A Python package for unwrapping ReLU Neural Networks
The following environments are required:
- Python 3.6, 3.7, 3.8, 3.9 (Try Google Colab)
- matplotlib>=3.1.3
- numpy>=1.17
- pandas>=1.1.2
- seaborn>=0.9.0
- scikit-learn>=0.23.0
- statsmodels>=0.12.2
pip install aletheia-dnnLoad data
importnumpyasnpimportpandasaspdimportmatplotlib.pyplotaspltfromsklearn.datasetsimportmake_circlesfromsklearn.model_selectionimporttrain_test_splitrandom_state=0x, y=make_circles(n_samples=2000, noise=0.1, random_state=random_state)
train_x, test_x, train_y, test_y=train_test_split(x, y, test_size=0.2, random_state=random_state)
plt.figure(figsize=(10,8))
scatter=plt.scatter(x[:, 0], x[:, 1], c=y)
plt.legend(*scatter.legend_elements(), loc="upper right")
plt.show()Train a ReLU Net
fromsklearn.neural_networkimportMLPClassifiermlp=MLPClassifier(hidden_layer_sizes=[40] *4, max_iter=2000, early_stopping=True, n_iter_no_change=100, validation_fraction=0.2,
solver='adam', activation="relu", random_state=random_state, learning_rate_init=0.001)
mlp.fit(train_x, train_y)UnwrapperClassifier
fromaletheiaimport*clf=UnwrapperClassifier(mlp.coefs_, mlp.intercepts_)
clf.fit(train_x, train_y)
clf.summary()
Partitioned regions
clf.visualize2D_regions(figsize=(8, 8), meshsize=300, show_label=False)
Simplification
fromsklearn.metricsimportmake_scorer, roc_auc_scorefromsklearn.model_selectionimportGridSearchCV, PredefinedSplitfromsklearn.linear_modelimportLogisticRegressionCV, LogisticRegressiondatanum=train_x.shape[0]
indices=np.arange(datanum)
idx1, idx2=train_test_split(indices, test_size=0.2, random_state=random_state)
val_fold=np.ones((len(indices)))
val_fold[idx1] =-1grid=GridSearchCV(MergerClassifier(unwrapper=None, weights=mlp.coefs_, biases=mlp.intercepts_,
min_samples=30,
n_neighbors=np.round(clf.nllms*0.01).astype(int),
refit_model=LogisticRegression()),
param_grid={"n_clusters": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 20]},
scoring={"auc": make_scorer(roc_auc_score, needs_proba=True)},
cv=PredefinedSplit(val_fold), refit="auc", n_jobs=10, error_score=np.nan)
grid.fit(train_x, train_y)
clf_merge=grid.best_estimator_clf_merge.summary()
Local Inference
tmpid=0clf_merge.visualize2D_one_line(tmpid, figsize=(8, 8))
clf_merge.local_inference_wald(tmpid).round(4)
Agus Sudjianto, William Knauth, Rahul Singh, Zebin Yang and Aijun Zhang. 2020. Unwrapping The Black Box of Deep ReLU Networks: Interpretability, Diagnostics, and Simplification. arXiv:2011.04041
@article{sudjianto2020unwrapping,
title={Unwrapping The Black Box of Deep ReLU Networks: Interpretability, Diagnostics, and Simplification},
author={Sudjianto, Agus and Knauth, William and Singh, Rahul and Yang, Zebin and Zhang, Aijun},
journal={arXiv:2011.04041},
year={2020}
}