cortex.experiment.train — thinktopic/experiment 0.9.22

backup-trained-network clj

(backup-trained-network network-filestem)

default-network-filestem clj

default-network-loss-eval-fn clj

(default-network-loss-eval-fn simple-loss-print? new-network test-ds batch-size)

Evaluate the network using its current loss terms

default-network-test-fn clj

(default-network-test-fn loss-val-fn
                         loss-compare-fn
                         {:keys [batch-size context]}
                         {:keys [new-network old-network test-ds]})

Test functions take two map arguments, one with global information and one with information local to the epoch. The job of a test function is to return a map indicating if the new network is indeed the best one and the network with enough information added to make comparing networks possible. {:best-network? boolean :network (assoc new-network :whatever information-needed-to-compare).}

load-network clj

(load-network network-filename)

Loads a map of {:cv-loss :network-description}.

print-trained-networks-summary clj

(print-trained-networks-summary
  &
  {:keys [network-filestem cv-loss->number cv-loss-display-precision extra-keys]
   :or {network-filestem default-network-filestem
        cv-loss->number (fn* [p1__67962#] (apply + (vals p1__67962#)))
        cv-loss-display-precision 3}})

Prints a summary of the different networks trained so far. Respects an (optional) network-filestem.

save-network clj

(save-network network network-filename)

Saves a trained network out to the filesystem.

train-n clj

(train-n network
         train-ds
         test-ds
         &
         {:keys [batch-size epoch-count network-filestem optimizer reset-score
                 force-gpu? simple-loss-print? test-fn context]
          :or {batch-size 128
               network-filestem default-network-filestem
               reset-score false}})

Given a network description, start training from scratch or given a trained network continue training. Keeps track of networks that are actually improving against a test-ds.

Networks are saved with a :cv-loss that is set to the best cv loss so far.

This system expects a dataset with online data augmentation so that it is effectively infinite although the cross-validation and holdout sets do not change. By default, the best network is saved to: trained-network.nippy

Note, we have to have enough memory to store the cross-validation dataset in memory while training.

Every epoch a test function is called with these 2 map arguments:

(test-fn global-context epoch-context)

It must return a map containing at least: {:best-network? true if this is the best network :network The new network with any extra information needed for comparison assoc'd onto it.}

If epoch-count is provided then we stop training after that many epochs else we continue to train forever.

trained-networks-folder clj