I'm trying to write a custom loss function for a Keras Model using TensorFlow 2.0. I followed the directions in a similar answer to get the input layer into the loss function like
here
Keras custom loss function: Accessing current input pattern
and here
but I also want to add the output of a second model to the loss function.
The error appears to be coming from v.predict(x). At first TensorFlow gives an error like
ValueError: When using data tensors as input to a model, you should specify the steps argument.
So I try adding the steps arg v.predict(x,steps=n) where n some integer and I get AttributeError: 'Tensor' object has no attribute 'numpy'
X = some np.random array Y = some function of X plus noise
def build_model():
il = tf.keras.Input(shape=(2,),dtype=tf.float32)
outl = kl.Dense(100,activation='relu')(il)
outl = kl.Dense(50,activation='relu')(outl)
outl = kl.Dense(1)(outl)
return il,outl
def f(X,a):
return (X[:,0:1] + theta*a)*a
def F(x,a):
eps = tf.random.normal(tf.shape(x),mean=loc,stddev=scale)[:,0:1]
z = tf.stack([x[:,0:1] + theta*a + eps,x[:,1:] - a],axis=1)[:,:,0]
return z
def c_loss(x=None,v=None):
def loss(y_true,y_pred):
xp = F(x,y_pred)
return kb.mean(f(x,y_pred) + v.predict(xp))
return loss
v_in,v_out = build_model()
v_model = tf.keras.Model(inputs=v_in,outputs=v_out)
v_model.compile(tf.keras.optimizers.Adam(),loss='mean_squared_error')
v_model.fit(x=X,y=Y)
c_in,c_out = build_model()
c_model = tf.keras.Model(inputs=c_in,outputs=c_out)
c_model.compile(tf.keras.optimizers.Adam(),loss=c_loss(x=c_in,v=v_model))
c_model.fit(x=X,y=Y_dummy)
Ideally I just expect the call of c_model.fit() to build a neural network to minimize the functional f(x,a) + v(x).
