a replication of Network in network in torch. original code : https://github.com/mavenlin/cuda-convnet
Learning rate 0.1 is used in the original code, but it doesn't make the learning happen in this script. Learning rate is set to 2e-3, and the accuracy gets to %87.6 in 500 epochs not %89.6 :(
Another difference is the original code runs on batches of 128, this is 64.
This script works with the last version of optim/sgd.lua Please update
luarocks install optim