-#!/usr/bin/env luajit
+
+--[[
+
+ 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'
-require 'image'
-require 'optim'
-
-----------------------------------------------------------------------
-local Graph, parent = torch.class('nn.Graph', 'nn.Container')
+local DAG, parent = torch.class('nn.DAG', '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)
- 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
+-- Apply f on t recursively; use the corresponding elements 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)
- end
- end
- else
- self:setOutput({ o })
+function DAG:createNode(nnm)
+ if not self.node[nnm] then
+ self:add(nnm) -- Add it to the object as a Container
+ local node = {}
+ node.succ = {}
+ node.pred = {}
+ node.index = #self.modules
+ self.node[nnm] = node
end
end
-function Graph:order()
- local distance = {}
-
- for _, a in pairs(self.inputModules) do
- distance[a] = 1
+function DAG:putInOrder()
+ if self.sorted then
+ return
end
- local nc
+ local distance = {}
+ self:nestedApply(
+ function(m) distance[m] = 1 end,
+ self.inputModules
+ )
+ local nc
+ local nl = 0
repeat
+ assert(nl < #self.modules, 'Cycle detected in the graph.')
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
end
+ nl = nl + 1
until nc == 0
- self.sorted = { }
- for i, d in pairs(distance) do
- table.insert(self.sorted, { d, i })
+ for _, nnm in pairs(self.modules) do
+ assert(distance[nnm], 'Some modules are not connected to inputs.')
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
+ self.sorted = {}
+ 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.distance < b.distance end)
+
+ for i, a in ipairs(self.sorted) do self.sorted[i] = a.nnm end
+end
+
+-- This accumulates x in a, where they are both nested tables of
+-- tensors with same structures / keys. If first is true, set a = x
+-- (in which case a can be nil) otherwise a = a + x. The behavior is
+-- undefined if a and x do not have the exact same structure.
+function DAG:nestedAccTensor(a, x, first)
+ if torch.type(x) == 'table' then
+ local b = {}
+ for i in pairs(x) do
+ b[i] = self:nestedAccTensor(a[i], x[i], first)
+ end
+ a = b
+ else
+ if first then
+ if a then
+ a:resizeAs(x):copy(x)
+ else
+ a = x:clone()
+ end
+ else
+ a:add(x)
+ end
+ end
+ return a
+end
+
+function DAG:updateGradOutput(node)
+ local gradInputSucc = node.gradInputSucc
+ if #gradInputSucc == 1 then
+ node.gradOutput = gradInputSucc[1]
+ elseif #gradInputSucc > 1 then
+ for k = 1, #gradInputSucc do
+ node.gradOutput = self:nestedAccTensor(node.gradOutput, gradInputSucc[k], k == 1)
+ end
+ end
+end
+
+----------------------------------------------------------------------
+
+-- Connect a sequence of modules
+function DAG:connect(...)
+ self.sorted = nil
+ local prev
+ for _, nnm in pairs({...}) do
+ self:createNode(nnm)
+ if prev then
+ table.insert(self.node[nnm].pred, prev)
+ table.insert(self.node[prev].succ, nnm)
+ end
+ prev = nnm
+ end
+end
+
+function DAG:setLabel(nnm, label)
+ self.node[nnm].label = label
end
-function Graph:print()
+function DAG:setInput(i)
+ self.sorted = nil
+ self.inputModules = i
+ self:nestedApply(
+ function(nnm)
+ assert(#self.node[nnm].succ > 0, 'Input modules must have outgoing edges.')
+ assert(#self.node[nnm].pred == 0, 'Input modules cannot have incoming edges.')
+ end,
+ self.inputModules
+ )
+end
+
+function DAG:setOutput(o)
+ self.sorted = nil
+ self.outputModules = o
+ self:nestedApply(
+ function(nnm)
+ assert(#self.node[nnm].pred > 0, 'Output module must have incoming edges.')
+ assert(#self.node[nnm].succ == 0, 'Output module cannot have outgoing edges.')
