X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=dagnn.lua;h=158ef785d446fb22f6226248c6a07c83cf47b6c3;hb=e6516772e13dd5424f0a1b7e2063a7417614844c;hp=1b467e720542469d45228b1dbc8a8fd0b021f6ad;hpb=e5030cca047eed4b8c5db172fc52e893b1b1d843;p=dagnn.git

diff --git a/dagnn.lua b/dagnn.lua
index 1b467e7..158ef78 100755
--- a/dagnn.lua
+++ b/dagnn.lua
@@ -1,4 +1,23 @@
 
+--[[
+
+   Copyright (c) 2016 Idiap Research Institute, http://www.idiap.ch/
+   Written by Francois Fleuret <francois.fleuret@idiap.ch>
+
+   This file is free software: you can redistribute it and/or modify
+   it under the terms of the GNU General Public License version 3 as
+   published by the Free Software Foundation.
+
+   It is distributed in the hope that it will be useful, but WITHOUT
+   ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+   or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
+   License for more details.
+
+   You should have received a copy of the GNU General Public License
+   along with this file.  If not, see <http://www.gnu.org/licenses/>.
+
+]]--
+
 require 'torch'
 require 'nn'
 
@@ -6,50 +25,51 @@ local DAG, parent = torch.class('nn.DAG', 'nn.Container')
 
 function DAG:__init()
    parent.__init(self)
-   self.pred = {}
-   self.succ = {}
+   -- Nodes are indexed by the module they contain
+   self.node = { }
 end
 
-function DAG:addEdge(a, b)
-   self.sorted = nil
-   local pred, succ = self.pred, self.succ
-   if not pred[a] and not succ[a] then
-      self:add(a)
-   end
-   if not pred[b] and not succ[b] then
-      self:add(b)
+function DAG:createNode(nnm)
+   if not self.node[nnm] then
+      self:add(nnm) -- Add it to the object as a Container
+      self.node[nnm] = {}
+      self.node[nnm].succ = {}
+      self.node[nnm].pred = {}
    end
-   pred[b] = pred[b] or {}
-   pred[b][#pred[b] + 1] = a
-   succ[a] = succ[a] or {}
-   succ[a][#succ[a] + 1] = b
 end
 
--- Apply f on t recursively; use the corresponding a1 and a2 elements
--- (i.e. same keys) as second and third parameters to f when
--- available; return the results from f, organized in a similarly
--- nested table.
-function DAG:applyOnModules(f, t, a1, a2)
+function DAG:addEdge(nnma, nnmb)
+   self.sorted = nil
+   self:createNode(nnma)
+   self:createNode(nnmb)
+   table.insert(self.node[nnmb].pred, nnma)
+   table.insert(self.node[nnma].succ, nnmb)
+end
+
+-- Apply f on t recursively; use the corresponding element from args
+-- (i.e. same keys) as second parameter to f when available; return
+-- the results from f, organized in a similarly nested table.
+function DAG:nestedApply(f, t, args)
    if torch.type(t) == 'table' then
       local result = {}
       for k, s in pairs(t) do
-         result[k] = self:applyOnModules(f, s, a1 and a1[k], a2 and a2[k])
+         result[k] = self:nestedApply(f, s, args and args[k])
       end
       return result
    else
-      return f(t, a1, a2)
+      return f(t, args)
    end
 end
 
 function DAG:setInput(i)
    self.sorted = nil
    self.inputModules = i
-   self:applyOnModules(
-      function(m)
-         if not self.succ[m] or #self.succ[m] == 0 then
+   self:nestedApply(
+      function(nnm)
+         if #self.node[nnm].succ == 0 then
             error('Input modules must have outgoing  edges.')
          end
-         if self.pred[m] and #self.pred[m] > 0 then
+         if #self.node[nnm].pred > 0 then
             error('Input modules cannog have incoming edges.')
          end
       end,
@@ -60,12 +80,12 @@ end
 function DAG:setOutput(o)
    self.sorted = nil
    self.outputModules = o
-   self:applyOnModules(
-      function(m)
-         if not self.pred[m] or #self.pred[m] == 0 then
+   self:nestedApply(
+      function(nnm)
+         if #self.node[nnm].pred == 0 then
             error('Output module must have incoming edges.')
          end
-         if self.succ[m] and #self.succ[m] > 0 then
+         if #self.node[nnm].succ > 0 then
             error('Output module cannot have outgoing edges.')
          end
       end,
@@ -73,23 +93,25 @@ function DAG:setOutput(o)
    )
 end
 
-function DAG:sort()
+function DAG:putInOrder()
    if self.sorted then
       return
    end
 
+   -- First, we sort the nodes according to the DAG order
+
    local distance = {}
 
-   self:applyOnModules(function(m) distance[m] = 1 end, self.inputModules)
+   self:nestedApply(function(m) distance[m] = 1 end, self.inputModules)
 
    local nc
 
    repeat
       nc = 0
-      for i, isucc in pairs(self.succ) do
-         for _, j in pairs(isucc) do
-            if distance[i] and (not distance[j] or distance[j] < distance[i] + 1) then
-               distance[j] = distance[i] + 1
+      for nnma, node in pairs(self.node) do
+         for _, nnmb in pairs(node.succ) do
+            if distance[nnma] and (not distance[nnmb] or distance[nnmb] < distance[nnma] + 1) then
+               distance[nnmb] = distance[nnma] + 1
                nc = nc + 1
             end
          end
@@ -97,16 +119,17 @@ function DAG:sort()
    until nc == 0
 
    self.sorted = { }
-   for i, d in pairs(distance) do
-      table.insert(self.sorted, { d, i })
+   for m, d in pairs(distance) do
+      table.insert(self.sorted, { distance = d, nnm = m })
    end
 
