diff options
| author | Clement Farabet <clement.farabet@gmail.com> | 2011-11-11 19:38:11 +0400 |
|---|---|---|
| committer | Clement Farabet <clement.farabet@gmail.com> | 2011-11-11 19:38:11 +0400 |
| commit | 98fb40a126afb2f33800af2f081d3119a8b55448 (patch) | |
| tree | b811fd62ac82766bc9d1a49f4d9e99cdc52ec8fa | |
| parent | a8ab27537e16f4440565fbebe6313b2172b489a1 (diff) | |
Upgraded all code to new torch master.torch-master
57 files changed, 210 insertions, 1142 deletions
diff --git a/Abs.lua b/Abs.lua deleted file mode 100644 index ca17acf..0000000 --- a/Abs.lua +++ /dev/null @@ -1,30 +0,0 @@ -local Abs, parent = torch.class('nn.Abs', 'nn.Module') - -function Abs:__init(args) - parent.__init(self) - if args then - error(xlua.usage('nn.Abs', - 'a simple component-wise mapping: abs()', - 'abs = nn.Abs()\n'.. - 'rectified = abs:forward(sometensor)', - {type='nil', help='no arg required'})) - end -end - -function Abs:forward(input) - input.nn.Abs_forward(self, input) - return self.output -end - -function Abs:backward(input, gradOutput) - input.nn.Abs_backward(self, input, gradOutput) - return self.gradInput -end - -function Abs:write(file) - parent.write(self, file) -end - -function Abs:read(file) - parent.read(self, file) -end diff --git a/BatchOptimization.lua b/BatchOptimization.lua index b0c2a12..032b61e 100644 --- a/BatchOptimization.lua +++ b/BatchOptimization.lua @@ -98,7 +98,6 @@ function Batch:forward_sequential(inputs, targets, options) -- estimate df/dW local df_do = self.criterion:backward(output, targets[i]) self.module:backward(inputs[i], df_do) - self.module:accGradParameters(inputs[i], df_do) -- user hook if self.posthook then self.posthook(self, {inputs[i], targets[i], options[i]}) @@ -118,7 +117,6 @@ function Batch:forward_sequential(inputs, targets, options) -- estimate df/dW local df_do = self.criterion:backward(output, targets) self.module:backward(inputs, df_do) - self.module:accGradParameters(inputs, df_do) -- update evaluation counter self.evalCounter = self.evalCounter + inputs:size(1) end @@ -369,7 +367,6 @@ function Batch:setup_mapreduce () -- estimate df/dW local df_do = criterion:backward(output, targets[i]) module:backward(inputs[i], df_do) - module:accGradParameters(inputs[i], df_do) -- user hook if posthook then posthook(optimizer, {inputs[i], targets[i], options[i]}) diff --git a/CAddTable.lua b/CAddTable.lua deleted file mode 100644 index 9603b4b..0000000 --- a/CAddTable.lua +++ /dev/null @@ -1,33 +0,0 @@ - -local CAddTable, parent = torch.class('nn.CAddTable', 'nn.Module') - -function CAddTable:__init() - parent.__init(self) - self.gradInput = {} -end - -function CAddTable:forward(input) - self.output:resizeAs(input[1]):copy(input[1]) - for i=2,#input do - self.output:add(input[i]) - end - return self.output -end - -function CAddTable:backward(input, gradOutput) - for i=1,#input do - self.gradInput[i] = self.gradInput[i] or torch.Tensor() - self.gradInput[i]:resizeAs(input[i]) - self.gradInput[i]:copy(gradOutput) - end - return self.gradInput -end - -function CAddTable:write(file) - parent.write(self, file) -end - -function CAddTable:read(file) - parent.read(self, file) - self.gradInput = {} -end diff --git a/CDivTable.lua b/CDivTable.lua deleted file mode 100644 index bae13c6..0000000 --- a/CDivTable.lua +++ /dev/null @@ -1,30 +0,0 @@ - -local CDivTable, parent = torch.class('nn.CDivTable', 'nn.Module') - -function CDivTable:__init() - parent.__init(self) - self.gradInput = {} -end - -function CDivTable:forward(input) - self.output:resizeAs(input[1]):copy(input[1]) - self.output:cdiv(input[2]) - return self.output -end - -function CDivTable:backward(input, gradOutput) - self.gradInput[1] = self.gradInput[1] or torch.Tensor() - self.gradInput[2] = self.gradInput[2] or torch.Tensor() - self.gradInput[1]:resizeAs(input[1]):copy(gradOutput):cdiv(input[2]) - self.gradInput[2]:resizeAs(input[2]):zero():addcdiv(-1,self.gradInput[1],input[2]):cmul(input[1]) - return self.gradInput -end - -function CDivTable:write(file) - parent.write(self, file) -end - -function CDivTable:read(file) - parent.read(self, file) - self.gradInput = {} -end diff --git a/CMulTable.lua b/CMulTable.lua deleted file mode 100644 index 500eea2..0000000 --- a/CMulTable.lua +++ /dev/null @@ -1,35 +0,0 @@ - -local CMulTable, parent = torch.class('nn.CMulTable', 'nn.Module') - -function CMulTable:__init() - parent.__init(self) - self.gradInput = {} -end - -function CMulTable:forward(input) - self.output:resizeAs(input[1]):copy(input[1]) - for i=2,#input do - self.output:cmul(input[i]) - end - return self.output -end - -function CMulTable:backward(input, gradOutput) - local tout = torch.Tensor():resizeAs(self.output) - for i=1,#input do - self.gradInput[i] = self.gradInput[i] or torch.Tensor() - self.gradInput[i]:resizeAs(input[i]):copy(gradOutput) - tout:copy(self.output):cdiv(input[i]) - self.gradInput[i]:cmul(tout) - end - return self.gradInput -end - -function CMulTable:write(file) - parent.write(self, file) -end - -function CMulTable:read(file) - parent.read(self, file) - self.gradInput = {} -end diff --git a/CSubTable.lua b/CSubTable.lua deleted file mode 100644 index ecdd378..0000000 --- a/CSubTable.lua +++ /dev/null @@ -1,30 +0,0 @@ - -local CSubTable, parent = torch.class('nn.CSubTable', 'nn.Module') - -function CSubTable:__init() - parent.