svgd 1.0.7

A flexible PyTorch library for SVGD and MET-SVGD, with support for custom distributions, kernels, step-sizes, and callbacks.

Metropolis-Hastings Stein Variational Gradient Descent (MET-SVGD)

Official implementation of the library accompanying the paper "Particles Don't Care About Z: Towards Scaling Entropy Estimation of Unnormalized Densities".

MET-SVGD is a principled extension of P-SVGD (Messaoud et al., 2024), a particle-based variational method built on Stein Variational Gradient Descent (SVGD). It integrates concepts from SVGD, parametric variational inference (P-VI), and Metropolis-Hastings (MH), leveraging their complementary strengths to provide an efficient and scalable framework for approximating the entropy of distributions known only up to a normalization constant.

Messaoud, S., Mokeddem, B., Xue, Z., Pang, L., An, B., Chen, H., and Chawla, S. S²AC: Energy-Based Reinforcement Learning With Stein Soft Actor Critic. ICLR 2024.

Table of Contents

1. Installation
2. Quickstart
3. The SVGD Class
4. Custom Distributions
5. Custom Kernel Bandwidths and Step-Sizes
6. Callbacks
7. Citation

Installation

Installing the library requires python>=3.9:

pip install svgd

Quickstart

The main svgd class is SVGD. After initialization, particles are sampled either with svgd.sample or svgd.sample_with_log_q, depending on the use case.

The main component classes are TargetDistribution, InitialDistribution, KP, and LR. Details in Custom Distributions and Custom Kernel Bandwidths and Step-Sizes.

These classes and example implementations can be found at:

• svgd/distributions
• svgd/kernels/parameters
• svgd/lrs

from svgd import SVGD
from svgd.distributions import TorchDistribution, Gaussian
from svgd.kernels import RBF
from svgd.kernels.parameters import ParameterKP
from svgd.lrs import ParameterLR

import torch
from torch.optim.adam import Adam
from torch.distributions import MultivariateNormal

from tqdm import tqdm
import matplotlib.pyplot as plt

from math import log, sqrt

device = "cuda:3"
d = 10

# initialize the target distribution
target_distribution = TorchDistribution(
    MultivariateNormal(torch.zeros(d).to(device), torch.eye(d).to(device))
)

# initialize the initial distribution
initial_distribution = Gaussian(torch.ones(d).mul(2), torch.ones(d))

# initialize the kernel
kernel = RBF(
    ParameterKP(
        torch.tensor(1.0).log(),
        lambda x: x.clamp(log(1e-2), log(sqrt(d / 2))).exp(),
    )
)

# initialize the learning rate
lr = ParameterLR(torch.tensor(0.1).log(), lambda x: x.exp())

# initialize the SVGD object
svgd = SVGD(
    target_distribution=target_distribution,
    initial_distribution=initial_distribution.requires_grad_(True),
    kernel=kernel.requires_grad_(True),
    lr=lr.requires_grad_(True),
    divergence_control="metropolis-hastings",
    bound_lr=True,
).to(device)

# initialize the optimizer
optimizer = Adam(svgd.parameters(), 5e-2)

# for statistic collection
loss_hist = []
entropy_hist = []

# main training loop
for _ in tqdm(range(200)):
    # get samples and their log densities
    x, mask, log_q = svgd.sample_with_log_q(n_particles=100, n_steps=100)

    # if only particles are required, call:
    # x, _, _ = svgd.sample(n_particles=100, n_steps=100)

    # estimate the entropy
    entropy = log_q.mul(mask).sum(-1).div(mask.sum(-1)).mul(-1)

    # compute the kld loss
    log_p = target_distribution.log_prob(x).mul(mask).sum(-1).div(mask.sum(-1))
    loss = entropy.add(log_p).mul(-1)

    # update parameters
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

    loss_hist.append(loss.item())
    entropy_hist.append(entropy.item())

The SVGD Class

The signature of the __init__ method of the SVGD class is:

def __init__(
    self: Self@SVGD,
    target_distribution: TargetDistribution,
    initial_distribution: InitialDistribution,
    kernel: Kernel,
    lr: LR,
    divergence_control: DivergenceControl = None,
    bound_lr: bool = False,
    track_convergence: bool = False,
    callbacks: List[Callback] = [],
    leaky_lr_clamp: bool = False
) -> None: ...

target_distribution and initial_distribution are explained in depth in the next section.

Using the RBF kernel is explained in the Quickstart example, and creating custom bandwidths and step-sizes is explained with examples here.

divergence_control can be either of None, or metropolis-hastings (details here).

bound_lr controls whether the step-size condition is applied.

track_convergence indicates whether or not to compute self.stein_identity. Useful with Callbacks.

callbacks is an array of Callbacks. This is explained in Callbacks.

leaky_lr_clamp indicates whether the lr_bound is a hard clamp or not. Hard clamping kills gradient flow at the extremes.

Custom Distributions

from svgd.distributions import TargetDistribution, InitialDistribution

import torch
from torch import Tensor
from torch.nn import Module, Parameter

import matplotlib.pyplot as plt

from math import log, pi

To create a custom target distribution, all that is required is to create a class that implements the TargetDistribution interface, which can be found in svgd.distributions. Similarly for initial distributions.

The following are example target and initial diagonal GMM distributions. I will assume that means and stdevs both have the shape (n_components, n_dimensions) and weights is (n_components,). But, of course, as long as your implementation adheres to the interface, your distributions can be as complicated as you require.

class TargetDiagonalGMM(TargetDistribution):
    def __init__(self, means: Tensor, stdevs: Tensor, weights: Tensor):
        assert len(means.shape) == 2
        assert len(weights.shape) == 1
        assert means.shape == stdevs.shape
        assert means.shape[-2] == weights.shape[-1]
        assert stdevs.gt(0.0).all()

        self.means = means
        self.stdevs = stdevs
        self.weights = weights.softmax(-1)

    def log_prob(self, x: Tensor):
        # the input tensor `x` has shape (..., n, d)

        vars = self.stdevs.pow(2)

        return (
            x.unsqueeze(-2)
            .sub(self.means)
            .pow(2)
            .div(vars)
            .add(vars.log())
            .add(log(2 * pi))
            .sum(-1)
            .div(-2)
            .add(self.weights.log())
            .logsumexp(-1)
        )

class InitialDiagonalGMM(InitialDistribution, Module):
    def __init__(self, means: Tensor, stdevs: Tensor, weights: Tensor):
        Module.__init__(self)

        assert len(means.shape) == 2
        assert len(weights.shape) == 1
        assert means.shape == stdevs.shape
        assert means.shape[-2] == weights.shape[-1]
        assert stdevs.gt(0.0).all()

        self.means = Parameter(means)
        self.log_stdevs = Parameter(stdevs.log())
        self.logits_weights = Parameter(weights.log_softmax(-1))

    @property
    def stdevs(self):
        return self.log_stdevs.clamp(-5, 2).exp()

    def rsample(self, n_particles: int) -> Tensor:
        stdevs = self.stdevs
        weights = self.logits_weights.softmax(-1)

        idx = torch.multinomial(weights, n_particles, True)

        return (
            torch.randn(n_particles, self.means.shape[-1], device=self.means.device)
            .mul(stdevs[idx])
            .add(self.means[idx])
        )

    def log_prob(self, x: Tensor):
        vars = self.stdevs.pow(2)
        log_weights = self.logits_weights.log_softmax(-1)

        return (
            x.unsqueeze(-2)
            .sub(self.means)
            .pow(2)
            .div(vars)
            .add(vars.log())
            .add(log(2 * pi))
            .sum(-1)
            .div(-2)
            .add(log_weights)
            .logsumexp(-1)
        )

Custom Kernel Bandwidths and Step-Sizes

from svgd.lrs import LR
from svgd.kernels.parameters import KP
from svgd.states import StateForKernel, StateForLR

from torch import Tensor
from torch.nn import Module, Parameter, Linear

from typing_extensions import Union, List

The RBF kernel is defined as $$k(x,y) = \exp\left(-\frac{1}{2\sigma^2}||x-y||^2\right).$$

In this library, $\sigma$ is defined as a class that implements the KP interface available at svgd.kernels.parameters. The interface only requires that a forward method be implemented. It takes state: StateForKernel and **kwargs as attributes. state lists the available quantities on the SVGD object.

sigma must be a (...,) tensor, where the batch dimensions correspond to the batch dimensions of the particles returned by InitialDistribution.rsample.

Suppose we decide that we are always going to perform L SVGD steps and that InitialDistribution.rsample returns a tensor whose shape is (n, d), then a straightforward sigma implementation would be a (L,) tensor learnable via gradient descent. At each step, we pick the corresponding $\sigma_l$, whose shape would be (,) (i.e. a scalar).

class CustomSigma(KP, Module):
    def __init__(self, sigma: Tensor):
        assert sigma.dim() == 1
        assert sigma.ge(0.0).all()

        Module.__init__(self)

        self.log_sigma = Parameter(sigma.add(1e-5).log())

    def forward(self, state: StateForKernel, **kwargs):
        # state.x has shape (n, d), so the returned tensor should have shape `(,)` (i.e. it should be a scalar)
        return self.log_sigma[state.step].clamp(-10, 10).exp()

