Support residual network architecture

[Zireae1]

Hi Avanti and others,

Thank you for the great tool! I would like to use your package to compute the feature contribution for the residual network, however it seems that this type of layers is not supported currently. Could you please adapt your tool to handle this type of networks?

My network architecture looks like this: ` import keras as ke PS = 6212
DR = 0.2

inputs = Input(shape=(PS,))

x = Dense(2000, activation='relu')(inputs)

x = Dense(1000, activation='relu')(x) x = Dropout(DR)(x)

y = Dense(1000, activation='relu')(x)

z = ke.layers.add([x,y]) z = Dropout(DR)(z)

y = Dense(1000, activation='relu')(z)

x = ke.layers.add([z,y]) x = Dropout(DR)(x)

y = Dense(1000, activation='relu')(x)

z = ke.layers.add([x,y]) z = Dropout(DR)(z)

y = Dense(1000, activation='relu')(z)

x = ke.layers.add([z,y]) x = Dropout(DR)(x)

x = Dense(500, activation='relu')(x) x = Dropout(DR)(x)

x = Dense(250, activation='relu')(x) x = Dropout(DR)(x)

x = Dense(125, activation='relu')(x) x = Dropout(DR)(x)

x = Dense(62, activation='relu')(x) x = Dropout(DR)(x)

x = Dense(30, activation='relu')(x) x = Dropout(DR)(x)

outputs = Dense(2, activation='softmax')(x)

model = Model(inputs=inputs, outputs=outputs)

model.summary()

model.compile(loss='categorical_crossentropy', optimizer=SGD(lr=0.001, momentum=0.9), metrics=['accuracy']) ` Thanks for help.

[AvantiShri]

Sorry I haven't gotten to this yet. It's on the to-do, I've just been sidetracked by other projects. In the meantime, if this is urgent for you, you could take a look at the DeepLIFT implementation in the DeepExplain repository - it only supports the Rescale rule of DeepLIFT (so it may not perform as well), but because it is implemented by overriding tensorflow gradient operators, it should support most architectures: https://github.com/kundajelab/deeplift#how-does-the-implementation-in-this-repository-compare-with-the-implementation-in-ancona-et-al-iclr-2018