__init(self) - self.gradInput = {} -end - -function CSubTable:forward(input) - self.output:resizeAs(input[1]):copy(input[1]) - self.output:add(-1,input[2]) - return self.output -end - -function CSubTable:backward(input, gradOutput) - self.gradInput[1] = self.gradInput[1] or torch.Tensor() - self.gradInput[2] = self.gradInput[2] or torch.Tensor() - self.gradInput[1]:resizeAs(input[1]):copy(gradOutput) - self.gradInput[2]:resizeAs(input[1]):copy(gradOutput):mul(-1) - return self.gradInput -end - -function CSubTable:write(file) - parent.write(self, file) -end - -function CSubTable:read(file) - parent.read(self, file) - self.gradInput = {} -end diff --git a/ConfusionMatrix.lua b/ConfusionMatrix.lua index b6c841d..f887dc5 100644 --- a/ConfusionMatrix.lua +++ b/ConfusionMatrix.lua @@ -92,21 +92,3 @@ function ConfusionMatrix:__tostring__() str = str .. ' + global correct: ' .. (self.totalValid*100) .. '%' return str end - -function ConfusionMatrix:write(file) - file:writeObject(self.mat) - file:writeObject(self.valids) - file:writeInt(self.nclasses) - file:writeInt(self.totalValid) - file:writeInt(self.averageValid) - file:writeObject(self.classes) -end - -function ConfusionMatrix:read(file) - self.mat = file:readObject() - self.valids = file:readObject() - self.nclasses = file:readInt() - self.totalValid = file:readInt() - self.averageValid = file:readInt() - self.classes = file:readObject() -end diff --git a/DataSet.lua b/DataSet.lua index 4efb85d..13a29f7 100644 --- a/DataSet.lua +++ b/DataSet.lua @@ -297,28 +297,6 @@ function lDataSet:useCacheFile(fileName) self.cacheFileName = fileName end - -function lDataSet:save(fileName) - local fileName = fileName or self.fileName - self.fileName = fileName - print('<DataSet> Saving DataSet to:',fileName) - local file = torch.DiskFile(fileName, 'w') - file:binary() - self:write(file) - file:close() -end - -function lDataSet:open(fileName) - local fileName = fileName or self.fileName - self.fileName = fileName - print('<DataSet> Loading DataSet from File:',fileName) - local file = torch.DiskFile(fileName, 'r') - file:binary() - self:read(file) - file:close() - print('<DataSet> '..self.nbSamples..' samples loaded') -end - function lDataSet:write(file) file:writeBool(self.resized) file:writeInt(self.nbSamples) diff --git a/DiagHessian.lua b/DiagHessian.lua index 26aed3e..2b9aca9 100644 --- a/DiagHessian.lua +++ b/DiagHessian.lua @@ -1,6 +1,6 @@ -- Module -function nn.Module.backwardDiagHessian(self, input, diagHessianOutput) +function nn.Module.updateDiagHessianInput(self, input, diagHessianOutput) self.diagHessianInput = self.diagHessianInput or diagHessianOutput return self.diagHessianInput end @@ -12,20 +12,20 @@ function nn.Module.initDiagHessianParameters(self) end -- Criterion -function nn.Criterion.backwardDiagHessian(self, input, diagHessianOutput) +function nn.Criterion.updateDiagHessianInput(self, input, diagHessianOutput) self.diagHessianInput = self.diagHessianInput or self.output.new() return self.diagHessianInput end -- MSECriterion -function nn.MSECriterion.backwardDiagHessian(self, input, diagHessianOutput) +function nn.MSECriterion.updateDiagHessianInput(self, input, diagHessianOutput) self.diagHessianInput = self.diagHessianInput or input.new() self.diagHessianInput:resizeAs(input):fill(1) return self.diagHessianInput end -- Linear -function nn.Linear.backwardDiagHessian(self, input, diagHessianOutput) +function nn.Linear.updateDiagHessianInput(self, input, diagHessianOutput) self.diagHessianInput = self.diagHessianInput or self.output.new() self.weightSq = self.weightSq or self.output.new():resizeAs(self.weight) self.weightSq:copy(self.weight):cmul(self.weightSq) @@ -60,7 +60,7 @@ function nn.Linear.accDiagHessianParameters(self, input, diagHessianOutput, scal end -- Tanh -function nn.Tanh.backwardDiagHessian(self, input, diagHessianOutput) +function nn.Tanh.updateDiagHessianInput(self, input, diagHessianOutput) self.diagHessianInput = self.diagHessianInput or self.output.new() self.derivativeSq = self.derivativeSq or self.output.new() self.derivativeSq:resizeAs(self.output):copy(self.output):cmul(self.output):mul(-1):add(1) @@ -70,15 +70,15 @@ function nn.Tanh.backwardDiagHessian(self, input, diagHessianOutput) end -- Sequential -function nn.Sequential.backwardDiagHessian(self, input, diagHessianOutput) +function nn.Sequential.updateDiagHessianInput(self, input, diagHessianOutput) local currentDiagHessianOutput = diagHessianOutput local currentModule = self.modules[#self.modules] for i=#self.modules-1,1,-1 do local previousModule = self.modules[i] - currentDiagHessianOutput = currentModule:backwardDiagHessian(previousModule.output, currentDiagHessianOutput) + currentDiagHessianOutput = currentModule:updateDiagHessianInput(previousModule.output, currentDiagHessianOutput) currentModule = previousModule end - currentDiagHessianOutput = currentModule:backwardDiagHessian(input, currentDiagHessianOutput) + currentDiagHessianOutput = currentModule:updateDiagHessianInput(input, currentDiagHessianOutput) self.diagHessianInput = currentDiagHessianOutput return currentDiagHessianOutput end @@ -101,29 +101,3 @@ function nn.Sequential.accDiagHessianParameters(self, input, diagHessianOutput, end currentModule:accDiagHessianParameters(input, currentDiagHessianOutput, scale) end - --- ConcatTable -function nn.ConcatTable.backwardDiagHessian(self, input, diagHessianOutput) - for i,module in ipairs(self.modules) do - local currentDiagHessianInput = module:backward(input, diagHessianOutput[i]) - if i == 1 then - self.diagHessianInput:resizeAs(currentDiagHessianInput):copy(currentDiagHessianInput) - else - self.diagHessianInput:add(currentDiagHessianInput) - end - end - return self.diagHessianInput -end - -function nn.ConcatTable.initDiagHessianParameters(self) - for i=1,#self.modules do - self.modules[i]:initDiagHessianParameters() - end -end - -function nn.ConcatTable.accDiagHessianParameters(self, input, diagHessianOutput, scale) - scale = scale or 1 - for i,module in ipairs(self.modules) do - module:accDiagHessianParameters(input, diagHessianOutput[i], scale) - end -end diff --git a/DistMarginCriterion.lua b/DistMarginCriterion.lua index 6b38771..acd943c 100644 --- a/DistMarginCriterion.lua +++ b/DistMarginCriterion.lua @@ -5,10 +5,10 @@ function DistMarginCriterion:__init() self.sizeAverage = true end -function DistMarginCriterion:forward(input, target) - return input.nn.DistMarginCriterion_forward(self, input, target) +function DistMarginCriterion:updateOutput(input, target) + return input.nn.DistMarginCriterion_updateOutput(self, input, target) end -function DistMarginCriterion:backward(input, target) - return input.nn.DistMarginCriterion_backward(self, input, target) +function DistMarginCriterion:updateGradInput(input, target) + return input.nn.DistMarginCriterion_updateGradInput(self, input, target) end diff --git a/DistNLLCriterion.lua b/DistNLLCriterion.lua index 01290c4..22204fc 100644 --- a/DistNLLCriterion.lua +++ b/DistNLLCriterion.lua @@ -17,7 +17,7 @@ end function DistNLLCriterion:normalize(input, target) -- normalize target if not self.targetIsProbability then - self.probTarget = self.targetSoftMax:forward(target) + self.probTarget = self.targetSoftMax:updateOutput(target) else self.probTarget = target end @@ -31,7 +31,7 @@ function DistNLLCriterion:normalize(input, target) -- normalize input if not self.inputIsLogProbability and not self.inputIsProbability then - self.logProbInput = self.inputLogSoftMax:forward(self.input) + self.logProbInput = self.inputLogSoftMax:updateOutput(self.input) elseif not self.inputIsLogProbability then print('TODO: implement nn.Log()') else @@ -42,7 +42,7 @@ end function DistNLLCriterion:denormalize() -- denormalize gradients if not self.inputIsLogProbability and not self.inputIsProbability then - self.gradInput = self.inputLogSoftMax:backward(self.input, self.gradLogInput) + self.gradInput = self.inputLogSoftMax:updateGradInput(self.input, self.gradLogInput) elseif not self.inputIsLogProbability then print('TODO: implement nn.Log()') else @@ -55,7 +55,7 @@ function DistNLLCriterion:denormalize() end end -function DistNLLCriterion:forward(input, target) +function DistNLLCriterion:updateOutput(input, target) self:normalize(input, target) self.output = 0 for i = 1,input:size(1) do @@ -64,7 +64,7 @@ function DistNLLCriterion:forward(input, target) return self.output end -function DistNLLCriterion:backward(input, target) +function DistNLLCriterion:updateGradInput(input, target) self:normalize(input, target) self.gradLogInput:resizeAs(input) for i = 1,input:size(1) do @@ -73,23 +73,3 @@ function DistNLLCriterion:backward(input, target) self:denormalize() return self.gradInput end - -function DistNLLCriterion:write(file) - parent.write(self, file) - file:writeBool(self.inputIsProbability) - file:writeBool(self.inputIsLogProbability) - file:writeBool(self.targetIsProbability) - file:writeObject(self.targetSoftMax) - file:writeObject(self.inputLogSoftMax) - file:writeObject(self.gradLogInput) -end - -function DistNLLCriterion:read(file) - parent.read(self, file) - self.inputIsProbability = file:readBool() - self.inputIsLogProbability = file:readBool() - self.targetIsProbability = file:readBool() - self.targetSoftMax = file:readObject() - self.inputLogSoftMax = file:readObject() - self.gradLogInput = file:readObject() -end diff --git a/GeneticSGDOptimization.lua b/GeneticSGDOptimization.lua index d7816f8..0ddced7 100644 --- a/GeneticSGDOptimization.lua +++ b/GeneticSGDOptimization.lua @@ -267,7 +267,6 @@ function GenSGD:setup_mapreduce () -- estimate df/dW local df_do = criterion:backward(output, targets[i]) module:backward(inputs[i], df_do) - module:accGradParameters(inputs[i], df_do) optimizer(module,opt_param) if opt_param.adaptive_batchSize and not opt_param.exact_batchSize then @@ -290,7 +289,6 @@ function GenSGD:setup_mapreduce () if opt_param.adaptive_batchSize and opt_param.exact_batchsize then local df_do = criterion:backward(output, targets[i]) module:backward(inputs[i], df_do) - module:accGradParameters(inputs[i], df_do) partialGrads[i]:copy(gradParameters) end end diff --git a/HardShrink.lua b/HardShrink.lua deleted file mode 100644 index b33fd6c..0000000 --- a/HardShrink.lua +++ /dev/null @@ -1,11 +0,0 @@ -local HardShrink = torch.class('nn.HardShrink', 'nn.Module') - -function HardShrink:forward(input) - input.nn.HardShrink_forward(self, input) - return self.output -end - -function HardShrink:backward(input, gradOutput) - input.nn.HardShrink_backward(self, input, gradOutput) - return self.gradInput -end diff --git a/KLDivCriterion.lua b/KLDivCriterion.lua index b4ce815..5dcf3b8 100644 --- a/KLDivCriterion.lua +++ b/KLDivCriterion.lua @@ -14,14 +14,14 @@ end function KLDivCriterion:normalize(input, target) -- normalize target if not self.targetIsProbability then - self.probTarget = self.targetSoftMax:forward(target) + self.probTarget = self.targetSoftMax:updateOutput(target) else self.probTarget = target end -- normalize input if not self.inputIsProbability then - self.probInput = self.inputSoftMax:forward(input) + self.probInput = self.inputSoftMax:updateOutput(input) else self.probInput = input end @@ -30,13 +30,13 @@ end function KLDivCriterion:denormalize(input) -- denormalize gradients if not self.inputIsProbability then - self.gradInput = self.inputSoftMax:backward(input, self.gradProbInput) + self.gradInput = self.inputSoftMax:updateGradInput(input, self.gradProbInput) else self.gradInput = self.gradProbInput end end -function KLDivCriterion:forward(input, target) +function KLDivCriterion:updateOutput(input, target) self:normalize(input, target) self.output = 0 for i = 1,input:size(1) do @@ -49,7 +49,7 @@ function KLDivCriterion:forward(input, target) return self.output end -function KLDivCriterion:backward(input, target) +function KLDivCriterion:updateGradInput(input, target) self:normalize(input, target) self.gradProbInput:resizeAs(input) for i = 1,input:size(1) do @@ -1,11 +1,11 @@ local Minus, parent = torch.class('nn.Minus', 'nn.Module') -function Minus:forward(input) +function Minus:updateOutput(input) self.output:resizeAs(input):copy(input):mul(-1) return self.output end -function Minus:backward(input, gradOutput) +function Minus:updateGradInput(input, gradOutput) self.gradInput:resizeAs(input):copy(gradOutput):mul(-1) return self.gradInput end diff --git a/Narrow.lua b/Narrow.lua deleted file mode 100644 index 5c6d07e..0000000 --- a/Narrow.lua +++ /dev/null @@ -1,70 +0,0 @@ -local Narrow, parent = torch.class('nn.Narrow', 'nn.Module') - -local help_desc = -[[Selects a subset of a dimension of a nxpxqx.. Tensor.]] - -local help_example = -[[mlp=nn.Sequential(); -mlp:add(nn.Narrow(1,3,2)) - -require "lab" -x=lab.randn(10,5) -print(x) -print(mlp:forward(x)) - --- gives the output: - 0.9720 -0.0836 0.0831 -0.2059 -0.0871 - 0.8750 -2.0432 -0.1295 -2.3932 0.8168 - 0.0369 1.1633 0.6483 1.2862 0.6596 - 0.1667 -0.5704 -0.7303 0.3697 -2.2941 - 0.4794 2.0636 0.3502 0.3560 -0.5500 --0.1898 -1.1547 0.1145 -1.1399 0.1711 --1.5130 1.4445 0.2356 -0.5393 -0.6222 --0.6587 0.4314 1.1916 -1.4509 1.9400 - 0.2733 1.0911 0.7667 0.4002 0.1646 - 0.5804 -0.5333 1.1621 1.5683 -0.1978 -[torch.Tensor of dimension 10x5] - - 0.0369 1.1633 0.6483 1.2862 0.6596 - 0.1667 -0.5704 -0.7303 0.3697 -2.2941 -[torch.Tensor of dimension 2x5] ]] - -function Narrow:__init(dimension,offset,length) - parent.