A rough benchmark script for depthwise convolution

measure-time-dw-conv.lua

require 'cunn'

torch.setdefaulttensortype('torch.FloatTensor')

torch.manualSeed(666)
cutorch.manualSeed(666)

-- change these as needed
-- no automated testing, sorry :(
local nInputPlane, nOutputPlane, inputHeight, inputWidth = 128, 1, 52, 52

input = torch.rand(160, nInputPlane, inputHeight, inputWidth):cuda()

-- *** depthwise ***
--require 'cudnn'; module = cudnn.SpatialConvolution(nInputPlane, nInputPlane, 3,3, 1,1, 1,1, nInputPlane)
module = nn.SpatialDepthWiseConvolution(nInputPlane, nOutputPlane, 3,3, 1,1, 1,1)

-- *** full ***
--require 'cudnn'; module = cudnn.SpatialConvolution(128, 128, 3,3, 1,1, 1,1)
--module = nn.SpatialConvolution(128, 128, 3,3, 1,1, 1,1)

module:cuda()

local nRepeats = 10 -- just to make sure GPU is still
local nRepeatsInner = 4
local times = torch.Tensor(nRepeats)

module:forward(input)

for k = 1,nRepeats do
   io.stdout:write(k .. ', '):flush()

   collectgarbage()
   cutorch.synchronize()
   local timer = torch.Timer()

   for _ = 1,nRepeatsInner do
      module:forward(input)
   end
   cutorch.synchronize()

   times[k] = timer:time().real
end

--print(times)
print('Forward average: ' .. times[{{2,-1}}]:mean() .. ' +/- ' .. times[{{2,-1}}]:std() * 2.5)

gradOutput = torch.rand(module.output:size())
module:backward(input, gradOutput)

for k = 1,nRepeats do
   io.stdout:write(k .. ', '):flush()

   collectgarbage()
   cutorch.synchronize()
   local timer = torch.Timer()

   for _ = 1,nRepeatsInner do
      --module:updateGradInput(input, gradOutput)
      --module:accGradParameters(input, gradOutput)
      module:backward(input, gradOutput)
   end
   cutorch.synchronize()

   times[k] = timer:time().real
end

--print(times)
print('Backward average: ' .. times[{{2,-1}}]:mean() .. ' +/- ' .. times[{{2,-1}}]:std() * 2.5)