Defining custom loss functions.

[agosztolai]

Hello,

I am trying to implement a simple custom loss function for the purpose of finding the least-squares projection ||f(x) - y||_2 to a pre-trained GP f defined by the model manifold_GP. I can provide an initial guess for x.

I have tried a million ways and failed. Below is my latest attempt.

Can you please tell me how to do this?

y = y_train[[1]]
def loss() -> tf.Tensor:
    Y_predicted = manifold_GP.predict_f(x_initial)
    squared_error = (Y_predicted - tf.convert_to_tensor(y)) ** 2
    return tf.reduce_mean(squared_error)

x_initial = np.array(x_test[[0]])  # Initial guess for x
opt = gpflow.optimizers.Scipy()
result = opt.minimize(loss, variables=[tf.Variable(x_initial)]) 

Documentation/tutorial notebooks

Is there anything missing in the docs?

Are there any mistakes in the docs?

Is there a feature that needs some example code in a notebook?

Do you know how to fix the docs? If so, it'd be amazing if you'd be willing to directly contribute a pull request :)

Links:

• latest GPflow documentation
• latest GPflow tutorial notebooks