Graphical Models

In the previous tutorials, we've learned how to train a family of models that ranged from shallow models (linear and logistic regression) to deep models (MLPs, ConvNets). We've seen that however the model was defined, the overall procedure was always the same:

  • define a model, and define a set of trainable/adjustable parameters
  • choose a loss function, an objective to minimize
  • pick an optimization algorithm, to minimize the loss function above, by modifying the trainable parameters

These three steps are really general, and can be found in many bodies of research: statistics, machine learning, computer vision, operating research, ...

In this tutorial, we're going to explore a different class of models, which also relies on the procedure above: undirected graphical models, with pairwise and unary potentials.

To describe graphical models, we use the package gm, which can be installed as any other package: torch-pkg install gmgm is based on UGM http://www.cs.ubc.ca/~schmidtm/Software/UGM.html , a really nice Matlab implementation of graphical models. This tutorial is for the most part based on Mark Schmidt's introduction to UGM and parameter estimation demos.

Although a bit out of scope in a deep-learning tutorial session, making the parallel between deep-learning models and graphical models is of interest. In fact, there are many tasks in signal processing that benefit from a structured prediction framework, in which a graphical model imposes a structure on the predictions, and the unary terms of the model are coming from some classifier. For tasks like scene understanding, or OCR, or in general, tasks where there's a need for jointly segmenting and classifying the input data, the combination of deep models and graphical models seems worth exploring.

The code associated with this tutorial is provided on GitHub, on this page.

Defining a Model

In pairwise undirected graphical models, the joint probability of a particular assignment to all the variables xixiis represented as a normalized product of a set of non-negative potential functions:


There is one potential function ϕϕ for each node ii, for each edge ee. Each edge connects two nodes, ejej and ekek. In a full graph, there is one edge between each possible pair of nodes. For common applications, such as computer vision, it's much more common to have locally connected graphs, i.e. graphs in which only (small) subsets of nodes are connected via edges.

The node potential function ϕiϕi gives a non-negative weight to each possible value of the random variable xixi. For example, we might set ϕi(x1=0)ϕi(x1=0) to 0.75 and ϕi(x1=1)ϕi(x1=1) to 0.25, which means means that node 11 has a higher potential of being in state 00 than state 11. Similarly, the edge potential function ϕeϕe gives a non-negative weight to all the combinations that xejxej and xekxek can take.

The normalization constant ZZ, or partition function, is a scalar value that forces the distribution to sum to one, over all possible joint configurations of the variables:


This normalizing constant ensures that the model defines a valid probability distribution.

Given the graph structure and the potentials, the gm package includes functions for four tasks:

  • Decoding: finding the joint configuration of the variables with the highest probability.
  • Inference: computing the normalization constant ZZ, as well as the probabilities of each variable taking each state.
  • Training (or parameter estimation): the task of computing the potentials that maximize the likelihood of set of data.

A Simple Model

Here we first show how to describe a very simple model. The first thing to define is the adjacency matrix of the graph, which by itself fully describes the architecture of the model. The package provides a couple of standard matrices.

-- define graph
nNodes = 10
adjacency = gm.adjacency.full(nNodes)

Once the adjacency matrix is defined, we create a graph, using the gm.graph method. The only information required is the number of possible (discrete) states per node, and the adjacency matrix. The maxIterparameter is used for loopy graphs, to inform the inference/decoding algorithms on how many iterations they should go through before stopping.

nStates = 2
g = gm.graph{adjacency=adjacency, nStates=nStates, maxIter=10, verbose=true}

In this first example, we are not going to train the potentials, but assume that we already have them. We need to define two types of potentials, the ϕiϕi potentials, and the ϕeϕe potentials. For each possible configuration, we need to provide a potential. Potentials are positively correlated with probabilities, so a high potential means a high probability.