-   table.sort(self.sorted, function(a, b) return a[1] < b[1] end)
-   for i, a in ipairs(self.sorted) do self.sorted[i] = a[2] end
+   table.sort(self.sorted, function(a, b) return a.distance < b.distance end)
+
+   for i, a in ipairs(self.sorted) do self.sorted[i] = a.nnm end
 end
 
 function DAG:print()
-   self:sort()
+   self:putInOrder()
 
    for i, d in ipairs(self.sorted) do
       print('#' .. i .. ' -> ' .. torch.type(d))
@@ -114,59 +137,120 @@ function DAG:print()
 end
 
 function DAG:updateOutput(input)
-   self:sort()
-
-   self:applyOnModules(function(m, i) m:updateOutput(i) end, self.inputModules, input)
-
-   for _, d in ipairs(self.sorted) do
-      if self.pred[d] then
-         if #self.pred[d] == 1 then
-            d:updateOutput(self.pred[d][1].output)
-         elseif #self.pred[d] > 1 then
-            local c = {}
-            for k = 1, #self.pred[d] do
-               c[k] = self.pred[d][k].output
+   self:putInOrder()
+
+   self:nestedApply(
+      function(nnm, i)
+         self.node[nnm].input = i
+         nnm:updateOutput(i)
+      end,
+      self.inputModules,
+      input
+   )
+
+   for _, nnm in ipairs(self.sorted) do
+      local node = self.node[nnm]
+      if #node.pred > 0 then
+         local i
+         if #node.pred == 1 then
+            i = node.pred[1].output
+         elseif #node.pred > 1 then
+            i = {}
+            for k = 1, #node.pred do
+               i[k] = node.pred[k].output
             end
-            d:updateOutput(c)
          end
+         node.input = i
+         nnm:updateOutput(i)
       end
    end
 
-   self.output = self:applyOnModules(function(m) return m.output end, self.outputModules)
+   self.output = self:nestedApply(
+      function(m) return m.output end,
+      self.outputModules
+   )
 
    return self.output
 end
 
+function DAG:computeGradInput(gradInputSucc)
+   local gi
+   if #gradInputSucc == 1 then
+      gi = gradInputSucc[1] -- we avoid a clone()
+   elseif #gradInputSucc > 1 then
+      for k = 1, #gradInputSucc do
+         if gi then
+            gi:add(gradInputSucc[k])
+         else
+            gi = gradInputSucc[k]:clone()
+         end
+      end
+   end
+   return gi
+end
+
 function DAG:updateGradInput(input, gradOutput)
-   self:sort()
+   self:putInOrder()
 
-   self:applyOnModules(
-      function(m, i, go) m:updateGradInput(i, go) end,
-      self.outputModules, input, gradOutput
+   self:nestedApply(
+      function(nnm, go) nnm:updateGradInput(self.node[nnm].input, go) end,
+      self.outputModules, gradOutput
    )
 
-   for k = self.sorted, 1, -1 do
-      local m = sorted[k]
-      if self.succ[d] then
-         if #self.succ[d] == 1 then
-            d:updateGradInput(self.succ[d][1].gradInput)
-         elseif #self.succ[d] > 1 then
-            local sum
-            for k = 1, #self.succ[d] do
-               if sum then
-                  sum:add(self.succ[d][k].gradInput)
-               else
-                  sum = self.succ[d][k].gradInput:clone()
-               end
-            end
-            d:updateGradInput(sum)
+   self:nestedApply(
+      function(nnm, i) self.node[nnm].input = i end,
+      self.inputModules, input
+   )
+
+   for _, node in pairs(self.node) do
+      node.gradInputSucc = {}
+   end
+
+   for k = #self.sorted, 1, -1 do
+      local nnm = self.sorted[k]
+      local node = self.node[nnm]
+      local pred, gradInputSucc = node.pred, node.gradInputSucc
+
+      if #gradInputSucc > 0 then
+         nnm:updateGradInput(node.input, self:computeGradInput(gradInputSucc))
+      end
+
+      -- We fill the gradInputSucc of our predecessors
+      if #pred == 1 then
+         table.insert(self.node[pred[1]].gradInputSucc, nnm.gradInput)
+      elseif #pred > 1 then
+         if not torch.type(nnm.gradInput) == 'table' then
+            error('Should have a table gradInput since it has multiple predecessors')
+         end
+         for n = 1, #pred do
+            table.insert(self.node[node.pred[n]].gradInputSucc, nnm.gradInput[n])
          end
       end
    end
 
-   self.gradInput = self:applyOnModules(function(m) return m.gradInput end, self.inputModules)
+   self.gradInput = self:nestedApply(function(m) return m.gradInput end, self.inputModules)
 
    return self.gradInput
 end
 
-return DAG
+function DAG:accGradParameters(input, gradOutput, scale)
+   scale = scale or 1
+
+   self:putInOrder()
+
+   self:nestedApply(
+      function(nnm, go) nnm:updateGradInput(self.node[nnm].input, go) end,
+      self.outputModules, gradOutput
+   )
+
+   self:nestedApply(
+      function(nnm, i) self.node[nnm].input = i end,
+      self.inputModules, input
+   )
+
+   for k = #self.sorted, 1, -1 do
+      local nnm = self.sorted[k]
+      local node = self.node[nnm]
+      nnm:accGradParameters(node.input, self:computeGradInput(node.gradInputSucc), scale)
+   end
+end