natgrad.py
# Copyright 2018-2020 The GPflow Contributors. 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 abc
import functools
from typing import Callable, Optional, Sequence, Tuple, Union
import numpy as np
import tensorflow as tf
from ..base import Parameter, _to_constrained
Scalar = Union[float, tf.Tensor, np.ndarray]
LossClosure = Callable[[], tf.Tensor]
NatGradParameters = Union[Tuple[Parameter, Parameter], Tuple[Parameter, Parameter, "XiTransform"]]
__all__ = [
"NaturalGradient",
"XiTransform",
"XiNat",
"XiSqrtMeanVar",
]
#
# Xi transformations necessary for natural gradient optimizer.
# Abstract class and two implementations: XiNat and XiSqrtMeanVar.
#
class XiTransform(metaclass=abc.ABCMeta):
"""
XiTransform is the base class that implements three transformations necessary
for the natural gradient calculation wrt any parameterization.
This class does not handle any shape information, but it is assumed that
the parameters pairs are always of shape (N, D) and (D, N, N).
"""
@staticmethod
@abc.abstractmethod
def meanvarsqrt_to_xi(mean, varsqrt):
"""
Transforms the parameter `mean` and `varsqrt` to `xi1`, `xi2`
:param mean: the mean parameter (N, D)
:param varsqrt: the varsqrt parameter (D, N, N)
:return: tuple (xi1, xi2), the xi parameters (N, D), (D, N, N)
"""
@staticmethod
@abc.abstractmethod
def xi_to_meanvarsqrt(xi1, xi2):
"""
Transforms the parameter `xi1`, `xi2` to `mean`, `varsqrt`
:param xi1: the ξ₁ parameter
:param xi2: the ξ₂ parameter
:return: tuple (mean, varsqrt), the meanvarsqrt parameters
"""
@staticmethod
@abc.abstractmethod
def naturals_to_xi(nat1, nat2):
"""
Applies the transform so that `nat1`, `nat2` is mapped to `xi1`, `xi2`
:param nat1: the θ₁ parameter
:param nat2: the θ₂ parameter
:return: tuple `xi1`, `xi2`
"""
class XiNat(XiTransform):
"""
This is the default transform. Using the natural directly saves the forward mode
gradient, and also gives the analytic optimal solution for gamma=1 in the case
of Gaussian likelihood.
"""
@staticmethod
def meanvarsqrt_to_xi(mean, varsqrt):
return meanvarsqrt_to_natural(mean, varsqrt)
@staticmethod
def xi_to_meanvarsqrt(xi1, xi2):
return natural_to_meanvarsqrt(xi1, xi2)
@staticmethod
def naturals_to_xi(nat1, nat2):
return nat1, nat2
class XiSqrtMeanVar(XiTransform):
"""
This transformation will perform natural gradient descent on the model parameters,
so saves the conversion to and from Xi.
"""
@staticmethod
def meanvarsqrt_to_xi(mean, varsqrt):
return mean, varsqrt
@staticmethod
def xi_to_meanvarsqrt(xi1, xi2):
return xi1, xi2
@staticmethod
def naturals_to_xi(nat1, nat2):
return natural_to_meanvarsqrt(nat1, nat2)
class NaturalGradient(tf.optimizers.Optimizer):
"""
Implements a natural gradient descent optimizer for variational models
that are based on a distribution q(u) = N(q_mu, q_sqrt q_sqrtᵀ) that is
parameterized by mean q_mu and lower-triangular Cholesky factor q_sqrt
of the covariance.
Note that this optimizer does not implement the standard API of
tf.optimizers.Optimizer. Its only public method is minimize(), which has
a custom signature (var_list needs to be a list of (q_mu, q_sqrt) tuples,
where q_mu and q_sqrt are gpflow.Parameter instances, not tf.Variable).
Note furthermore that the natural gradients are implemented only for the
full covariance case (i.e., q_diag=True is NOT supported).
