research/rebar/README.md
# REINFORCing Concrete with REBAR
*Implemention of REBAR (and other closely related methods) as described
in "REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models" by
George Tucker, Andriy Mnih, Chris J. Maddison, Dieterich Lawson, Jascha Sohl-Dickstein [(https://arxiv.org/abs/1703.07370)](https://arxiv.org/abs/1703.07370).*
Learning in models with discrete latent variables is challenging due to high variance gradient estimators. Generally, approaches have relied on control variates to reduce the variance of the REINFORCE estimator. Recent work ([Jang et al. 2016](https://arxiv.org/abs/1611.01144); [Maddison et al. 2016](https://arxiv.org/abs/1611.00712)) has taken a different approach, introducing a continuous relaxation of discrete variables to produce low-variance, but biased, gradient estimates. In this work, we combine the two approaches through a novel control variate that produces low-variance, unbiased gradient estimates. Then, we introduce a novel continuous relaxation and show that the tightness of the relaxation can be adapted online, removing it as a hyperparameter. We show state-of-the-art variance reduction on several benchmark generative modeling tasks, generally leading to faster convergence to a better final log likelihood.
REBAR applied to multilayer sigmoid belief networks is implemented in rebar.py and rebar_train.py provides a training/evaluation setup. As a comparison, we also implemented the following methods:
* [NVIL](https://arxiv.org/abs/1402.0030)
* [MuProp](https://arxiv.org/abs/1511.05176)
* [Gumbel-Softmax](https://arxiv.org/abs/1611.01144)
The code is not optimized and some computation is repeated for ease of
implementation. We hope that this code will be a useful starting point for future research in this area.
## Errata
11/27/2019
The _generator_network function has separate paths for the unconditional and conditional generative models. In the conditional generative models code path, the generative model does not have multiple stochastic layers even when n_layers is > 1. My intention was to have multiple stochastic layers in the conditional generative model, however, due to a bug this is not how it was implemented. As the code is currently, with the conditional generative model and n_layers > 1, the recognition network has multiple stochastic layers, but the generative model has a single stochastic layer.
Hai-Tao Yu (yuhaitao@slis.tsukuba.ac.jp) discovered this issue.
## Quick Start:
Requirements:
* TensorFlow (see tensorflow.org for how to install)
* MNIST dataset
* Omniglot dataset
First download datasets by selecting URLs to download the data from. Then
fill in the download_data.py script like so:
```
MNIST_URL = 'http://yann.lecun.com/exdb/mnist'
MNIST_BINARIZED_URL = 'http://www.cs.toronto.edu/~larocheh/public/datasets/binarized_mnist'
OMNIGLOT_URL = 'https://github.com/yburda/iwae/raw/master/datasets/OMNIGLOT/chardata.mat'
```
Then run the script to download the data:
```
python download_data.py
```
Then run the training script:
```
python rebar_train.py --hparams="model=SBNDynamicRebar,learning_rate=0.0003,n_layer=2,task=sbn"
```
and you should see something like:
```
Step 2084: [-231.026474 0.3711713 1. 1.06934261 1.07023323
1.02173257 1.02171052 1. 1. 1. 1. ]
-3.6465678215
Step 4168: [-156.86795044 0.3097114 1. 1.03964758 1.03936625
1.02627242 1.02629256 1. 1. 1. 1. ]
-4.42727231979
Step 6252: [-143.4650116 0.26153237 1. 1.03633797 1.03600132
1.02639604 1.02639794 1. 1. 1. 1. ]
-4.85577583313
Step 8336: [-137.65275574 0.22313026 1. 1.03467286 1.03428006
1.02336085 1.02335203 0.99999988 1. 0.99999988
1. ]
-4.95563364029
```
The first number in the list is the log likelihood lower bound and the number
after the list is the log of the variance of the gradient estimator. The rest of
the numbers are for debugging.
We can also compare the variance between methods:
```
python rebar_train.py \
--hparams="model=SBNTrackGradVariances,learning_rate=0.0003,n_layer=2,task=omni"
```
and you should see something like:
```
Step 959: [ -2.60478699e+02 3.84281784e-01 6.31126612e-02 3.27319391e-02
6.13379292e-03 1.98278503e-04 1.96425783e-04 8.83973844e-04
8.70995224e-04 -inf]
('DynamicREBAR', -3.725339889526367)
('MuProp', -0.033569782972335815)
('NVIL', 2.7640280723571777)
('REBAR', -3.539274215698242)
('SimpleMuProp', -0.040744658559560776)
Step 1918: [ -2.06948471e+02 3.35904926e-01 5.20901568e-03 7.81541676e-05
2.06885766e-03 1.08521657e-04 1.07351625e-04 2.30646547e-04
2.26554010e-04 -8.22885323e+00]
('DynamicREBAR', -3.864381790161133)
('MuProp', -0.7183765172958374)
('NVIL', 2.266523599624634)
('REBAR', -3.662022113800049)
('SimpleMuProp', -0.7071359157562256)
```
where the tuples show the log of the variance of the gradient estimators.
The training script has a number of hyperparameter configuration flags:
* task (sbn): one of {sbn, sp, omni} which correspond to MNIST generative
modeling, structured prediction on MNIST, and Omniglot generative modeling,
respectively
* model (SBNGumbel) : one of {SBN, SBNNVIL, SBNMuProp, SBNSimpleMuProp,
SBNRebar, SBNDynamicRebar, SBNGumbel SBNTrackGradVariances}. DynamicRebar automatically
adjusts the temperature, whereas Rebar and Gumbel-Softmax require tuning the
temperature. The ones named after
methods uses that method to estimate the gradients (SBN refers to
REINFORCE). SBNTrackGradVariances runs multiple methods and follows a single
optimization trajectory
* n_hidden (200): number of hidden nodes per layer
* n_layer (1): number of layers in the model
* nonlinear (false): if true use 2 x tanh layers between each stochastic layer,
otherwise use a linear layer
* learning_rate (0.001): learning rate
* temperature (0.5): temperature hyperparameter (for DynamicRebar, this is the initial
value of the temperature)
* n_samples (1): number of samples used to compute the gradient estimator (for the
experiments in the paper, set to 1)
* batch_size (24): batch size
* muprop_relaxation (true): if true use the new relaxation described in the paper,
otherwise use the Concrete/Gumbel softmax relaxation
* dynamic_b (false): if true dynamically binarize the training set. This
increases the effective training dataset size and reduces overfitting, though
it is not a standard dataset
Maintained by George Tucker (gjt@google.com, github user: gjtucker).