+ end,
+ self.outputModules
+ )
+end
+
+function DAG:print()
+ self:putInOrder()
+
for i, d in ipairs(self.sorted) do
- print('#' .. i .. ' -> ' .. torch.type(d))
+ local decoration = ''
+ if self.node[d].label then
+ decoration = ' [' .. self.node[d].label .. ']'
+ end
+ print('#' .. i .. ' -> ' .. torch.type(d) .. decoration)
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])
+----------------------------------------------------------------------
+
+function DAG:saveDot(filename)
+ local file = (filename and io.open(filename, 'w')) or io.stdout
+
+ local function writeNestedCluster(prefix, list, indent)
+ local indent = indent or ''
+ if torch.type(list) == 'table' then
+ file:write(indent .. ' subgraph cluster_' .. prefix .. ' {\n');
+ for k, x in pairs(list) do
+ writeNestedCluster(prefix .. '_' .. k, x, ' ' .. indent)
+ end
+ file:write(indent .. ' }\n');
+ else
+ file:write(indent .. ' ' .. self.node[list].index .. ' [color=red]\n')
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
- end
- d:updateOutput(c)
+ file:write('digraph {\n')
+
+ file:write('\n')
+
+ writeNestedCluster('input', self.inputModules)
+ writeNestedCluster('output', self.outputModules)
+
+ file:write('\n')
+
+ for nnmb, node in pairs(self.node) do
+ file:write(
+ ' '
+ .. node.index
+ .. ' [shape=box,label=\"' .. (self.node[nnmb].label or torch.type(nnmb)) .. '\"]'
+ .. '\n'
+ )
+
+ for i, nnma in pairs(node.pred) do
+ local decoration = ''
+ if #node.pred > 1 then
+ -- decoration = ' [headlabel=\"' .. i .. '\"]'
+ decoration = ' [label=\"' .. i .. '\"]'
end
+ file:write(
+ ' '
+ .. self.node[nnma].index
+ .. ' -> '
+ .. self.node[nnmb].index
+ .. decoration
+ .. '\n'
+ )
end
+
+ file:write('\n')
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
+ file:write('}\n')
+
+end
+
+----------------------------------------------------------------------
+
+function DAG:updateOutput(input)
+ self:putInOrder()
+
+ self:nestedApply(
+ function(nnm, i)
+ local node = self.node[nnm]
+ node.input = i
+ self:rethrowErrors(nnm, node.index, 'updateOutput', i)
+ end,
+ self.inputModules,
+ input
+ )
+
+ for _, nnm in ipairs(self.sorted) do
+ local node = self.node[nnm]
+ local pred = node.pred
+ if #pred > 0 then
+ local i
+ if #pred == 1 then
+ i = pred[1].output
+ elseif #pred > 1 then
+ i = {}
+ for k = 1, #pred do
+ i[k] = pred[k].output
+ end
+ end
+ node.input = i
+ self:rethrowErrors(nnm, node.index, 'updateOutput', i)
end
end
+ self.output = self:nestedApply(
+ function(m) return m.output end,
+ self.outputModules
+ )
+
return self.output
end
-----------------------------------------------------------------------
+function DAG:updateGradInput(input, gradOutput)
+ assert(self.sorted, 'There has been a structure change before a DAG:updateGradInput.')
-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)
+ self:nestedApply(
+ function(nnm, go)
+ local node = self.node[nnm]
+ node.gradOutput = go
+ self:rethrowErrors(nnm, node.index, 'updateGradInput', node.input, go)
+ end,
+ self.outputModules, gradOutput
+ )
---[[
+ self:nestedApply(
+ function(nnm, i) self.node[nnm].input = i end,
+ self.inputModules, input
+ )
- a -----> b ---> c ----> e ---
- \ /
- \--> d ---/
- \
- \---> f ---
-]]--
+ 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 = node.pred
-g = Graph:new()
+ if #node.gradInputSucc > 0 then
+ self:updateGradOutput(node)
+ self:rethrowErrors(nnm, node.index, 'updateGradInput', node.input, node.gradOutput)
+ 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
+ assert(torch.type(nnm.gradInput) == 'table',
+ 'Should have a table gradInput since it has multiple predecessors.')
+ for n = 1, #pred do
+ table.insert(self.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
-g:print(graph)
+function DAG:accGradParameters(input, gradOutput, scale)
+ assert(self.sorted, 'There has been a structure change before a DAG:accGradParameters.')
-input = torch.Tensor(3, 10):uniform()
+ self:nestedApply(
+ function(nnm, go) self.node[nnm].gradOutput = go end,
+ self.outputModules, gradOutput
+ )
-output = g:updateOutput(input)
+ self:nestedApply(
+ function(nnm, i) self.node[nnm].input = i end,
+ self.inputModules, input
+ )
-print(output[1])
-print(output[2])
+ for k = 1, #self.modules do
+ local nnm = self.modules[k]
+ local node = self.node[nnm]
+ self:rethrowErrors(nnm, k, 'accGradParameters', node.input, node.gradOutput, scale)
+ end
+end
+
+function DAG:clearState()
+ self.sorted = nil
+ for _, node in pairs(self.node) do
+ node.input = nil
+ node.gradInputSucc = nil
+ node.gradOutput = nil
+ end
+ return parent.clearState(self)
+end