# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. # # 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. import enum import math import typing import paddle import paddle.nn.functional as F from paddle.distribution import ( constraint, distribution, transformed_distribution, variable, ) __all__ = [ # noqa 'Transform', 'AbsTransform', 'AffineTransform', 'ChainTransform', 'ExpTransform', 'IndependentTransform', 'PowerTransform', 'ReshapeTransform', 'SigmoidTransform', 'SoftmaxTransform', 'StackTransform', 'StickBreakingTransform', 'TanhTransform', ] class Type(enum.Enum): """Mapping type of a transformation.""" BIJECTION = 'bijection' # bijective(injective and surjective) INJECTION = 'injection' # injective-only SURJECTION = 'surjection' # surjective-only OTHER = 'other' # general, neither injective nor surjective @classmethod def is_injective(cls, _type): """Both bijection and injection are injective mapping.""" return _type in (cls.BIJECTION, cls.INJECTION) class Transform: r"""Base class for the transformations of random variables. ``Transform`` can be used to represent any differentiable and injective function from the subset of :math:`R^n` to subset of :math:`R^m`, generally used for transforming a random sample generated by ``Distribution`` instance. Suppose :math:`X` is a K-dimensional random variable with probability density function :math:`p_X(x)`. A new random variable :math:`Y = f(X)` may be defined by transforming :math:`X` with a suitably well-behaved funciton :math:`f`. It suffices for what follows to note that if `f` is one-to-one and its inverse :math:`f^{-1}` have a well-defined Jacobian, then the density of :math:`Y` is .. math:: p_Y(y) = p_X(f^{-1}(y)) |det J_{f^{-1}}(y)| where det is the matrix determinant operation and :math:`J_{f^{-1}}(y)` is the Jacobian matrix of :math:`f^{-1}` evaluated at :math:`y`. Taking :math:`x = f^{-1}(y)`, the Jacobian matrix is defined by .. math:: J(y) = \begin{bmatrix} {\frac{\partial x_1}{\partial y_1}} &{\frac{\partial x_1}{\partial y_2}} &{\cdots} &{\frac{\partial x_1}{\partial y_K}} \\ {\frac{\partial x_2}{\partial y_1}} &{\frac{\partial x_2} {\partial y_2}}&{\cdots} &{\frac{\partial x_2}{\partial y_K}} \\ {\vdots} &{\vdots} &{\ddots} &{\vdots}\\ {\frac{\partial x_K}{\partial y_1}} &{\frac{\partial x_K}{\partial y_2}} &{\cdots} &{\frac{\partial x_K}{\partial y_K}} \end{bmatrix} A ``Transform`` can be characterized by three operations: #. forward Forward implements :math:`x \rightarrow f(x)`, and is used to convert one random outcome into another. #. inverse Undoes the transformation :math:`y \rightarrow f^{-1}(y)`. #. log_det_jacobian The log of the absolute value of the determinant of the matrix of all first-order partial derivatives of the inverse function. Subclass typically implement follow methods: * _forward * _inverse * _forward_log_det_jacobian * _inverse_log_det_jacobian (optional) If the transform changes the shape of the input, you must also implemented: * _forward_shape * _inverse_shape """ _type = Type.INJECTION def __init__(self): super().__init__() @classmethod def _is_injective(cls): """Is the transformation type one-to-one or not. Returns: bool: ``True`` denotes injective. ``False`` denotes non-injective. """ return Type.is_injective(cls._type) def __call__(self, input): """Make this instance as a callable object. The return value is depening on the input type. * If the input is a ``Tensor`` instance, return ``self.forward(input)`` . * If the input is a ``Distribution`` instance, return ``TransformedDistribution(base=input, transforms=[self])`` . * If the input is a ``Transform`` instance, return ``ChainTransform([self, input])`` . Args: input (Tensor|Distribution|Transform): The input value. Returns: [Tensor|TransformedDistribution|ChainTransform]: The return value. """ if isinstance(input, distribution.Distribution): return transformed_distribution.TransformedDistribution( input, [self] ) if isinstance(input, Transform): return ChainTransform([self, input]) return self.forward(input) def forward(self, x): """Forward transformation with mapping :math:`y = f(x)`. Useful for turning one random outcome into another. Args: x (Tensos): Input parameter, generally is a sample generated from ``Distribution``. Returns: Tensor: Outcome of forward transformation. """ if not isinstance(x, paddle.fluid.framework.Variable): raise TypeError( f"Expected 'x' is a Tensor or Real, but got {type(x)}." ) if x.dim() < self._domain.event_rank: raise ValueError( f'The dimensions of x({x.dim()}) should be ' f'grater than or equal to {self._domain.event_rank}' ) return self._forward(x) def inverse(self, y): """Inverse transformation :math:`x = f^{-1}(y)`. It's useful for "reversing" a transformation to compute one probability in terms of another. Args: y (Tensor): Input parameter for inverse transformation. Returns: Tensor: Outcome of inverse transform. """ if not isinstance(y, paddle.fluid.framework.Variable): raise TypeError( f"Expected 'y' is a Tensor or Real, but got {type(y)}." ) if y.dim() < self._codomain.event_rank: raise ValueError( f'The dimensions of y({y.dim()}) should be ' f'grater than or equal to {self._codomain.event_rank}' ) return self._inverse(y) def forward_log_det_jacobian(self, x): """The log of the absolute value of the determinant of the matrix of all first-order partial derivatives of the inverse function. Args: x (Tensor): Input tensor, generally is a sample generated from ``Distribution`` Returns: Tensor: The log of the absolute value of Jacobian determinant. """ if not isinstance(x, paddle.fluid.framework.Variable): raise TypeError( f"Expected 'y' is a Tensor or Real, but got {type(x)}." ) if ( isinstance(x, paddle.fluid.framework.Variable) and x.dim() < self._domain.event_rank ): raise ValueError( f'The dimensions of x({x.dim()}) should be ' f'grater than or equal to {self._domain.event_rank}' ) if not self._is_injective(): raise NotImplementedError( "forward_log_det_jacobian can't be implemented for non-injective" "transforms." ) return self._call_forward_log_det_jacobian(x) def inverse_log_det_jacobian(self, y): """Compute :math:`log|det J_{f^{-1}}(y)|`. Note that ``forward_log_det_jacobian`` is the negative of this function, evaluated at :math:`f^{-1}(y)`. Args: y (Tensor): The input to the ``inverse`` Jacobian determinant evaluation. Returns: Tensor: The value of :math:`log|det J_{f^{-1}}(y)|`. """ if not isinstance(y, paddle.fluid.framework.Variable): raise TypeError(f"Expected 'y' is a Tensor, but got {type(y)}.") if y.dim() < self._codomain.event_rank: raise ValueError( f'The dimensions of y({y.dim()}) should be ' f'grater than or equal to {self._codomain.event_rank}' ) return self._call_inverse_log_det_jacobian(y) def forward_shape(self, shape): """Infer the shape of forward transformation. Args: shape (Sequence[int]): The input shape. Returns: Sequence[int]: The output shape. """ if not isinstance(shape, typing.Sequence): raise TypeError( f"Expected shape is Sequence[int] type, but got {type(shape)}." ) return self._forward_shape(shape) def inverse_shape(self, shape): """Infer the shape of inverse transformation. Args: shape (Sequence[int]): The input shape of inverse transformation. Returns: Sequence[int]: The output shape of inverse transformation. """ if not isinstance(shape, typing.Sequence): raise TypeError( f"Expected shape is Sequence[int] type, but got {type(shape)}." ) return self._inverse_shape(shape) @property def _domain(self): """The domain of this transformation""" return variable.real @property def _codomain(self): """The codomain of this transformation""" return variable.real def _forward(self, x): """Inner method for publid API ``forward``, subclass should overwrite this method for supporting forward transformation. """ raise NotImplementedError('Forward not implemented') def _inverse(self, y): """Inner method of public API ``inverse``, subclass should overwrite this method for supporting inverse transformation. """ raise NotImplementedError('Inverse not implemented') def _call_forward_log_det_jacobian(self, x): """Inner method called by ``forward_log_det_jacobian``.""" if hasattr(self, '_forward_log_det_jacobian'): return self._forward_log_det_jacobian(x) if hasattr(self, '_inverse_log_det_jacobian'): return -self._inverse_log_det_jacobian(self.forward(x)) raise NotImplementedError( 'Neither _forward_log_det_jacobian nor _inverse_log_det_jacobian' 'is implemented. One of them is required.' ) def _call_inverse_log_det_jacobian(self, y): """Inner