bayestorch 0.0.3

Lightweight Bayesian deep learning library for fast prototyping based on PyTorch

BayesTorch

Welcome to bayestorch, a lightweight Bayesian deep learning library for fast prototyping based on PyTorch. It provides the basic building blocks for the following Bayesian inference algorithms:

• Bayes by Backprop (BBB)
• Markov chain Monte Carlo (MCMC)
• Stein variational gradient descent (SVGD)

---

💡 Key features

• Low-code definition of Bayesian (or partially Bayesian) models
• Support for custom neural network layers
• Support for custom prior/posterior distributions
• Support for layer/parameter-wise prior/posterior distributions
• Support for composite prior/posterior distributions
• Highly modular object-oriented design
• User-friendly and easily extensible APIs
• Detailed API documentation

---

🛠️️ Installation

Using Pip

First of all, install Python 3.6 or later. Open a terminal and run:

pip install bayestorch

From source

First of all, install Python 3.6 or later. Clone or download and extract the repository, navigate to <path-to-repository>, open a terminal and run:

pip install -e .

---

▶️ Quickstart

Here are a few code snippets showcasing some key features of the library. For complete training loops, please refer to examples/mnist and examples/regression.

Bayesian model trainable via Bayes by Backprop

from torch.nn import Linear

from bayestorch.distributions import (
    get_mixture_log_scale_normal,
    get_softplus_inv_scale_normal,
)
from bayestorch.nn import VariationalPosteriorModule


# Define model
model = Linear(5, 1)

# Define log scale normal mixture prior over the model parameters
prior_builder, prior_kwargs = get_mixture_log_scale_normal(
    model.parameters(),
    weights=[0.75, 0.25],
    locs=(0.0, 0.0),
    log_scales=(-1.0, -6.0)
)

# Define inverse softplus scale normal posterior over the model parameters
posterior_builder, posterior_kwargs = get_softplus_inv_scale_normal(
    model.parameters(), loc=0.0, softplus_inv_scale=-7.0, requires_grad=True,
)

# Define Bayesian model trainable via Bayes by Backprop
model = VariationalPosteriorModule(
    model, prior_builder, prior_kwargs, posterior_builder, posterior_kwargs
)

Partially Bayesian model trainable via Bayes by Backprop

from torch.nn import Linear

from bayestorch.distributions import (
    get_mixture_log_scale_normal,
    get_softplus_inv_scale_normal,
)
from bayestorch.nn import VariationalPosteriorModule


# Define model
model = Linear(5, 1)

# Define log scale normal mixture prior over `model.weight`
prior_builder, prior_kwargs = get_mixture_log_scale_normal(
    [model.weight],
    weights=[0.75, 0.25],
    locs=(0.0, 0.0),
    log_scales=(-1.0, -6.0)
)

# Define inverse softplus scale normal posterior over `model.weight`
posterior_builder, posterior_kwargs = get_softplus_inv_scale_normal(
    [model.weight], loc=0.0, softplus_inv_scale=-7.0, requires_grad=True,
)

# Define partially Bayesian model trainable via Bayes by Backprop
model = VariationalPosteriorModule(
    model, prior_builder, prior_kwargs,
    posterior_builder, posterior_kwargs, [model.weight],
)

Composite prior

from torch.distributions import Independent
from torch.nn import Linear

from bayestorch.distributions import (
    CatDistribution,
    get_laplace,
    get_normal,
    get_softplus_inv_scale_normal,
)
from bayestorch.nn import VariationalPosteriorModule


# Define model
model = Linear(5, 1)

# Define normal prior over `model.weight`
weight_prior_builder, weight_prior_kwargs = get_normal(
    [model.weight],
    loc=0.0,
    scale=1.0,
    prefix="weight_",
)

# Define Laplace prior over `model.bias`
bias_prior_builder, bias_prior_kwargs = get_laplace(
    [model.bias],
    loc=0.0,
    scale=1.0,
    prefix="bias_",
)

# Define composite prior over the model parameters
prior_builder = (
    lambda **kwargs: CatDistribution([
        Independent(weight_prior_builder(**kwargs), 1),
        Independent(bias_prior_builder(**kwargs), 1),
    ])
)
prior_kwargs = {**weight_prior_kwargs, **bias_prior_kwargs}

# Define inverse softplus scale normal posterior over the model parameters
posterior_builder, posterior_kwargs = get_softplus_inv_scale_normal(
    model.parameters(), loc=0.0, softplus_inv_scale=-7.0, requires_grad=True,
)

# Define Bayesian model trainable via Bayes by Backprop
model = VariationalPosteriorModule(
    model, prior_builder, prior_kwargs, posterior_builder, posterior_kwargs,
)

---

📧 Contact

[e-mail]

---