Function suggest

hyperopt/tpe.py:830–958

Given previous trials and the domain, suggest the best expected hp point according to the TPE-EI algo Args: prior_weight( n_startup_jobs: n_EI_candidates: gamma: verbose: Returns:

(
    new_ids,
    domain,
    trials,
    seed,
    prior_weight=_default_prior_weight,
    n_startup_jobs=_default_n_startup_jobs,
    n_EI_candidates=_default_n_EI_candidates,
    gamma=_default_gamma,
    verbose=True,
)

def suggest(
    new_ids,
    domain,
    trials,
    seed,
    prior_weight=_default_prior_weight,
    n_startup_jobs=_default_n_startup_jobs,
    n_EI_candidates=_default_n_EI_candidates,
    gamma=_default_gamma,
    verbose=True,
):
    """
    Given previous trials and the domain, suggest the best expected hp point
    according to the TPE-EI algo


    Args:
        prior_weight(
        n_startup_jobs:
        n_EI_candidates:
        gamma:
        verbose:

    Returns:

    """

    t0 = time.time()
    # use build_posterior_wrapper to create the pyll nodes
    observed, observed_loss, posterior = build_posterior_wrapper(
        domain, prior_weight, gamma
    )
    tt = time.time() - t0
    if verbose:
        logger.info("build_posterior_wrapper took %f seconds" % tt)

    # Loop over previous trials to collect best_docs and best_docs_loss
    best_docs = dict()
    best_docs_loss = dict()
    for doc in trials.trials:

        # get either these docs own tid or the one that it's from
        tid = doc["misc"].get("from_tid", doc["tid"])

        # associate infinite loss to new/running/failed jobs
        loss = doc["result"].get("loss")
        loss = float("inf") if loss is None else float(loss)

        # if set, update loss for this tid if it's higher than current loss
        # otherwise, set it
        best_docs_loss.setdefault(tid, loss)
        if loss <= best_docs_loss[tid]:
            best_docs_loss[tid] = loss
            best_docs[tid] = doc

    # -- sort docs by order of suggestion
    #    so that linear_forgetting removes the oldest ones
    tid_docs = sorted(best_docs.items())