Fix critical bugs and improve code quality from code review#12
Conversation
- Fix off-by-one bin index clipping in _array.py (len(edges) -> len(edges) - 1) - Add epsilon guard to softmax denominator in _loss.py to prevent division by zero - Replace silent `except Exception: pass` with warnings.warn() for CUDA kernel compilation failures across _loss.py, _predict.py, _core/_primitives.py, _models/_gam.py - Add security warning to joblib deserialization docstring in _persistence.py - Fix ruff target-version mismatch (py311 -> py310) to match requires-python - Align dependency version conflicts in pyproject.toml (xgboost, pytest) - Add ruff lint step to CI workflow - Add warnings for silent n_bins truncation and unseen categories in _array.py - Add DeprecationWarning for NGBoost* aliases via module __getattr__ https://claude.ai/code/session_013WmZYijJFKvq2YM3sHZJmD
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6163f6f32c
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| bin_idx = np.searchsorted(edges, col[~nan_mask], side='right') | ||
| # Clip to valid range (in case test values exceed training range) | ||
| bin_idx = np.clip(bin_idx, 0, len(edges)) | ||
| bin_idx = np.clip(bin_idx, 0, len(edges) - 1) |
There was a problem hiding this comment.
np.digitize in training produces bin indices in [0, len(edges)], where values above the largest edge map to len(edges). Clipping transform-time indices to len(edges) - 1 remaps all out-of-range high values into the second-highest bin, so inference no longer matches training bin semantics and can change split routing/predictions whenever test data exceeds the training range.
Useful? React with 👍 / 👎.
2 participants
except Exception: passwith warnings.warn() for CUDA kernelcompilation failures across _loss.py, _predict.py, _core/_primitives.py, _models/_gam.py
https://claude.ai/code/session_013WmZYijJFKvq2YM3sHZJmD