github/microsoft-SkillOpt
1
0
Fork 0
mirror of https://github.com/microsoft/SkillOpt.git synced 2026-07-07 16:49:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #20 from 1s1x/fix-continuous-reward-scores

Browse Source 
fix: support continuous reward scores (int truncation + falsy float)
This commit is contained in:
Yif Yang
2026-05-30 15:30:15 +08:00
committed by GitHub
parent dbc90bd755 41be2f1803
commit cf287cb608
3 changed files with 2504 additions and 2505 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3824
skillopt/engine/trainer.py
View File

File diff suppressed because it is too large Load Diff

1176
skillopt/gradient/reflect.py
View File

File diff suppressed because it is too large Load Diff

9
skillopt/utils/scoring.py
View File

@@ -7,17 +7,16 @@ import hashlib
def compute_score(results: list) -> tuple[float, float]:
"""Compute hard and soft accuracy from a list of episode results.
Accepts both plain dicts and :class:`~skillopt.types.RolloutResult`
instances.
Accepts both plain dicts and :class: instances. hard may be continuous (0.0-1.0) when using smoothed reward.
"""
if not results:
return 0.0, 0.0
def _hard(r: object) -> int:
return int(r.hard if hasattr(r, "hard") else r.get("hard", 0)) # type: ignore[union-attr]
def _hard(r: object) -> float:
return float(r.hard if hasattr(r, "hard") else r.get("hard", 0))
def _soft(r: object) -> float:
return float(r.soft if hasattr(r, "soft") else r.get("soft", 0.0)) # type: ignore[union-attr]
return float(r.soft if hasattr(r, "soft") else r.get("soft", 0.0))
hard = sum(_hard(r) for r in results) / len(results)
soft = sum(_soft(r) for r in results) / len(results)