Advanced Example: Simulate from a Bayesian Neural Network – Storage Constraints

Jack Baker

Often big datasets and high dimensionality go hand in hand. Sometimes the dimensionality is so high that storage of the full MCMC chain in memory becomes an issue. There are a number of ways around this, including: calculating the Monte Carlo estimate on the fly; reducing the dimensionality of the chain using a test function; or just periodically saving a the chain to the hard disk and starting from scratch. To give you more flexibility we allow an SGMCMC algorithm to be run step by step. This allows you to do what you want with the output of the chain. This guide goes into more detail about how to do this, but it needs more TensorFlow knowledge, such as knowledge of TensorFlow sessions and how to build your own placeholders. For more details on these see the TensorFlow for R documentation.

To demonstrate this concept we fit a two layer Bayesian neural network to the MNIST dataset. The MNIST dataset consists of \(28 \times 28\) pixel images of handwritten digits from zero to nine. The images are flattened to be a vector of length 784. The dataset is available as a standard dataset from the TensorFlow library, with a matrix of 55000 training vectors and 10000 test vectors, each with their corresponding labels. First, let’s construct the dataset and a testset. We assume you’ve read some of the earlier vignettes, so are familiar with how to do this. The MNIST dataset can be downloaded using the sgmcmc function getDataset as follows:

# Download and load MNIST dataset
mnist = getDataset("mnist")
# Build dataset list and testset list
dataset = list("X" = mnist$train$images, "y" = mnist$train$labels)
testset = list("X" = mnist$test$images, "y" = mnist$test$labels)

We’ll build the same neural network model as in the original SGHMC paper (Chen et. al 2014). Suppose \(Y_i\) takes values in \(\{0,\dots,9\}\), so is the output label of a digit, and \(\mathbf x_i\) is the input vector, with \(\mathbf X\) the full \(N \times 784\) dataset, where \(N\) is the number of observations. Then we model as follows \[ Y_i | \theta, \mathbf x_i \sim \text{Categorical}( \beta(\theta, \mathbf x_i) ), \\ \beta(\theta, \mathbf x_i) = \sigma \left( \sigma \left( \mathbf x_i^T B + b \right) A + a \right). \] Here \(A\), \(B\), \(a\), \(b\) are parameters to be inferred with \(\theta = (A, B, a, b)\); \(\sigma(.)\) is the softmax function. \(A\), \(B\), \(a\) and \(b\) are matrices with dimensions: \(100 \times 10\), \(784 \times 100\), \(1 \times 10\) and \(1 \times 100\) respectively. Each element of these parameters is given a Normal prior to give \[ A_{kl} | \lambda_A \sim N(0, \lambda_A^{-1}), \quad B_{jk} | \lambda_B \sim N(0, \lambda_B^{-1}), \\ a_l | \lambda_a \sim N(0, \lambda_a^{-1}), \quad b_k | \lambda_b \sim N(0, \lambda_b^{-1}), \\ j = 1,\dots,784; \quad k = 1,\dots,100; \quad l = 1,\dots,10; \] where \(\lambda_A\), \(\lambda_B\), \(\lambda_a\) and \(\lambda_b\) are hyperparameters. Finally we assume \[ \lambda_A, \lambda_B, \lambda_a, \lambda_b \sim \text{Gamma}(1, 1). \]

As you can see this is a lot of high dimensional parameters, and unless you have a lot of RAM to hand, a standard chain of length \(10^4\) will not fit into memory. First let’s create the params dictionary, and then we can code the logLik and logPrior functions. We’ll sample initial \(\lambda\) parameters from a standard Gamma, and the rest from a standard Normal as follows

# Sample initial weights from standard Normal
d = ncol(dataset$X) # dimension of chain
params = list()
params$A = matrix( rnorm(10*100), ncol = 10 )
params$B = matrix(rnorm(d*100), ncol = 100)
# Sample initial bias parameters from standard Normal
params$a = rnorm(10)
params$b = rnorm(100)
# Sample initial precision parameters from standard Gamma
params$lambdaA = rgamma(1, 1)
params$lambdaB = rgamma(1, 1)
params$lambdaa = rgamma(1, 1)
params$lambdab = rgamma(1, 1)