In other cases, your $\sigma$ will be parametrized by a neural network (e.g. reinforcement learning). In this case, we can't just inspect the weights to get the current sigma values as we would do with a Parameter $\sigma$, since every evaluation gives a different one, but we can track the values the network outputted throughout the SVGD steps. To that end, you could implement a logger wrapper class.

class SigmaNN(KP, Module):
    def __init__(self, obs_dim: int, hidden_dim: int):
        Module.__init__(self)

        self.obs: Union[None, Tensor] = None  # (..., obs_dim)

        self.l1 = Linear(obs_dim, hidden_dim)
        self.l2 = Linear(hidden_dim, hidden_dim)
        self.log_sigma = Linear(hidden_dim, 1)

    def set_observation(self, obs: Tensor):
        self.obs = obs

    def forward(self, state: StateForKernel, **kwargs):
        if self.obs is None:
            raise Exception()
        
        # in this context, for each `obs`, we have `n` particles
        # therefore, `state.x`'s shape is `(..., n, d)

        l1 = self.l1.forward(self.obs).relu()
        l2 = self.l2.forward(l1).relu()
        sigma = self.log_sigma.forward(l2).clamp(-10, 10).exp()
        
        self.obs = None

        return sigma.squeeze(-1)  # must be (...)

class SigmaWithHist(KP, Module):
    def __init__(self, sigma: KP):
        Module.__init__(self)

        self.sigma = sigma
        self.log_sigma = False

    def set_default_state(self):
        self.sigma_hist: List[Tensor] = []

    def forward(self, state: StateForKernel, **kwargs):
        sigma = self.sigma.forward(state, **kwargs)

        if not self.log_sigma:
            return sigma

        if state.step == 0:
            self.set_default_state()

        self.sigma_hist.append(sigma)

This can then be used as follows:

sigma = SigmaWithHist(SigmaNN(100, 64))
# ...
svgd = SVGD(kernel=RBF(sigma), ...)

for i in range(n_iter):
    sigma.log_sigma = i % 100 == 0
    obs, ... = ...
    sigma.set_observation(obs)
    x, mask, log_q = svgd.sample_with_log_q(n_particles, n_steps)
    # ... etc

Working with the learning rate is exactly the same. The interface that should be implemented is LR, which can be found at gamma.svgd.lrs. The learning rate, too, must be a (...) tensor.

class CustomLR(LR, Module):
    def __init__(self, lr: Tensor):
        assert lr.dim() == 1
        assert lr.ge(0.0).all()

        Module.__init__(self)

        self.log_lr = Parameter(lr.add(1e-5).log())

    def forward(self, state: StateForLR):
        return self.log_lr[state.step].clamp(-10, state.x.shape[-2]).exp()

Callbacks

Callback is an abstract class defined as:

class Callback:
    def on_initialization_done(self, state: StateOnInitializationDone):
        pass

    def on_sampling_iteration_started(self, state: StateOnSamplingIterationStarted):
        pass

    def on_score_computed(self, state: StateOnScoreComputed):
        pass

    def on_proposal_computed(self, state: StateOnProposalComputed):
        pass

    def on_sampling_iteration_done(self, state: StateOnSamplingIterationDone):
        pass

    def on_sampling_done(self, state: StateOnSamplingDone):
        pass

Callbacks are executed at specified stages identified by states during a given iteration in the following order:

1. on_initialization_done
2. on_sampling_iteration_started
3. on_score_computed
4. on_proposal_computed
5. on_sampling_iteration_done
6. on_sampling_done

For example, the state StateOnProposalComputed means that the loop is at the stage where the proposal has been computed and is available in the state variable.

Importantly, we expose f_attach_iteration, which controls whether or not to detach the gradients during a given iteration (it is set to False by default), and f_break_sampling_loop, which controls whether to break the sampling loop (e.g. based on the stein identity).

f_attach_iteration should be set in on_sampling_iteration_started, and f_break_sampling_loop before on_sampling_iteration_done.

All states are listed under svgd.states. Note that they build upon one another.

We provide three utility callbacks:

• svgd.callbacks.Breaker: A callback that halts SVGD sampling when Stein’s identity drops below a threshold or keeps increasing for too many consecutive iterations, signaling convergence or divergence.
• svgd.callbacks.Detacher: A callback that selectively attaches sampling iterations to the computational graph based on a fixed frequency and optionally the final iteration.
• svgd.callbacks.Logger: A callback that periodically logs selected SVGD diagnostics, statistics, and particle states during initialization and sampling.

Citation

If you found this library useful in your research, please consider citing the following works:

@inproceedings{
messaoud2024s2ac,
title={S\$2\${AC}: Energy-Based Reinforcement Learning with Stein Soft Actor Critic},
author={Safa Messaoud and Billel Mokeddem and Zhenghai Xue and Linsey Pang and Bo An and Haipeng Chen and Sanjay Chawla},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
}