Keras Add_loss 抛出运算符在图形错误中不允许

Question

I created a basic loss function that takes the CDF (cumsum of pdf) and does a mean_squared error between the two.

Here is the code:

def tuner_loss(y_true, y_pred):
    y_actual=K.cumsum(y_true)
    y_pred=K.cumsum(y_pred)
    return K.mean(K.square(y_actual-y_pred))

I tried model.add_loss(loss_model) with loss="None" then tried loss=loss_model , then commented out the model.add_loss and loss_model and directly put the tuner_loss function into loss.

None of the them worked.

Here is that code:

input_layer= Input(179,)
y_actual=Input(199,)
x=Dense(32, activation='relu')(input_layer)
x=Dense(64, activation='relu')(x)
x=Dense(32, activation='relu')(x)
x=Dropout(0.2)(x)
x=Dense(64)(x)
output_layer = Dense(199, activation='softmax')(x)
model=Model(inputs=[input_layer, y_actual], outputs=output_layer)
#loss_model= tuner_loss(y_actual, output_layer)
#model.add_loss(loss_model)
model.compile(optimizer='adam', loss= tuner_loss(y_actual,output_layer), metrics=['accuracy']) 

Any help would be greatly appreciated, I figure it has to be something with how I'm manipulating the tf.tensors in the tuner_loss .
Here is error message I got running last code:

"OperatorNotAllowedInGraphError using a tf.Tensor as a Python bool is not allowed in Graph execution. Use Eager execution or decorate this function with @tf.function.

I tried the decorator on the tuner, that also threw an error, so that didn't work.

Neither did enabling eager execution...

Answer 1

Why not use the standard?

model = Model(input_layer, output_layer)
model.compile(loss = tuner_loss, ...)
model.fit(input_data, output_data, ...)