mixle.program module¶
Compatibility shim: mixle.program moved to mixle.experimental.program.
The optimization-program approach (moves + combinators) was a reasonable exploration but not a mature surface –
its closure-taking API (minimize(lambda: loss, over=params)) is the PyTorch-style jank it meant to avoid –
so it now lives under mixle.experimental. For the common cases prefer the declarative neural surface:
from mixle.ppl import Categorical, Normal, Net, free
Categorical(logits=Net(out=10)).fit(y, given={"x": X}) # neural classification, zero closures
Normal(Net(out=1), free).fit(y, given={"x": X}) # neural mean + learned noise (the blend)
This module re-exports the old API so existing imports keep working; new code should import from
mixle.experimental.program (or use the declarative surface above).