X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=dagnn.lua;h=0f93d95f63a87b320d9e6d261a89c9798b4d9b55;hb=d0743d66135ed7cedcb3777cfa5dda883cbeadb3;hp=1ec9b4ea86f40600f5ff3ac70ed7bfa926841dff;hpb=c8a895aa5f221f0de11733e8d05373e89ae9476e;p=dagnn.git

diff --git a/dagnn.lua b/dagnn.lua
index 1ec9b4e..0f93d95 100755
--- a/dagnn.lua
+++ b/dagnn.lua
@@ -1,75 +1,98 @@
-#!/usr/bin/env luajit
 
 require 'torch'
 require 'nn'
-require 'image'
-require 'optim'
 
-----------------------------------------------------------------------
+local DAG, parent = torch.class('nn.DAG', 'nn.Container')
 
-local Graph, parent = torch.class('nn.Graph', 'nn.Container')
-
-function Graph:__init()
+function DAG:__init()
    parent.__init(self)
-   self.pred = {}
-   self.succ = {}
+   -- Nodes are indexed by the module they contain
+   self.node = { }
 end
 
-function Graph:addEdge(a, b)
-   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
 
-function Graph:setInput(i)
-   if torch.type(i) == 'table' then
-      self.inputModules = i
-      for _, m in ipairs(i) do
-         if not self.pred[m] and not self.succ[m] then
-            self:add(m)
-         end
+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:nestedApply(f, s, args and args[k])
       end
+      return result
    else
-      self:setInput({ i })
+      return f(t, args)
    end
 end
 
-function Graph:setOutput(o)
-   if torch.type(o) == 'table' then
-      self.outputModules = o
-      for _, m in ipairs(o) do
-         if not self.pred[m] and not self.succ[m] then
-            self:add(m)
+function DAG:setInput(i)
+   self.sorted = nil
+   self.inputModules = i
+   self:nestedApply(
+      function(nnm)
+         if #self.node[nnm].succ == 0 then
+            error('Input modules must have outgoing  edges.')
          end
-      end
-   else
-      self:setOutput({ o })
-   end
+         if #self.node[nnm].pred > 0 then
+            error('Input modules cannog have incoming edges.')
+         end
+      end,
+      self.inputModules
+   )
 end
 
-function Graph:order()
-   local distance = {}
+function DAG:setOutput(o)
+   self.sorted = nil
+   self.outputModules = o
+   self:nestedApply(
+      function(nnm)
+         if #self.node[nnm].pred == 0 then
+            error('Output module must have incoming edges.')
+         end
+         if #self.node[nnm].succ > 0 then
+            error('Output module cannot have outgoing edges.')
+         end
+      end,
+      self.outputModules
+   )
+end
 
-   for _, a in pairs(self.inputModules) do
-      distance[a] = 1
+function DAG:putInOrder()
+   if self.sorted then
+      return
    end
 
+   -- First, we sort the nodes according to the DAG order
+
+   local distance = {}
+
+   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
@@ -77,91 +100,140 @@ function Graph:order()
    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 Graph:print()
+function DAG:print()
+   self:putInOrder()
+
    for i, d in ipairs(self.sorted) do
       print('#' .. i .. ' -> ' .. torch.type(d))
    end
 end
 
-function Graph:updateOutput(input)
-   if #self.inputModules == 1 then
-      self.inputModules[1]:updateOutput(input)
-   else
-      for i, d in ipairs(self.inputModules) do
-         d:updateOutput(input[i])
-      end
-   end
-
-   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
+function DAG:updateOutput(input)
+   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
 
-   if #self.outputModules == 1 then
-      self.output = self.outputModules[1].output
-   else
-      self.output = { }
-      for i, d in ipairs(self.outputModules) do
-         self.output[i] = d.output
-      end
-   end
+   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: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
+   )
 
-a = nn.Linear(10, 10)
-b = nn.ReLU()
-c = nn.Linear(10, 3)
-d = nn.Linear(10, 3)
-e = nn.CMulTable()
-f = nn.Linear(3, 2)
+   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
 
-   a -----> b ---> c ----> e ---
-             \           /
-              \--> d ---/
-                    \
-                     \---> f ---
-]]--
+      if #gradInputSucc > 0 then
+         nnm:updateGradInput(node.input, self:computeGradInput(gradInputSucc))
+      end
 
-g = Graph:new()
+      -- 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
 
-g:setInput(a)
-g:setOutput({ e, f })
-g:addEdge(c, e)
-g:addEdge(a, b)
-g:addEdge(d, e)
-g:addEdge(b, c)
-g:addEdge(b, d)
-g:addEdge(d, f)
+   self.gradInput = self:nestedApply(function(m) return m.gradInput end, self.inputModules)
 
-g:order()
+   return self.gradInput
+end
+
+function DAG:accGradParameters(input, gradOutput, scale)
+   scale = scale or 1
+
+   self:putInOrder()
 
-g:print(graph)
+   self:nestedApply(
+      function(nnm, go) nnm:updateGradInput(self.node[nnm].input, go) end,
+      self.outputModules, gradOutput
+   )
 
-input = torch.Tensor(3, 10):uniform()
+   self:nestedApply(
+      function(nnm, i) self.node[nnm].input = i end,
+      self.inputModules, input
+   )
 
-output = g:updateOutput(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
 
-print(output[1])
-print(output[2])
+return DAG