Potential bug in SGPR with changepoint kernel

[jl139277]

Im trying to create a sparse model with changepoint kernel. The general setup I use works fine when using regular kernels:

model = gpflow.models.SGPR(
(X, Y),
kernel=kernel1,
inducing_variable=X[::100],
)

This works as expected.
When I set it up as follows I get errors:

model = gpflow.models.SGPR(
(X, Y),
kernel=gpflow.kernels.ChangePoints(
kernels=[
kernel1,
kernel2,
],
locations=[x_0],
steepness=0.01,
),
inducing_variable=X[::100],
)

Specifically:

ValueError: Dimension 1 in both shapes must be equal, but are 1000 and 10. Shapes are [10,10] and [10,100]. for '{{node concat_2}} = ConcatV2[N=2, T=DT_DOUBLE, Tidx=DT_INT32](ones_1, mul_7, concat_2/axis)' with input shapes: [10,10,1], [10,1000,0], [] and with computed input tensors: input[2] = <2>.

As far as I can tell this is a bug but I am fairly new to GPflow so I might be wrong about that. After looking around I did find this previous issue with the changepoint kernel which seemed similar. #1440

[uri-granta]

Thanks for raising an issue! Could you please include more of your example (in particular, what are X, Y, kernel1 and kernel2)? The best would be to include a short self-contained example that can be run by anyone to demonstrate the bug. (Also FYI here is how to format a code block.)

[jl139277]

Thank you for your reply and the information! Here is the additional information you requested please let me know if you need anything else.

X, Y and the kernels used did not seem to cause the issue as I could run the same model with a different kernels and inputs without issue. This minimal example replicates the error for me:

import numpy as np
import gpflow

X = np.array(
    [
        [0.865], [0.666], [0.804], [0.771], [0.147], [0.866], [0.007], [0.026],
        [0.171], [0.889], [0.243], [0.028],
    ]
)
Y = np.array(
    [
        [1.57], [3.48], [3.12], [3.91], [3.07], [1.35], [3.80], [3.82], [3.49],
        [1.30], [4.00], [3.82],
    ]
)

inducing_points = np.array([[0.125], [0.375], [0.625], [0.875]])
kernel1 = gpflow.kernels.SquaredExponential()
kernel2 = gpflow.kernels.SquaredExponential()

model = gpflow.models.SGPR(
    (X, Y),
    kernel=gpflow.kernels.ChangePoints(
        kernels=[
            kernel1,
            kernel2,
        ],
        locations=[0.5],
        steepness=0.1,
    ),
    inducing_variable=inducing_points,
)

opt = gpflow.optimizers.Scipy()
opt.minimize(model.training_loss, model.trainable_variables)

Specifically the final error reads:
ValueError: Dimension 1 in both shapes must be equal, but are 4 and 12. Shapes are [4,4] and [4,12]. for '{{node concat_2}} = ConcatV2[N=2, T=DT_DOUBLE, Tidx=DT_INT32](ones_1, mul_7, concat_2/axis)' with input shapes: [4,4,1], [4,12,1], [] and with computed input tensors: input[2] = <2>.

[uri-granta]

Thanks. That code example actually runs fine for me without giving any errors!

Could you please include which versions of gpflow, tensorflow and python you're using? (You can get these by running pip show gpflow, pip show tensorflow and python --version.)

[jl139277]

Huh, strange.
The versions I'm using are:

Python: 3.9.13,
Tensorflow: 2.15.0
Gpflow: 2.9.0

[uri-granta]

Thanks! It seems to work with tensorflow 2.12 but not with 2.15. I will investigate (and ensure our tests are also run against 2.15).

[jl139277]

Thank you very much mine works too now!
Incase somebody else runs into the same issue, I did have to update python to v 3.11 as well. Only changing the version of tensorflow to 2.12 left me with the same error.