gpflow model prediction says input was expected to be a double tensor but is a float tensor

Question

I am trying to run the code from the gpflow tutorial: https://gpflow.readthedocs.io/en/stable/notebooks/regression.html However, it doesn't work.

The following code:

N = 12
X = np.random.rand(N,1)
Y = np.sin(12*X) + 0.66*np.cos(25*X) + np.random.randn(N,1)*0.1 + 3
plt.plot(X, Y, 'kx', mew=2)

k = gpflow.kernels.Matern52(variance=1.0, lengthscale=1.0)
m = gpflow.models.GPR((X, Y), k, mean_function=None, noise_variance=1.0)
m.likelihood.variance = 0.01

def plot(m):
    xx = np.linspace(-0.2, 1.2, 141)[:,None]
    xx=tf.convert_to_tensor(xx,dtype=tf.float64)
    mean, var = m.predict_y(xx)
    plt.figure(figsize=(12, 6))
    plt.plot(X, Y, 'kx', mew=2)
    plt.plot(xx, mean, 'b', lw=2)
    plt.fill_between(xx[:,0], mean[:,0] - 2*np.sqrt(var[:,0]), mean[:,0] + 2*np.sqrt(var[:,0]), color='blue', alpha=0.2)
    plt.xlim(-0.1, 1.1)
plot(m)

returns the following error:

InvalidArgumentError: cannot compute AddV2 as input #1(zero-based) was expected to be a double tensor but is a float tensor [Op:AddV2] name: add/

I have windows 10, python 3.6, tensorflow 2.0, tensorflow probability 0.9, and gpflow was installed with pip install -e . command on 21st of february 2020.

Could you help me with this? I do transform the input to double so I think it could be that gpflow updated the code but not the tutorial.

Answer 1

The problem you encounter is due to new way we update parameter values in GPflow. Instead of doing model.parameter = value you should use assign:

m.likelihood.variance.assign(0.01) 

This makes sure the type of the parameters is not changed.

I was able to get the following plot, after setting the lengthscale of the kernel to 0.25.