So, back to decorating the classes and not the objects so that torch.save() does...

[profiler-torch.git] / profiler.lua

diff --git a/profiler.lua b/profiler.lua
index 34e180b..4e45787 100644 (file)
--- 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