zeiler.lua

-- Zeiler network
local function createModel(backend)
  local matio = require 'matio'
  
  local maxpooling
  local spatialconv
  local relu
  
  if backend == 'nn' then
    maxpooling = nn.SpatialMaxPooling
    spatialconv = nn.SpatialConvolution
    relu = nn.ReLU
  elseif backend == 'cudnn' then
    maxpooling = cudnn.SpatialMaxPooling
    spatialconv = cudnn.SpatialConvolution
    relu = cudnn.ReLU
  end

  local features = nn.Sequential()
  
  local fS =     {96, 256, 384, 384, 256}
  fS[0] = 3
  local ks =     {7,5,3,3,3}
  local stride = {2,2,1,1,1}
  local pad =    {1,0,1,1,1}
  local lrn =    {true,true,false,false,false}
  local pool =   {true,true,false,false,false}
  
  for i=1,#fS do
    features:add(spatialconv(fS[i-1],fS[i],ks[i],ks[i],stride[i],stride[i],pad[i],pad[i]))
    features:add(relu(true))
    if lrn[i] then
      features:add(inn.SpatialSameResponseNormalization(3,0.00005,0.75))
    end
    if pool[i] then
      features:add(maxpooling(3,3,2,2):ceil())
    end
  end
  
  local classifier = nn.Sequential()  
  local fS = {4096,4096,1000}
  fS[0] = 256*50
  
  for i=1,#fS do
    classifier:add(nn.Linear(fS[i-1],fS[i]))
    if i < #fS then
      classifier:add(relu(true))
      classifier:add(nn.Dropout(0.5,true))
    end
  end

  local modelsdir = 'data/models'
  local modelpath = paths.concat(modelsdir,'Zeiler_imagenet_weights.mat')
  assert(paths.filep(modelpath),
    'Parameters file not found: '..modelpath)
  
  local mat = matio.load(modelpath)
  local idx = 1
  for i=1,features:size() do
    if torch.typename(features:get(i))=='nn.SpatialConvolutionMM' or
      torch.typename(features:get(i))=='nn.SpatialConvolution' or
      torch.typename(features:get(i))=='cudnn.SpatialConvolution' then
      features:get(i).weight:copy(mat['conv'..idx..'_w']:transpose(1,4):transpose(2,3))
      features:get(i).bias:copy(mat['conv'..idx..'_b'])
      idx = idx + 1
    end
  end

  local idx = 6
  
  for i=1,classifier:size() do
    if torch.typename(classifier:get(i))=='nn.Linear' then
      classifier:get(i).weight:copy(mat['fc'..idx..'_w']:transpose(1,2))
      classifier:get(i).bias:copy(mat['fc'..idx..'_b'])
      idx = idx + 1
    end
  end
  
  local model = nn.Sequential()
  model:add(features)
  model:add(inn.SpatialPyramidPooling({{1,1},{2,2},{3,3},{6,6}}))
  model:add(classifier)
  return model
end

return createModel