Support for DaskXGBClassifier to PMML

[apoorv22]

Is there a way currently or a possible plan in future to convert a DaskXBGClassifier to PMML

[vruusmann]

Where does this class come from? Is it from the Dask-ML project, or the Dask-XGBoost project?

In either case, there's probably a thin Python wrapper around the native XGBoost binary model file. If the former can be saved to the pickle file, then conversion should be trivial.

@apoorv22 Can you share a reproducible example about how you're training an XGBoost model using Dask? Specifically, how is the feature matrix defined/composed - I assume there's no Scikit-Learn pipeline involved.

[apoorv22]

Hi @vruusmann ,
This class is from dask-ml Refer this

Here is a reproducible example

from distributed import Client
from sklearn2pmml import PMMLPipeline, make_pmml_pipeline, sklearn2pmml
from xgboost.dask import DaskXGBClassifier
import pandas as pd
import dask.dataframe as dd

client = Client('localhost:9787')
df = pd.read_csv('/home/user/gender_voice_scikit_label_dataset.csv')

col_names = list(df)
dependant_var = 'label'
df = df[col_names]
col_names.remove(dependant_var)
ddf = dd.from_pandas(df,npartitions=-1)
xgb_model = DaskXGBClassifier(max_depth=1, learning_rate=0.1, n_estimators=100,
                              verbosity=1, objective='binary:logistic', booster='gbtree', n_jobs=1,
                              nthread=None, gamma=0, min_child_weight=1, max_delta_step=0,
                              colsample_bytree=1, colsample_bylevel=1, subsample=1,
                              reg_alpha=1, reg_lambda=0,
                              random_state=29, seed=29, missing=None)

# Training
model = xgb_model.fit(X=ddf, y=ddf[dependant_var])
ppl = PMMLPipeline([
    ('classifier', model)])

try:
    pipeline = make_pmml_pipeline(ppl, active_fields=col_names, target_fields=[dependant_var])
    sklearn2pmml(pipeline, '/home/user/scr.pmml', debug=True)
except Exception as e:
    print(e)

Dataset
gender_voice_scikit_label_dataset.zip

[noahisch]

Have you tried converting the DaskXGBClassifier to an XGBClassifier and then converting?
dmlc/xgboost#6547