Probabilistic Torch is library for deep generative models that extends PyTorch

Probabilistic Torch is library for deep generative models that extends PyTorch. It is similar in spirit and design goals to Edward and Pyro, sharing many design characteristics with the latter.

The design of Probabilistic Torch is intended to be as PyTorch-like as possible. Probabilistic Torch models are written just like you would write any PyTorch model, but make use of three additional constructs:

1. A library of reparameterized distributions that implement methods for sampling and evaluation of the log probability mass and density functions (now available in PyTorch)

2. A Trace data structure, which is both used to instantiate and store random variables.

3. Objective functions that approximate the lower bound on the log marginal likelihood using Monte Carlo and Importance-weighted estimators.

This repository accompanies the NIPS 2017 paper:

@inproceedings{siddharth2017learning,
    title = {Learning Disentangled Representations with Semi-Supervised Deep Generative Models},
    author = {Siddharth, N. and Paige, Brooks and van de Meent, Jan-Willem and Desmaison, Alban and Goodman, Noah D. and Kohli, Pushmeet and Wood,
Frank and Torr, Philip},
    booktitle = {Advances in Neural Information Processing Systems 30},
    editor = {I. Guyon and U. V. Luxburg and S. Bengio and H. Wallach and R. Fergus and S. Vishwanathan and R. Garnett},
    pages = {5927--5937},
    year = {2017},
    publisher = {Curran Associates, Inc.},
    url = {http://papers.nips.cc/paper/7174-learning-disentangled-representations-with-semi-supervised-deep-generative-models.pdf}
}

Contributors

(in order of joining)

• Jan-Willem van de Meent
• Siddharth Narayanaswamy
• Brooks Paige
• Alban Desmaison
• Alican Bozkurt
• Amirsina Torfi
• Babak Esmaeli
• Eli Sennesh

Installation

1. Install PyTorch [instructions]

2. Install this repository from source

pip install git+git://github.com/probtorch/probtorch

3. Refer to the examples/ subdirectory for Jupyter notebooks that illustrate usage.

4. To build and read the API documentation, please do the following

git clone git://github.com/probtorch/probtorch
cd probtorch/docs
pip install -r requirements.txt
make html
open build/html/index.html

Mini-Tutorial: Semi-supervised MNIST

Models in Probabilistic Torch define variational autoencoders. Both the encoder and the decoder model can be implemented as standard PyTorch models that subclass nn.Module.

In the __init__ method we initialize network layers, just as we would in a PyTorch model. In the forward method, we additionally initialize a Trace variable, which is a write-once dictionary-like object. The Trace data structure implements methods for instantiating named random variables, whose values and log probabilities are stored under the specifed key.

Here is an implementation for the encoder of a standard semi-supervised VAE, as introduced by Kingma and colleagues [1]

import torch
import torch.nn as nn
import probtorch

class Encoder(nn.Module):
    def __init__(self, num_pixels=784, num_hidden=50, num_digits=10, num_style=2):
        super(self.__class__, self).__init__()
        self.h = nn.Sequential(
                    nn.Linear(num_pixels, num_hidden),
                    nn.ReLU())
        self.y_log_weights = nn.Linear(num_hidden, num_digits)
        self.z_mean = nn.Linear(num_hidden + num_digits, num_style)
        self.z_log_std = nn.Linear(num_hidden + num_digits, num_style)

    def forward(self, x, y_values=None, num_samples=10):
        q = probtorch.Trace()
        x = x.expand(num_samples, *x.size())
        if y_values is not None:
            y_values = y_values.expand(num_samples, *y_values.size())
        h = self.h(x)
        y = q.concrete(logits=self.y_log_weights(h), temperature=0.66,
                       value=y_values, name='y')
        h2 = torch.cat([y, h], -1)
        z = q.normal(loc=self.z_mean(h2),
                     scale=torch.exp(self.z_log_std(h2)),
                     name='z')
        return q

In the code above, the method q.concrete samples or observes from a Concrete/Gumbel-Softmax relaxation of the discrete distribution, depending on whether supervision values y_values are provided. The method q.normal samples from a univariate normal.

The resulting trace q now contains two entries q['y'] and q['z'], which are instances of a RandomVariable class, which stores both the value and the log probability associated with the variable. The stored values are now used to condition execution of the decoder model:

def binary_cross_entropy(x_mean, x, EPS=1e-9):
    return - (torch.log(x_mean + EPS) * x +
              torch.log(1 - x_mean + EPS) * (1 - x)).sum(-1)

class Decoder(nn.Module):
    def __init__(self, num_pixels=784, num_hidden=50, num_digits=10, num_style=2):
        super(self.__class__, self).__init__()
        self.num_digits = num_digits
        self.h = nn.Sequential(
                   nn.Linear(num_style + num_digits, num_hidden),
                   nn.ReLU())
        self.x_mean = nn.Sequential(
                        nn.Linear(num_hidden, num_pixels),
                        nn.Sigmoid())

    def forward(self, x, q=None):
        if q is None:
            q = probtorch.Trace()
        p = probtorch.Trace()
        y = p.concrete(logits=torch.zeros(x.size(0), self.num_digits),
                       temperature=0.66,
                       value=q['y'], name='y')
        z = p.normal(loc=0.0, scale=1.0, value=q['z'], name='z')
        h = self.h(torch.cat([y, z], -1))
        p.loss(binary_cross_entropy, self.x_mean(h), x, name='x')
        return p

The model above can be used both for conditioned forward execution, but also for generation. The reason for this is that q[k] returns None for variable names k that have not been instantiated.

To train the model components above, probabilistic Torch provides objectives that compute an estimate of a lower bound on the log marginal likelihood, which can now be maximized with standard PyTorch optimizers

from probtorch.objectives.montecarlo import elbo
from random import rand
# initialize model and optimizer
enc = Encoder()
dec = Decoder()
optimizer =  torch.optim.Adam(list(enc.parameters())
                              + list(dec.parameters()))
# define subset of batches that will be supervised
supervise = [rand() < 0.01 for _ in data]
# train model for 10 epochs
for epoch in range(10):
    for b, (x, y) in data:
        x = Variable(x)
        if supervise[b]:
            y = Variable(y)
            q = enc(x, y)
        else:
            q = enc(x)
        p = dec(x, q)
        loss = -elbo(q, p, sample_dim=0, batch_dim=1)
        loss.backward()
        optimizer.step()

For a more details, see the Jupyter notebooks in the examples/ subdirectory.

References

[1] Kingma, Diederik P, Danilo J Rezende, Shakir Mohamed, and Max Welling. 2014. “Semi-Supervised Learning with Deep Generative Models.” http://arxiv.org/abs/1406.5298.