The goal of training is to create a classifier (in this case a neural network) that can be used to classify cat head images.

After a final round of training you should have the JSON state of a neural network in the file "network.json", which can be imported and used by kittydar.

First you need to collect positive and negative images to train the network with. See the collection directory for more information.

You can train a network with:

node train-network.js POSITIVES NEGATIVES

where POSITIVES is the directory of positive images (cat head crops), and NEGATIVES is a directory of samples from non-cat images.

This will write the network to "network.json".

After training the network you can test the network on a set of test positive and negative images (different from the ones that trained it):

node test-network.js POSITIVES_TEST NEGATIVES_TEST --network ./network.json

This will report the neural network error, as well as binary classification statistics like precision and recall.

Find the best parameters for the feature extraction and classifier with cross-validation. Edit the combos object to add a combination and run with:

node cross-validate.js POSITIVES NEGATIVES

This will cross-validate on each combination of parameteres and report statistics on each combination, including the precision, recall, accuracy, and error of the test set.

After you've trained a classifier, you can test the classifier on a different set of negative images and save any false positives as "hard negatives". You can take the hard negatives and the positives and train a new (more precise) classifier.

node mine-negatives.js NEGATIVES_EXTRA HARD --samples 1 --network ./network.json

where HARD is a new directory to hold the mined negatives. The threshold param determines when a negative is classified as hard. It's a number from 0.5 to 1.0 (from "leaning positive" to very false positive).

samples is the number of times to sample each negative image. It can take a lot of images to find a few hard negatives if you're classifier is good enough, so specifying a higher value will mine more hard negatives in the end.

You can then train a new classifier with:

node train-network.js POSITIVES HARD