
The current epoch.
- train_collect_stat: CollectStatsBase#
The statistics of the last call to the training collector.
- test_collect_stat: CollectStats | None#
The statistics of the last call to the test collector.
- training_stat: TrainingStats | None#
The statistics of the last model update step. Can be None if no model update is performed, typically in the last training iteration.