tpe.py 32 KB
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961
"""
Graphical model (GM)-based optimization algorithm using Theano
"""
from __future__ import print_function
from __future__ import absolute_import
from __future__ import division
from builtins import str
from builtins import map
from builtins import zip
from builtins import range
from past.utils import old_div
import logging
import time

import numpy as np
from scipy.special import erf
from . import pyll
from .pyll import scope
from .pyll.stochastic import implicit_stochastic

from .base import miscs_to_idxs_vals
from .base import miscs_update_idxs_vals

# from .base import Trials
from . import rand

__authors__ = "James Bergstra"
__license__ = "3-clause BSD License"
__contact__ = "github.com/jaberg/hyperopt"
logger = logging.getLogger(__name__)

EPS = 1e-12

# -- default linear forgetting. don't try to change by writing this variable
# because it's captured in function default args when this file is read
DEFAULT_LF = 25


adaptive_parzen_samplers = {}


# a decorator to register functions to the dict `adaptive_parzen_samplers`
def adaptive_parzen_sampler(name):
    def wrapper(f):
        assert name not in adaptive_parzen_samplers
        adaptive_parzen_samplers[name] = f
        return f

    return wrapper


#
# These are some custom distributions
# that are used to represent posterior distributions.
#

# -- Categorical


@scope.define
def categorical_lpdf(sample, p):
    """
    """
    if sample.size:
        return np.log(np.asarray(p)[sample])
    return np.asarray([])


@scope.define
def randint_via_categorical_lpdf(sample, p):
    if sample.size:
        return np.log(np.asarray(p)[sample])
    return np.asarray([])


# -- Bounded Gaussian Mixture Model (BGMM)


@implicit_stochastic
@scope.define
def GMM1(weights, mus, sigmas, low=None, high=None, q=None, rng=None, size=()):
    """Sample from truncated 1-D Gaussian Mixture Model"""
    weights, mus, sigmas = list(map(np.asarray, (weights, mus, sigmas)))
    assert len(weights) == len(mus) == len(sigmas)
    n_samples = int(np.prod(size))
    # n_components = len(weights)
    if low is None and high is None:
        # -- draw from a standard GMM
        active = np.argmax(rng.multinomial(1, weights, (n_samples,)), axis=1)
        samples = rng.normal(loc=mus[active], scale=sigmas[active])
    else:
        # -- draw from truncated components, handling one-sided truncation
        low = float(low) if low is not None else -float("Inf")
        high = float(high) if high is not None else float("Inf")
        if low >= high:
            raise ValueError("low >= high", (low, high))
        samples = []
        while len(samples) < n_samples:
            active = np.argmax(rng.multinomial(1, weights))
            draw = rng.normal(loc=mus[active], scale=sigmas[active])
            if low <= draw < high:
                samples.append(draw)
    samples = np.reshape(np.asarray(samples), size)
    if q is None:
        return samples
    else:
        return np.round(old_div(samples, q)) * q


@scope.define
def normal_cdf(x, mu, sigma):
    top = x - mu
    bottom = np.maximum(np.sqrt(2) * sigma, EPS)
    z = old_div(top, bottom)
    return 0.5 * (1 + erf(z))


@scope.define
def GMM1_lpdf(samples, weights, mus, sigmas, low=None, high=None, q=None):
    verbose = 0
    samples, weights, mus, sigmas = list(
        map(np.asarray, (samples, weights, mus, sigmas))
    )
    if samples.size == 0:
        return np.asarray([])
    if weights.ndim != 1:
        raise TypeError("need vector of weights", weights.shape)
    if mus.ndim != 1:
        raise TypeError("need vector of mus", mus.shape)
    if sigmas.ndim != 1:
        raise TypeError("need vector of sigmas", sigmas.shape)
    assert len(weights) == len(mus) == len(sigmas)
    _samples = samples
    samples = _samples.flatten()

    if verbose:
        print("GMM1_lpdf:samples", set(samples))
        print("GMM1_lpdf:weights", weights)
        print("GMM1_lpdf:mus", mus)
        print("GMM1_lpdf:sigmas", sigmas)
        print("GMM1_lpdf:low", low)
        print("GMM1_lpdf:high", high)
        print("GMM1_lpdf:q", q)