When using in your work, please cite
@inproceedings{salimbeni18,
title={Natural Gradients in Practice: Non-Conjugate Variational Inference in Gaussian Process Models},
author={Salimbeni, Hugh and Eleftheriadis, Stefanos and Hensman, James},
booktitle={AISTATS},
year={2018}
"""
def __init__(self, gamma: Scalar, xi_transform: XiTransform = XiNat(), name=None):
"""
:param gamma: natgrad step length
:param xi_transform: default ξ transform (can be overridden in the call to minimize())
The XiNat default choice works well in general.
"""
name = self.__class__.__name__ if name is None else name
super().__init__(name)
self.gamma = gamma
self.xi_transform = xi_transform
def minimize(
self,
loss_fn: LossClosure,
var_list: Sequence[NatGradParameters],
):
"""
Minimizes objective function of the model.
Natural Gradient optimizer works with variational parameters only.
:param loss_fn: Loss function.
:param var_list: List of pair tuples of variational parameters or
triplet tuple with variational parameters and ξ transformation.
If ξ is not specified, will use self.xi_transform.
For example, `var_list` could be
```
var_list = [
(q_mu1, q_sqrt1),
(q_mu2, q_sqrt2, XiSqrtMeanVar())
]
```
GPflow implements the `XiNat` (default) and `XiSqrtMeanVar` transformations
for parameters. Custom transformations that implement the `XiTransform`
interface are also possible.
"""
parameters = [(v[0], v[1], (v[2] if len(v) > 2 else None)) for v in var_list]
self._natgrad_steps(loss_fn, parameters)
def _natgrad_steps(
self,
loss_fn: LossClosure,
parameters: Sequence[Tuple[Parameter, Parameter, Optional[XiTransform]]],
):
"""
Computes gradients of loss_fn() w.r.t. q_mu and q_sqrt, and updates
these parameters using the natgrad backwards step, for all sets of
variational parameters passed in.
:param loss_fn: Loss function.
:param parameters: List of tuples (q_mu, q_sqrt, xi_transform)
"""
q_mus, q_sqrts, xis = zip(*parameters)
q_mu_vars = [p.unconstrained_variable for p in q_mus]
q_sqrt_vars = [p.unconstrained_variable for p in q_sqrts]
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(q_mu_vars + q_sqrt_vars)
loss = loss_fn()
q_mu_grads, q_sqrt_grads = tape.gradient(loss, [q_mu_vars, q_sqrt_vars])
# NOTE that these are the gradients in *unconstrained* space
with tf.name_scope(f"{self._name}/natural_gradient_steps"):
for q_mu_grad, q_sqrt_grad, q_mu, q_sqrt, xi_transform in zip(
q_mu_grads, q_sqrt_grads, q_mus, q_sqrts, xis
):
self._natgrad_apply_gradients(q_mu_grad, q_sqrt_grad, q_mu, q_sqrt, xi_transform)
def _assert_shapes(self, q_mu, q_sqrt):
tf.debugging.assert_shapes(
[
(q_mu, ["M", "L"]),
(q_sqrt, ["L", "M", "M"]),
]
)
def _natgrad_apply_gradients(
self,
q_mu_grad: tf.Tensor,
q_sqrt_grad: tf.Tensor,
q_mu: Parameter,
q_sqrt: Parameter,
xi_transform: Optional[XiTransform] = None,
):
"""
This function does the backward step on the q_mu and q_sqrt parameters,
given the gradients of the loss function with respect to their unconstrained
variables. I.e., it expects the arguments to come from
with tf.GradientTape() as tape:
loss = loss_function()
q_mu_grad, q_mu_sqrt = tape.gradient(loss, [q_mu, q_sqrt])
(Note that tape.gradient() returns the gradients in *unconstrained* space!)
