myAlexModel.lua

require 'cudnn'
require 'cunn'
local SpatialConvolution = cudnn.SpatialConvolution--lib[1]
local SpatialMaxPooling = cudnn.SpatialMaxPooling--lib[2]

-- from https://code.google.com/p/cuda-convnet2/source/browse/layers/layers-imagenet-1gpu.cfg
-- this is AlexNet that was presented in the One Weird Trick paper. http://arxiv.org/abs/1404.5997
local features = nn.Sequential()
features:add(SpatialConvolution(1,192,11,11,4,4,0,0))       -- 224 -> 55
features:add(SpatialMaxPooling(3,3,2,2))                   -- 55 ->  27
features:add(cudnn.ReLU(true))
features:add(nn.SpatialBatchNormalization(192,nil,nil,false))

features:add(SpatialConvolution(192,512,5,5,1,1,2,2))       --  27 -> 27
features:add(SpatialMaxPooling(3,3,2,2))                   --  27 ->  13
features:add(cudnn.ReLU(true))
features:add(nn.SpatialBatchNormalization(512,nil,nil,false))

features:add(SpatialConvolution(512,768,3,3,1,1,1,1))      --  13 ->  13
features:add(cudnn.ReLU(true))
features:add(nn.SpatialBatchNormalization(768,nil,nil,false))
features:add(SpatialConvolution(768,768,3,3,1,1,1,1))      --  13 ->  13
features:add(cudnn.ReLU(true))
features:add(nn.SpatialBatchNormalization(768,nil,nil,false))
features:add(SpatialConvolution(768,512,3,3,1,1,1,1))      --  13 ->  13
features:add(SpatialMaxPooling(3,3,2,2))                   -- 13 -> 6
features:add(cudnn.ReLU(true))
features:add(nn.SpatialBatchNormalization(512,nil,nil,false))

local classifier = nn.Sequential()
classifier:add(nn.View(512*5*5))
classifier:add(nn.Dropout(0.5))
classifier:add(nn.Linear(512*5*5, 4096))
classifier:add(nn.Threshold(0, 1e-6))
classifier:add(nn.BatchNormalization(4096,nil,nil,false))
classifier:add(nn.Dropout(0.5))
classifier:add(nn.Linear(4096, 4096))
classifier:add(nn.Threshold(0, 1e-6))
classifier:add(nn.BatchNormalization(4096,nil,nil,false))
classifier:add(nn.Linear(4096, 352))
classifier:add(nn.LogSoftMax())

local model = nn.Sequential()

function fillBias(m)
for i=1, #m.modules do
    if m:get(i).bias then
        m:get(i).bias:fill(0.1)
    end
end
end

--fillBias(features)
--fillBias(classifier)
model:add(features):add(classifier)

return model