    if low is None and high is None:
        p_accept = 1
    else:
        p_accept = np.sum(
            weights * (normal_cdf(high, mus, sigmas) - normal_cdf(low, mus, sigmas))
        )

    if q is None:
        dist = samples[:, None] - mus
        mahal = (old_div(dist, np.maximum(sigmas, EPS))) ** 2
        # mahal shape is (n_samples, n_components)
        Z = np.sqrt(2 * np.pi * sigmas ** 2)
        coef = weights / Z / p_accept
        rval = logsum_rows(-0.5 * mahal + np.log(coef))
    else:
        prob = np.zeros(samples.shape, dtype="float64")
        for w, mu, sigma in zip(weights, mus, sigmas):
            if high is None:
                ubound = samples + old_div(q, 2.0)
            else:
                ubound = np.minimum(samples + old_div(q, 2.0), high)
            if low is None:
                lbound = samples - old_div(q, 2.0)
            else:
                lbound = np.maximum(samples - old_div(q, 2.0), low)
            # -- two-stage addition is slightly more numerically accurate
            inc_amt = w * normal_cdf(ubound, mu, sigma)
            inc_amt -= w * normal_cdf(lbound, mu, sigma)
            prob += inc_amt
        rval = np.log(prob) - np.log(p_accept)

    if verbose:
        print("GMM1_lpdf:rval:", dict(list(zip(samples, rval))))

    rval.shape = _samples.shape
    return rval


# -- Mixture of Log-Normals


@scope.define
def lognormal_cdf(x, mu, sigma):
    # wikipedia claims cdf is
    # .5 + .5 erf( log(x) - mu / sqrt(2 sigma^2))
    #
    # the maximum is used to move negative values and 0 up to a point
    # where they do not cause nan or inf, but also don't contribute much
    # to the cdf.
    if len(x) == 0:
        return np.asarray([])
    if x.min() < 0:
        raise ValueError("negative arg to lognormal_cdf", x)
    olderr = np.seterr(divide="ignore")
    try:
        top = np.log(np.maximum(x, EPS)) - mu
        bottom = np.maximum(np.sqrt(2) * sigma, EPS)
        z = old_div(top, bottom)
        return 0.5 + 0.5 * erf(z)
    finally:
        np.seterr(**olderr)


@scope.define
def lognormal_lpdf(x, mu, sigma):
    # formula copied from wikipedia
    # http://en.wikipedia.org/wiki/Log-normal_distribution
    assert np.all(sigma >= 0)
    sigma = np.maximum(sigma, EPS)
    Z = sigma * x * np.sqrt(2 * np.pi)
    E = 0.5 * (old_div((np.log(x) - mu), sigma)) ** 2
    rval = -E - np.log(Z)
    return rval


@scope.define
def qlognormal_lpdf(x, mu, sigma, q):
    # casting rounds up to nearest step multiple.
    # so lpdf is log of integral from x-step to x+1 of P(x)

    # XXX: subtracting two numbers potentially very close together.
    return np.log(lognormal_cdf(x, mu, sigma) - lognormal_cdf(x - q, mu, sigma))


@implicit_stochastic
@scope.define
def LGMM1(weights, mus, sigmas, low=None, high=None, q=None, rng=None, size=()):
    weights, mus, sigmas = list(map(np.asarray, (weights, mus, sigmas)))
    n_samples = np.prod(size)
    # n_components = len(weights)
    if low is None and high is None:
        active = np.argmax(rng.multinomial(1, weights, (n_samples,)), axis=1)
        assert len(active) == n_samples
        samples = np.exp(rng.normal(loc=mus[active], scale=sigmas[active]))
    else:
        # -- draw from truncated components
        # TODO: one-sided-truncation
        low = float(low)
        high = float(high)
        if low >= high:
            raise ValueError("low >= high", (low, high))
        samples = []
        while len(samples) < n_samples:
            active = np.argmax(rng.multinomial(1, weights))
            draw = rng.normal(loc=mus[active], scale=sigmas[active])
            if low <= draw < high:
                samples.append(np.exp(draw))
        samples = np.asarray(samples)

    samples = np.reshape(np.asarray(samples), size)
    if q is not None:
        samples = np.round(old_div(samples, q)) * q
    return samples


def logsum_rows(x):
    m = x.max(axis=1)
    return np.log(np.exp(x - m[:, None]).sum(axis=1)) + m


