Out of bound access when dataset in continual training has fewer features than in the loaded model #5156
Labels
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Description
When the training dataset has fewer features than in the loaded model, out of bound access can happen at least in one place here
LightGBM/src/boosting/gbdt_model_text.cpp
Line 642 in fc0c8fd
where
feature_importancehas the size of the feature number in the training dataset in continual training, while the feature indices in trees come from the loaded model and can cause out of bound access.Reproducible example
A reproducible example will be added in the PR fixing this bug, also as a test case.
Environment info
LightGBM version or commit hash:
LightGBM master branch
Additional Comments
When the input dataset is from LibSVM file, this can be a bug because it is possible that in the dataset for continual training all values of a feature are missing, and thus when loading the dataset, the found number of features is fewer than before.
However, in terms of other input formats, this problem should be classified as a misusing case. And proper warning, or fatal message should be provided.
The text was updated successfully, but these errors were encountered: