This package allows pruning Flux models into sparse arrays.

An iterative pruning strategy first prunes the trained model and then retrains it (which is called fine-tuning) until the stop condition is achieved.
A pruning schedule defines the sequence of pruning and fine tuning steps.

You can create an iterative scheduler using

schedule = [
    (PruneByPercentage(0.50), TuneByEpochs(1)),
    (PruneByPercentage(0.75), TuneByEpochs(3)),
    (PruneByPercentage(0.90), TuneByEpochs(5))
]

Then, you can run the schedule with

sparsemodel = scheduledpruning(trainedmodel, schedule, lossfunction, optiser, data)

A working example, training a model in the MNIST dataset, can be found in the examples folder.

• Percentage of randomly chosen weights
• Quantity of randomly chosen weights
• All weights within a given magnitude
• Percentage of the weights with the smallest magnitude
• Quantity of the weights with the smallest magnitude
• Percentage of the weights according to a given sensitivity function
• Quantity of the weights according to a given sensitivity function

• Schedule by epochs
• Schedule by absolute loss value
• Schedule by loss difference from the previous epoch