Tensorboard logging fixes #544
Conversation
trieste/models/gpflow/interface.py
Outdated
| for i, lengthscale in enumerate(lengthscales): | ||
| logging.scalar(f"kernel.lengthscale[{i}]", lengthscale) | ||
| kernel = self.get_kernel() | ||
| if isinstance(kernel, gpflow.kernels.Stationary): |
There was a problem hiding this comment.
hard to say if this would be a good enough check, I don't know GPflow well enough - perhaps add a test for this?
it might be good enough to support only the basic case, but I think we probably can do something more generic, like fetching all trainable variables? I think they also come with names so we can still make it descriptive enough, no?
There was a problem hiding this comment.
The check should be sufficient: gpflow.kernels.Stationary is where the lengthscales and variance parameters are set.
Makes sense to add a test. I'll see how much we can extract generically, but it might be tricky inferring types, shapes, etc. If it's too much work this may have to do for now (though I can add it to a WIBNI issue).
| for i, lengthscale in enumerate(lengthscales): | ||
| logging.scalar(f"kernel.lengthscale[{i}]", lengthscale) | ||
| kernel = self.get_kernel() | ||
| components = _merge_leaf_components(leaf_components(kernel)) |
There was a problem hiding this comment.
wasnt aware of this, great that you have worked it out!
No description provided.