How to use different activation functions in one Keras layer?

Question

I am working on Keras in Python and I have a neural network (see code below). Currently it works with only a ReLu activation.

For experimental reasons I would like to have some neurons on ReLu and some on softmax (or any other activation function). for example in a Layer with 20 neurons, I would like to have 10 with ReLu and 10 with Softmax.

I have tried some different ways, but always failed to get an output.

Would you know how I should do this?

# - Libraries
from keras.layers import Dense
from keras.models import Sequential
from keras.callbacks import EarlyStopping
early_spotting_monitor = EarlyStopping(patience=2)
layers = 4
neurons = 20
act = "ReLu"

# - Create Neural Network
model = Sequential()
model.add(Dense(neurons,activation=act,input_dim=X_train.shape[1]))

layers -= 1
while layers > 0:
    model.add(Dense(neurons,activation=act))  
    layers -= 1
model.add(Dense(n_months))
model.compile(optimizer="adam",loss="mean_absolute_error")

model.fit(X_train,Y_train,validation_split=0.10,epochs=13,callbacks=[early_spotting_monitor])

EDIT: this is my (working) code now:

# - Libraries
from keras.callbacks import EarlyStopping
early_spotting_monitor = EarlyStopping(patience=2)
from keras.layers import Input, Dense
from keras.models import Model       
from keras.layers.merge import concatenate

# input layer
visible = Input(shape=(X_train.shape[1],))

hidden11 = Dense(14, activation='relu')(visible)
hidden12 = Dense(3, activation='softplus')(visible)
hidden13 = Dense(2, activation='linear')(visible)    
hidden13 = Dense(2, activation='selu')(visible)  
merge1 = concatenate([hidden11, hidden12, hidden13])

hidden21 = Dense(14, activation='relu')(merge1)
hidden22 = Dense(3, activation='softplus')(merge1)
hidden23 = Dense(2, activation='linear')(merge1)    
hidden13 = Dense(2, activation='selu')(visible) 
merge2 = concatenate([hidden21, hidden22, hidden23])

hidden3 = Dense(20, activation='relu')(merge2)

output = Dense(Y_train.shape[1],activation="linear")(hidden3)
model = Model(inputs=visible, outputs=output)

model.compile(optimizer="adam",loss="mean_absolute_error")
model.fit(X_train,Y_train,validation_split=0.10,epochs=13,callbacks=[early_spotting_monitor])  # starts training

return model

Steamer

slice the layer x into two (or more) x_1,x_2,x_3..., then use your different activation functions y_n = act_n(x_n), then concat y_ns. – Suk

That is exactly what I fail to do. Would you have an example of this? – Steamer

How do you want to split the neurons for the two activation functions? Just have the first ten passed through softmax and the other ten through ReLU? – Suk

Exactly, the layer would be 20 neurons where 10 are Relu and 10 are softmax (for example) – Steamer

Answer 1

You have to use the Functional API to do this, for example:

input = Input(shape = (X_train.shape[1]))
branchA = Dense(neuronsA, activation = "relu")(input)
branchB = Dense(neuronsB, activation = "sigmoid")(input)

out = concatenate([branchA, branchB])

You cannot do it with the Sequential API, so I recommend you move your code to the functional API.

Maryland

Answer 2

So this is something I have been trying to do recently and so far this is what I have done. I think it's working, but I would appreciate if anyone tells me what I'm doing wrong here. I'm doing this only on the output layer and my output layer has two units:

def activations(l):
    l_0 = tf.keras.activations.exponential(l[...,0])
    l_1 = tf.keras.activations.elu(l[...,1])
    lnew = tf.stack([l_0, l_1], axis = 1)
    return lnew

model = tf.keras.Sequential([..., Dense(2, activation = activations)])

Garik