-- unary potentials
nodePot = Tensor{{1,3}, {9,1}, {1,3}, {9,1}, {1,1},
                 {1,3}, {9,1}, {1,3}, {9,1}, {1,1}}
-- joint potentials: these are Potts potentials
edgePot = Tensor(g.nEdges,nStates,nStates)
basic = Tensor{{2,1}, {1,2}}
for e = 1,g.nEdges do
  edgePot[e] = basic
-- set potentials

At this stage, we have a fully parametrized graphical model, in which we can do inference and decoding.

But first a few definitions:

  • the decoding task is to find the most likely configuration of the variables.
  • the inference task is to find the normalizing constant ZZ, as well as all the marginal probabilities of individual nodes taking each state.

There are multiple methods to do inference and decoding. Exact methods involve an exhaustive search through all possible combinations of states.

optimal = g:decode('exact')
print('<gm.testme> maximum likelihood configuration:')
local nodeBel,edgeBel,logZ = g:infer('exact')
print('<gm.testme> node beliefs:')
print('<gm.testme> edge beliefs:')
print('<gm.testme> log(Z):')

Exact inference is of course impossible for any reasonably-sized model. A very well known inference technique is belief propagation, which is also exact in the case of chain-shaped or tree-shaped graphs. In the case of loopy graphs, the algorithm becomes approximate, and needs to be run iteratively, until convergence, or until a max number of iterations has been reached.

Same code using belief propagation:

optimal = g:decode('bp')
print('<gm.testme> optimal config with belief propagation:')
local nodeBel,edgeBel,logZ = g:infer('bp')
print('<gm.testme> node beliefs:')
print('<gm.testme> edge beliefs:')
print('<gm.testme> log(Z):')

So that's basically all there is to describing arbitrary graphs, thanks to the adjacency matrix; doing inference; and decoding the optimal joint configuration of states.

Conditional Random Field for Images

Let's now move on to a slightly more complex scenario: instead of defining the potential functions by hand, we're now going to train a set of parameters to produce optimal potential functions, i.e. potential functions that maximize the likelihood of each sample.

In this example, we will try to denoise an image, which was corrupted by some Gaussian noise. In this case, we have input variables (the observed pixels), and output variables (the clean labels). This is a typical scenario in which we can use conditional random fields (CRFs).

In CRFs, we have two types of variables: (i) the features XX are treated as fixed non-random variables (observed), and (ii) the labels yy are treated as random variables in a graph, where the parameters of the graph depend on the features.

I invite you to take a look at this great tutorial on CRFs, for a more formal introduction.

In our case, the labels will be the 32×3232×32 clean predictions, which can take binary states, and the features will be the 32×3232×32 pixels, which are continuous grayscale values.

The first step, as before, is to define the graph structure. We use a 4-connexity lattice, which is rather typical in the image world:

-- assuming images of geometry:
nRow = 32
nCol = 32
nNodes = nRows*nCols
-- define adjacency matrix (4-connexity lattice)
local adj = gm.adjacency.lattice2d(nRows,nCols,4)
-- create graph
nStates = 2
g = gm.graph{adjacency=adj, nStates=nStates, verbose=true, type='crf', maxIter=10}

Note the type flag, which is set to crf. This defines how the loss function will be computed. We can now move on to the next section: training this CRF.

Training our Model (CRF)

We then need to generate some training data:

-- load a clean, black and white image of an X
sample = image.load('some_32x32_image.png')
-- how many instances to generate in the training data:
nInstances = 100
-- generate some labels (y), which correspond to the MAP solution to the CRF:
y = tensor(nInstances,nRows*nCols)
for i = 1,nInstances do
  y[i] = sample
y = y + 1  -- Lua is 1-based
-- generate an ensemble of noisy versions of that image:
X = tensor(nInstances,1,nRows*nCols)
for i = 1,nInstances do
  X[i] = sample
X = X + randn(X:size())/2

Here are some examples of the noisy input samples:

Once we have training data, we need to generate all the features XX, on which the CRF is conditioned. As for our previous work in supervised learning, it is key to properly whiten / preprocess the training data:

