Implement truncated variables

aeppl/logprob.py

@@ -45,6 +45,12 @@ def xlogy0(m, x):
     return at.switch(at.eq(x, 0), at.switch(at.eq(m, 0), 0.0, -np.inf), m * at.log(x))
 
 
+def logdiffexp(a, b):
+    """log(exp(a) - exp(b))"""
+    # TODO: This should be a basic Aesara stabilization
+    return a + at.log1mexp(b - a)
+
+
 def logprob(rv_var, *rv_values, **kwargs):
     """Create a graph for the log-probability of a ``RandomVariable``."""
     logprob = _logprob(rv_var.owner.op, rv_values, *rv_var.owner.inputs, **kwargs)

aeppl/truncation.py

@@ -1,18 +1,33 @@
 import warnings
+from functools import singledispatch
 from typing import List, Optional
 
 import aesara.tensor as at
+import aesara.tensor.random.basic as arb
 import numpy as np
+from aesara import scan, shared
+from aesara.compile.builders import OpFromGraph
+from aesara.graph import Op
 from aesara.graph.basic import Node
 from aesara.graph.fg import FunctionGraph
 from aesara.graph.opt import local_optimizer
+from aesara.raise_op import Assert
 from aesara.scalar.basic import Clip
 from aesara.scalar.basic import clip as scalar_clip
+from aesara.scan import until
 from aesara.tensor.elemwise import Elemwise
-from aesara.tensor.var import TensorConstant
+from aesara.tensor.random.op import RandomVariable
+from aesara.tensor.var import TensorConstant, TensorVariable
 
 from aeppl.abstract import MeasurableVariable, assign_custom_measurable_outputs
-from aeppl.logprob import CheckParameterValue, _logcdf, _logprob
+from aeppl.logprob import (
+    CheckParameterValue,
+    _logcdf,
+    _logprob,
+    icdf,
+    logcdf,
+    logdiffexp,
+)
 from aeppl.opt import rv_sinking_db
 
 
@@ -123,3 +138,223 @@ def censor_logprob(op, values, base_rv, lower_bound, upper_bound, **kwargs):
         )
 
