How to integrate FastAI classifier into sklearn VotingClassifier?

I have a bunch of tabular data and I managed to train a RandomForestClassifier, a GradientBoostingClassifier and a deep learning model (the tabular learner from fastai) with them. I noticed in the results that every model does better then the others in a particular label, different for every model. I was wondering if I could put all the models into a VotingClassifier (the one from sklearn). I have no problem with the RandomForestClassifier and the GradientBoostingClassifier but I didn’t find anything about putting the tabular learner inside the VotingClassifier. Is it possible to do that?

Should be possible with a wrapper which defines __init__, fit and predict_proba which are required by the VotingClassifier:

from fastai.tabular.all import *
from sklearn.base import BaseEstimator, ClassifierMixin

class FastAITabularClassifier(BaseEstimator, ClassifierMixin):
    def __init__(self, dls, layers, metrics):
        self.dls = dls
        self.layers = layers
        self.metrics = metrics
        self.learn = None

    def fit(self, X, y):
        dls = self.dls.new(X, y)
        self.learn = tabular_learner(dls, layers=self.layers, metrics=self.metrics)
        self.learn.fit_one_cycle(5)
        return self

    def predict_proba(self, X):
        dl = self.dls.test_dl(X, with_labels=False)
        preds, _ = self.learn.get_preds(dl=dl)
        return preds.numpy()

Then you should be able to use it with VotingClassifier:

from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier, VotingClassifier

rf = RandomForestClassifier(n_estimators=100)
gbm = GradientBoostingClassifier(n_estimators=100)

fastai_model = FastAITabularClassifier(dls, layers=[200,100], metrics=accuracy)

voting_clf = VotingClassifier(estimators=[
    ('rf', rf),
    ('gbm', gbm),
    ('fastai', fastai_model)
], voting='soft')

voting_clf.fit(X_train, y_train)