subset function cause corruption using 2.3.2

[aprilffff]

I have updated to version 2.3.2. During dataset construction, espectially when I use the function dataset.subset, It caused corruption occasionally. Corruptions not happened every time, but mostly on small datasets(less than 50k samples).

also，dataset.add_features_from caused corruptions at the same time.

Any idea on that？

[guolinke]

which code version you used?

[aprilffff]

which code version you used?

2.3.2

[guolinke]

I mean the git commit id, as we didn't officially release 2.3.2 yet.

[aprilffff]

dbb804f

[guolinke]

Thanks, I will investigate that. BTW, could you provide the example that can cause corruption? As I didn't see corruption before, it will be helpful for debugging.

[aprilffff]

pls use attached files below.
testfile.zip

import pandas as pd
import lightgbm as lgb
ref=lgb.Dataset('test_ref.bin')
data=pd.read_pickle('test_data.pkl').values
ds=lgb.Dataset(data,reference=ref).construct()

##will cause corruption below##
ds.subset(np.arange(10)).construct()  

I though it might be related to the reference. If no ref, it wont crash.

[guolinke]

is "test_ref.bin" generated by the same code version?
The binary format of Dataset is changed in recent commits.

[aprilffff]

same version for sure