Source code for pyro.distributions.transforms.sylvester

# Copyright (c) 2017-2019 Uber Technologies, Inc.
# SPDX-License-Identifier: Apache-2.0

import torch
import torch.nn as nn

from .. import constraints
from ..torch_transform import TransformModule
from ..util import copy_docs_from
from .householder import Householder


[docs]@copy_docs_from(TransformModule) class Sylvester(Householder): r""" An implementation of the Sylvester bijective transform of the Householder variety (Van den Berg Et Al., 2018), :math:`\mathbf{y} = \mathbf{x} + QR\tanh(SQ^T\mathbf{x}+\mathbf{b})` where :math:`\mathbf{x}` are the inputs, :math:`\mathbf{y}` are the outputs, :math:`R,S\sim D\times D` are upper triangular matrices for input dimension :math:`D`, :math:`Q\sim D\times D` is an orthogonal matrix, and :math:`\mathbf{b}\sim D` is learnable bias term. The Sylvester transform is a generalization of :class:`~pyro.distributions.transforms.Planar`. In the Householder type of the Sylvester transform, the orthogonality of :math:`Q` is enforced by representing it as the product of Householder transformations. Together with :class:`~pyro.distributions.TransformedDistribution` it provides a way to create richer variational approximations. Example usage: >>> base_dist = dist.Normal(torch.zeros(10), torch.ones(10)) >>> transform = Sylvester(10, count_transforms=4) >>> pyro.module("my_transform", transform) # doctest: +SKIP >>> flow_dist = dist.TransformedDistribution(base_dist, [transform]) >>> flow_dist.sample() # doctest: +SKIP tensor([-0.4071, -0.5030, 0.7924, -0.2366, -0.2387, -0.1417, 0.0868, 0.1389, -0.4629, 0.0986]) The inverse of this transform does not possess an analytical solution and is left unimplemented. However, the inverse is cached when the forward operation is called during sampling, and so samples drawn using the Sylvester transform can be scored. References: [1] Rianne van den Berg, Leonard Hasenclever, Jakub M. Tomczak, Max Welling. Sylvester Normalizing Flows for Variational Inference. UAI 2018. """ domain = constraints.real_vector codomain = constraints.real_vector bijective = True def __init__(self, input_dim, count_transforms=1): super().__init__(input_dim, count_transforms) # Create parameters for Sylvester transform self.R_dense = nn.Parameter(torch.Tensor(input_dim, input_dim)) self.S_dense = nn.Parameter(torch.Tensor(input_dim, input_dim)) self.R_diag = nn.Parameter(torch.Tensor(input_dim)) self.S_diag = nn.Parameter(torch.Tensor(input_dim)) self.b = nn.Parameter(torch.Tensor(input_dim)) # Register masks and indices triangular_mask = torch.triu(torch.ones(input_dim, input_dim), diagonal=1) self.register_buffer("triangular_mask", triangular_mask) self._cached_logDetJ = None self.tanh = nn.Tanh() self.reset_parameters2() # Derivative of hyperbolic tan
[docs] def dtanh_dx(self, x): return 1.0 - self.tanh(x).pow(2)
# Construct upper diagonal R matrix
[docs] def R(self): return self.R_dense * self.triangular_mask + torch.diag(self.tanh(self.R_diag))
# Construct upper diagonal S matrix
[docs] def S(self): return self.S_dense * self.triangular_mask + torch.diag(self.tanh(self.S_diag))
# Construct orthonomal matrix using Householder flow
[docs] def Q(self, x): u = self.u() partial_Q = torch.eye( self.input_dim, dtype=x.dtype, layout=x.layout, device=x.device ) - 2.0 * torch.ger(u[0], u[0]) for idx in range(1, self.u_unnormed.size(-2)): partial_Q = torch.matmul( partial_Q, torch.eye(self.input_dim) - 2.0 * torch.ger(u[idx], u[idx]) ) return partial_Q
# Self.u_unnormed is initialized in parent class
[docs] def reset_parameters2(self): for v in [self.b, self.R_diag, self.S_diag, self.R_dense, self.S_dense]: v.data.uniform_(-0.01, 0.01)
def _call(self, x): """ :param x: the input into the bijection :type x: torch.Tensor Invokes the bijection x=>y; in the prototypical context of a :class:`~pyro.distributions.TransformedDistribution` `x` is a sample from the base distribution (or the output of a previous transform) """ Q = self.Q(x) R = self.R() S = self.S() A = torch.matmul(Q, R) B = torch.matmul(S, Q.t()) preactivation = torch.matmul(x, B) + self.b y = x + torch.matmul(self.tanh(preactivation), A) self._cached_logDetJ = torch.log1p( self.dtanh_dx(preactivation) * R.diagonal() * S.diagonal() + 1e-8 ).sum(-1) return y def _inverse(self, y): """ :param y: the output of the bijection :type y: torch.Tensor Inverts y => x. As noted above, this implementation is incapable of inverting arbitrary values `y`; rather it assumes `y` is the result of a previously computed application of the bijector to some `x` (which was cached on the forward call) """ raise KeyError( "Sylvester object expected to find key in intermediates cache but didn't" )
[docs] def log_abs_det_jacobian(self, x, y): """ Calculates the elementwise determinant of the log Jacobian """ x_old, y_old = self._cached_x_y if x is not x_old or y is not y_old: # This call to the parent class Transform will update the cache # as well as calling self._call and recalculating y and log_detJ self(x) return self._cached_logDetJ
[docs]def sylvester(input_dim, count_transforms=None): """ A helper function to create a :class:`~pyro.distributions.transforms.Sylvester` object for consistency with other helpers. :param input_dim: Dimension of input variable :type input_dim: int :param count_transforms: Number of Sylvester operations to apply. Defaults to input_dim // 2 + 1. :type count_transforms: int """ if count_transforms is None: count_transforms = input_dim // 2 + 1 return Sylvester(input_dim, count_transforms=count_transforms)