I want to use LIME for explaining my models results.

I have a DNN model ending with 1 neuron output layer with sigmoid function.

Nearly ALL examples I came across the internet had examples either on regression, or DT models.
I couldn’t find any matching my case.

DNN = tf.keras.models.Sequential([
  #Input layer with 121 (nb of features) neurons
  tf.keras.layers.Input(121),

  #1st hidden layer with 182 neurons
  tf.keras.layers.Dense(182, activation='relu'),

  #Dropout layer with rate = 0.2
  tf.keras.layers.Dropout(0.2),

  #2nd hidden layer with 182 neurons
  tf.keras.layers.Dense(182, activation='relu'),

  #Ouput layer with 1 neuron
  tf.keras.layers.Dense(1,activation='sigmoid')
])

DNN.compile(optimizer='adam',
              loss="binary_crossentropy",
              metrics=['accuracy'])

DNN_trained=DNN.fit(nsl_kdd_train_X2, nsl_kdd_train_Y2, epochs=5)


#lime
from lime import lime_tabular
explainer = lime_tabular.LimeTabularExplainer( training_data = np.array(nsl_kdd_train_X2),
                                              feature_names = nsl_kdd_data_X2.columns,
                                              class_names = ['normal', 'anomaly'],
                                              mode="classification"
                                                )


explainer = lime_tabular.LimeTabularExplainer(nsl_kdd_test_X2, mode='classification',feature_names=nsl_kdd_data_X2.columns, verbose=True) #,training_labels=nsl_kdd_data['class']



explanation = explainer.explain_instance(nsl_kdd_test_X2[1], DNN.predict_proba, num_features=len(nsl_kdd_data_X2.columns))

it will give error: “AttributeError: ‘Sequential’ object has no attribute ‘predict_proba'”

so predict_proba is not available, and only predict function is, which outputs 0 or 1 only.

I tried to borrow this part of code:

def prob(data):
    return np.array(list(zip(1-model.predict(data),model.predict(data))))

from

https://towardsdatascience.com/decrypting-your-machine-learning-model-using-lime-5adc035109b5

but it transformed the data into 3D array, which LIME explain_instance function complained from. I’m not sure why it became 3D.