@scope.define
def LGMM1_lpdf(samples, weights, mus, sigmas, low=None, high=None, q=None):
    samples, weights, mus, sigmas = list(
        map(np.asarray, (samples, weights, mus, sigmas))
    )
    assert weights.ndim == 1
    assert mus.ndim == 1
    assert sigmas.ndim == 1
    _samples = samples
    if samples.ndim != 1:
        samples = samples.flatten()

    if low is None and high is None:
        p_accept = 1
    else:
        p_accept = np.sum(
            weights * (normal_cdf(high, mus, sigmas) - normal_cdf(low, mus, sigmas))
        )

    if q is None:
        # compute the lpdf of each sample under each component
        lpdfs = lognormal_lpdf(samples[:, None], mus, sigmas)
        rval = logsum_rows(lpdfs + np.log(weights))
    else:
        # compute the lpdf of each sample under each component
        prob = np.zeros(samples.shape, dtype="float64")
        for w, mu, sigma in zip(weights, mus, sigmas):
            if high is None:
                ubound = samples + old_div(q, 2.0)
            else:
                ubound = np.minimum(samples + old_div(q, 2.0), np.exp(high))
            if low is None:
                lbound = samples - old_div(q, 2.0)
            else:
                lbound = np.maximum(samples - old_div(q, 2.0), np.exp(low))
            lbound = np.maximum(0, lbound)
            # -- two-stage addition is slightly more numerically accurate
            inc_amt = w * lognormal_cdf(ubound, mu, sigma)
            inc_amt -= w * lognormal_cdf(lbound, mu, sigma)
            prob += inc_amt
        rval = np.log(prob) - np.log(p_accept)
    rval.shape = _samples.shape
    return rval


#
# This is the weird heuristic ParzenWindow estimator used for continuous
# distributions in various ways.
#


@scope.define_info(o_len=3)
def adaptive_parzen_normal_orig(mus, prior_weight, prior_mu, prior_sigma):
    """
    A heuristic estimator for the mu and sigma values of a GMM
    TODO: try to find this heuristic in the literature, and cite it - Yoshua
    mentioned the term 'elastic' I think?

    mus - matrix (N, M) of M, N-dimensional component centers
    """
    mus_orig = np.array(mus)
    mus = np.array(mus)
    assert str(mus.dtype) != "object"

    if mus.ndim != 1:
        raise TypeError("mus must be vector", mus)
    if len(mus) == 0:
        mus = np.asarray([prior_mu])
        sigma = np.asarray([prior_sigma])
    elif len(mus) == 1:
        mus = np.asarray([prior_mu] + [mus[0]])
        sigma = np.asarray([prior_sigma, prior_sigma * 0.5])
    elif len(mus) >= 2:
        order = np.argsort(mus)
        mus = mus[order]
        sigma = np.zeros_like(mus)
        sigma[1:-1] = np.maximum(mus[1:-1] - mus[0:-2], mus[2:] - mus[1:-1])
        if len(mus) > 2:
            lsigma = mus[2] - mus[0]
            usigma = mus[-1] - mus[-3]
        else:
            lsigma = mus[1] - mus[0]
            usigma = mus[-1] - mus[-2]

        sigma[0] = lsigma
        sigma[-1] = usigma

        # XXX: is sorting them necessary anymore?
        # un-sort the mus and sigma
        mus[order] = mus.copy()
        sigma[order] = sigma.copy()

        if not np.all(mus_orig == mus):
            print("orig", mus_orig)
            print("mus", mus)
        assert np.all(mus_orig == mus)

