pystacknet is a light python version of StackNet which was originally made in Java.
It supports many of the original features, with some new elements.
git clone https://github.com/h2oai/pystacknet
cd pystacknet
python setup.py install
pystacknet's main object is a 2-dimensional list of sklearn type of models. This list defines the StackNet structure. This is the equivalent of parameters in the Java version. A representative example could be:
from sklearn.ensemble import RandomForestClassifier, ExtraTreesClassifier, GradientBoostingClassifier
from sklearn.linear_model import LogisticRegression
models=[
######## First level ########
[RandomForestClassifier (n_estimators=100, criterion="entropy", max_depth=5, max_features=0.5, random_state=1),
ExtraTreesClassifier (n_estimators=100, criterion="entropy", max_depth=5, max_features=0.5, random_state=1),
GradientBoostingClassifier(n_estimators=100, learning_rate=0.1, max_depth=5, max_features=0.5, random_state=1),
LogisticRegression(random_state=1)
],
######## Second level ########
[RandomForestClassifier (n_estimators=200, criterion="entropy", max_depth=5, max_features=0.5, random_state=1)]
]pystacknet is not as strict as in the Java version and can allow Regressors, Classifiers or even Transformers at any level of StackNet. In other words the following could work just fine:
from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor, ExtraTreesClassifier, ExtraTreesRegressor, GradientBoostingClassifier,GradientBoostingRegressor
from sklearn.linear_model import LogisticRegression, Ridge
from sklearn.decomposition import PCA
models=[
[RandomForestClassifier (n_estimators=100, criterion="entropy", max_depth=5, max_features=0.5, random_state=1),
ExtraTreesRegressor (n_estimators=100, max_depth=5, max_features=0.5, random_state=1),
GradientBoostingClassifier(n_estimators=100, learning_rate=0.1, max_depth=5, max_features=0.5, random_state=1),
LogisticRegression(random_state=1),
PCA(n_components=4,random_state=1)
],
[RandomForestClassifier (n_estimators=200, criterion="entropy", max_depth=5, max_features=0.5, random_state=1)]
]Note that not all transformers are meaningful in this context and you should use it at your own risk.
A typical usage for classification could be :
from pystacknet.pystacknet import StackNetClassifier
model=StackNetClassifier(models, metric="auc", folds=4,
restacking=False,use_retraining=True, use_proba=True,
random_state=12345,n_jobs=1, verbose=1)
model.fit(x,y)
preds=model.predict_proba(x_test)
Where :
| Command | Explanation |
|---|---|
| models | List of models. This should be a 2-dimensional list . The first level hould defice the stacking level and each entry is the model. |
| metric | Can be "auc","logloss","accuracy","f1","matthews" or your own custom metric as long as it implements (ytrue,ypred,sample_weight=) |
| folds | This can be either integer to define the number of folds used in StackNet or an iterable yielding train/test splits. |
| restacking | True for restacking else False |
| use_proba | When evaluating the metric, it will use probabilities instead of class predictions if use_proba==True |
| use_retraining | If True it does one model based on the whole training data in order to score the test data. Otherwise it takes the average of all models used in the folds ( however this takes more memory and there is no guarantee that it will work better.) |
| random_state | Integer for randomised procedures |
| n_jobs | Number of models to run in parallel. This is independent of any extra threads allocated |
| n_jobs | Number of models to run in parallel. This is independent of any extra threads allocated from the selected algorithms. e.g. it is possible to run 4 models in parallel where one is a randomforest that runs on 10 threads (it selected). |
| verbose | Integer value higher than zero to allow printing at the console. |