cortex.experiment.classification — thinktopic/experiment 0.9.22

create-listener clj

(create-listener observation->image-fn class-mapping argmap)

initializes any prerequisites for listening functions, and returns a listener function. Arguments:

• observation->image-fn: A function that can take observation data and return png data for web display.
  • class-mapping: A map with two entries
    • :class-name->index a map from class name strings to softmax indexes
    • :index->class-name a map from softmax indexes to class name strings Trains the net indefinitely on the provided training data, evaluating against the test data, and gives live updates on a local webserver hosted at http://localhost:8091.

create-tensorboard-listener clj

(create-tensorboard-listener {:keys [file-path]})

initializes any prerequisites for listening functions, and returns a listener function. Takes a file-path argument where the events are logged to.

log-weights clj

(log-weights network)

Given a network, writes all the weights from different layers to the tensorboard event file

perform-experiment clj

(perform-experiment initial-description train-ds test-ds listener)

(perform-experiment initial-description train-ds test-ds listener train-args)

Main entry point:

• initial-description: A cortex neural net description to train.
• train-ds: A dataset (sequence of maps) with keys :data, :labels, used for training.
• test-ds: A dataset (sequence of maps) with keys :data, :labels, used for testing.
• listener: a function which takes 3 arguments: initial-description, train-ds and test-ds and returns a function that is executed per epoch. It could be used to evaluate status of training (e.g. for early stopping) or to save the network.
• train-args: a map of optional arguments such as a force-gpu?. See cortex.experiment.train/train-n for the full list of arguments

tensorboard-log clj

(tensorboard-log file-path
                 {:keys [batch-size context]}
                 {:keys [new-network old-network test-ds train-ds]})

Given the context, old network, the new network and a test dataset, return a map with the updated network. As a side-effect, stream the train/test loss as well as all buffers as tensorboard events, appended to the file-path argument