[FEA] cuML estimators should support non-consecutive labels outside of [0, n) where appropriate #4478
Comments
beckernick
added
feature request
|
This issue has been labeled |
|
I ran into this, I'm working with Pandas dataframes. After using imblearn RandomUnderSampler, my binary valued labels were in order, i.e. all 0's then all 1's. I used pd.sample to shuffle the data, and I had to reset indices to make the error go away. Here is the snippet: |
|
Generalizing from @jhancock1975 's code snippet, this error can occur even with correctly formatted data during cross validation (if a fold doesn't get the right subset of labels). import pandas as pd
import cuml
from sklearn.model_selection import cross_val_score
df = pd.DataFrame({
"x1": [0.0,1,2,3,3,4],
"x2": [-3,2.0,5,5,3,2],
"y": [0,1,1,1,2,2]
})
clf = cuml.ensemble.RandomForestClassifier()
cross_val_score(
clf,
df[["x1", "x2"]],
df["y"],
cv=2,
error_score="raise"
)
# Throws the error linked above |
|
Scikit-learn does this under the hood here: |
|
This issue has been labeled |
|
It should be possible to do this using This function looks like it might have an expensive JIT cost, though. Perhaps cuml/cpp/src_prims/label/classlabels.cuh Line 164 in 768a4ed |
|
@beckernick, yep, this is indeed something we had written originally in C++ in cuml for DBSCAN and have since moved to RAFT (pending removal from cuml, the raft version is more up to date). If desired, this could also be a good reason to expose through pylibraft and use in cuml (reusable and very clean separation of implementation details). |
|
That sounds like it could be a good solution. I suspect this non-consecutive label issue will keep popping up. Will file a new issue on RAFT, cross-link, and mark it as a good first issue based on your description. |
beckernick
changed the title
|
This issue has been labeled |
…ive labels where appropriate (#4780) This PR closes #4478 by transforming non-consecutive labels outside of [0,n) to consecutive labels inside [0,n) similar to what Scikit-learn does under the hood. Closes #691 Authors: - https://github.com/VamsiTallam95 Approvers: - Micka (https://github.com/lowener) - Dante Gama Dessavre (https://github.com/dantegd) - Corey J. Nolet (https://github.com/cjnolet) URL: #4780
…ive labels where appropriate (rapidsai#4780) This PR closes rapidsai#4478 by transforming non-consecutive labels outside of [0,n) to consecutive labels inside [0,n) similar to what Scikit-learn does under the hood. Closes rapidsai#691 Authors: - https://github.com/VamsiTallam95 Approvers: - Micka (https://github.com/lowener) - Dante Gama Dessavre (https://github.com/dantegd) - Corey J. Nolet (https://github.com/cjnolet) URL: rapidsai#4780
As noted in rapidsai/cudf#10024 , cuML RandomForestClassifier will throw an error if the target column has non-consecutive labels outside of the [0, n) range. This does not occur in scikit-learn, perhaps due to label encoding happening under the hood.
This may occur with other estimators as well.
The text was updated successfully, but these errors were encountered: