Can I train a Tensorflow keras model with complex input/output?

Question

I am trying to train a very simple model which only have one convolution layer.

 def kernel_model(filters=1, kernel_size=3):
    input_layer = Input(shape=(250,1))
    conv_layer = Conv1D(filters=filters,kernel_size=kernel_size,padding='same',use_bias = False)(input_layer)
    model = Model(inputs=input_layer,output=conv_layer)
    return model 

But the input(X), prediction output(y_pred) and true_output(y_true) are all complex number. When I call the function model.fit(X,y_true)

There is the error TypeError: Gradients of complex tensors must set grad_ys (y.dtype = tf.complex64)

Does that means I have to write the back-propagation by hand?
What should I do to solve this problem? thanks

Answer 1

Your DNN needs to mininimize the Loss-function through back-propagation. To minimize something, it naturally needs to have an ordering. Complex numbers are not ordered, while Reals are. So you generally need a loss function L: Complex -> Reals

Change your complex-valued loss function from simple square:

error = K.cast(K.mean(K.square(y_pred_propgation - y_true)),tf.complex64)

to a real-valued magnitude ||.||^2 of the complex number:

error = K.mean(K.square(K.abs(y_true-y_pred)))