method called by ``inverse_log_det_jacobian``""" if hasattr(self, '_inverse_log_det_jacobian'): return self._inverse_log_det_jacobian(y) if hasattr(self, '_forward_log_det_jacobian'): return -self._forward_log_det_jacobian(self._inverse(y)) raise NotImplementedError( 'Neither _forward_log_det_jacobian nor _inverse_log_det_jacobian ' 'is implemented. One of them is required' ) def _forward_shape(self, shape): """Inner method called by ``forward_shape``, which is used to infer the forward shape. Subclass should overwrite this method for supporting ``forward_shape``. """ return shape def _inverse_shape(self, shape): """Inner method called by ``inverse_shape``, whic is used to infer the invese shape. Subclass should overwrite this method for supporting ``inverse_shape``. """ return shape class AbsTransform(Transform): r"""Absolute transformation with formula :math:`y = f(x) = abs(x)`, element-wise. This non-injective transformation allows for transformations of scalar distributions with the absolute value function, which maps ``(-inf, inf)`` to ``[0, inf)`` . * For ``y`` in ``(0, inf)`` , ``AbsTransform.inverse(y)`` returns the set invese ``{x in (-inf, inf) : |x| = y}`` as a tuple, ``-y, y`` . * For ``y`` equal ``0`` , ``AbsTransform.inverse(0)`` returns ``0, 0``, which is not the set inverse (the set inverse is the singleton {0}), but "works" in conjunction with ``TransformedDistribution`` to produce a left semi-continuous pdf. * For ``y`` in ``(-inf, 0)`` , ``AbsTransform.inverse(y)`` returns the wrong thing ``-y, y``. This is done for efficiency. Examples: .. code-block:: python import paddle abs = paddle.distribution.AbsTransform() print(abs.forward(paddle.to_tensor([-1., 0., 1.]))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 0., 1.]) print(abs.inverse(paddle.to_tensor([1.]))) # (Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [-1.]), Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1.])) # The |dX/dY| is constant 1. So Log|dX/dY| == 0 print(abs.inverse_log_det_jacobian(paddle.to_tensor(1.))) # (Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # 0.), Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # 0.)) #Special case handling of 0. print(abs.inverse(paddle.to_tensor([0.]))) # (Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.]), Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.])) print(abs.inverse_log_det_jacobian(paddle.to_tensor(0.))) # (Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # 0.), Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # 0.)) """ _type = Type.SURJECTION def _forward(self, x): return x.abs() def _inverse(self, y): return -y, y def _inverse_log_det_jacobian(self, y): zero = paddle.zeros([], dtype=y.dtype) return zero, zero @property def _domain(self): return variable.real @property def _codomain(self): return variable.positive class AffineTransform(Transform): r"""Affine transformation with mapping :math:`y = \text{loc} + \text{scale} \times x`. Args: loc (Tensor): The location parameter. scale (Tensor): The scale parameter. Examples: .. code-block:: python import paddle x = paddle.to_tensor([1., 2.]) affine = paddle.distribution.AffineTransform(paddle.to_tensor(0.), paddle.to_tensor(1.)) print(affine.forward(x)) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 2.]) print(affine.inverse(affine.forward(x))) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 2.]) print(affine.forward_log_det_jacobian(x)) # Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # 0.) """ _type = Type.BIJECTION def __init__(self, loc, scale): if not isinstance(loc, paddle.fluid.framework.Variable): raise TypeError(f"Expected 'loc' is a Tensor, but got {type(loc)}") if not isinstance(scale, paddle.fluid.framework.Variable): raise TypeError( f"Expected scale is a Tensor, but got {type(scale)}" ) self._loc = loc self._scale = scale super().__init__() @property def loc(self): return self._loc @property def scale(self): return self._scale def _forward(self, x): return self._loc + self._scale * x def _inverse(self, y): return (y - self._loc) / self._scale def _forward_log_det_jacobian(self, x): return paddle.abs(self._scale).log() def _forward_shape(self, shape): return tuple( paddle.broadcast_shape( paddle.broadcast_shape(shape, self._loc.shape), self._scale.shape, ) ) def _inverse_shape(self, shape): return tuple( paddle.broadcast_shape( paddle.broadcast_shape(shape, self._loc.shape), self._scale.shape, ) ) @property def _domain(self): return variable.real @property def _codomain(self): return variable.real class ChainTransform(Transform): r"""Composes multiple transforms in a chain. Args: transforms (Sequence[Transform]): A sequence of transformations. Examples: .. code-block:: python import paddle x = paddle.to_tensor([0., 1., 2., 3.]) chain = paddle.distribution.ChainTransform(( paddle.distribution.AffineTransform( paddle.to_tensor(0.), paddle.to_tensor(1.)), paddle.distribution.ExpTransform() )) print(chain.forward(x)) # Tensor(shape=[4], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1. , 2.71828175 , 7.38905621 , 20.08553696]) print(chain.inverse(chain.forward(x))) # Tensor(shape=[4], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0., 1., 2., 3.]) print(chain.forward_log_det_jacobian(x)) # Tensor(shape=[4], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0., 1., 2., 3.]) print(chain.inverse_log_det_jacobian(chain.forward(x))) # Tensor(shape=[4], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [ 0., -1., -2., -3.]) """ def __init__(self, transforms): if not isinstance(transforms, typing.Sequence): raise TypeError( f"Expected type of 'transforms' is Sequence, but got {type(transforms)}" ) if not all(isinstance(t, Transform) for t in transforms): raise TypeError( "All elements of transforms should be Transform type." ) self.transforms = transforms super().__init__() def _is_injective(self): return all(t._is_injective() for t in self.transforms) def _forward(self, x): for transform in self.transforms: x = transform.forward(x) return x def _inverse(self, y): for transform in reversed(self.transforms): y = transform.inverse(y) return y def _forward_log_det_jacobian(self, x): value = 0.0 event_rank = self._domain.event_rank for t in self.transforms: value += self._sum_rightmost( t.forward_log_det_jacobian(x), event_rank - t._domain.event_rank ) x = t.forward(x) event_rank += t._codomain.event_rank - t._domain.event_rank return value def _forward_shape(self, shape): for transform in self.transforms: shape = transform.forward_shape(shape) return shape def _inverse_shape(self, shape): for transform in self.transforms: shape = transform.inverse_shape(shape) return shape def _sum_rightmost(self, value, n): """sum value along rightmost n dim""" return value.sum(list(range(-n, 0))) if n > 0 else value @property def _domain(self): domain = self.transforms[0]._domain # Compute the lower bound of input dimensions for chain transform. # # Suppose the dimensions of input tensor is N, and chain [t0,...ti,...tm], # ti(in) denotes ti.domain.event_rank, ti(out) denotes ti.codomain.event_rank, # delta(ti) denotes (ti(out) - ti(in)). # For transform ti, N shoud satisfy the constraint: # N + delta(t0) + delta(t1)...delta(t(i-1)) >= ti(in) # So, for all transform in chain, N shoud satisfy follow constraints: # t0: N >= t0(in) # t1: N >= t1(in) - delta(t0) # ... # tm: N >= tm(in) - ... - delta(ti) - ... - delta(t0) # # Above problem can be solved more effectively use dynamic programming. # Let N(i) denotes lower bound of transform ti, than the state # transition equation is: # N(i) = max{N(i+1)-delta(ti), ti(in)} event_rank = self.transforms[-1]._codomain.event_rank for t in reversed(self.transforms): event_rank -= t._codomain.event_rank - t._domain.event_rank event_rank = max(event_rank, t._domain.event_rank) return variable.Independent(domain, event_rank - domain.event_rank) @property def _codomain(self): codomain = self.transforms[-1]._codomain event_rank = self.transforms[0]._domain.event_rank for t in self.transforms: event_rank += t._codomain.event_rank - t._domain.event_rank event_rank = max(event_rank, t._codomain.event_rank) return variable.Independent(codomain, event_rank - codomain.event_rank) class ExpTransform(Transform): r"""Exponent transformation with mapping :math:`y = \exp(x)`. Examples: .. code-block:: python import paddle exp = paddle.distribution.ExpTransform() print(exp.forward(paddle.to_tensor([1., 2., 3.]))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [2.71828175 , 7.38905621 , 20.08553696]) print(exp.inverse(paddle.to_tensor([1., 2., 3.]))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0. , 0.69314718, 1.09861231]) print(exp.forward_log_det_jacobian(paddle.to_tensor([1., 2., 3.]))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 2., 3.]) print(exp.inverse_log_det_jacobian(paddle.to_tensor([1., 2., 3.]))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [ 0. , -0.69314718, -1.09861231]) """ _type = Type.BIJECTION def __init__(self): super().__init__() @property def _domain(self): return variable.real @property def _codomain(self): return variable.positive def _forward(self, x): return x.exp() def _inverse(self, y): return y.log() def _forward_log_det_jacobian(self, x): return x class IndependentTransform(Transform): r""" ``IndependentTransform`` wraps a base transformation, reinterprets some of the rightmost batch axes as event axes. Generally, it is used to expand the event axes. This has no effect on the forward or inverse transformaion, but does sum out the ``reinterpretd_bach_rank`` rightmost dimensions in computing the determinant of Jacobian matrix. To see this, consider the ``ExpTransform`` applied to a Tensor which has sample, batch, and event ``(S,B,E)`` shape semantics. Suppose the Tensor's paritioned-shape is ``(S=[4], B=[2, 2], E=[3])`` , reinterpreted_batch_rank is 1. Then the reinterpreted Tensor's shape is ``(S=[4], B=[2], E=[2, 3])`` . The shape returned by ``forward`` and ``inverse`` is unchanged, ie, ``[4,2,2,3]`` . However the shape returned by ``inverse_log_det_jacobian`` is ``[4,2]``, because the Jacobian determinant is a reduction over the event dimensions. Args: base (Transform): The base transformation. reinterpreted_batch_rank (int): The num of rightmost batch rank that will be reinterpreted as event rank. Examples: .. code-block:: python import paddle x = paddle.to_tensor([[1., 2., 3.], [4., 5., 6.]]) # Exponential transform with event_rank = 1 multi_exp = paddle.distribution.IndependentTransform( paddle.distribution.ExpTransform(), 1) print(multi_exp.forward(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[2.71828175 , 7.38905621 , 20.08553696 ], # [54.59814835 , 148.41316223, 403.42880249]]) print(multi_exp.forward_log_det_jacobian(x)) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [6. , 15.]) """ def __init__(self, base, reinterpreted_batch_rank): if not isinstance(base, Transform): raise TypeError( f"Expected 'base' is Transform type, but get {type(base)}" ) if reinterpreted_batch_rank <= 0: raise ValueError( f"Expected 'reinterpreted_batch_rank' is grater than zero, but got {reinterpreted_batch_rank}" ) self._base = base self._reinterpreted_batch_rank = reinterpreted_batch_rank super().__init__() def _is_injective(self): return self._base._is_injective() def _forward(self, x): if x.dim() < self._domain.event_rank: raise ValueError("Input dimensions is less than event dimensions.") return self._base.forward(x) def _inverse(self, y): if y.dim() < self._codomain.event_rank: raise ValueError("Input dimensions is less than event dimensions.") return self._base.inverse(y) def _forward_log_det_jacobian(self, x): return self._base.forward_log_det_jacobian(x).sum( list(range(-self._reinterpreted_batch_rank, 0)) ) def _forward_shape(self, shape): return self._base.forward_shape(shape) def _inverse_shape(self, shape): return self._base.inverse_shape(shape) @property def _domain(self): return variable.Independent( self._base._domain, self._reinterpreted_batch_rank ) @property def _codomain(self): return variable.Independent( self._base._codomain, self._reinterpreted_batch_rank ) class PowerTransform(Transform): r""" Power transformation with mapping :math:`y = x^{\text{power}}`. Args: power (Tensor): The power parameter. Examples: .. code-block:: python import paddle x = paddle.to_tensor([1., 2.]) power = paddle.distribution.PowerTransform(paddle.to_tensor(2.)) print(power.forward(x)) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 4.]) print(power.inverse(power.forward(x))) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [1., 2.]) print(power.forward_log_det_jacobian(x)) # Tensor(shape=[2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.69314718, 1.38629436]) """ _type = Type.BIJECTION def __init__(self, power): if not isinstance(power, paddle.fluid.framework.Variable): raise TypeError( f"Expected 'power' is a tensor, but got {type(power)}" ) self._power = power super().