trainer.md

Trainer

This module allows to perform training from a given network, dataset and options.

Initialization

trainer.initialize(network, criterion, opt)

This functions initializes the basic states of the optimizer. It must be called before any other method on the trainer. It takes as input the network, the criterion and an option table. The possible options are:

• opt.lr: the learning rate
• opt.lrd: the learning rate decay
• opt.mom: the momentum
• opt.wd: the weight decay
• opt.no_cuda: to disable the use of cuda
• opt.bs: the minibatch size

Training

trainer.train(dataset)

This function performs one epoch of training using the given dataset. It returns the total error on the dataset.

Testing

trainer.test(dataset)

This function evaluates the current network on the given dataset. It returns both the total error and accuracy on the dataset.

Example

local trainer = require 'trainer'

local train_dataset, test_dataset, network, criterion
------
-- Initialize all these elements with the other modules

local opt = {}
opt.lr = 0.01

trainer.initialize(network, criterion, opt)

local accuracy
for epoch=1,10 do
  trainer.train(train_dataset)
  _, accuracy = trainer.test(test_dataset)
end

print("Final accuracy: "..accuracy)