Implements equation [10] from
@inproceedings{salimbeni18,
title={Natural Gradients in Practice: Non-Conjugate Variational Inference in Gaussian Process Models},
author={Salimbeni, Hugh and Eleftheriadis, Stefanos and Hensman, James},
booktitle={AISTATS},
year={2018}
In addition, for convenience with the rest of GPflow, this code computes ∂L/∂η using
the chain rule (the following assumes a numerator layout where the gradient is a row
vector; note that TensorFlow actually returns a column vector), where L is the loss:
∂L/∂η = (∂L / ∂[q_mu, q_sqrt])(∂[q_mu, q_sqrt] / ∂η)
In total there are three derivative calculations:
natgrad of L w.r.t ξ = (∂ξ / ∂θ) [(∂L / ∂[q_mu, q_sqrt]) (∂[q_mu, q_sqrt] / ∂η)]ᵀ
Note that if ξ = θ (i.e. [q_mu, q_sqrt]) some of these calculations are the identity.
In the code η = eta, ξ = xi, θ = nat.
:param q_mu_grad: gradient of loss w.r.t. q_mu (in unconstrained space)
:param q_sqrt_grad: gradient of loss w.r.t. q_sqrt (in unconstrained space)
:param q_mu: parameter for the mean of q(u) with shape [M, L]
:param q_sqrt: parameter for the square root of the covariance of q(u)
with shape [L, M, M] (the diagonal parametrization, q_diag=True, is NOT supported)
:param xi_transform: the ξ transform to use (self.xi_transform if not specified)
"""
self._assert_shapes(q_mu, q_sqrt)
if xi_transform is None:
xi_transform = self.xi_transform
# 1) the ordinary gpflow gradient
dL_dmean = _to_constrained(q_mu_grad, q_mu.transform)
dL_dvarsqrt = _to_constrained(q_sqrt_grad, q_sqrt.transform)
with tf.GradientTape(persistent=True, watch_accessed_variables=False) as tape:
tape.watch([q_mu.unconstrained_variable, q_sqrt.unconstrained_variable])
# the three parameterizations as functions of [q_mu, q_sqrt]
eta1, eta2 = meanvarsqrt_to_expectation(q_mu, q_sqrt)
# we need these to calculate the relevant gradients
meanvarsqrt = expectation_to_meanvarsqrt(eta1, eta2)
if not isinstance(xi_transform, XiNat):
nat1, nat2 = meanvarsqrt_to_natural(q_mu, q_sqrt)
xi1_nat, xi2_nat = xi_transform.naturals_to_xi(nat1, nat2)
dummy_tensors = tf.ones_like(xi1_nat), tf.ones_like(xi2_nat)
with tf.GradientTape(watch_accessed_variables=False) as forward_tape:
forward_tape.watch(dummy_tensors)
dummy_gradients = tape.gradient(
[xi1_nat, xi2_nat], [nat1, nat2], output_gradients=dummy_tensors
)
# 2) the chain rule to get ∂L/∂η, where η (eta) are the expectation parameters
dL_deta1, dL_deta2 = tape.gradient(
meanvarsqrt, [eta1, eta2], output_gradients=[dL_dmean, dL_dvarsqrt]
)
if not isinstance(xi_transform, XiNat):
nat_dL_xi1, nat_dL_xi2 = forward_tape.gradient(
dummy_gradients, dummy_tensors, output_gradients=[dL_deta1, dL_deta2]
)
else:
nat_dL_xi1, nat_dL_xi2 = dL_deta1, dL_deta2
del tape # Remove "persistent" tape
xi1, xi2 = xi_transform.meanvarsqrt_to_xi(q_mu, q_sqrt)
xi1_new = xi1 - self.gamma * nat_dL_xi1
xi2_new = xi2 - self.gamma * nat_dL_xi2
# Transform back to the model parameters [q_mu, q_sqrt]
mean_new, varsqrt_new = xi_transform.xi_to_meanvarsqrt(xi1_new, xi2_new)
q_mu.assign(mean_new)
q_sqrt.assign(varsqrt_new)
def get_config(self):
config = super().get_config()
config.update({"gamma": self._serialize_hyperparameter("gamma")})
