deebn.rbm — deebn 0.1.0

build-jd-rbm clj

(build-jd-rbm visible hidden classes)

Factory function to build a joint density RBM for testing purposes.

This RBM has two sets of visible units - the typical set representing each observation in the data set, and a softmax unit representing the label for each observation. These are combined, and the label becaomes part of the input vector.

build-rbm clj

(build-rbm visible hidden)

Factory function to produce an RBM record.

check-overfitting clj

(check-overfitting rbm train-sample validations)

Given an rbm, a sample from the training set, and a validation set, determine if the model is starting to overfit the data. This is measured by a difference in the average free energy over the training set sample and the validation set.

edn->CRBM clj

(edn->CRBM data)

The default map->RBM function provided by the defrecord doesn't provide us with the performant implementation (i.e. matrices and arrays from core.matrix), so this function adds a small step to ensure that.

edn->RBM clj

(edn->RBM data)

The default map->RBM function provided by the defrecord doesn't provide us with the performant implementation (i.e. matrices and arrays from core.matrix), so this function adds a small step to ensure that.

free-energy clj

(free-energy x rbm)

Compute the free energy of a given visible vector and RBM. Lower is better.

get-prediction clj

(get-prediction x rbm num-classes labeled?)

For a given observation and RBM, return the predicted class.

load-rbm clj

(load-rbm filepath)

Load a RBM from disk.

save-rbm clj

(save-rbm rbm filepath)

Save a RBM to disk.

select-overfitting-sets clj

(select-overfitting-sets dataset)

Given a dataset, attempt to choose reasonable validation and test sets to monitor overfitting.

test-rbm clj

(test-rbm rbm dataset num-classes)

Test a joint density RBM trained on a data set. Returns an error percentage.

dataset should have the label as the last entry in each observation.

train-epoch clj

(train-epoch rbm dataset learning-rate momentum batch-size)

Train a single epoch

train-rbm clj

(train-rbm rbm dataset params)

Given a training set, train an RBM

params is a map with various options: learning-rate: defaults to 0.1 initial-momentum: starting momentum. Defaults to 0.5 momentum: momentum after momentum-delay epochs have passed. Defaults to 0.9 momentum-delay: epochs after which momentum is used instead of initial-momentum. Defaults to 3 batch-size: size of each mini-batch. Defaults to 10 epochs: number of times to train the model over the entire training set. Defaults to 100 gap-delay: number of epochs elapsed before early stopping is considered gap-stop-delay: number of sequential epochs where energy gap is increasing before stopping

update-rbm clj

(update-rbm batch rbm learning-rate momentum)

Single batch step update of RBM parameters

update-weights clj

(update-weights ph ph2 batch pv)

Determine the weight gradient from this batch