Torch7学习（七）——Neural-Style代码解析

torch7学习（一）——Tensor
Torch7学习(二) —— Torch与Matlab的语法对比
Torch7学习（三）——学习神经网络包的用法（1）
Torch7学习（四）——学习神经网络包的用法（2）
Torch7学习（五）——学习神经网路包的用法（3）
Torch7学习（六）——学习神经网络包的用法（4）——利用optim进行训练
Torch7学习（七）——Neural-Style代码解析

Neural-style用的可能是最容易入门的代码之一吧。比较简单，写的很清晰。只涉及到最简单的网络构建方式，同时也是最基本最重要的写法。直接看neural-style的代码吧。代码附有大量的注释，而在最后也有一定的分析。
值得注意的是，最新的该论文的代码已经是多GPU版本的了。可以指定不同层放入不同的GPU进行训练。，主要是加入了 Controlling Perceptual Factors in Neural Style Transfer代码的支持。个人觉得下面的版本比较简单，看懂了再看最新的版本的吧。https://github.com/jcjohnson/neural-style

代码

require 'torch'
require 'nn'
require 'image'
require 'optim'

require 'loadcaffe'

--------------------------------------------------------------------------------

local cmd = torch.CmdLine()

-- Basic options
cmd:option('-style_image', 'examples/inputs/starry_night.jpg',
           'Style target image')
cmd:option('-style_blend_weights', 'nil')
cmd:option('-content_image', 'examples/inputs/5.jpg',
           'Content target image')
cmd:option('-image_size', 512, 'Maximum height / width of generated image')
cmd:option('-gpu', 0, 'Zero-indexed ID of the GPU to use; for CPU mode set -gpu = -1')

-- Optimization options
cmd:option('-content_weight', 5e0)
cmd:option('-style_weight', 1e2)
cmd:option('-tv_weight', 1e-3)
cmd:option('-num_iterations', 1000)
cmd:option('-normalize_gradients', false)
cmd:option('-init', 'random', 'random|image')
cmd:option('-optimizer', 'lbfgs', 'lbfgs|adam')
cmd:option('-learning_rate', 1e1)

-- Output options
cmd:option('-print_iter', 50)
cmd:option('-save_iter', 100)
cmd:option('-output_image', 'out.png')

-- Other options
cmd:option('-style_scale', 1.0)
cmd:option('-pooling', 'max', 'max|avg')
cmd:option('-proto_file', 'models/VGG_ILSVRC_19_layers_deploy.prototxt')
cmd:option('-model_file', 'models/VGG_ILSVRC_19_layers.caffemodel')
cmd:option('-backend', 'nn', 'nn|cudnn|clnn')
cmd:option('-cudnn_autotune', false)
cmd:option('-seed', -1)

cmd:option('-content_layers', 'relu4_2', 'layers for content')
cmd:option('-style_layers', 'relu1_1,relu2_1,relu3_1,relu4_1,relu5_1', 'layers for style')

local function main(params)
  if params.gpu >= 0 then
    if params.backend ~= 'clnn' then
      require 'cutorch'
      require 'cunn'
      cutorch.setDevice(params.gpu + 1)
    else
      require 'clnn'
      require 'cltorch'
      cltorch.setDevice(params.gpu + 1)
    end
  else
    params.backend = 'nn'
  end

  if params.backend == 'cudnn' then
    require 'cudnn'
    if params.cudnn_autotune then
      cudnn.benchmark = true
    end
    cudnn.SpatialConvolution.accGradParameters = nn.SpatialConvolutionMM.accGradParameters -- ie: nop
  end

  local loadcaffe_backend = params.backend
  if params.backend == 'clnn' then loadcaffe_backend = 'nn' end
  local cnn = loadcaffe.load(params.proto_file, params.model_file, loadcaffe_backend):float()
  if params.gpu >= 0 then
    if params.backend ~= 'clnn' then
      cnn:cuda()
    else
      cnn:cl()
    end
  end

  local content_image = image.load(params.content_image, 3)
  content_image = image.scale(content_image, params.image_size, 'bilinear')
  local content_image_caffe = preprocess(content_image):float()

  local style_size = math.ceil(params.style_scale * params.image_size)
  local style_image_list = params.style_image:split(',')
  local style_images_caffe = {}
  for _, img_path in ipairs(style_image_list) do
    local img = image.load(img_path, 3)
    img = image.scale(img, style_size, 'bilinear')
    local img_caffe = preprocess(img):float()
    table.insert(style_images_caffe, img_caffe)
  end

