You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It seems that the ComponentArray constructor is not differentiable.
Context: I have a two-step loss function, where I do some upfront work to estimate some parameters from the data, then predict using those parameters and others, so I'm trying to build a single parameter array combining the sets of parameters.
Oh interesting. I'll have to give that one some thought.
It's also worth noting that the normal ComponentArray(a=something, b=something_else, ...) style of constructor is pretty slow because it has to recurse through the structure and build up the inner array as it goes. It wasn't really intended to be used in a hot loops. The other style of constructor, ComponentArray(data, axes) is both differentiable and very fast, but right now it's kinda hard to add fields this way because the Axis interface is so clunky.
I've had it on the back of my mind for a while to add a ComponentArray(θ; other=other_params, ...) style method for quickly creating new ComponentArrays with additional component fields. This gives me good reason to finally implement that.
I try to take Saturdays off, so either tomorrow night or the Monday I'll try to get to it.
It seems that the
ComponentArray
constructor is not differentiable.Context: I have a two-step loss function, where I do some upfront work to estimate some parameters from the data, then predict using those parameters and others, so I'm trying to build a single parameter array combining the sets of parameters.
Reproducible code sample, adapted from docs
Error message:
The text was updated successfully, but these errors were encountered: