Kernel dimensionality bug in examples. about pydeepgp HOT 4 CLOSED

Hi,

When I run the latest code I get this result:

# RMSE DGP               : 2.5763593393648176
# RMSE DGP (with samples): 2.577221154102472
# RMSE GP                : 3.033963097677277

From my training log, it is only the 2nd optimization that differs, getting a better objective f after optimization. The rest of the objectives after optimization are similar to the ones of your log.

The performance gets almost identical to the regular GP when I set back_constraint=True (it's False by default), since it's an overkill of a model for this particular data.

In any case, this example is meant to show how one can use deepgp for regression, as opposed to demonstrate that it beats GP. For the kinds of data that a GP can already model well (smooth, stationary etc) one should expect it to have similar performance with deepgp.

from pydeepgp.

What I meant is that RBF kernel dimension number for the "hidden" layer (i.e. kern2 variable), which is currently equals to Q, actually should be X_tr.shape[1]. It seems like a bug in the example code, or I may completely misunderstand how DGPs work.

In any case, it makes sense for GP-LVM and DGP to yield similar results on the example data. Thanks!

from pydeepgp.

Oh I see, you're right, thanks for catching this.

When setting the dimensionality in kern2 as Q, being incompatible with input dimensionality, GPy automatically sets the active_dims for this kernel to be the first Q (you can see that by typing m.layers[1].kern.rbf.active_dims) meaning that this kernel is forced to only look at a small subset of dimensions and the rest are handled by the Bias kernel. While this is a valid model, it is certainly not what this code was intended to do :) Although it seemed that this simpler mode did give better results in this particular case!

from pydeepgp.

Thank you for clarification!

from pydeepgp.