Revision 291ae6c7dbfcbded27c604f136982a5067d14b8e authored by thevincentadam on 20 January 2020, 12:17:20 UTC, committed by thevincentadam on 20 January 2020, 12:17:20 UTC
1 parent 5dc31b8
Raw File
import functools
from typing import List, Optional, Tuple, Union

import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp
from tensorflow.python.ops import array_ops

from .config import default_float

DType = Union[np.dtype, tf.DType]
VariableData = Union[List, Tuple, np.ndarray, int, float]
TensorLike = object  # Union[tf.Tensor, tf.Variable, np.ndarray], but doesn't work with multipledispatch
Transform = tfp.bijectors.Bijector
Prior = tfp.distributions.Distribution

    return isinstance(o, Parameter)

    return (_IS_PARAMETER(o) and o.trainable)

class Module(tf.Module):
    def parameters(self):
        return tuple(self._flatten(predicate=_IS_PARAMETER))

    def trainable_parameters(self):
        return tuple(self._flatten(predicate=_IS_TRAINABLE_PARAMETER))

class Parameter(tf.Module):
    def __init__(self,
                 transform: Optional[Transform] = None,
                 prior: Optional[Prior] = None,
                 trainable: bool = True,
                 dtype: Optional[DType] = None,
                 name: Optional[str] = None):
        Unconstrained parameter representation.
        According to standard terminology `y` is always transformed representation or,
        in other words, it is constrained version of the parameter. Normally, it is hard
        to operate with unconstrained parameters. For e.g. `variance` cannot be negative,
        therefore we need positive constraint and it is natural to use constrained values.

        self._transform = transform

        if isinstance(value, tf.Variable):
            self._unconstrained = value
            unconstrained_value = self.validate_unconstrained_value(value, dtype)
            self._unconstrained = tf.Variable(unconstrained_value,
                                              dtype=dtype, name=name, trainable=trainable)

        self.prior = prior

    def log_prior(self):
        x = self.read_value()
        y = self._unconstrained

        if self.prior is not None:
            out = tf.reduce_sum(self.prior.log_prob(x))
            if self.transform is not None:
                log_det_jacobian = self.transform.forward_log_det_jacobian(y, y.shape.ndims)
                out += tf.reduce_sum(log_det_jacobian)
            return out
            return tf.convert_to_tensor(0., dtype=self.dtype)

    def value(self):
        return _to_constrained(self._unconstrained.value(), self.transform)

    def read_value(self):
        return _to_constrained(self._unconstrained.read_value(), self.transform)

    def experimental_ref(self):
        return self

    def deref(self):
        return self

    def unconstrained_variable(self):
        return self._unconstrained

    def transform(self):
        return self._transform

    def transform(self, new_transform):
        constrained_value = self.read_value()
        self._transform = new_transform

    def trainable(self):
        return self._unconstrained.trainable

    def trainable(self, flag: Union[bool, int]):
        self._unconstrained._trainable = bool(flag)

    def initial_value(self):
        return self._unconstrained.initial_value

    def validate_unconstrained_value(self, value: tf.Tensor, dtype: DType) -> tf.Tensor:
        value = _cast_to_dtype(value, dtype)
        unconstrained_value = _to_unconstrained(value, self.transform)
        message = "gpflow.Parameter: unconstrained value of passed value " \
                  "has NaN or Inf and cannot be assigned."
        return tf.debugging.assert_all_finite(unconstrained_value, message=message)

    def assign(self, value: tf.Tensor, use_locking=False, name=None, read_value=True) -> tf.Variable:
        Assigns constrained `value` to the unconstrained parameter's variable.
        It passes constrained value through parameter's transform first.

            a = Parameter(2.0, transform=tfp.bijectors.Softplus())
            b = Parameter(3.0)

            a.assign(4.0)               # `a` parameter to `2.0` value.
            a.assign(tf.constant(5.0))  # `a` parameter to `5.0` value.
            a.assign(b)                 # `a` parameter to constrained value of `b`.