Now let’s declare the logLik and logPrior functions. Remember that for ease of use, all distribution functions implemented in the TensorFlow Probability package are located at tf$distributions (for more details see the Get Started page).

logLik = function(params, dataset) {
    # Calculate estimated probabilities
    beta = tf$nn$softmax(tf$matmul(dataset$X, params$B) + params$b)
    beta = tf$nn$softmax(tf$matmul(beta, params$A) + params$a)
    # Calculate log likelihood of categorical distn with probabilities beta
    logLik = tf$reduce_sum(dataset$y * tf$log(beta))

logPrior = function(params) {
    distLambda = tf$distributions$Gamma(1, 1)
    distA = tf$distributions$Normal(0, tf$rsqrt(params$lambdaA))
    logPriorA = tf$reduce_sum(distA$log_prob(params$A)) + distLambda$log_prob(params$lambdaA)
    distB = tf$distributions$Normal(0, tf$rsqrt(params$lambdaB))
    logPriorB = tf$reduce_sum(distB$log_prob(params$B)) + distLambda$log_prob(params$lambdaB)
    dista = tf$distributions$Normal(0, tf$rsqrt(params$lambdaa))
    logPriora = tf$reduce_sum(dista$log_prob(params$a)) + distLambda$log_prob(params$lambdaa)
    distb = tf$distributions$Normal(0, tf$rsqrt(params$lambdab))
    logPriorb = tf$reduce_sum(distb$log_prob(params$b)) + distLambda$log_prob(params$lambdab)
    logPrior = logPriorA + logPriorB + logPriora + logPriorb

Now suppose we want to make inference using stochastic gradient Langevin dynamics (SGLD). If we do this in the normal way then we will most likely run out of memory when the function builds the array to store the output. So instead we just initialize an sgld object using sgldSetup. Similarly we could build an sgldcv object using sgldcvSetup or an sgnht object using sgnhtSetup. Then we can run the SGLD algorithm one step at a time and decide what to do with the output at each iteration ourselves. We’ll just set our stepsize to 1e-4 for this example. To make the results reproducible we’ll set the seed to 13.

stepsize = 1e-4
sgld = sgldSetup(logLik, dataset, params, stepsize, logPrior = logPrior, 
        minibatchSize = 500, seed = 13)

This sgld object is a type of sgmcmc object, which is basically just a list with a number of entries. The most important of these entries to us is called params, which holds a list, with the same names as you had in the params you fed to sgld, but this list contains tf$Variable objects. This is how you access the tensors which hold your current parameter values in the chain. For more details on the attributes of these objects, see the documentation for sgldSetup, sgldcvSetup etc.

Now that we have created the sgld object, you want to initialise the TensorFlow graph and the sgmcmc algorithm you’ve chosen. If you are using a standard algorthm, this will just initialise the TensorFlow graph and all the tensors that were created. If you’re using an algorithm with control variates (e.g. sgldcv), then this will also find the MAP estimates of the parameter and calculate the full log posterior gradient at that point. The function we use to do this is initSess as follows

sess = initSess(sgld)

The sess returned by initSess is the current TensorFlow session, which is needed to run the SGMCMC algorithm of choice, and to access any of the tensors you need, such as sgld$params.

Now we have everything to run an SGLD algorithm step by step as follows

for (i in 1:10^3L) {
    sgmcmcStep(sgld, sess)
    currentState = getParams(sgld, sess)

Here the function sgmcmcStep will update sgld$params using a single update of SGLD, or whichever SGMCMC algorithm you chose. The function getParams will return a list of parameters as R objects rather than as tensors to make life easier for you.