__init(self) - self.dimension=dimension - self.index=offset - self.length=length or 1 - if not dimension or not offset then - error(xlua.usage('nn.Narrow', help_desc, help_example, - {type='number', help='dimension', req=true}, - {type='number', help='offset', req=true}, - {type='number', help='length', default=1})) - end -end - -function Narrow:forward(input) - local output=input:narrow(self.dimension,self.index,self.length); - self.output:resizeAs(output) - return self.output:copy(output) -end - -function Narrow:backward(input, gradOutput) - self.gradInput:resizeAs(input) - self.gradInput:zero(); - self.gradInput:narrow(self.dimension,self.index,self.length):copy(gradOutput) - return self.gradInput -end - -function Narrow:write(file) - parent.write(self, file) - file:writeInt(self.dimension) - file:writeLong(self.index) - file:writeLong(self.length) -end - -function Narrow:read(file, version) - parent.read(self, file) - self.dimension = file:readInt() - self.index = file:readLong() - self.length = file:readLong() -end diff --git a/OnlineTrainer.lua b/OnlineTrainer.lua index dc6e860..920a523 100644 --- a/OnlineTrainer.lua +++ b/OnlineTrainer.lua @@ -49,9 +49,7 @@ function OnlineTrainer:log() end end print('<trainer> saving network to '..filename) - local file = torch.DiskFile(filename,'w') - self.module:write(file) - file:close() + torch.save(filename, self.module) end function OnlineTrainer:train(dataset) @@ -173,15 +171,3 @@ function OnlineTrainer:test(dataset) return self.currentError end - -function OnlineTrainer:write(file) - parent.write(self,file) - file:writeObject(self.module) - file:writeObject(self.criterion) -end - -function OnlineTrainer:read(file) - parent.read(self,file) - self.module = file:readObject() - self.criterion = file:readObject() -end diff --git a/Power.lua b/Power.lua deleted file mode 100644 index c0b60a1..0000000 --- a/Power.lua +++ /dev/null @@ -1,37 +0,0 @@ - -local Power, parent = torch.class('nn.Power','nn.Module') - -function Power:__init(p) - parent.__init(self) - self.pow = p - if args then - error(xlua.usage('nn.Power', - 'a simple component-wise mapping: power(p)', - 'pow = nn.Power(p)\n'.. - 'powered = pow:forward(sometensor)', - {type='nil', help='no arg required'})) - end -end - -function Power:forward(input) - self.output:resizeAs(input):copy(input) - self.output:pow(self.pow) - return self.output -end - -function Power:backward(input, gradOutput) - self.gradInput:resizeAs(input):copy(gradOutput) - self.gradInput:cmul(self.output):cdiv(input):mul(self.pow) - return self.gradInput -end - - -function Power:write(file) - parent.write(self,file) - file:writeDouble(self.pow) -end - -function Power:read(file) - parent.read(self,file) - self.pow = file:readDouble() -end @@ -11,11 +11,11 @@ function Probe:__init(...) {arg='backw', type='boolean', help='activates probe for backward()', default=false}) end -function Probe:forward(input) +function Probe:updateOutput(input) self.output = input if self.size or self.content then print('') - print('<probe::' .. self.name .. '> forward()') + print('<probe::' .. self.name .. '> updateOutput()') if self.content then print(input) elseif self.size then print(#input) end @@ -26,12 +26,12 @@ function Probe:forward(input) return self.output end -function Probe:backward(input, gradOutput) +function Probe:updateGradInput(input, gradOutput) self.gradInput = gradOutput if self.backw then if self.size or self.content then print('') - print('<probe::' .. self.name .. '> backward()') + print('<probe::' .. self.name .. '> updateGradInput()') if self.content then print(gradOutput) elseif self.size then print(#gradOutput) end @@ -42,21 +42,3 @@ function Probe:backward(input, gradOutput) end return self.gradInput end - -function Probe:write(file) - parent.write(self, file) - file:writeObject(self.name) - file:writeBool(self.content) - file:writeBool(self.display) - file:writeBool(self.size) - file:writeBool(self.backw) -end - -function Probe:read(file) - parent.read(self, file) - self.name = file:readObject() - self.content = file:readBool() - self.display = file:readBool() - self.size = file:readBool() - self.backw = file:readBool() -end diff --git a/Replicate.lua b/Replicate.lua deleted file mode 100644 index 2f8db97..0000000 --- a/Replicate.lua +++ /dev/null @@ -1,40 +0,0 @@ - -local Replicate, parent = torch.class('nn.Replicate','nn.Module') - -function Replicate:__init(nf) - parent.__init(self) - self.nfeatures = nf -end - -function Replicate:forward(input) - local sz = torch.LongStorage(input:dim()+1) - sz[1] = self.nfeatures - for i = 1,input:dim() do - sz[i+1] = input:size(i) - end - local st = torch.LongStorage(input:dim()+1) - st[1] = 0 - for i = 1,input:dim() do - st[i+1] = input:stride(i) - end - self.output = torch.Tensor(input:storage(),input:storageOffset(),sz,st) - return self.output -end - -function Replicate:backward(input, gradOutput) - self.gradInput:resizeAs(input):zero() - for k = 1,gradOutput:size(1) do - self.gradInput:add(gradOutput[k]) - end - return self.gradInput -end - -function Replicate:write(file) - parent.write(self,file) - file:writeInt(self.nfeatures) -end - -function Replicate:read(file) - parent.read(self,file) - self.nfeatures = file:readInt() -end diff --git a/SGDOptimization.lua b/SGDOptimization.lua index 1309ae9..3e4a1e9 100644 --- a/SGDOptimization.lua +++ b/SGDOptimization.lua @@ -108,16 +108,16 @@ function SGD:diagHessian(inputs, targets) for i = 1,#inputs do local output = self.module:forward(inputs[i]) local critDiagHessian = - self.criterion:backwardDiagHessian(output, targets[i]) - self.module:backwardDiagHessian(inputs[i], critDiagHessian) + self.criterion:updateDiagHessianInput(output, targets[i]) + self.module:updateDiagHessianInput(inputs[i], critDiagHessian) self.module:accDiagHessianParameters(inputs[i], critDiagHessian) end self.diagHessianParameters:div(#inputs) else local