  -- Handle style blending weights for multiple style inputs
  local style_blend_weights = nil
  if params.style_blend_weights == 'nil' then
    -- Style blending not specified, so use equal weighting
    style_blend_weights = {}
    -- #表示长度
    for i = 1, #style_image_list do
      table.insert(style_blend_weights, 1.0)
    end
  else
    style_blend_weights = params.style_blend_weights:split(',')
    assert(#style_blend_weights == #style_image_list,
      '-style_blend_weights and -style_images must have the same number of elements')
  end
  -- Normalize the style blending weights so they sum to 1
  local style_blend_sum = 0
  for i = 1, #style_blend_weights do
    style_blend_weights[i] = tonumber(style_blend_weights[i])
    style_blend_sum = style_blend_sum + style_blend_weights[i]
  end
  for i = 1, #style_blend_weights do
    style_blend_weights[i] = style_blend_weights[i] / style_blend_sum
  end


  if params.gpu >= 0 then
    if params.backend ~= 'clnn' then
      content_image_caffe = content_image_caffe:cuda()
      for i = 1, #style_images_caffe do
        style_images_caffe[i] = style_images_caffe[i]:cuda()
      end
    else
      content_image_caffe = content_image_caffe:cl()
      for i = 1, #style_images_caffe do
        style_images_caffe[i] = style_images_caffe[i]:cl()
      end
    end
  end

  local content_layers = params.content_layers:split(",")
  local style_layers = params.style_layers:split(",")

  -- Set up the network, inserting style and content loss modules
  local content_losses, style_losses = {}, {}
  local next_content_idx, next_style_idx = 1, 1
  local net = nn.Sequential()
  if params.tv_weight > 0 then
    local tv_mod = nn.TVLoss(params.tv_weight):float()
    if params.gpu >= 0 then
      if params.backend ~= 'clnn' then
        tv_mod:cuda()
      else
        tv_mod:cl()
      end
    end
    net:add(tv_mod)
  end
  for i = 1, #cnn do
    if next_content_idx <= #content_layers or next_style_idx <= #style_layers then
      local layer = cnn:get(i)
      local name = layer.name
      local layer_type = torch.type(layer)
      local is_pooling = (layer_type == 'cudnn.SpatialMaxPooling' or layer_type == 'nn.SpatialMaxPooling')
      if is_pooling and params.pooling == 'avg' then
        assert(layer.padW == 0 and layer.padH == 0)
        local kW, kH = layer.kW, layer.kH
        local dW, dH = layer.dW, layer.dH
        local avg_pool_layer = nn.SpatialAveragePooling(kW, kH, dW, dH):float()
        if params.gpu >= 0 then
          if params.backend ~= 'clnn' then
            avg_pool_layer:cuda()
          else
            avg_pool_layer:cl()
          end
        end
        local msg = 'Replacing max pooling at layer %d with average pooling'
        print(string.format(msg, i))
        net:add(avg_pool_layer)
      else
        -- 如果不是pooling层，直接add这一层
        net:add(layer)
      end
      if name == content_layers[next_content_idx] then
        print("Setting up content layer", i, ":", layer.name)
        -- 如果这一层是content的话，那么就要加入loss_module，而loss_module则需要：content_weight, target, norm来进行初始化类。
        -- target就是“输入”经过前面网络所有层得到的输出。因此target = net:forward(content_image_caffe):clone.
        local target = net:forward(content_image_caffe):clone()
        local norm = params.normalize_gradients
        local loss_module = nn.ContentLoss(params.content_weight, target, norm):float()
        if params.gpu >= 0 then
          if params.backend ~= 'clnn' then
            loss_module:cuda()
          else
            loss_module:cl()
          end
        end
        net:add(loss_module)
        table.insert(content_losses, loss_module)
        next_content_idx = next_content_idx + 1
      end

      if name == style_layers[next_style_idx] then
        print("Setting up style layer  ", i, ":", layer.name)
        local gram = GramMatrix():float()
        if params.gpu >= 0 then
          if params.backend ~= 'clnn' then
            gram = gram:cuda()
          else
            gram = gram:cl()
          end
        end
        local target = nil
        -- style_images_caffe是众多style_images的Tensor组成的table。
        -- 因此要将每个style_images_caffe[i]送入net中得到相应的输出。每个输出要经过gram处理，
        -- 得到相应的grams值，每张style图片所附的权值得到target
        for i = 1, #style_images_caffe do
          local target_features = net:forward(style_images_caffe[i]):clone()
          local target_i = gram:forward(target_features):clone()
          target_i:div(target_features:nElement())
          target_i:mul(style_blend_weights[i])
          if i == 1 then
            target = target_i
          else
            target:add(target_i)
          end
        end

        l