deebn.dnn — deebn 0.1.0

classify-obv clj

(classify-obv dnn obv)

Given a DNN and a single observation, return the model's prediction.

dbn->dnn clj

(dbn->dnn dbn classes)

Given a pretrained Deep Belief Network, use the trained weights and biases to build a Deep Neural Network.

edn->DNN clj

(edn->DNN data)

The default map->DNN 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.

feed-forward clj

(feed-forward batch dnn)

Given an initial input batch and a DNN, feed the batch through the net, retaining the output of each layer.

layer-error clj

(layer-error weights next-error output)

Calculate the error for a particular layer in a net, given the weights for the next layer, the error for the next layer, and the output for the current layer.

load-dnn clj

(load-dnn filepath)

Load a DNN from disk.

net-output clj

(net-output net input)

Propagate an input matrix through the network.

prop-up clj

(prop-up input weights bias)

Given an input matrix, weight matrix, and bias vector, propagate the signal through the layer.

save-dnn clj

(save-dnn dnn filepath)

Save a DNN to disk.

softmax->class clj

(softmax->class x)

Get the predicted class from a softmax output.

test-dnn clj

(test-dnn dnn dataset)

Test a Deep Neural Network on a dataset. Returns an error percentage.

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

train-batch clj

(train-batch batch dnn observations learning-rate lambda)

Given a batch of training data and a DNN, update the weights and biases accordingly.

train-dnn clj

(train-dnn dnn dataset params)

Given a labeled dataset, train a DNN.

The dataset should have the label as the last element of each input vector.

params is a map that may have the following keys: batch-size: default 100 epochs: default 100 learning-rate: default 0.5 lambda: default 0.1

train-epoch clj

(train-epoch net dataset observations learning-rate lambda batch-size)

Given a training dataset and a net, train it for one epoch (one pass over the dataset).

train-top-layer clj

(train-top-layer dnn
                 dataset
                 observations
                 batch-size
                 epochs
                 learning-rate
                 lambda)

Pre-train the top logistic regression layer before moving to fine-tuning.

update-layer clj

(update-layer weights
              biases
              input
              error
              learning-rate
              lambda
              batch-size
              observations)

Update the weights and biases of a layer, given the previous weights and biases, input coming into the weights, the error for the layer, the learning rate, and the batch size.