How do you change the number of units in a layer when using Keras?

Question

The below code works perfectly okay. If I try to change all the 64s to 128s then I get an error about shape. Do I need to change the input data shape if I change the number of layers in an artificial neural network when using Keras? I didn't think so because it asks for input_dim which is correct.

Works:

model = Sequential()
model.add(Dense(64, input_dim=14, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(64, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(64, init='uniform'))
model.add(Activation('softmax'))

sgd3 = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy', optimizer=sgd3)
model.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose = 2)

Doesn't Work:

model = Sequential()
model.add(Dense(128, input_dim=14, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(128, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(128, init='uniform'))
model.add(Activation('softmax'))

sgd3 = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy', optimizer=sgd3)
model.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose = 2)

Answer 1

You can use a different number of hidden layers and a different number of units/neurons for each of them on the same input.

Each Dense except the last one can be seen as a hidden layer. The last Dense should have a number of outputs equal to your desired output dimension (in your case the dimension of y seems to be 64).

Try this:

model = Sequential()
model.add(Dense(128, input_dim=14, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(128, init='uniform'))
model.add(Activation('tanh'))
model.add(Dropout(0.5))
model.add(Dense(64, init='uniform'))
model.add(Activation('softmax'))

sgd3 = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy', optimizer=sgd3)
model.fit(X_train, y_train, nb_epoch=20, batch_size=16, show_accuracy=True, validation_split=0.2, verbose = 2)