        # put the prior back in
        mus = np.asarray([prior_mu] + list(mus))
        sigma = np.asarray([prior_sigma] + list(sigma))

    maxsigma = prior_sigma
    # -- magic formula:
    minsigma = old_div(prior_sigma, np.sqrt(1 + len(mus)))

    sigma = np.clip(sigma, minsigma, maxsigma)

    weights = np.ones(len(mus), dtype=mus.dtype)
    weights[0] = prior_weight

    weights = old_div(weights, weights.sum())

    return weights, mus, sigma


@scope.define
def linear_forgetting_weights(N, LF):
    assert N >= 0
    assert LF > 0
    if N == 0:
        return np.asarray([])
    elif N < LF:
        return np.ones(N)
    else:
        ramp = np.linspace(old_div(1.0, N), 1.0, num=N - LF)
        flat = np.ones(LF)
        weights = np.concatenate([ramp, flat], axis=0)
        assert weights.shape == (N,), (weights.shape, N)
        return weights


# XXX: make TPE do a post-inference pass over the pyll graph and insert
# non-default LF argument


@scope.define_info(o_len=3)
def adaptive_parzen_normal(mus, prior_weight, prior_mu, prior_sigma, LF=DEFAULT_LF):
    """
    mus - matrix (N, M) of M, N-dimensional component centers
    """
    mus = np.array(mus)
    assert str(mus.dtype) != "object"

    if mus.ndim != 1:
        raise TypeError("mus must be vector", mus)
    if len(mus) == 0:
        srtd_mus = np.asarray([prior_mu])
        sigma = np.asarray([prior_sigma])
        prior_pos = 0
    elif len(mus) == 1:
        if prior_mu < mus[0]:
            prior_pos = 0
            srtd_mus = np.asarray([prior_mu, mus[0]])
            sigma = np.asarray([prior_sigma, prior_sigma * 0.5])
        else:
            prior_pos = 1
            srtd_mus = np.asarray([mus[0], prior_mu])
            sigma = np.asarray([prior_sigma * 0.5, prior_sigma])
    elif len(mus) >= 2:

        # create new_mus, which is sorted, and in which
        # the prior has been inserted
        order = np.argsort(mus)
        prior_pos = np.searchsorted(mus[order], prior_mu)
        srtd_mus = np.zeros(len(mus) + 1)
        srtd_mus[:prior_pos] = mus[order[:prior_pos]]
        srtd_mus[prior_pos] = prior_mu
        srtd_mus[prior_pos + 1 :] = mus[order[prior_pos:]]
        sigma = np.zeros_like(srtd_mus)
        sigma[1:-1] = np.maximum(
            srtd_mus[1:-1] - srtd_mus[0:-2], srtd_mus[2:] - srtd_mus[1:-1]
        )
        lsigma = srtd_mus[1] - srtd_mus[0]
        usigma = srtd_mus[-1] - srtd_mus[-2]
        sigma[0] = lsigma
        sigma[-1] = usigma

    if LF and LF < len(mus):
        unsrtd_weights = linear_forgetting_weights(len(mus), LF)
        srtd_weights = np.zeros_like(srtd_mus)
        assert len(unsrtd_weights) + 1 == len(srtd_mus)
        srtd_weights[:prior_pos] = unsrtd_weights[order[:prior_pos]]
        srtd_weights[prior_pos] = prior_weight
        srtd_weights[prior_pos + 1 :] = unsrtd_weights[order[prior_pos:]]

    else:
        srtd_weights = np.ones(len(srtd_mus))
        srtd_weights[prior_pos] = prior_weight

    # -- magic formula:
    maxsigma = old_div(prior_sigma, 1.0)
    minsigma = old_div(prior_sigma, min(100.0, (1.0 + len(srtd_mus))))

    sigma = np.clip(sigma, minsigma, maxsigma)

    sigma[prior_pos] = prior_sigma
    assert prior_sigma > 0
    assert maxsigma > 0
    assert minsigma > 0
    assert np.all(sigma > 0), (sigma.min(), minsigma, maxsigma)

    srtd_weights /= srtd_weights.sum()

    return srtd_weights, srtd_mus, sigma


#
# Adaptive Parzen Samplers
# These produce conditional estimators for various prior distributions
#
# NOTE: These are actually used in a fairly complicated way.
# They are actually returning pyll.Apply AST (Abstract Syntax Tree) objects.
# This AST is then manipulated and the corresponding _lpdf function is called
# (e.g  GMM1_lpdf)
#
# Please see the build_posterior function for details