output = self.module:forward(inputs) -- not sure if we can do the fast version yet - local critDiagHessian = criterion:backwardDiagHessian(output, targets) - module:backwardDiagHessian(inputs, critDiagHessian) + local critDiagHessian = criterion:updateDiagHessianInput(output, targets) + module:updateDiagHessianInput(inputs, critDiagHessian) module:accDiagHessianParameters(inputs, critDiagHessian) self.diagHessianParameters:div(inputs:size(1)) end @@ -189,7 +189,6 @@ function SGD:optimalLearningRate(inputs, targets) -- estimate df/dW local df_do = self.criterion:backward(output, targets[i]) self.module:backward(inputs[i], df_do) - self.module:accGradParameters(inputs[i], df_do) end -- normalize gradients -- self.gradParameters:div(#inputs) @@ -223,7 +222,6 @@ function SGD:optimalLearningRate(inputs, targets) -- estimate df/dW df_do = self.criterion:backward(output, targets[i]) self.module:backward(inputs[i], df_do) - self.module:accGradParameters(inputs[i], df_do) end -- normalize gradients -- self.gradParameters:div(#inputs) @@ -255,7 +253,6 @@ function SGD:optimalLearningRate(inputs, targets) -- estimate df/dW local df_do = self.criterion:backward(output, targets) self.module:backward(inputs, df_do) - self.module:accGradParameters(inputs, df_do) -- backup gradient and weights self.param_bkup:copy(self.parameters) self.grad_bkup:copy(self.gradParameters) @@ -279,7 +276,6 @@ function SGD:optimalLearningRate(inputs, targets) -- re-estimate df/dW df_do = self.criterion:backward(output, targets) self.module:backward(inputs, df_do) - self.module:accGradParameters(inputs, df_do) -- self.gradParameters:div(inputs:size(1)) -- (3) phi - 1/alpha(dE/dw(w + alpha * oldphi / || oldphi ||) - dE/dw(w)) diff --git a/SparseCriterion.lua b/SparseCriterion.lua index ddaa75c..d086f28 100644 --- a/SparseCriterion.lua +++ b/SparseCriterion.lua @@ -5,22 +5,12 @@ function SparseCriterion:__init() self.sizeAverage = true end -function SparseCriterion:forward(input) - input.nn.SparseCriterion_forward(self, input) +function SparseCriterion:updateOutput(input) + input.nn.SparseCriterion_updateOutput(self, input) return self.output end -function SparseCriterion:backward(input) - input.nn.SparseCriterion_backward(self, input) +function SparseCriterion:updateGradInput(input) + input.nn.SparseCriterion_updateGradInput(self, input) return self.gradInput end - -function SparseCriterion:write(file) - parent.write(self, file) - file:writeBool(self.sizeAverage) -end - -function SparseCriterion:read(file) - parent.read(self, file) - self.sizeAverage = file:readBool() -end diff --git a/SpatialClassNLLCriterion.lua b/SpatialClassNLLCriterion.lua index b8a4c87..71e6320 100644 --- a/SpatialClassNLLCriterion.lua +++ b/SpatialClassNLLCriterion.lua @@ -42,7 +42,7 @@ function SpatialClassNLLCriterion:adjustTarget(input, target) return target end -function SpatialClassNLLCriterion:forward(input,target) +function SpatialClassNLLCriterion:updateOutput(input,target) -- (1) adjust target: class -> distributions of classes -- compensate for convolution losses -- compensate for striding effects @@ -54,7 +54,7 @@ function SpatialClassNLLCriterion:forward(input,target) self.fullOutput = self.fullOutput or torch.Tensor() self.fullOutput:resizeAs(target) -- (3) compute the dense errors: - input.nn.SpatialClassNLLCriterion_forward(self,input,target) + input.nn.SpatialClassNLLCriterion_updateOutput(self,input,target) -- (4) prune the errors, either by averaging, or accumulation: if self.sizeAverage then self.output = self.fullOutput:mean() @@ -64,7 +64,7 @@ function SpatialClassNLLCriterion:forward(input,target) return self.output end -function SpatialClassNLLCriterion:backward(input,target) +function SpatialClassNLLCriterion:updateGradInput(input,target) -- (1) retrieve adjusted target target = self.target -- (2) resize input gradient map @@ -72,12 +72,12 @@ function SpatialClassNLLCriterion:backward(input,target) -- (3) compute input gradients, based on the nbGradients param if self.nbGradients == -1 then -- dense gradients - input.nn.SpatialClassNLLCriterion_backward(self,input,target,self.gradInput) + input.nn.SpatialClassNLLCriterion_updateGradInput(self,input,target,self.gradInput) elseif self.nbGradients == 1 then -- only 1 gradient is computed, sampled in the center self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialClassNLLCriterion_backward(self,input,target,self.fullGradInput) + input.nn.SpatialClassNLLCriterion_updateGradInput(self,input,target,self.fullGradInput) local y = math.ceil(self.gradInput:size(2)/2) local x = math.ceil(self.gradInput:size(3)/2) self.gradInput:select(3,x):select(2,y):copy(self.fullGradInput:select(3,x):select(2,y)) @@ -85,7 +85,7 @@ function SpatialClassNLLCriterion:backward(input,target) -- only N gradients are computed, sampled in random locations self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialClassNLLCriterion_backward(self,input,target,self.fullGradInput) + input.nn.SpatialClassNLLCriterion_updateGradInput(self,input,target,self.fullGradInput) for i = 1,self.nbGradients do local x = math.random(1,self.gradInput:size(1)) local y = math.random(1,self.gradInput:size(2)) @@ -94,18 +94,3 @@ function SpatialClassNLLCriterion:backward(input,target) end return self.gradInput end - -function SpatialClassNLLCriterion:write(file) - parent.write(self, file) - file:writeDouble(self.resampleTarget) - file:writeInt(self.nbGradients) - file:writeBool(self.sizeAverage) -end - -function SpatialClassNLLCriterion:read(file) - parent.read(self, file) - self.resampleTarget= file:readDouble() - self.nbGradients = file:readInt() - self.fullOutput = torch.Tensor() - self.sizeAverage = file:readBool() -end diff --git a/SpatialClassifier.lua b/SpatialClassifier.lua index 9d39f80..da15852 100644 --- a/SpatialClassifier.lua +++ b/SpatialClassifier.lua @@ -22,7 +22,7 @@ function Classifier:add(module) self.classifier:add(module) end -function Classifier:forward(input) +function Classifier:updateOutput(input) -- get dims: if input:nDimension() ~= 3 then error('<nn.SpatialClassifier> input should be 3D: KxHxW') @@ -37,7 +37,7 @@ function