/highwaytest

Primary LanguageJupyter Notebook

This is my attempt to recreate the results in http://arxiv.org/pdf/1505.00387v2.pdf

Basically just run

Usage: th main.lua [options] 
Testing the HighwayLayers

Options
  -json       json output file [/dev/null]
  -type       layer type: should be vanilla or highway [vanilla]
  -set        input set: should be mnist or xor [mnist]
  -layers     number of layers [2]
  -size       hidden layer size [71]
  -max_epochs number of full passes through the training data [200]
  -seed       torch manual random number generator seed [12345]
  -gpuid      which gpu to use. -1 = use CPU [-1]
  -cudnn      Use CUDNN [0]

So, I plan to do this:

for layers in 10 20 50 100
do
	th main.lua -gpuid 0 -cudnn 1 -json "results/"vanilla-${layers}.json" -type vanilla -set mnist -layers "${layers}" -size 71 -max_epochs 400 
	th main.lua -gpuid 1 -cudnn 1 -json "results/"highway-${layers}.json" -type highway -set mnist -layers "${layers}" -size 50 -max_epochs 400 
done

(Tho, obviously, I'll run these in parallel on multiple shells.)