matlab code for exponential family harmoniums, RBMs, DBNs, and relata

To run an EFH on your own data, you basically do three things:

1. Write four functions, one for each of:

   • generating the latent variables of the training data (for real-world data, this will probably be null)
   • generating the observed variables of the training data
   • generating the latent and observed variables of the testing data
   • computing some kind of scalar-valued learning measure (for the main code to plot; this can be omitted)

2. Add a new case to setParams.m specifying these functions and the hyperparameters for the EFH you'd like to train;

3. Change the call to setParams in DBNtrain.m (~line 23) to specify your case:

   params = setParams('datatype','my_new_case');

   Then you can just call DBNtrain from the command line.

What exactly has to be specified in your new case in setParams.m? A good case to copy is 'polyphonicmusic'. You don't have to set everything set in that case, but you will certainly want to set the following:

• numsUnits, a cell array of integers/vectors of integers

• typeUnits, a cell array of cell arrays of strings

• Ncases, cases/(mini)batch

• Nbatches, batches/epoch

• NepochsMax,

• setLearningSchedules, a function that sets the learning rates and how they change over time

• getLatents: a function that generates the data latent variables

  • inputs:
    • Nexamples, an integer specifying the number of examples per epoch
    • dataclass, a string specifying either 'double' or 'gpuArray' (in which case the code will execute the gpu-version)
  • outputs:
    • X a matrix of data latent variables of size [Nexamples, ...]
    • Q a structure for holding other miscellaneous parameters required to generate the training data

  These are really just suggested outputs, since they will only ever be passed to params.getData, which the user must also write (see next entry). For real (non-synthetic) data, you probably won't set these outputs to anything at all--just set this to a null function in setParams.m like this:

   params.getLatents = @(Nexamples,yrclass,varargin)(getLatentsNull(Nexamples,yrclass,T,varargin{:}));
  

  This will set X to an empty tensor whose first dimension has size Nexamples, allowing that parameter to be passed on to getData.

• getData: a function generating the data you will train on.

  • inputs: the outputs of getLatents, sc.,
    • X, a tensor of latent variables (i.e. of the true generative process, not the hidden vector of the EFH),
    • Q, a structure with whatever else you need to generate the data.
  • outputs:
    • R, a tensor of data (observed) variables, of size [Nexamples, sum(params.numsUnits{1})]
    • Q, the input structure of whatever else was needed to generate the data, possibly modified to reflect something about the data generation During training, X and Q will be returned as outputs from getLatents and passed as inputs to getData. So, if you're going to train on synthetic data, getLatents basically corresponds to the prior, and getData to the emission. If you're training on real data stored somewhere, you can just write the function you assign to getData so that it ignores its inputs.

• testEFH: a function for evaluating your EFH

  • inputs:
    • R, the (observed) testing data,
    • X, the latent variables of the testing data, if available
    • Q, the catch-all data-related structure (see above)
    • wts, the weights for the EFH as produced by EFH.m, a cell array of matrices
    • params, the master parameters returned by setParams.m.
  • outputs:
    • e, some scalar measuring how well or badly your model is doing. Don't set this to be the n-step reconstruction error, because that will be reported independently in any case. Instead, e.g., if you're using synthetic data, this function could infer the latent variables of the EFH and compare (somehow) to the "true" latent variables X. Or in sequential data (synthetic or not), one could infer the hidden vector of the EFH and then use this to predict the next observable data in the sequence (see testEFHNextFrameError.m, e.g.), and then report the mean error. Or etc.

• getTestData: A function for generating the data on which the EFH will be tested by params.testEFH.

  • inputs:
    • dataclass: either 'double' or 'gpuArray'. The latter will instruct this function to place data on the GPU.
  • outputs:
    • R, the observed variables,
    • X, the latent variables (if they're available)
    • Q, a catch-all structure for anything else.

  But since the outputs of this function only ever get used by params.testEFH, the user is really free to set them to whatever he likes, as long as they are compatible with the function he has assigned to params.testEFH.

...