fit_ensemble

Contents

fit_ensemble#

hybrid_aif.train.fit_ensemble(
system,
params,
learn_state,
batch,
*,
row=None,
member_weights=None,
lr=0.0001,
lr_by_slot=None,
clip=100.0,
adam_eps=1e-05,
)[source]#

One gradient step of ensemble learning: (params, learn_state, losses).

Pure, jit-able, resumable; steps only its row’s scoring:<row> address and threads every other LearnState entry through. row names which of ROWS to step; None resolves the one SystemParams carries and raises if it carries both. member_weights goes to ensemble_loss(). clip is a global-norm clip over all members (None disables it); adam_eps builds the optimizer. lr_by_slot is the rate grammar, a static 0.0 skipping the update at trace time.

Raises ValueError when learn_state is None or lacks this row’s optimizer entry (never re-seeded), and when row cannot be resolved.