__init__() @property def power(self): return self._power @property def _domain(self): return variable.real @property def _codomain(self): return variable.positive def _forward(self, x): return x.pow(self._power) def _inverse(self, y): return y.pow(1 / self._power) def _forward_log_det_jacobian(self, x): return (self._power * x.pow(self._power - 1)).abs().log() def _forward_shape(self, shape): return tuple(paddle.broadcast_shape(shape, self._power.shape)) def _inverse_shape(self, shape): return tuple(paddle.broadcast_shape(shape, self._power.shape)) class ReshapeTransform(Transform): r"""Reshape the event shape of a tensor. Note that ``in_event_shape`` and ``out_event_shape`` must have the same number of elements. Args: in_event_shape(Sequence[int]): The input event shape. out_event_shape(Sequence[int]): The output event shape. Examples: .. code-block:: python import paddle x = paddle.ones((1,2,3)) reshape_transform = paddle.distribution.ReshapeTransform((2, 3), (3, 2)) print(reshape_transform.forward_shape((1,2,3))) # (5, 2, 6) print(reshape_transform.forward(x)) # Tensor(shape=[1, 3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[[1., 1.], # [1., 1.], # [1., 1.]]]) print(reshape_transform.inverse(reshape_transform.forward(x))) # Tensor(shape=[1, 2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[[1., 1., 1.], # [1., 1., 1.]]]) print(reshape_transform.forward_log_det_jacobian(x)) # Tensor(shape=[1], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.]) """ _type = Type.BIJECTION def __init__(self, in_event_shape, out_event_shape): if not isinstance(in_event_shape, typing.Sequence) or not isinstance( out_event_shape, typing.Sequence ): raise TypeError( f"Expected type of 'in_event_shape' and 'out_event_shape' is " f"Squence[int], but got 'in_event_shape': {in_event_shape}, " f"'out_event_shape': {out_event_shape}" ) in_size = 1 for e in in_event_shape: in_size *= e out_size = 1 for e in out_event_shape: out_size *= e if in_size != out_size: raise ValueError( f"The numel of 'in_event_shape' should be 'out_event_shape', " f"but got {in_size}!={out_size}" ) self._in_event_shape = tuple(in_event_shape) self._out_event_shape = tuple(out_event_shape) super().__init__() @property def in_event_shape(self): return self._in_event_shape @property def out_event_shape(self): return self._out_event_shape @property def _domain(self): return variable.Independent(variable.real, len(self._in_event_shape)) @property def _codomain(self): return variable.Independent(variable.real, len(self._out_event_shape)) def _forward(self, x): return x.reshape( tuple(x.shape)[: x.dim() - len(self._in_event_shape)] + self._out_event_shape ) def _inverse(self, y): return y.reshape( tuple(y.shape)[: y.dim() - len(self._out_event_shape)] + self._in_event_shape ) def _forward_shape(self, shape): if len(shape) < len(self._in_event_shape): raise ValueError( f"Expected length of 'shape' is not less than {len(self._in_event_shape)}, but got {len(shape)}" ) if tuple(shape[-len(self._in_event_shape) :]) != tuple( self._in_event_shape ): raise ValueError( f"Event shape mismatch, expected: {self._in_event_shape}, but got {shape[-len(self._in_event_shape):]}" ) return ( tuple(shape[: -len(self._in_event_shape)]) + self._out_event_shape ) def _inverse_shape(self, shape): if len(shape) < len(self._out_event_shape): raise ValueError( f"Expected 'shape' length is not less than {len(self._out_event_shape)}, but got {len(shape)}" ) if tuple(shape[-len(self._out_event_shape) :]) != tuple( self._out_event_shape ): raise ValueError( f"Event shape mismatch, expected: {self._out_event_shape}, but got {shape[-len(self._out_event_shape):]}" ) return ( tuple(shape[: -len(self._out_event_shape)]) + self._in_event_shape ) def _forward_log_det_jacobian(self, x): # TODO(zhouwei): should not set shape to [1], which is [] shape = x.shape[: x.dim() - len(self._in_event_shape)] or [1] return paddle.zeros(shape, dtype=x.dtype) class SigmoidTransform(Transform): r"""Sigmoid transformation with mapping :math:`y = \frac{1}{1 + \exp(-x)}` and :math:`x = \text{logit}(y)`. Examples: .. code-block:: python import paddle x = paddle.ones((2,3)) t = paddle.distribution.SigmoidTransform() print(t.forward(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[0.73105860, 0.73105860, 0.73105860], # [0.73105860, 0.73105860, 0.73105860]]) print(t.inverse(t.forward(x))) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[1.00000012, 1.00000012, 1.00000012], # [1.00000012, 1.00000012, 1.00000012]]) print(t.forward_log_det_jacobian(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[-1.62652326, -1.62652326, -1.62652326], # [-1.62652326, -1.62652326, -1.62652326]]) """ @property def _domain(self): return variable.real @property def _codomain(self): return variable.Variable(False, 0, constraint.Range(0.0, 1.0)) def _forward(self, x): return F.sigmoid(x) def _inverse(self, y): return y.log() - (-y).log1p() def _forward_log_det_jacobian(self, x): return -F.softplus(-x) - F.softplus(x) class SoftmaxTransform(Transform): r"""Softmax transformation with mapping :math:`y=\exp(x)` then normalizing. It's generally used to convert unconstrained space to simplex. This mapping is not injective, so ``forward_log_det_jacobian`` and ``inverse_log_det_jacobian`` are not implemented. Examples: .. code-block:: python import paddle x = paddle.ones((2,3)) t = paddle.distribution.SoftmaxTransform() print(t.forward(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[0.33333334, 0.33333334, 0.33333334], # [0.33333334, 0.33333334, 0.33333334]]) print(t.inverse(t.forward(x))) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[-1.09861231, -1.09861231, -1.09861231], # [-1.09861231, -1.09861231, -1.09861231]]) """ _type = Type.OTHER @property def _domain(self): return variable.Independent(variable.real, 1) @property def _codomain(self): return variable.Variable(False, 1, constraint.simplex) def _forward(self, x): x = (x - x.max(-1, keepdim=True)[0]).exp() return x / x.sum(-1, keepdim=True) def _inverse(self, y): return y.log() def _forward_shape(self, shape): if len(shape) < 1: raise ValueError( f"Expected length of shape is grater than 1, but got {len(shape)}" ) return shape def _inverse_shape(self, shape): if len(shape) < 1: raise ValueError( f"Expected length of shape is grater than 1, but got {len(shape)}" ) return shape class StackTransform(Transform): r"""``StackTransform`` applies a sequence of transformations along the specific axis. Args: transforms (Sequence[Transform]): The sequence of transformations. axis (int, optional): The axis along which will be transformed. default value is 0. Examples: .. code-block:: python import paddle x = paddle.stack( (paddle.to_tensor([1., 2., 3.]), paddle.to_tensor([1, 2., 3.])), 1) t = paddle.distribution.StackTransform( (paddle.distribution.ExpTransform(), paddle.distribution.PowerTransform(paddle.to_tensor(2.))), 1 ) print(t.forward(x)) # Tensor(shape=[3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[2.71828175 , 1. ], # [7.38905621 , 4. ], # [20.08553696, 9. ]]) print(t.inverse(t.forward(x))) # Tensor(shape=[3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[1., 1.], # [2., 2.], # [3., 3.]]) print(t.forward_log_det_jacobian(x)) # Tensor(shape=[3, 2], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[1. , 0.69314718], # [2. , 1.38629436], # [3. , 1.79175949]]) """ def __init__(self, transforms, axis=0): if not transforms or not isinstance(transforms, typing.Sequence): raise TypeError( f"Expected 'transforms' is Sequence[Transform], but got {type(transforms)}." ) if not all(isinstance(t, Transform) for t in transforms): raise TypeError( 'Expected all element in transforms is Transform Type.' ) if not isinstance(axis, int): raise TypeError(f"Expected 'axis' is int, but got{type(axis)}.") self._transforms = transforms self._axis = axis def _is_injective(self): return all(t._is_injective() for t in self._transforms) @property def transforms(self): return self._transforms @property def axis(self): return self._axis def _forward(self, x): self._check_size(x) return paddle.stack( [ t.forward(v) for v, t in zip(paddle.unstack(x, self._axis), self._transforms) ], self._axis, ) def _inverse(self, y): self._check_size(y) return paddle.stack( [ t.inverse(v) for v, t in