        :param value: Constrained tensor-like value.
        :param use_locking: If `True`, use locking during the assignment.
        :param name: The name of the operation to be created.
        :param read_value: if True, will return something which evaluates to the new
            value of the variable; if False will return the assign op.
        unconstrained_value = self.validate_unconstrained_value(value, self.dtype)
        return self._unconstrained.assign(unconstrained_value,
                use_locking=use_locking, name=name, read_value=read_value)

    def is_tensor_like(self):
        This method means that TensorFlow's `tensor_util.is_tensor` function
        will return `True`
        return True

    def name(self):

    def initializer(self):
        return self._unconstrained.initializer

    def device(self):
        return self._unconstrained.device

    def dtype(self):
        return self._unconstrained.dtype

    def op(self):
        return self._unconstrained.op

    def shape(self):
        if self.transform is not None:
            return self.transform.forward_event_shape(self._unconstrained.shape)
        return self._unconstrained.shape

    def numpy(self):
        return self.read_value().numpy()

    def get_shape(self):
        return self.shape

    def _should_act_as_resource_variable(self):
        # needed so that Parameters are correctly identified by TensorFlow's
        # is_resource_variable() in
        pass  # only checked by TensorFlow using hasattr()

    def handle(self):
        return self._unconstrained.handle

    def __repr__(self):
        unconstrained = self.unconstrained_variable
        constrained = self.read_value()
        info = f"dtype={} " \
               f"unconstrained-shape={unconstrained.shape} " \
               f"unconstrained-numpy={unconstrained.numpy()} " \
               f"constrained-shape={constrained.shape} " \

        return f"<gpflow.Parameter {!r} {info}>"

    # Below
    # TensorFlow copy-paste code to make variable-like object to work

    def _OverloadAllOperators(cls):  # pylint: disable=invalid-name
        """Register overloads for all operators."""
        for operator in tf.Tensor.OVERLOADABLE_OPERATORS:
        # For slicing, bind getitem differently than a tensor (use SliceHelperVar
        # instead)
        # pylint: disable=protected-access
        setattr(cls, "__getitem__", array_ops._SliceHelperVar)

    def _OverloadOperator(cls, operator):  # pylint: disable=invalid-name
        """Defer an operator overload to `ops.Tensor`.

        We pull the operator out of ops.Tensor dynamically to avoid ordering issues.

            operator: string. The operator name.
        tensor_oper = getattr(tf.Tensor, operator)

        def _run_op(a, *args, **kwargs):
            # pylint: disable=protected-access
            return tensor_oper(a.read_value(), *args, **kwargs)

        functools.update_wrapper(_run_op, tensor_oper)
        setattr(cls, operator, _run_op)

    # NOTE(mrry): This enables the Variable's overloaded "right" binary
    # operators to run when the left operand is an ndarray, because it
    # accords the Variable class higher priority than an ndarray, or a
    # numpy matrix.
    # TODO(mrry): Convert this to using numpy's __numpy_ufunc__
    # mechanism, which allows more control over how Variables interact
    # with ndarrays.
    __array_priority__ = 100

tf.register_tensor_conversion_function(Parameter, lambda x, *args, **kwds: x.read_value())

def _cast_to_dtype(value: VariableData, dtype: Optional[DType] = None) -> tf.Tensor:
    if dtype is None:
        dtype = default_float()
    if tf.is_tensor(value):
        return tf.cast(value, dtype)
        return tf.convert_to_tensor(value, dtype)

def _to_constrained(value: VariableData, transform: Transform) -> tf.Tensor:
    if transform is not None:
        return transform.forward(value)
    return value

def _to_unconstrained(value: VariableData, transform: Transform) -> tf.Tensor:
    if transform is not None:
        return transform.inverse(value)
    return value
back to top