xgbclassifier gives error with gridsearchcv and dask dataframes

import dask.dataframe as dd
import numpy as np
import pandas as pd
from dask_ml.model_selection import GridSearchCV
from dask_ml.xgboost import XGBClassifier
from distributed import Client
from sklearn.datasets import load_iris

if __name__ == '__main__':

    client = Client()

    data = load_iris()

    x = pd.DataFrame(data=data['data'], columns=data['feature_names'])
    x = dd.from_pandas(x, npartitions=2)

    y = pd.Series(data['target'])
    y = dd.from_pandas(y, npartitions=2)

    estimator = XGBClassifier(objective='multi:softmax', num_class=4)
    grid_search = GridSearchCV(
        estimator,
        param_grid={
            'n_estimators': np.arange(15, 105, 15)
        },
        scheduler='threads'
    )

    grid_search.fit(x, y)
    results = pd.DataFrame(grid_search.cv_results_)
    print(results.to_string())

gives this

Traceback (most recent call last): File “d.py”, line 30, in <module> grid_search.fit(x, y) File “/usr/local/lib/python3.7/site-packages/dask_ml/model_selection/_search.py”, line 1233, in fit cache_cv=self.cache_cv, File “/usr/local/lib/python3.7/site-packages/dask_ml/model_selection/_search.py”, line 203, in build_cv_graph X_name, y_name, groups_name = to_keys(dsk, X, y, groups) File “/usr/local/lib/python3.7/site-packages/dask_ml/model_selection/utils.py”, line 85, in to_keys assert not is_dask_collection(x) AssertionError

[xiaozhongtian]

@TomAugspurger Hello, I have met the same problem here that is exactly what you summarized above. Dask dataframes should be accepted in GridSearchCV. It’s really important to have a complete chain with this component. I have re-implement some estimators of sklearn for dask dataframe and I found that It has a great problmem that it’s not supported by the dask dataframe in GridSearch with the pipeline of these estimaors.

[mrocklin]

@stsievert , who works on dask-ml a bit, will also be around.