zip(paddle.unstack(y, self._axis), self._transforms) ], self._axis, ) def _forward_log_det_jacobian(self, x): self._check_size(x) return paddle.stack( [ t.forward_log_det_jacobian(v) for v, t in zip(paddle.unstack(x, self._axis), self._transforms) ], self._axis, ) def _check_size(self, v): if not (-v.dim() <= self._axis < v.dim()): raise ValueError( f'Input dimensions {v.dim()} should be grater than stack ' f'transform axis {self._axis}.' ) if v.shape[self._axis] != len(self._transforms): raise ValueError( f'Input size along {self._axis} should be equal to the ' f'length of transforms.' ) @property def _domain(self): return variable.Stack([t._domain for t in self._transforms], self._axis) @property def _codomain(self): return variable.Stack( [t._codomain for t in self._transforms], self._axis ) class StickBreakingTransform(Transform): r"""Convert an unconstrained vector to the simplex with one additional dimension by the stick-breaking construction. Examples: .. code-block:: python import paddle x = paddle.to_tensor([1.,2.,3.]) t = paddle.distribution.StickBreakingTransform() print(t.forward(x)) # Tensor(shape=[4], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.47536686, 0.41287899, 0.10645414, 0.00530004]) print(t.inverse(t.forward(x))) # Tensor(shape=[3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [0.99999988, 2. , 2.99999881]) print(t.forward_log_det_jacobian(x)) # Tensor(shape=[], dtype=float32, place=Place(gpu:0), stop_gradient=True, # -9.10835075) """ _type = Type.BIJECTION def _forward(self, x): offset = x.shape[-1] + 1 - paddle.ones([x.shape[-1]]).cumsum(-1) z = F.sigmoid(x - offset.log()) z_cumprod = (1 - z).cumprod(-1) return F.pad(z, [0] * 2 * (len(x.shape) - 1) + [0, 1], value=1) * F.pad( z_cumprod, [0] * 2 * (len(x.shape) - 1) + [1, 0], value=1 ) def _inverse(self, y): y_crop = y[..., :-1] offset = y.shape[-1] - paddle.ones([y_crop.shape[-1]]).cumsum(-1) sf = 1 - y_crop.cumsum(-1) x = y_crop.log() - sf.log() + offset.log() return x def _forward_log_det_jacobian(self, x): y = self.forward(x) offset = x.shape[-1] + 1 - paddle.ones([x.shape[-1]]).cumsum(-1) x = x - offset.log() return (-x + F.log_sigmoid(x) + y[..., :-1].log()).sum(-1) def _forward_shape(self, shape): if not shape: raise ValueError(f"Expected 'shape' is not empty, but got {shape}") return shape[:-1] + (shape[-1] + 1,) def _inverse_shape(self, shape): if not shape: raise ValueError(f"Expected 'shape' is not empty, but got {shape}") return shape[:-1] + (shape[-1] - 1,) @property def _domain(self): return variable.Independent(variable.real, 1) @property def _codomain(self): return variable.Variable(False, 1, constraint.simplex) class TanhTransform(Transform): r"""Tanh transformation with mapping :math:`y = \tanh(x)`. Examples: .. code-block:: python import paddle tanh = paddle.distribution.TanhTransform() x = paddle.to_tensor([[1., 2., 3.], [4., 5., 6.]]) print(tanh.forward(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[0.76159418, 0.96402758, 0.99505478], # [0.99932933, 0.99990922, 0.99998772]]) print(tanh.inverse(tanh.forward(x))) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[1.00000012, 2. , 3.00000286], # [4.00002146, 5.00009823, 6.00039864]]) print(tanh.forward_log_det_jacobian(x)) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[-0.86756170 , -2.65000558 , -4.61865711 ], # [-6.61437654 , -8.61379623 , -10.61371803]]) print(tanh.inverse_log_det_jacobian(tanh.forward(x))) # Tensor(shape=[2, 3], dtype=float32, place=Place(gpu:0), stop_gradient=True, # [[0.86756176 , 2.65000558 , 4.61866283 ], # [6.61441946 , 8.61399269 , 10.61451530]]) """ _type = Type.BIJECTION @property def _domain(self): return variable.real @property def _codomain(self): return variable.Variable(False, 0, constraint.Range(-1.0, 1.0)) def _forward(self, x): return x.tanh() def _inverse(self, y): return y.atanh() def _forward_log_det_jacobian(self, x): """We implicitly rely on _forward_log_det_jacobian rather than explicitly implement ``_inverse_log_det_jacobian`` since directly using ``-tf.math.log1p(-tf.square(y))`` has lower numerical precision. See details: https://github.com/tensorflow/probability/blob/master/tensorflow_probability/python/bijectors/tanh.py#L69-L80 """ return 2.0 * (math.log(2.0) - x - F.softplus(-2.0 * x))