     return logprob
+
+
+class TruncatedRV(OpFromGraph):
+    """An `Op` constructed from an Aesara graph that represents a truncated univariate RV."""
+
+    default_output = 1
+    base_rv_op = None
+
+    def __init__(self, base_rv_op: Op, *args, **kwargs):
+        self.base_rv_op = base_rv_op
+        super().__init__(*args, **kwargs)
+
+
+@singledispatch
+def _truncated(op: Op, lower, upper, *params):
+    """Return the truncated equivalent of another ``RandomVariable``."""
+    raise NotImplementedError(
+        f"{op} does not have an equivalent truncated version implemented"
+    )
+
+
+def truncate(
+    rv: TensorVariable, lower=None, upper=None, max_n_steps: int = 10_000, rng=None
+):
+    """Truncate a univariate RandomVariable between lower and upper.
+
+    If lower or upper is ``None``, the variable is not truncated on that side.
+
+    Depending on dispatched implementations, this function returns either a specialized
+    `Op`, or equivalent graph representing the truncation process, via inverse CDF
+    sampling, or rejection sampling.
+
+    The argument `max_n_steps` controls the maximum number of resamples that are
+    attempted when performing rejection sampling. An Error is raised if convergence is
+    not reached after that many steps.
+
+    TODO: Add Note about updates
+    """
+
+    lower = at.as_tensor_variable(lower) if lower is not None else at.constant(-np.inf)
+    upper = at.as_tensor_variable(upper) if upper is not None else at.constant(np.inf)
+    if lower is None and upper is None:
+        raise ValueError("lower and upper cannot both be None")
+
+    if not isinstance(rv.owner.op, RandomVariable):
+        raise ValueError(f"truncation not implemented for Op {rv.owner.op}")
+
+    if rv.owner.op.ndim_supp > 0:
+        raise NotImplementedError(
+            "truncation not implemented for multivariate variables"
+        )
+
+    if rng is None:
+        rng = shared(np.random.RandomState(), borrow=True)
+
+    # Try to use specialized Op
+    try:
+        truncated_rv = _truncated(rv.owner.op, lower, upper, rng, *rv.owner.inputs[1:])
+        truncated_rv.update = (
+            truncated_rv.owner.inputs[0],
+            truncated_rv.owner.outputs[0],
+        )
+        return truncated_rv
+    except NotImplementedError:
+        pass
+
+    # Variables with `_` suffix identify dummy inputs for the OpFromGraph
+    graph_inputs = [rng, *rv.owner.inputs[1:], lower, upper]
+    graph_inputs_ = [inp.type() for inp in graph_inputs]
+    *rv_inputs_, lower_, upper_ = graph_inputs_
+
+    # Try to use inverted cdf sampling
+    try:
+        rv_ = rv.owner.op.make_node(*rv_inputs_).default_output()
+        # For left truncated discrete RVs, we need to include the whole lower bound.
+        # This may result in draws below the truncation range, if any uniform == 0
+        lower_value = lower_ - 1 if rv.owner.op.dtype.startswith("int") else lower_
+        cdf_lower_ = at.exp(logcdf(rv_, lower_value))
+        cdf_upper_ = at.exp(logcdf(rv_, upper_))
+        uniform_ = at.random.uniform(
+            cdf_lower_,
+            cdf_upper_,
+            rng=rv_inputs_[0],
+            size=rv_inputs_[1],
+        )
+        truncated_rv_ = icdf(rv_, uniform_)
+        truncated_rv = TruncatedRV(
+            base_rv_op=rv.owner.op,
+            inputs=graph_inputs_,
+            outputs=[uniform_.owner.outputs[0], truncated_rv_],
+            inline=True,
+        )(*graph_inputs)
+        truncated_rv.update = (
+            truncated_rv.owner.inputs[0],
+            truncated_rv.owner.outputs[0],
+        )
+        return truncated_rv
+    except NotImplementedError:
+        pass
+
+    # Fallback to rejection sampling
+    # TODO: Handle potential broadcast by lower / upper
+
+    # Scan forces us to use a shared variable for the RNG
+    graph_inputs = graph_inputs[1:]
+    graph_inputs_ = graph_inputs_[1:]
+    *rv_inputs_, lower_, upper_ = (rng, *graph_inputs_)
+
+    rv_ = rv.owner.op.make_node(*rv_inputs_).default_output()
+
+    def loop_fn(truncated_rv, reject_draws, lower, upper, rng, *rv_inputs):
+        # We need to set default_update for scan to generate updates
+        next_rng, new_truncated_rv = rv.owner.op.make_node(rng, *rv_inputs).outputs
+        rng.default_update = next_rng
+
+        truncated_rv = at.set_subtensor(
+            truncated_rv[reject_draws],
+            new_truncated_rv[reject_draws],
+        )
+        reject_draws = at.or_((truncated_rv < lower), (truncated_rv > upper))
+
+        return (truncated_rv, reject_draws), until(~at.any(reject_draws))
+
+    (truncated_rv_, reject_draws_), updates = scan(
+        loop_fn,
+        outputs_info=[
+            at.empty_like(rv_),
+            at.ones_like(rv_, dtype=bool),
+        ],
+        non_sequences=[lower_, upper_, *rv_inputs_],
+        n_steps=max_n_steps,
+        strict=True,
+    )
+
+    truncated_rv_ = truncated_rv_[-1]
+    convergence_ = ~at.any(reject_draws_[-1])
+    truncated_rv_ = Assert(
+        f"truncation did not converge in predefined {max_n_steps} steps"
+    )(truncated_rv_, convergence_)
+
+    # TODO: Scan does not return updates when a single step is performed, so this
+    #  will fail with max_n_steps = 1
+    truncated_rv = TruncatedRV(
+        base_rv_op=rv.owner.op,
+        inputs=graph_inputs_,
+        outputs=[tuple(updates.values())[0], truncated_rv_],
+        inline=True,
+    )(*graph_inputs)
+    truncated_rv.update = (truncated_rv.owner.inputs[-1], truncated_rv.owner.outputs[0])
+    return truncated_rv
+
+
+@_logprob.register(TruncatedRV)
+def truncated_logprob(op, values, *inputs, **kwargs):
+    (value,) = values
+
+    # rng shows up as the last input when using rejection sampling
+    if op.shared_inputs:
+        *rv_inputs, lower_bound, upper_bound, rng = inputs
+        rv_inputs = [rng, *rv_inputs]
+    else:
+        *rv_inputs, lower_bound, upper_bound = inputs
+
+    base_rv_op = op.base_rv_op
+    logp = _logprob(base_rv_op, (value,), *rv_inputs, **kwargs)
+    # For left truncated RVs, we don't want to include the lower bound in the
+    # normalization term
+    lower_bound_value = (
+        lower_bound - 1 if base_rv_op.dtype.startswith("int") else lower_bound
+    )
+    lower_logcdf = _logcdf(base_rv_op, lower_bound_value, *rv_inputs, **kwargs)
+    upper_logcdf = _logcdf(base_rv_op, upper_bound, *rv_inputs, **kwargs)
+
+    if base_rv_op.name:
+        logp.name = f"{base_rv_op}_logprob"
+        lower_logcdf.name = f"{base_rv_op}_lower_logcdf"
+        upper_logcdf.name = f"{base_rv_op}_upper_logcdf"
+
+    is_lower_bounded = not (
+        isinstance(lower_bound, TensorConstant)
+        and np.all(np.isneginf(lower_bound.value))
+    )
+    is_upper_bounded = not (
+        isinstance(upper_bound, TensorConstant) and np.all(np.isinf(upper_bound.value))
+    )
+
+    if is_lower_bounded and is_upper_bounded:
+        lognorm = logdiffexp(upper_logcdf, lower_logcdf)
+    elif is_lower_bounded:
+        lognorm = at.log1mexp(lower_logcdf)
+    elif is_upper_bounded:
+        lognorm = upper_logcdf
+    else:
+        lognorm = 0
+
+    logp = logp - lognorm
+
+    if is_lower_bounded:
+        logp = at.switch(value < lower_bound, -np.inf, logp)
+
+    if is_upper_bounded:
+        logp = at.switch(value <= upper_bound, logp, -np.inf)
+
+    if is_lower_bounded and is_upper_bounded:
+        logp = CheckParameterValue("lower_bound <= upper_bound")(
+            logp, at.all(at.le(lower_bound, upper_bound))
+        )
+
+    return logp
+
+
+@_truncated.register(arb.UniformRV)
+def uniform_truncated(op, lower, upper, rng, size, dtype, lower_orig, upper_orig):
+    return at.random.uniform(
+        at.max((lower_orig, lower)),
+        at.min((upper_orig, upper)),
+        rng=rng,
+        size=size,
+        dtype=dtype,
+    )