From b8c7166b9123735e8226d34b717d3cbc2dc1fa02 Mon Sep 17 00:00:00 2001
From: Francois Fleuret <francois@fleuret.org>
Date: Tue, 6 Dec 2016 09:07:06 +0100
Subject: [PATCH] So, back to decorating the classes and not the objects so
 that torch.save() does not complain with SpatialConvolution. Added the
 possibility to pass the total time to profiler.print() so that the fraction
 of time used by the different functions can be displayed.

---
 profiler.lua      | 32 ++++++++++++++++++++++----------
 test-profiler.lua | 11 ++++++++---
 2 files changed, 30 insertions(+), 13 deletions(-)

diff --git a/profiler.lua b/profiler.lua
index 34e180b..4e45787 100644
--- a/profiler.lua
+++ b/profiler.lua
@@ -50,9 +50,18 @@ function profiler.decorate(model, functionsToDecorate)
 
       local nameOrig = name .. '__orig'
 
-      if model[name] and not model[nameOrig] then
-         model[nameOrig] = model[name]
-         model[name] = function(self, ...)
+      -- We decorate the class and not the object, otherwise we cannot
+      -- save models anymore.
+
+      if rawget(model, name) then
+         error('We decorate the class, not the objects, and there is a ' .. name .. ' in ' .. model)
+      end
+
+      local toDecorate = getmetatable(model)
+
+      if toDecorate[name] and not toDecorate[nameOrig] then
+         toDecorate[nameOrig] = toDecorate[name]
+         toDecorate[name] = function(self, ...)
             local startTime = sys.clock()
             local result = { self[nameOrig](self, unpack({...})) }
             local endTime = sys.clock()
@@ -71,24 +80,27 @@ function profiler.decorate(model, functionsToDecorate)
 
 end
 
-function profiler.print(model, nbSamples, indent)
+function profiler.print(model, nbSamples, totalTime, indent)
    local indent = indent or ''
 
    print(string.format('%s* %s', indent, model.__typename))
 
    for l, t in pairs(model.accTime) do
-      local s
+      local s = string.format('%s  %s %.02fs', indent, l, t)
+      if totalTime then
+         s = s .. string.format(' [%.02f%%]', 100 * t / totalTime)
+      end
       if nbSamples then
-         s = string.format(' (%.01fmus/sample)', 1e6 * t / nbSamples)
-      else
-         s = ''
+         s = s .. string.format(' (%.01fmus/sample)', 1e6 * t / nbSamples)
       end
-      print(string.format('%s  %s %.02fs%s', indent, l, t, s))
+      print(s)
    end
 
+   print()
+
    if torch.isTypeOf(model, nn.Container) then
       for _, m in ipairs(model.modules) do
-         profiler.print(m, nbSamples, indent .. '  ')
+         profiler.print(m, nbSamples, totalTime, indent .. '  ')
       end
    end
 end
diff --git a/test-profiler.lua b/test-profiler.lua
index a78c944..18677ec 100755
--- a/test-profiler.lua
+++ b/test-profiler.lua
@@ -39,9 +39,14 @@ require 'profiler'
 
 -- Create a model
 
+local w, h, fs = 50, 50, 3
+local nhu =  (w - fs + 1) * (h - fs + 1)
+
 local model = nn.Sequential()
    :add(nn.Sequential()
-           :add(nn.Linear(1000, 1000))
+           :add(nn.SpatialConvolution(1, 1, fs, fs))
+           :add(nn.Reshape(nhu))
+           :add(nn.Linear(nhu, 1000))
            :add(nn.ReLU())
        )
    :add(nn.Linear(1000, 100))
@@ -55,7 +60,7 @@ torch.save('model.t7', model)
 
 -- Create the data and criterion
 
-local input = torch.Tensor(1000, 1000)
+local input = torch.Tensor(1000, 1, h, w)
 local target = torch.Tensor(input:size(1), 100)
 local criterion = nn.MSECriterion()
 
@@ -88,7 +93,7 @@ end
 
 -- Print the accumulated timings
 
-profiler.print(model, nbSamples)
+profiler.print(model, nbSamples, modelTime)
 -- profiler.print(model)
 
 print()
-- 
2.20.1