Our simple example is fine, but we really would like to calculate a Monte Carlo average of the parameters on the fly. Also with these large examples, they take a long time to run, so it’s useful to check how the algorithm is doing every once in a while. This is especially useful when tuning by trial and error as you can stop an algorithm early if it’s doing badly. This is why we let you declare the TensorFlow session yourself: it lets you create your custom tensors to print algorithm progress, or to create your own test functions to reduce the chain dimensionality (they have to be declared before the TensorFlow session starts).

Let’s delete everything after we created our sgld object. Now we’re going to demonstrate a more complicated step by step example where we print performance and calulate the Monte Carlo estimate on the fly. Suppose we have test data \(X^*\), and test labels \(y^*\), and at some iteration \(i\) our SGMCMC algorithm outputs values for all the parameters \(\theta_t\). Then the probability that our neural network model will classify a given test observation to class \(k\) is given by \(\beta_k(\theta_t, \mathbf x_i^*)\); i.e. the \(k^{th}\) element of \(\beta(\theta_t, \mathbf x_i^*)\), which was defined earlier. A common performance measure for a classifier is the log loss, defined by \[ ll = - \frac{1}{N} \sum_{i=1}^{N_{\text{test}}} \sum_{k=1}^K y^*_{i,k} \log \beta( \theta_t, \mathbf x_i^* ). \]

This is also \(-\frac{1}{N}\) times the log likelihood at the current parameter, given the test set. So it’s very easy for us to calculate this in practice and output it. We’ll do this every 100 iterations to check the algorithm’s performance and check for convergence. To do this, we need to create a new placeholder to hold the test set, and then create a tensor that will calculate the log loss, which can easily be done using the logLik function already declared. First we’ll create a placeholder for both X and y in the test set, and make sure these have the same dimensions so can hold the full test set.

testPlaceholder = list()
testPlaceholder[["X"]] = tf$placeholder(tf$float32, dim(testset[["X"]]))
testPlaceholder[["y"]] = tf$placeholder(tf$float32, dim(testset[["y"]]))

Now we can create a tensor that calculates the log loss. We’ll link this to our testPlaceholder and the current parameter values, located at sgld$params.

# Get number of observations in test set, ensuring it's a double (R equivalent of float)
Ntest = as.double(nrow(testset[["X"]]))
logLoss = - logLik(sgld$params, testPlaceholder) / Ntest

Now we’ll declare the TensorFlow session, and run the chain step by step, calculating an average parameter estimate and printing the log loss of the current state every 100 iterations

sess = initSess(sgld)
# Fill a feed dict with full test set (used to calculate log loss)
feedDict = dict()
feedDict[[testPlaceholder[["X"]]]] = testset[["X"]]
feedDict[[testPlaceholder[["y"]]]] = testset[["y"]]
# Burn-in chain
message("Burning-in chain...")
message("iteration\tlog loss")
for (i in 1:10^3) {
    # Print progress
    if (i %% 100 == 0) {
        progress = sess$run(logLoss, feed_dict = feedDict)
        message(paste0(i, "\t", progress))
    sgmcmcStep(sgld, sess)
# Initialise Monte Carlo estimate using value after burn-in
avParams = getParams(sgld, sess)
# Run chain
message("Running SGMCMC...")
for (i in 1:10^4) {
    sgmcmcStep(sgld, sess)
    # Update av Params
    currentState = getParams(sgld, sess)
    for (paramName in names(avParams)) {
        avParams[[paramName]] = (avParams[[paramName]] * i + currentState[[paramName]]) / (i + 1)
    # Print progress
    if (i %% 100 == 0) {
        progress = sess$run(logLoss, feed_dict = feedDict)
        message(paste0(i, "\t", progress))

Obviously calculating the log loss is costly to do every 100 iterations as the test set has \(10^4\) observations itself, this was just for demonstration purposes. In practice we’d recommend subsampling this test set when calculating the log loss, or leaving more iterations until it’s calculated.