Shape mismatch when running predictions with Gaussian Processes

Questions version agnostic
3

Hi, from reading your post, it looks like you may be after some sort of post training validation-f if i misread i apologize

One thing that has helped me tremendously in my own workflow is to do something very similar to @Dekermanjian’s recommendation: Name your spatial and temporal variables, train your model on your ‘train set’-and then after performing your inference/checks, extend your original data. This notebook: Gaussian Processes: HSGP Advanced Usage — PyMC example gallery has an example of doing that. Here’s the most relevant piece of code after you have trained your original model:

with model:
    pm.set_data({"X": x_full[:, None]})

    idata.extend(
        pm.sample_posterior_predictive(
            idata,
            var_names=["f_mu", "f"],
            predictions=True,
            compile_kwargs={"mode": "NUMBA"},
            random_seed=rng,
        ),
    )

Here, you can set the training data container 'X" to contain all of your points

I realize I’m abstracting away probably the most important bits like defining your kronecker structure, defining hyperparamters ect etc, but if you’re after some code to help you work backwards-this block might serve a good reference!

1 Like