I'd like to export a 'lightweight' version of my GP model. The use case would be to make an interactive online version. Right now if I pickle my GP instance it's around 8mb, which is a little too large.
My first thought is to create a new lightweight predict method that just accepts the conditioning data and the new points and returns the predicted mean (ignoring the covariance matrix). As far as I can tell from the existing predict method, the main components required to produce a new arbitrary prediction are:
the conditioning data (values and coordinates)
the solver
the kernel
some other lightweight methods in the GP class (e.g. _call_mean)
Anything else I'm missing? Or would you recommend another approach?
I'd like to export a 'lightweight' version of my GP model. The use case would be to make an interactive online version. Right now if I pickle my
GPinstance it's around 8mb, which is a little too large.My first thought is to create a new lightweight
predictmethod that just accepts the conditioning data and the new points and returns the predicted mean (ignoring the covariance matrix). As far as I can tell from the existingpredictmethod, the main components required to produce a new arbitrary prediction are:GPclass (e.g._call_mean)Anything else I'm missing? Or would you recommend another approach?