mixle.experimental.wake_sleep module

P12 (experimental) – wake-sleep library learning over a model-structure grammar.

DreamCoder’s loop, in miniature, for mixle’s structure grammar:

  • WAKE – solve each modeling task by greedy search over a library of structure atoms (each atom is a group of basis components), scoring candidates by description length (MDL).

  • SLEEP-ABSTRACTION – compress the solutions: a set of primitives that recurs across many solved tasks is anti-unified into a single reusable library fragment.

  • Re-solving tasks that need that motif now reaches it in ONE search step (pick the fragment) instead of composing it primitive by primitive – so search cost drops.

The card’s claim: on a corpus of related tasks sharing a latent motif, the wake-sleep loop discovers the motif as a fragment and cuts held-out search cost, with the fragment reused across the corpus. The kill criterion is measured, not assumed: if no fragment is reused enough, or the speedup does not materialize, the receipt says so.

Exploratory mixle.experimental code (P12 card).

class Atom(name, cols)[source]

Bases: object

A library atom: a named group of feature-column indices (primitive = 1 col; fragment = many).

Parameters:
class SearchResult(selected: 'list[str]', used_cols: 'tuple[int, ...]', n_evals: 'int')[source]

Bases: object

Parameters:
greedy_search(y, features, library, *, penalty=2.0)[source]

Forward-selection structure search; returns the chosen atoms and the number of fit evaluations.

Parameters:
Return type:

SearchResult

primitive_library(n_primitives)[source]

One atom per basis component.

Parameters:

n_primitives (int)

Return type:

list[Atom]

make_task(rng, features, motif, n_specific)[source]

A regression target built from the shared motif plus a few task-specific components.

The motif components carry a strong coefficient (so greedy reliably recovers all of them), the task-specific ones a weaker coefficient.

Parameters:
Return type:

ndarray

abstract_fragment(solutions, *, min_support=0.6, min_size=2)[source]

Return the largest exact itemset jointly present in at least min_support of solutions.

Ties are resolved by larger exact support and then lexicographically. Every returned fragment is therefore a subset of at least the required number of observed solutions; individually frequent columns are never unioned into a fragment unless that exact union is also jointly frequent.

Parameters:
Return type:

Atom | None

class WakeSleepReport(fragment: 'Atom | None', flat_evals: 'float', library_evals: 'float', speedup: 'float', fragment_reuse: 'int', history: 'list' = <factory>)[source]

Bases: object

Parameters:

  • fragment (Atom | None)

  • flat_evals (float)

  • library_evals (float)

  • speedup (float)

  • fragment_reuse (int)

  • history (list)

wake_sleep(*, n_train=30, n_heldout=30, n_primitives=16, n_t=160, motif=(2, 5, 8, 11, 3, 13, 6), n_specific=1, seed=0)[source]

Run one wake-sleep round and measure held-out search cost with vs without the learned fragment.

Parameters:
Return type:

WakeSleepReport