# -- Uniform


@adaptive_parzen_sampler("uniform")
def ap_uniform_sampler(obs, prior_weight, low, high, size=(), rng=None):
    prior_mu = 0.5 * (high + low)
    prior_sigma = 1.0 * (high - low)
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        obs, prior_weight, prior_mu, prior_sigma
    )
    return scope.GMM1(
        weights, mus, sigmas, low=low, high=high, q=None, size=size, rng=rng
    )


@adaptive_parzen_sampler("quniform")
def ap_quniform_sampler(obs, prior_weight, low, high, q, size=(), rng=None):
    prior_mu = 0.5 * (high + low)
    prior_sigma = 1.0 * (high - low)
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        obs, prior_weight, prior_mu, prior_sigma
    )
    return scope.GMM1(weights, mus, sigmas, low=low, high=high, q=q, size=size, rng=rng)


@adaptive_parzen_sampler("loguniform")
def ap_loguniform_sampler(obs, prior_weight, low, high, size=(), rng=None):
    prior_mu = 0.5 * (high + low)
    prior_sigma = 1.0 * (high - low)
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        scope.log(obs), prior_weight, prior_mu, prior_sigma
    )
    rval = scope.LGMM1(weights, mus, sigmas, low=low, high=high, size=size, rng=rng)
    return rval


@adaptive_parzen_sampler("qloguniform")
def ap_qloguniform_sampler(obs, prior_weight, low, high, q, size=(), rng=None):
    prior_mu = 0.5 * (high + low)
    prior_sigma = 1.0 * (high - low)
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        scope.log(
            # -- map observations that were quantized to be below exp(low)
            #    (particularly 0) back up to exp(low) where they will
            #    interact in a reasonable way with the AdaptiveParzen
            #    thing.
            scope.maximum(
                obs,
                scope.maximum(  # -- protect against exp(low) underflow
                    EPS, scope.exp(low)
                ),
            )
        ),
        prior_weight,
        prior_mu,
        prior_sigma,
    )
    return scope.LGMM1(weights, mus, sigmas, low, high, q=q, size=size, rng=rng)


# -- Normal


@adaptive_parzen_sampler("normal")
def ap_normal_sampler(obs, prior_weight, mu, sigma, size=(), rng=None):
    weights, mus, sigmas = scope.adaptive_parzen_normal(obs, prior_weight, mu, sigma)
    return scope.GMM1(weights, mus, sigmas, size=size, rng=rng)


@adaptive_parzen_sampler("qnormal")
def ap_qnormal_sampler(obs, prior_weight, mu, sigma, q, size=(), rng=None):
    weights, mus, sigmas = scope.adaptive_parzen_normal(obs, prior_weight, mu, sigma)
    return scope.GMM1(weights, mus, sigmas, q=q, size=size, rng=rng)


@adaptive_parzen_sampler("lognormal")
def ap_loglognormal_sampler(obs, prior_weight, mu, sigma, size=(), rng=None):
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        scope.log(obs), prior_weight, mu, sigma
    )
    rval = scope.LGMM1(weights, mus, sigmas, size=size, rng=rng)
    return rval


@adaptive_parzen_sampler("qlognormal")
def ap_qlognormal_sampler(obs, prior_weight, mu, sigma, q, size=(), rng=None):
    log_obs = scope.log(scope.maximum(obs, EPS))
    weights, mus, sigmas = scope.adaptive_parzen_normal(
        log_obs, prior_weight, mu, sigma
    )
    rval = scope.LGMM1(weights, mus, sigmas, q=q, size=size, rng=rng)
    return rval


# -- Categorical


@adaptive_parzen_sampler("randint")
def ap_randint_sampler(
    obs, prior_weight, low, high=None, size=(), rng=None, LF=DEFAULT_LF
):
    # randint can be seen as a categorical with high - low categories
    weights = scope.linear_forgetting_weights(scope.len(obs), LF=LF)
    # if high is None, then low represents high and there is no offset
    domain_size = low if high is None else high - low
    offset = pyll.Literal(0) if high is None else low
    counts = scope.bincount(obs, offset=offset, minlength=domain_size, weights=weights)
    # -- add in some prior pseudocounts
    pseudocounts = counts + prior_weight
    random_variable = scope.randint_via_categorical(
        old_div(pseudocounts, scope.sum(pseudocounts)), size=size, rng=rng
    )
    return random_variable


