Source code for pymc.step_methods.slicer

#   Copyright 2024 The PyMC Developers
#
#   Licensed under the Apache License, Version 2.0 (the "License");
#   you may not use this file except in compliance with the License.
#   You may obtain a copy of the License at
#
#       http://www.apache.org/licenses/LICENSE-2.0
#
#   Unless required by applicable law or agreed to in writing, software
#   distributed under the License is distributed on an "AS IS" BASIS,
#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#   See the License for the specific language governing permissions and
#   limitations under the License.

# Modified from original implementation by Dominik Wabersich (2013)


import numpy as np
import numpy.random as nr

from pymc.blocking import RaveledVars, StatsType
from pymc.model import modelcontext
from pymc.pytensorf import compile_pymc, join_nonshared_inputs, make_shared_replacements
from pymc.step_methods.arraystep import ArrayStepShared
from pymc.step_methods.compound import Competence
from pymc.util import get_value_vars_from_user_vars
from pymc.vartypes import continuous_types

__all__ = ["Slice"]

LOOP_ERR_MSG = "max slicer iters %d exceeded"




class Slice(ArrayStepShared):
    """
    Univariate slice sampler step method.

    Parameters
    ----------
    vars: list
        List of value variables for sampler.
    w: float
        Initial width of slice (Defaults to 1).
    tune: bool
        Flag for tuning (Defaults to True).
    model: PyMC Model
        Optional model for sampling step. Defaults to None (taken from context).

    """

    name = "slice"
    default_blocked = False
    stats_dtypes_shapes = {
        "tune": (bool, []),
        "nstep_out": (int, []),
        "nstep_in": (int, []),
    }



    def __init__(self, vars=None, w=1.0, tune=True, model=None, iter_limit=np.inf, **kwargs):
        model = modelcontext(model)
        self.w = np.asarray(w).copy()
        self.tune = tune
        self.n_tunes = 0.0
        self.iter_limit = iter_limit

        if vars is None:
            vars = model.continuous_value_vars
        else:
            vars = get_value_vars_from_user_vars(vars, model)

        point = model.initial_point()
        shared = make_shared_replacements(point, vars, model)
        [logp], raveled_inp = join_nonshared_inputs(
            point=point, outputs=[model.logp()], inputs=vars, shared_inputs=shared
        )
        self.logp = compile_pymc([raveled_inp], logp)
        self.logp.trust_input = True

        super().__init__(vars, shared)




    def astep(self, apoint: RaveledVars) -> tuple[RaveledVars, StatsType]:
        # The arguments are determined by the list passed via `super().__init__(..., fs, ...)`
        q0_val = apoint.data

        if q0_val.shape != self.w.shape:
            self.w = np.resize(self.w, len(q0_val))  # this is a repmat

        nstep_out = nstep_in = 0

        q = np.copy(q0_val)
        ql = np.copy(q0_val)  # l for left boundary
        qr = np.copy(q0_val)  # r for right boundary

        logp = self.logp
        for i, wi in enumerate(self.w):
            # uniformly sample from 0 to p(q), but in log space
            y = logp(q) - nr.standard_exponential()

            # Create initial interval
            ql[i] = q[i] - nr.uniform() * wi  # q[i] + r * w
            qr[i] = ql[i] + wi  # Equivalent to q[i] + (1-r) * w

            # Stepping out procedure
            cnt = 0
            while y <= logp(ql):  # changed lt to leq  for locally uniform posteriors
                ql[i] -= wi
                cnt += 1
                if cnt > self.iter_limit:
                    raise RuntimeError(LOOP_ERR_MSG % self.iter_limit)
            nstep_out += cnt

            cnt = 0
            while y <= logp(qr):
                qr[i] += wi
                cnt += 1
                if cnt > self.iter_limit:
                    raise RuntimeError(LOOP_ERR_MSG % self.iter_limit)
            nstep_out += cnt

            cnt = 0
            q[i] = nr.uniform(ql[i], qr[i])
            while y > logp(q):  # Changed leq to lt, to accommodate for locally flat posteriors
                # Sample uniformly from slice
                if q[i] > q0_val[i]:
                    qr[i] = q[i]
                elif q[i] < q0_val[i]:
                    ql[i] = q[i]
                q[i] = nr.uniform(ql[i], qr[i])
                cnt += 1
                if cnt > self.iter_limit:
                    raise RuntimeError(LOOP_ERR_MSG % self.iter_limit)
            nstep_in += cnt

            if self.tune:
                # I was under impression from MacKays lectures that slice width can be tuned without
                # breaking markovianness. Can we do it regardless of self.tune?(@madanh)
                self.w[i] = wi * (self.n_tunes / (self.n_tunes + 1)) + (qr[i] - ql[i]) / (
                    self.n_tunes + 1
                )

            # Set qr and ql to the accepted points (they matter for subsequent iterations)
            qr[i] = ql[i] = q[i]

        if self.tune:
            self.n_tunes += 1

        stats = {
            "tune": self.tune,
            "nstep_out": nstep_out,
            "nstep_in": nstep_in,
        }

        return RaveledVars(q, apoint.point_map_info), [stats]




    @staticmethod
    def competence(var, has_grad):
        if var.dtype in continuous_types:
            if not has_grad and var.ndim == 0:
                return Competence.PREFERRED
            return Competence.COMPATIBLE
        return Competence.INCOMPATIBLE