Rational Quadratic Kernels gives wrong results!

[tirthasheshpatel]

Here's the output of RatQuad kernel on np.array([[1., 2.], [3., 4.]])

>>> x = np.array([[1., 2.], [3., 4.]])
>>> k = pm.gp.cov.RatQuad(1, 1., 1.)
>>> k(x, x).eval()
array([[1.        , 0.33333333],
       [0.33333333, 1.        ]])

This is a wrong result. Comparing with tensorflow_probability and custom implementation, the result should be np.array([[1., 0.2], [0.2, 1.]])

>>> x = np.array([[1., 2.], [3., 4.]])
>>> k = tfp.python.math.psd_kernels.RationalQuadratic(1., 1., 1.)
>>> k.matrix(x, x)
<tf.Tensor: shape=(2, 2), dtype=float32, numpy=
array([[1.        , 0.19999999],
       [0.19999999, 1.        ]], dtype=float32)>

I was not able to identify what the problem is by looking briefly at the source. This came up while I was testing some PyMC4 kernels.

Versions and main components

• PyMC3 Version: 3.8
• Theano Version: 1.0.4
• Python Version: 3.6.9
• Operating system: Linux
• How did you install PyMC3: git source

[junpenglao]

It should be

k = pm.gp.cov.RatQuad(input_dim=2, alpha=1., ls=1.)

otherwise the 2nd dim is ignored.

[junpenglao]

FWIW tho, I am not sure if the slicing in https://github.com/pymc-devs/pymc3/blob/683faaa9d7e58701f0689b1a1fd4080151f7e057/pymc3/gp/cov.py#L396 is obvious to user - maybe a warning should be raised if the part of the input is ignored?

[tirthasheshpatel]

Yeah! My bad... I got confused between the active_dims and input_dims arguments!

[tirthasheshpatel]

Will try to work on it today!