@scope.define
def tpe_cat_pseudocounts(counts, prior_weight, p, size):
    if np.prod(size) == 0:
        return []
    if p.ndim == 2:
        assert np.all(p == p[0])
        p = p[0]
    pseudocounts = counts + p.size * (prior_weight * p)
    return old_div(pseudocounts, np.sum(pseudocounts))


@adaptive_parzen_sampler("categorical")
def ap_categorical_sampler(obs, prior_weight, p, size=(), rng=None, LF=DEFAULT_LF):
    weights = scope.linear_forgetting_weights(scope.len(obs), LF=LF)
    # in order to support pchoice here, we need to find the size of p,
    # but p can have p.ndim == 2, so we pass p to bincount and unpack it
    # (if required) there
    counts = scope.bincount(obs, p=p, weights=weights)
    pseudocounts = scope.tpe_cat_pseudocounts(counts, prior_weight, p, size)
    return scope.categorical(pseudocounts, size=size, rng=rng)


#
# Posterior clone performs symbolic inference on the pyll graph of priors.
#


@scope.define_info(o_len=2)
def ap_split_trials(o_idxs, o_vals, l_idxs, l_vals, gamma, gamma_cap=DEFAULT_LF):
    """Split the elements of `o_vals` (observations values) into two groups: those for
    trials whose losses (`l_vals`) were above gamma, and those below gamma. Note that
    only unique elements are returned, so the total number of returned elements might
    be lower than `len(o_vals)`
    """
    o_idxs, o_vals, l_idxs, l_vals = list(
        map(np.asarray, [o_idxs, o_vals, l_idxs, l_vals])
    )

    # XXX if this is working, refactor this sort for efficiency

    # Splitting is done this way to cope with duplicate loss values.
    n_below = min(int(np.ceil(gamma * np.sqrt(len(l_vals)))), gamma_cap)
    l_order = np.argsort(l_vals)

    keep_idxs = set(l_idxs[l_order[:n_below]])
    below = [v for i, v in zip(o_idxs, o_vals) if i in keep_idxs]

    keep_idxs = set(l_idxs[l_order[n_below:]])
    above = [v for i, v in zip(o_idxs, o_vals) if i in keep_idxs]

    return np.asarray(below), np.asarray(above)


@scope.define
def broadcast_best(samples, below_llik, above_llik):
    if len(samples):
        score = below_llik - above_llik
        if len(samples) != len(score):
            raise ValueError()
        best = np.argmax(score)
        return [samples[best]] * len(samples)
    else:
        return []


def build_posterior(
    specs,
    prior_idxs,
    prior_vals,
    obs_idxs,
    obs_vals,
    obs_loss_idxs,
    obs_loss_vals,
    oloss_gamma,
    prior_weight,
):
    """
    This method clones a posterior inference graph by iterating forward in
    topological order, and replacing prior random-variables (prior_idxs, prior_vals)
    with new posterior distributions (post_specs, post_idxs, post_vals) that make use
    of observations (obs_idxs, obs_vals).

    """
    assert all(
        isinstance(arg, pyll.Apply)
        for arg in [obs_loss_idxs, obs_loss_vals, oloss_gamma]
    )
    assert set(prior_idxs.keys()) == set(prior_vals.keys())

    expr = pyll.as_apply([specs, prior_idxs, prior_vals])
    nodes = pyll.dfs(expr)

