# This is a more complicated digit recognition task that data from actual
# handwritten digits.
# The training and test set come from different sets of people.  Sure you can
# get 100% accuracy on the training set, but can you break 98% on the test set?
# Try improving generalization with some weight decay (maybe around 0.001) or
# by changing the number of hidden units.

addNet "Handwritten Digits" 64 20 10 SOFT_MAX MAX_CRIT
loadExamples hand-digits.trn.bex -e PERMUTED
loadExamples hand-digits.tst.bex

setObj learningRate   0.2
setObj reportInterval 1
setObj randRange      0.1
setObj testGroupCrit  1.0
setObj numUpdates     50
resetNet

setObj input.numColumns 8
autoPlot 10

graphObj error