BUG when judging ndim of a series in dataframe

[huyp182]

File: qlib/contrib/model/gbdt.py Line 43: x, y = df["feature"], df["label"] if y.values.ndim == 2 and y.values.shape[1] == 1: y = np.squeeze(y.values) else: raise ValueError("LightGBM doesn't support multi-label training")

y is a Series type from a Dataframe, so the ndim of y.value can only be 1, and is impossible to be 2, then error raises.

[SunsetWolf]

Hi, @huyp182 Can you tell me how to reproduce this?

[SunsetWolf]

Hi, @huyp182

Thank you for your attention to qlib. The issue you mentioned is not a bug, because:

• The DataFrame has multi-level indices (datetime, instrument);
• The DataFrame has multi-level columns (feature and label);
• There is only one subcolumn LABEL0 under the label column.

Your observation is based on the fact that “Series values must be 1D,” which is correct. However, in this DataFrame, df[“label”] returns a DataFrame, not a Series. Therefore, y.values.ndim == 2 is reasonable.

Reference code: LGBModel._prepare_data -> DatasetH.prepare -> DatasetH._prepare_seg -> DataHandler.fetch -> DataHandler._fetch_data