return config
#
# Auxiliary gaussian parameter conversion functions.
#
# The following functions expect their first and second inputs to have shape
# [D, N, 1] and [D, N, N], respectively. Return values are also of shapes [D, N, 1] and [D, N, N].
def swap_dimensions(method):
"""
Converts between GPflow indexing and tensorflow indexing
`method` is a function that broadcasts over the first dimension (i.e. like all tensorflow matrix ops):
`method` inputs [D, N, 1], [D, N, N]
`method` outputs [D, N, 1], [D, N, N]
:return: Function that broadcasts over the final dimension (i.e. compatible with GPflow):
inputs: [N, D], [D, N, N]
outputs: [N, D], [D, N, N]
"""
@functools.wraps(method)
def wrapper(a_nd, b_dnn, swap=True):
if swap:
if a_nd.shape.ndims != 2: # pragma: no cover
raise ValueError("The mean parametrization must have 2 dimensions.")
if b_dnn.shape.ndims != 3: # pragma: no cover
raise ValueError("The covariance parametrization must have 3 dimensions.")
a_dn1 = tf.linalg.adjoint(a_nd)[:, :, None]
A_dn1, B_dnn = method(a_dn1, b_dnn)
A_nd = tf.linalg.adjoint(A_dn1[:, :, 0])
return A_nd, B_dnn
else:
return method(a_nd, b_dnn)
return wrapper
@swap_dimensions
def natural_to_meanvarsqrt(nat1: tf.Tensor, nat2: tf.Tensor):
var_sqrt_inv = tf.linalg.cholesky(-2 * nat2)
var_sqrt = _inverse_lower_triangular(var_sqrt_inv)
S = tf.linalg.matmul(var_sqrt, var_sqrt, transpose_a=True)
mu = tf.linalg.matmul(S, nat1)
# We need the decomposition of S as L L^T, not as L^T L,
# hence we need another cholesky.
return mu, tf.linalg.cholesky(S)
@swap_dimensions
def meanvarsqrt_to_natural(mu: tf.Tensor, s_sqrt: tf.Tensor):
s_sqrt_inv = _inverse_lower_triangular(s_sqrt)
s_inv = tf.linalg.matmul(s_sqrt_inv, s_sqrt_inv, transpose_a=True)
return tf.linalg.matmul(s_inv, mu), -0.5 * s_inv
@swap_dimensions
def natural_to_expectation(nat1: tf.Tensor, nat2: tf.Tensor):
args = natural_to_meanvarsqrt(nat1, nat2, swap=False)
return meanvarsqrt_to_expectation(*args, swap=False)
@swap_dimensions
def expectation_to_natural(eta1: tf.Tensor, eta2: tf.Tensor):
args = expectation_to_meanvarsqrt(eta1, eta2, swap=False)
return meanvarsqrt_to_natural(*args, swap=False)
@swap_dimensions
def expectation_to_meanvarsqrt(eta1: tf.Tensor, eta2: tf.Tensor):
var = eta2 - tf.linalg.matmul(eta1, eta1, transpose_b=True)
return eta1, tf.linalg.cholesky(var)
@swap_dimensions
def meanvarsqrt_to_expectation(m: tf.Tensor, v_sqrt: tf.Tensor):
v = tf.linalg.matmul(v_sqrt, v_sqrt, transpose_b=True)
return m, v + tf.linalg.matmul(m, m, transpose_b=True)
def _inverse_lower_triangular(M):
"""
Take inverse of lower triangular (e.g. Cholesky) matrix. This function
broadcasts over the first index.
:param M: Tensor with lower triangular structure of shape [D, N, N]
:return: The inverse of the Cholesky decomposition. Same shape as input.
"""
if M.shape.ndims != 3: # pragma: no cover
raise ValueError("Number of dimensions for input is required to be 3.")
D, N = tf.shape(M)[0], tf.shape(M)[1]
I_dnn = tf.eye(N, dtype=M.dtype)[None, :, :] * tf.ones((D, 1, 1), dtype=M.dtype)
return tf.linalg.triangular_solve(M, I_dnn)