Classifier:forward(input) self.inputT:resize(HW, K):copy(self.inputF:t()) -- classify all locations: - self.outputT = self.classifier:forward(self.inputT) + self.outputT = self.classifier:updateOutput(self.inputT) if self.spatialOutput then -- transpose output: @@ -51,7 +51,7 @@ function Classifier:forward(input) return self.output end -function Classifier:backward(input, gradOutput) +function Classifier:updateGradInput(input, gradOutput) -- get dims: local K = input:size(1) local H = input:size(2) @@ -72,7 +72,7 @@ function Classifier:backward(input, gradOutput) end -- backward through classifier: - self.gradInputT = self.classifier:backward(self.inputT, self.gradOutputT) + self.gradInputT = self.classifier:updateGradInput(self.inputT, self.gradOutputT) -- transpose gradInput self.gradInputF:resize(K, HW):copy(self.gradInputT:t()) diff --git a/SpatialColorTransform.lua b/SpatialColorTransform.lua index a06efd4..941a391 100644 --- a/SpatialColorTransform.lua +++ b/SpatialColorTransform.lua @@ -89,9 +89,9 @@ function SpatialColorTransform:__init(type) end end -function SpatialColorTransform:forward(input) +function SpatialColorTransform:updateOutput(input) if self.islinear then - self.output = self.linear:forward(input) + self.output = self.linear:updateOutput(input) else if self.transform == 'rgb2hsl' then self.output = image.rgb2hsl(input, self.output) @@ -112,12 +112,12 @@ function SpatialColorTransform:forward(input) return self.output end -function SpatialColorTransform:backward(input, gradOutput) +function SpatialColorTransform:updateGradInput(input, gradOutput) if self.islinear then - self.gradInput = self.linear:backward(input, gradOutput) + self.gradInput = self.linear:updateGradInput(input, gradOutput) else - xlua.error('backward not implemented for non-linear transforms', - 'SpatialColorTransform.backward') + xlua.error('updateGradInput not implemented for non-linear transforms', + 'SpatialColorTransform.updateGradInput') end return self.gradInput end @@ -128,22 +128,3 @@ function SpatialColorTransform:type(type) self.linear:type(type) end end - -function SpatialColorTransform:write(file) - parent.write(self, file) - file:writeObject(self.transform) - file:writeBool(self.islinear) - if self.islinear then - file:writeObject(self.linear) - end -end - -function SpatialColorTransform:read(file) - parent.read(self, file) - self.transform = file:readObject() - self.islinear = file:readBool() - if self.islinear then - self.linear = file:readObject() - end - self.type = nil -end diff --git a/SpatialFovea.lua b/SpatialFovea.lua index c4ebcb8..dc03301 100644 --- a/SpatialFovea.lua +++ b/SpatialFovea.lua @@ -124,7 +124,7 @@ function SpatialFovea:configure(width,height) end end -function SpatialFovea:forward(input) +function SpatialFovea:updateOutput(input) -- input must be 3D if input:nDimension() ~= 3 then xerror('input must be 3d','nn.SpatialFovea') @@ -163,13 +163,13 @@ function SpatialFovea:forward(input) if not retrieved then -- (1) generate pyramid for idx = 1,nscales do - self.pyramid[idx] = self.downsamplers[idx]:forward(input) + self.pyramid[idx] = self.downsamplers[idx]:updateOutput(input) end -- (2) preprocess for idx = 1,nscales do if self.preProcessors[idx] then - self.preProcessed[idx] = self.preProcessors[idx]:forward(self.pyramid[idx]) + self.preProcessed[idx] = self.preProcessors[idx]:updateOutput(self.pyramid[idx]) else self.preProcessed[idx] = self.pyramid[idx] end @@ -177,7 +177,7 @@ function SpatialFovea:forward(input) -- (3) pad inputs for idx = 1,nscales do - self.padded[idx] = self.padders[idx]:forward(self.preProcessed[idx]) + self.padded[idx] = self.padders[idx]:updateOutput(self.preProcessed[idx]) end -- store preprocessed input for future use @@ -205,7 +205,7 @@ function SpatialFovea:forward(input) -- (5) apply processors to pyramid for idx = 1,nscales do - self.processed[idx] = self.processors[idx]:forward(self.narrowed[idx]) + self.processed[idx] = self.processors[idx]:updateOutput(self.narrowed[idx]) end -- (6) upscale, only if fovea is not focused @@ -215,7 +215,7 @@ function SpatialFovea:forward(input) end else for idx = 1,nscales do - self.upsampled[idx] = self.upsamplers[idx]:forward(self.processed[idx]) + self.upsampled[idx] = self.upsamplers[idx]:updateOutput(self.processed[idx]) end end @@ -234,7 +234,7 @@ function SpatialFovea:forward(input) return self.output end -function SpatialFovea:backward(input, gradOutput) +function SpatialFovea:updateGradInput(input, gradOutput) -- nb of scales local nscales = #self.ratios @@ -252,13 +252,13 @@ function SpatialFovea:backward(input, gradOutput) end else for idx = 1,nscales do - self.gradProcessed[idx] = self.upsamplers[idx]:backward(self.processed[idx], self.gradUpsampled[idx]) + self.gradProcessed[idx] = self.upsamplers[idx]:updateGradInput(self.processed[idx], self.gradUpsampled[idx]) end end -- (5) bprop through processors for idx = 1,nscales do - self.gradNarrowed[idx] = self.processors[idx]:backward(self.narrowed[idx], self.gradProcessed[idx]) + self.gradNarrowed[idx] = self.processors[idx]:updateGradInput(self.narrowed[idx], self.gradProcessed[idx]) end -- (beta) if caching preprocessed input, no need to compute @@ -285,13 +285,13 @@ function SpatialFovea:backward(input, gradOutput) -- (3) bprop through padders for idx = 1,nscales do - self.gradPreProcessed[idx] = self.padders[idx]:backward(self.preProcessed[idx], self.gradPadded[idx]) + self.gradPreProcessed[idx] = self.padders[idx]:updateGradInput(self.preProcessed[idx], self.gradPadded[idx]) end -- (2) bprop through preProcessors for idx = 1,nscales do if self.preProcessors[idx] then - self.gradPyramid[idx] = self.preProcessors[idx]:backward(self.pyramid[idx], self.gradPreProcessed[idx]) + self.gradPyramid[idx] = self.preProcessors[idx]:updateGradInput(self.pyramid[idx], self.gradPreProcessed[idx]) else self.gradPyramid[idx] = self.gradPreProcessed[idx] end @@ -300,7 +300,7 @@ function SpatialFovea:backward(input, gradOutput) -- (1) bprop through pyramid self.gradInput:resizeAs(self.gradPyramid[1]):zero() for idx = 1,nscales do - self.gradInput:add( self.downsamplers[idx]:backward(input, self.gradPyramid[idx]) ) + self.gradInput:add( self.downsamplers[idx]:updateGradInput(input, self.gradPyramid[idx]) ) end