-- create node features (normalized X and a bias)
Xnode = tensor(nInstances,2,nNodes)
Xnode[{ {},1 }] = 1 -- bias
-- normalize features:
nFeatures = X:size(2)
for f = 1,nFeatures do
  local Xf = X[{ {},f }]
  local mu = Xf:mean()
  local sigma = Xf:std()
Xnode[{ {},2 }] = X -- features (simple normalized grayscale)
nNodeFeatures = Xnode:size(2)
-- create edge features
nEdges = g.edgeEnds:size(1)
nEdgeFeatures = nNodeFeatures*2-1 -- sharing bias, but not grayscale features
Xedge = zeros(nInstances,nEdgeFeatures,nEdges)
for i = 1,nInstances do
  for e =1,nEdges do
     local n1 = g.edgeEnds[e][1]
     local n2 = g.edgeEnds[e][2]
     for f = 1,nNodeFeatures do
        -- get all features from node1
        Xedge[i][f][e] = Xnode[i][f][n1]
     for f = 1,nNodeFeatures-1 do
        -- get all features from node1, except bias (shared)
        Xedge[i][nNodeFeatures+f][e] = Xnode[i][f+1][n2]

The last step before training is to decide what the trainable parameters are going to be. To do so, we create a map that associates trainable parameters (weights) to nodes and edges in the graph. Here is the procedure:

-- tie node potentials to parameter vector
nodeMap = zeros(nNodes,nStates,nNodeFeatures)
for f = 1,nNodeFeatures do
  nodeMap[{ {},1,f }] = f
-- tie edge potentials to parameter vector
local f = nodeMap:max()
edgeMap = zeros(nEdges,nStates,nStates,nEdgeFeatures)
for ef = 1,nEdgeFeatures do
  edgeMap[{ {},1,1,ef }] = f+ef
  edgeMap[{ {},2,2,ef }] = f+ef
-- initialize parameters

And voila!, we can now train the CRF using arbitrary optimization techniques, as we did with the MLPs, ConvNets, and autoencoders.

In this case, we minimize the negative log-likelihood in order to maximize the likelihood. The negative log-likelihood (and its) gradient are computed by the nll() method of our graph (remember that we create the graph with the flag type='crf', which produces the correct nll() method).

With log-linear CRFs, the negative log-likelihood is still differentiable and jointly convex in the node and edge parameters, so we could compute its optimal value using a batch algorithm, such as L-BFGS. As usual though, if the training data is quite large, starting with an epoch or two of SGD can be significantly faster, and then turning to simple averaging SGD might be an option, as shown by Leon Bottou.

-- and train on 30 samples
sgdconf = {learningRate=1e-3}
for iter = 1,100 do
  local i = floor(uniform(1,nInstances)+0.5)
  local feval = function()
     return g:nll(Xnode[i],Xedge[i],y[i],'bp')

Note/Exercise: the trainable parameters are availabel in g.w. Can you infer how many trainable parameters are there in this example? Since you have access to the outer loop of the SGD trainer, can you easily implement an L2 regularization?

The results of the training procedure are shown here:

t7> gm.examples.trainCRF()
SGD @ iteration 1: objective = 709.78271289324
SGD @ iteration 2: objective = 86.476069966592
SGD @ iteration 3: objective = 22.048118240666
SGD @ iteration 4: objective = 10.34447115497
SGD @ iteration 5: objective = 12.507434214913
SGD @ iteration 6: objective = -0.24103941715157
SGD @ iteration 7: objective = 13.513625673658
SGD @ iteration 8: objective = 23.680651547275
SGD @ iteration 9: objective = 13.759790868401
SGD @ iteration 10: objective = 24.419772339165
SGD @ iteration 11: objective = 19.331090159347

This is inference (i.e. the marginal probabilities for each pixel):

And this is the decoding (i.e. the optimal joint configuration):

tutorial_graphicalmodels.txt · Last modified: 2012/10/02 13:28 (external edit)