local layerSize = {inputSize, 67, 67, 100}
model = nn.Sequential()
model:add(nn.View(28 * 28)) --reshapes the image into a vector without copy
for i=1, #layerSize-1 do
model:add(nn.Linear(layerSize[i], layerSize[i+1]))
model:add(nn.ReLU())
end
criterion = nn.ClassNLLCriterion():cuda()
optim.sgd(feval, w, optimState)
epochs = 200 batchSize = 128
optimState = {
learningRate = 0.35,
momentum = 0.7,
weightDecay = 5e-05
}
lr:[0.35] , momentum:[0.7] , wDecay:[5e-05]