1_resnet-34-export.md

Export ResNet-34

Trained by facebook using https://github.com/facebook/fb.resnet.torch

The model was converted to nn backend and BatchNorm folded into convolutional layers with this script https://github.com/szagoruyko/imagine-nn/blob/utils/utils.lua

gradWeight and gradBias were removed from convolutional layers.

top-1 central crop accuracy: 72.94%

Download link: https://www.dropbox.com/s/1pvw1xa5726iqe9/resnet-34-export.t7 (84MB)

Preprocessing

Separate mean std per channel is saved with the network:

> print(net.transform)
{
  mean :
    {
      1 : 0.48462227599918
      2 : 0.45624044862054
      3 : 0.40588363755159
    }
  std :
    {
      1 : 0.22889466674951
      2 : 0.22446679341259
      3 : 0.22495548344775
    }
}

Can be loaded without CUDA support.

2_convert.lua

require 'cudnn'
require 'cunn'
require 'inn'
local utils = require 'inn.utils'
local iterm = require 'iterm'
local generateGraph = require 'optnet.graphgen'

local name = 'resnet-34.t7'

local net = torch.load(name)
net:cuda():evaluate()

local input = torch.CudaTensor(10,3,224,224)

local function restoreBias(net)
   net:apply(function(x)
      if torch.typename(x):find'SpatialConvolution' then
         if not x.bias then
            x.bias = x.weight.new():resize(x.weight:size(1)):zero()
            x:resetWeightDescriptors()
         end
      end
   end)
end

print(utils.testSurgery(input, utils.foldBatchNorm, net))
print(utils.testSurgery(input, restoreBias, net))
print(utils.testSurgery(input, cudnn.convert, net, nn))

iterm.dot(generateGraph(net, input), name..'.pdf')

net.transform = {
   mean = {
      0.48462227599918,
      0.45624044862054,
      0.40588363755159,
   },
   std = {
      0.22889466674951,
      0.22446679341259,
      0.22495548344775,
   }
}

net:apply(function(x)
   if x.weight then
      x.gradWeight = nil
      x.gradBias = nil
   end
end)

net:clearState()
net:float()

torch.save('resnet-34-export.t7', net)

Does this conversion tool only support ResNet and Inception models? Or does it also support AlexNet?