Linear Machines and Trainers

Machines are one of the core components of Bob. They represent statistical models or other functions defined by parameters that can be learnt or manually set. The simplest of Bob‘s machines is a bob.learn.linear.Machine. This package contains the definition of this class as well as trainers that can learn linear machine parameters from data.

Linear machines

Linear machines execute the simple operation y = \mathbf{W} x, where y is the output vector, x is the input vector and W is a matrix (2D array) stored in the machine. The input vector x should be composed of double-precision floating-point elements. The output will also be in double-precision. Here is how to use a bob.learn.linear.Machine:

>>> W = numpy.array([[0.5, 0.5], [1.0, 1.0]], 'float64')
>>> W
array([[ 0.5,  0.5],
       [ 1. ,  1. ]])
>>> machine = bob.learn.linear.Machine(W)
>>> machine.shape
(2, 2)
>>> x = numpy.array([0.3, 0.4], 'float64')
>>> y = machine(x)
>>> y
array([ 0.55,  0.55])

As was shown in the above example, the way to pass data through a machine is to call its bob.learn.linear.Machine.forward() method, for which the __call__ method is an alias.

The first thing to notice about machines is that they can be stored and retrieved in bob.io.base.HDF5File. To save the before metioned machine to a file, just use the machine’s bob.learn.linear.Machine.save() command. Because several machines can be stored on the same bob.io.base.HDF5File, we let the user open the file and set it up before the machine can write to it:

>>> myh5_file = bob.io.base.HDF5File('linear.hdf5', 'w')
>>> #do other operations on myh5_file to set it up, optionally
>>> machine.save(myh5_file)
>>> del myh5_file #close

You can load the machine again in a similar way:

>>> myh5_file = bob.io.base.HDF5File('linear.hdf5')
>>> reloaded = bob.learn.linear.Machine(myh5_file)
>>> numpy.array_equal(machine.weights, reloaded.weights)
True

The shape of a bob.learn.linear.Machine (see bob.learn.linear.Machine.shape) indicates the size of the input vector that is expected by this machine and the size of the output vector it produces, in a tuple format like (input_size, output_size):

>>> machine.shape
(2, 2)

A bob.learn.linear.Machine also supports pre-setting normalization vectors that are applied to every input x. You can set a subtraction factor and a division factor, so that the actual input x' that is fed to the matrix W is x' = (x - s) ./ d. The variables s and d are vectors that have to have the same size as the input vector x. The operator ./ indicates an element-wise division. By default, s := 0.0 and d := 1.0.

>>> machine.input_subtract
array([ 0.,  0.])
>>> machine.input_divide
array([ 1.,  1.])

To set a new value for s or d just assign the desired machine property:

>>> machine.input_subtract = numpy.array([0.5, 0.8])
>>> machine.input_divide = numpy.array([2.0, 4.0])
>>> y = machine(x)
>>> y
array([-0.15, -0.15])

Note

In the event you save a machine that has the subtraction and/or a division factor set, the vectors are saved and restored automatically w/o user intervention.

Linear machine trainers

Next, we examine available ways to train a bob.learn.linear.Machine so they can do something useful for you.

Principal component analysis

PCA [1] is one way to train a bob.learn.linear.Machine. The associated Bob class is bob.learn.linear.PCATrainer as the training procedure mainly relies on a singular value decomposition.

PCA belongs to the category of unsupervised learning algorithms, which means that the training data is not labelled. Therefore, the training set can be represented by a set of features stored in a container. Using Bob, this container is a 2D numpy.ndarray.

>>> data = numpy.array([[3,-3,100], [4,-4,50], [3.5,-3.5,-50], [3.8,-3.7,-100]], dtype='float64')
>>> print(data)
[[   3.    -3.   100. ]
 [   4.    -4.    50. ]
 [   3.5   -3.5  -50. ]
 [   3.8   -3.7 -100. ]]

Once the training set has been defined, the overall procedure to train a bob.learn.linear.Machine with a bob.learn.linear.PCATrainer is simple and shown below. Please note that the concepts remains very similar for most of the other trainers and machines.

>>> trainer = bob.learn.linear.PCATrainer() # Creates a PCA trainer
>>> [machine, eig_vals] = trainer.train(data)  # Trains the machine with the given data
>>> print(machine.weights)  # The weights of the returned (linear) Machine after the training procedure
[[ 0.002 -0.706 -0.708]
 [-0.002  0.708 -0.706]
 [-1.    -0.003 -0.   ]]

Next, input data can be projected using this learned projection matrix W.

>>> e = numpy.array([3.2,-3.3,-10], 'float64')
>>> print(machine(e))
[ 9.999 0.47 0.092]

Linear discriminant analysis

LDA [2] is another way to train a bob.learn.linear.Machine. The associated Bob class is bob.learn.linear.FisherLDATrainer.

In contrast to PCA [1], LDA [2] is a supervised technique. Furthermore, the training data should be organized differently. It is indeed required to be a list of 2D numpy.ndarray‘s, one for each class.

>>> data1 = numpy.array([[3,-3,100], [4,-4,50], [40,-40,150]], dtype='float64')
>>> data2 = numpy.array([[3,6,-50], [4,8,-100], [40,79,-800]], dtype='float64')
>>> data = [data1,data2]

Once the training set has been defined, the procedure to train the bob.learn.linear.Machine with LDA is very similar to the one for PCA. This is shown below.

>>> trainer = bob.learn.linear.FisherLDATrainer()
>>> [machine,eig_vals] = trainer.train(data)  # Trains the machine with the given data
>>> print(eig_vals)  
[ 13.10097786 0. ]
>>> machine.resize(3,1)  # Make the output space of dimension 1
>>> print(machine.weights)  # The new weights after the training procedure
[[ 0.609]
 [ 0.785]
 [ 0.111]]

Whitening

This is generally used for i-vector preprocessing.

Let’s consider a 2D array of data used to train the withening, and a sample to be whitened:

>>> data = numpy.array([[ 1.2622, -1.6443, 0.1889], [ 0.4286, -0.8922, 1.3020], [-0.6613,  0.0430, 0.6377], [-0.8718, -0.4788, 0.3988], [-0.0098, -0.3121,-0.1807],  [ 0.4301,  0.4886, -0.1456]])
>>> sample = numpy.array([1, 2, 3.])

The initialisation of the trainer and the machine:

>>> t = bob.learn.linear.WhiteningTrainer()

Then, the training and projection are done as follows:

>>> m = t.train(data)
>>> withened_sample = m.forward(sample)

Within-Class Covariance Normalisation

This can also be used for i-vector preprocessing. Let’s first put the training data into list of numpy arrays.

>>> data = [numpy.array([[ 1.2622, -1.6443, 0.1889], [ 0.4286, -0.8922, 1.3020]]), numpy.array([[-0.6613,  0.0430, 0.6377], [-0.8718, -0.4788, 0.3988]]), numpy.array([[-0.0098, -0.3121,-0.1807],  [ 0.4301,  0.4886, -0.1456]])]

The initialisation of the trainer is done as follows:

>>> t = bob.learn.linear.WCCNTrainer()

Then, the training and projection are done as follows:

>>> m = t.train(data)
>>> wccn_sample = m.forward(sample)
[1](1, 2) http://en.wikipedia.org/wiki/Principal_component_analysis
[2](1, 2) http://en.wikipedia.org/wiki/Linear_discriminant_analysis