Currently there is no way for the user to pass a RNG to the builder for weight initialisation purposes, which makes reproducibility limited to reseeding the global RNG.
Also, one cannot parallelise multiple learning curve computations with MLJ.learning_curve for this reason (this also requires a user-specifiable RNG).
Breaking the builder signature is will be breaking and I don't really see a way around this.
edit
Actually, we could have two signatures for backwards-compatibility; if the new signature is not implemented, it falls back to the old one, ie the rng specified by the user plays no role in weight intialization.
ablaom
commented
3 years ago
Currently there is no way for the user to pass a RNG to the builder for weight initialisation purposes, which makes reproducibility limited to reseeding the global RNG.
Also, one cannot parallelise multiple learning curve computations with
MLJ.learning_curvefor this reason (this also requires a user-specifiable RNG).Breaking thebuildersignature is will be breaking and I don't really see a way around this.edit Actually, we could have two signatures for backwards-compatibility; if the new signature is not implemented, it falls back to the old one, ie the
rngspecified by the user plays no role in weight intialization.cc @bkamins