    # build the joint posterior distribution as the values in this memo
    memo = {}
    # map prior RVs to observations
    obs_memo = {}

    for nid in prior_vals:
        # construct the leading args for each call to adaptive_parzen_sampler
        # which will permit the "adaptive parzen samplers" to adapt to the
        # correct samples.
        obs_below, obs_above = scope.ap_split_trials(
            obs_idxs[nid], obs_vals[nid], obs_loss_idxs, obs_loss_vals, oloss_gamma
        )
        obs_memo[prior_vals[nid]] = [obs_below, obs_above]
    for node in nodes:
        if node not in memo:
            new_inputs = [memo[arg] for arg in node.inputs()]
            if node in obs_memo:
                # -- this case corresponds to an observed Random Var
                # node.name is a distribution like "normal", "randint", etc.
                obs_below, obs_above = obs_memo[node]
                aa = [memo[a] for a in node.pos_args]
                fn = adaptive_parzen_samplers[node.name]
                b_args = [obs_below, prior_weight] + aa
                named_args = {kw: memo[arg] for (kw, arg) in node.named_args}
                b_post = fn(*b_args, **named_args)
                a_args = [obs_above, prior_weight] + aa
                a_post = fn(*a_args, **named_args)

                # fn is a function e.g ap_uniform_sampler, ap_normal_sampler, etc
                # b_post and a_post are pyll.Apply objects that are
                # AST (Abstract Syntax Trees).  They create the distribution,
                # (e.g. using adaptive_parzen_normal), and then
                # call a function to sample randomly from that distribution
                # (e.g. using scope.GMM1) which return those samples.
                #
                # However we are only interested in using the samples from b_post.
                # This code looks at the AST and grabs the function name that we used
                # for sampling (e.g. scope.GMM1)   and modifies it, e.g. to
                # "scope.GMM1_lpdf". It then calls this function, passing in the
                # samples as the first parameter.a_args
                #
                # The result is that we are effectively calling, for example:
                # below_llik = GMM1_lpdf( b_post, *adaptive_parzen_normal(obs_below, ...))
                # above_llik = GMM1_lpdf( b_post, *adaptive_parzen_normal(obs_above, ...))

                assert a_post.name == b_post.name
                fn_lpdf = getattr(scope, a_post.name + "_lpdf")
                a_kwargs = {
                    n: a for n, a in a_post.named_args if n not in ("rng", "size")
                }
                b_kwargs = {
                    n: a for n, a in b_post.named_args if n not in ("rng", "size")
                }

                # calculate the log likelihood of b_post under both distributions
                below_llik = fn_lpdf(*([b_post] + b_post.pos_args), **b_kwargs)
                above_llik = fn_lpdf(*([b_post] + a_post.pos_args), **a_kwargs)
                # compute new_node based on below & above log likelihood
                new_node = scope.broadcast_best(b_post, below_llik, above_llik)
            elif hasattr(node, "obj"):
                # -- keep same literals in the graph
                new_node = node
            else:
                # -- this case is for all the other stuff in the graph
                new_node = node.clone_from_inputs(new_inputs)
            memo[node] = new_node

    post_idxs = {nid: memo[idxs] for nid, idxs in prior_idxs.items()}
    post_vals = {nid: memo[vals] for nid, vals in prior_vals.items()}
    return post_idxs, post_vals


# TODO: is this used?
# @scope.define
# def idxs_prod(full_idxs, idxs_by_label, llik_by_label):
#     """Add all of the  log-likelihoods together by id.
#
#     Example arguments:
#     full_idxs = [0, 1, ... N-1]
#     idxs_by_label = {'node_a': [1, 3], 'node_b': [3]}
#     llik_by_label = {'node_a': [0.1, -3.3], node_b: [1.0]}
#
#     This would return N elements: [0, 0.1, 0, -2.3, 0, 0, ... ]
#     """
#     assert len(set(full_idxs)) == len(full_idxs)
#     full_idxs = list(full_idxs)
#     rval = np.zeros(len(full_idxs))
#     pos_of_tid = dict(list(zip(full_idxs, list(range(len(full_idxs))))))
#     assert set(idxs_by_label.keys()) == set(llik_by_label.keys())
#     for nid in idxs_by_label:
#         idxs = idxs_by_label[nid]
#         llik = llik_by_label[nid]
#         assert np.all(np.asarray(idxs) > 1)
#         assert len(set(idxs)) == len(idxs)
#         assert len(idxs) == len(llik)
#         for ii, ll in zip(idxs, llik):
#             rval[pos_of_tid[ii]] += ll
#     return rval


_default_prior_weight = 1.0

# -- suggest best of this many draws on every iteration
_default_n_EI_candidates = 24

# -- gamma * sqrt(n_trials) is fraction of to use as good
_default_gamma = 0.25

_default_n_startup_jobs = 20

_default_linear_forgetting = DEFAULT_LF


def build_posterior_wrapper(domain, prior_weight, gamma):
    """
    Calls build_posterior
    Args:
        domain (hyperopt.base.Domain): contains info about the obj function and the hp
            space passed to fmin
        prior_weight (float): smoothing factor for counts, to avoid having 0 prob
        # TODO: consider renaming or improving documentation for suggest
        gamma (float): the threshold to split between l(x) and g(x), see eq. 2 in
            https://papers.nips.cc/paper/4443-algorithms-for-hyper-parameter-optimization.pdf

    Returns:

    """

    # -- these dummy values will be replaced in build_posterior() and never used
    observed = {"idxs": pyll.Literal(), "vals": pyll.Literal()}
    observed_loss = {"idxs": pyll.Literal(), "vals": pyll.Literal()}

    posterior = build_posterior(
        # -- vectorized clone of bandit template
        domain.vh.v_expr,
        # -- this dict and next represent prior dists
        domain.vh.idxs_by_label(),
        domain.vh.vals_by_label(),
        observed["idxs"],
        observed["vals"],
        observed_loss["idxs"],
        observed_loss["vals"],
        pyll.Literal(gamma),
        pyll.Literal(float(prior_weight)),
    )

    return observed, observed_loss, posterior


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())
    losses = [best_docs_loss[tid] for tid, doc in tid_docs]
    tids, docs = list(zip(*tid_docs)) if tid_docs else ([], [])

    if verbose:
        if docs:
            s = "%i/%i trials with best loss %f" % (len(docs), len(trials), min(losses))
        else:
            s = "0 trials"
        logger.info("TPE using %s" % s)

    if len(docs) < n_startup_jobs:
        # N.B. THIS SEEDS THE RNG BASED ON THE new_id
        return rand.suggest(new_ids, domain, trials, seed)

    # Sample and compute log-probability.
    first_new_id = new_ids[0]
    if tids:
        # -- the +2 co-ordinates with an assertion above
        #    to ensure that fake ids are used during sampling
        #    TODO: not sure what assertion this refers to...
        fake_id_0 = max(max(tids), first_new_id) + 2
    else:
        # -- weird - we're running the TPE algo from scratch
        assert n_startup_jobs <= 0
        fake_id_0 = first_new_id + 2

    fake_ids = list(range(fake_id_0, fake_id_0 + n_EI_candidates))

    # -- this dictionary will map pyll nodes to the values
    #    they should take during the evaluation of the pyll program
    memo = {domain.s_new_ids: fake_ids, domain.s_rng: np.random.RandomState(seed)}

    memo[observed_loss["idxs"]] = tids
    memo[observed_loss["vals"]] = losses

    observed_idxs_dict, observed_vals_dict = miscs_to_idxs_vals(
        [doc["misc"] for doc in docs], keys=list(domain.params.keys())
    )
    memo[observed["idxs"]] = observed_idxs_dict
    memo[observed["vals"]] = observed_vals_dict

    # evaluate `n_EI_candidates` pyll nodes in `posterior` using `memo`
    # TODO: it seems to return idxs, vals, all the same. Is this correct?
    idxs, vals = pyll.rec_eval(posterior, memo=memo, print_node_on_error=False)

    # hack to add offset again for randint params
    for label, param in domain.params.items():
        if param.name == "randint" and len(param.pos_args) == 2:
            offset = param.pos_args[0].obj
            vals[label] = [val + offset for val in vals[label]]

    # -- retrieve the best of the samples and form the return tuple

    rval_specs = [None]  # specs are deprecated since build_posterior makes all the same
    rval_results = [domain.new_result()]
    rval_miscs = [{"tid": first_new_id, "cmd": domain.cmd, "workdir": domain.workdir}]

    miscs_update_idxs_vals(
        rval_miscs,
        idxs,
        vals,
        idxs_map={fake_ids[0]: first_new_id},
        assert_all_vals_used=False,
    )
    # return the doc for the best new trial
    return trials.new_trial_docs([first_new_id], rval_specs, rval_results, rval_miscs)