return self.gradInput end @@ -345,52 +345,3 @@ function SpatialFovea:type(type) end return self end - -function SpatialFovea:write(file) - parent.write(self, file) - file:writeInt(self.nInputPlane) - file:writeInt(self.padding) - file:writeInt(self.fov) - file:writeInt(self.sub) - file:writeBool(self.bilinear) - file:writeObject(self.ratios) - file:writeObject(self.downsamplers) - file:writeObject(self.padders) - file:writeObject(self.upsamplers) - file:writeObject(self.processors) - file:writeObject(self.preProcessors) - file:writeObject(self.pyramid) - file:writeObject(self.preProcessed) - file:writeObject(self.padded) - file:writeObject(self.narrowed) - file:writeObject(self.processed) - file:writeObject(self.upsampled) -end - -function SpatialFovea:read(file) - parent.read(self, file) - self.nInputPlane = file:readInt() - self.padding = file:readInt() - self.fov = file:readInt() - self.sub = file:readInt() - self.bilinear = file:readBool() - self.ratios = file:readObject() - self.downsamplers = file:readObject() - self.padders = file:readObject() - self.upsamplers = file:readObject() - self.processors = file:readObject() - self.preProcessors = file:readObject() - self.pyramid = file:readObject() - self.preProcessed = file:readObject() - self.padded = file:readObject() - self.narrowed = file:readObject() - self.processed = file:readObject() - self.upsampled = file:readObject() - self.gradUpsampled = {} - self.gradProcessed = {} - self.gradNarrowed = {} - self.gradPadded = {} - self.gradPreProcessed = {} - self.gradPyramid = {} - self.modules = self.processors -end diff --git a/SpatialGraph.lua b/SpatialGraph.lua index 253a800..6501313 100644 --- a/SpatialGraph.lua +++ b/SpatialGraph.lua @@ -42,28 +42,14 @@ function SpatialGraph:__init(...) end end -function SpatialGraph:forward(input) +function SpatialGraph:updateOutput(input) self.output:resize(self.connex / 2, input:size(2), input:size(3)) - input.nn.SpatialGraph_forward(self, input) + input.nn.SpatialGraph_updateOutput(self, input) return self.output end -function SpatialGraph:backward(input, gradOutput) +function SpatialGraph:updateGradInput(input, gradOutput) self.gradInput:resizeAs(input) - input.nn.SpatialGraph_backward(self, input, gradOutput) + input.nn.SpatialGraph_updateGradInput(self, input, gradOutput) return self.gradInput end - -function SpatialGraph:write(file) - parent.write(self, file) - file:writeInt(self.connex) - file:writeInt(self.dist) - file:writeInt(self.normalize) -end - -function SpatialGraph:read(file) - parent.read(self, file) - self.connex = file:readInt() - self.dist = file:readInt() - self.normalize = file:readInt() -end diff --git a/SpatialLinear.lua b/SpatialLinear.lua index 1cf850e..4fb12f9 100644 --- a/SpatialLinear.lua +++ b/SpatialLinear.lua @@ -52,36 +52,14 @@ function SpatialLinear:decayParameters(decay) self.bias:add(-decay, self.bias) end -function SpatialLinear:forward(input) +function SpatialLinear:updateOutput(input) self.output:resize(self.fanout, input:size(2), input:size(3)) - input.nn.SpatialLinear_forward(self, input) + input.nn.SpatialLinear_updateOutput(self, input) return self.output end -function SpatialLinear:backward(input, gradOutput) +function SpatialLinear:updateGradInput(input, gradOutput) self.gradInput:resize(self.fanin, input:size(2), input:size(3)) - input.nn.SpatialLinear_backward(self, input, gradOutput) + input.nn.SpatialLinear_updateGradInput(self, input, gradOutput) return self.gradInput end - -function SpatialLinear:write(file) - parent.write(self, file) - file:writeInt(self.fanin) - file:writeInt(self.fanout) - file:writeDouble(self.weightDecay) - file:writeObject(self.weight) - file:writeObject(self.bias) - file:writeObject(self.gradWeight) - file:writeObject(self.gradBias) -end - -function SpatialLinear:read(file) - parent.read(self, file) - self.fanin = file:readInt() - self.fanout = file:readInt() - self.weightDecay = file:readDouble() - self.weight = file:readObject() - self.bias = file:readObject() - self.gradWeight = file:readObject() - self.gradBias = file:readObject() -end diff --git a/SpatialMSECriterion.lua b/SpatialMSECriterion.lua index 4c2ced8..634bfc5 100644 --- a/SpatialMSECriterion.lua +++ b/SpatialMSECriterion.lua @@ -58,7 +58,7 @@ function SpatialMSECriterion:adjustTarget(input, target) return target end -function SpatialMSECriterion:forward(input,target) +function SpatialMSECriterion:updateOutput(input,target) -- (1) adjust target: class -> distributions of classes -- compensate for convolution losses -- compensate for striding effects @@ -70,7 +70,7 @@ function SpatialMSECriterion:forward(input,target) self.fullOutput = self.fullOutput or torch.Tensor() self.fullOutput:resizeAs(input) -- (3) compute the dense errors: - input.nn.SpatialMSECriterion_forward(self, input, target) + input.nn.SpatialMSECriterion_updateOutput(self, input, target) -- (4) prune the errors, either by averaging, or accumulation: if self.sizeAverage then self.output = self.fullOutput:mean() @@ -80,7 +80,7 @@ function SpatialMSECriterion:forward(input,target) return self.output end -function SpatialMSECriterion:backward(input,target) +function SpatialMSECriterion:updateGradInput(input,target) -- (1) retrieve adjusted target target = self.target -- (2) resize input gradient map @@ -88,12 +88,12 @@ function SpatialMSECriterion:backward(input,target) -- (3) compute input gradients, based on the nbGradients param if self.nbGradients == -1 then -- dense gradients - input.nn.SpatialMSECriterion_backward(self, input, target, self.gradInput) + input.nn.SpatialMSECriterion_updateGradInput(self, input, target, self.gradInput) elseif self.nbGradients == 1 then -- only 1 gradient is computed, sampled in the center self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialMSECriterion_backward(self, input, target, self.fullGradInput) + input.nn.SpatialMSECriterion_updateGradInput(self, input, target, self.fullGradInput) local y = math.ceil(self.gradInput:size(2)/2) local x = math.ceil(self.gradInput:size(3)/2) self.gradInput:select(3,x):select(2,y):copy(self.fullGradInput:select(3,x):select(2,y)) @@ -101,7 +101,7 @@ function SpatialMSECriterion:backward(input,target) -- only N gradients are computed, sampled in random locations self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialMSECriterion_backward(self, input, target, self.fullGradInput) + input.nn.SpatialMSECriterion_updateGradInput(self, input, target, self.fullGradInput) for i = 1,self.nbGradients do local x = math.random(1,self.gradInput:size(1)) local y = math.random(1,self.gradInput:size(2)) @@ -110,23 +110,3 @@ function SpatialMSECriterion:backward(input,target) end return self.gradInput end - -function SpatialMSECriterion:write(file) - parent.write(self, file) - file:writeDouble(self.resampleTarget) - file:writeInt(self.nbGradients) - if not self.ignoreClass then - file:writeInt(-1) - end -end - -function SpatialMSECriterion:read(file) - parent.read(self, file) - self.resampleTarget= file:readDouble() - self.nbGradients = file:readInt() - self.ignoreClass = file:readInt() - if self.ignoreClass == -1 then - self.ignoreClass = false - end - self.fullOutput = torch.Tensor() -end diff --git a/SpatialMaxSampling.lua b/SpatialMaxSampling.lua index aa16701..d6dc512 100644 --- a/SpatialMaxSampling.lua +++ b/SpatialMaxSampling.lua @@ -11,12 +11,12 @@ function SpatialMaxSampling:__init(...) self.indices = torch.Tensor() end -function SpatialMaxSampling:forward(input) - input.nn.SpatialMaxSampling_forward(self, input) +function SpatialMaxSampling:updateOutput(input) + input.nn.SpatialMaxSampling_updateOutput(self, input) return self.output end -function SpatialMaxSampling:backward(input, gradOutput) - input.nn.SpatialMaxSampling_backward(self, input, gradOutput) +function SpatialMaxSampling:updateGradInput(input, gradOutput) + input.nn.SpatialMaxSampling_updateGradInput(self, input, gradOutput) return self.gradInput end diff --git a/SpatialNormalization.lua b/SpatialNormalization.lua index e0e5c60..012b609 100644 --- a/SpatialNormalization.lua +++ b/SpatialNormalization.lua @@ -182,7 +182,7 @@ function SpatialNormalization:__init(...) -- kernel for weighted mean | nb of fe self.thstd = torch.Tensor() end -function SpatialNormalization:forward(input) +function SpatialNormalization:updateOutput(input) -- auto switch to 3-channel self.input = input if (input:nDimension() == 2) then @@ -192,32 +192,32 @@ function SpatialNormalization:forward(input) -- recompute coef only if necessary if (self.input:size(3) ~= self.coef:size(2)) or (self.input:size(2) ~= self.coef:size(1)) then local intVals = self.input.new(self.nfeatures,self.input:size(2),self.input:size(3)):fill(1) - self.coef = self.convo:forward(intVals) + self.coef = self.convo:updateOutput(intVals) self.coef = self.coef:clone() end -- compute mean - self.inConvo = self.convo:forward(self.input) - self.inMean = self.meanDiviseMod:forward{self.inConvo,self.coef} - self.inputZeroMean = self.subtractMod:forward{self.input,self.inMean} + self.inConvo = self.convo:updateOutput(self.input) + self.inMean = self.meanDiviseMod:updateOutput{self.inConvo,self.coef} + self.inputZeroMean = self.subtractMod:updateOutput{self.input,self.inMean} -- compute std dev - self.inputZeroMeanSq = self.squareMod:forward(self.inputZeroMean) - self.inConvoVar = self.convostd:forward(self.inputZeroMeanSq) - self.inStdDevNotUnit = self.sqrtMod:forward(self.inConvoVar) - self.inStdDev = self.stdDiviseMod:forward({self.inStdDevNotUnit,self.coef}) + self.inputZeroMeanSq = self.squareMod:updateOutput(self.inputZeroMean) + self.inConvoVar = self.convostd:updateOutput(self.inputZeroMeanSq) + self.inStdDevNotUnit = self.sqrtMod:updateOutput(self.inConvoVar) + self.inStdDev = self.stdDiviseMod:updateOutput({self.inStdDevNotUnit,self.coef}) local meanstd = self.inStdDev:mean() self.thresMod.threshold = self.fixedThres or math.max(meanstd,1e-3) self.thresMod.val = self.fixedThres or math.max(meanstd,1e-3) - self.stdDev = self.thresMod:forward(self.inStdDev) + self.stdDev = self.thresMod:updateOutput(self.inStdDev) --remove std dev - self.diviseMod:forward{self.inputZeroMean,self.stdDev} + self.diviseMod:updateOutput{self.inputZeroMean,self.stdDev} self.output = self.diviseMod.output return self.output end -function SpatialNormalization:backward(input, gradOutput) +function SpatialNormalization:updateGradInput(input, gradOutput) -- auto switch to 3-channel self.input = input if (input:nDimension() == 2) then @@ -226,23 +226,23 @@ function SpatialNormalization:backward(input, gradOutput) self.gradInput:resizeAs(self.input):zero() -- backprop all - local gradDiv = self.diviseMod:backward({self.inputZeroMean,self.stdDev},gradOutput) + local gradDiv = self.diviseMod:updateGradInput({self.inputZeroMean,self.stdDev},gradOutput) local gradThres = gradDiv[2] local gradZeroMean = gradDiv[1] - local gradinStdDev = self.thresMod:backward(self.inStdDev,gradThres) - local gradstdDiv = self.stdDiviseMod:backward({self.inStdDevNotUnit,self.coef},gradinStdDev) + local gradinStdDev = self.thresMod:updateGradInput(self.inStdDev,gradThres) + local gradstdDiv = self.stdDiviseMod:updateGradInput({self.inStdDevNotUnit,self.coef},gradinStdDev) local gradinStdDevNotUnit = gradstdDiv[1] - local gradinConvoVar = self.sqrtMod:backward(self.inConvoVar,gradinStdDevNotUnit) - local gradinputZeroMeanSq = self.convostd:backward(self.inputZeroMeanSq,gradinConvoVar) - gradZeroMean:add(self.squareMod:backward(self.inputZeroMean,gradinputZeroMeanSq)) - local gradDiff = self.subtractMod:backward({self.input,self.inMean},gradZeroMean) + local gradinConvoVar = self.sqrtMod:updateGradInput(self.inConvoVar,gradinStdDevNotUnit) + local gradinputZeroMeanSq = self.convostd:updateGradInput(self.inputZeroMeanSq,gradinConvoVar) + gradZeroMean:add(self.squareMod:updateGradInput(self.inputZeroMean,gradinputZeroMeanSq)) + local gradDiff = self.subtractMod:updateGradInput({self.input,self.inMean},gradZeroMean) local gradinMean = gradDiff[2] - local gradinConvoNotUnit = self.meanDiviseMod:backward({self.inConvo,self.coef},gradinMean) + local gradinConvoNotUnit = self.meanDiviseMod:updateGradInput({self.inConvo,self.coef},gradinMean) local gradinConvo = gradinConvoNotUnit[1] -- first part of the gradInput self.gradInput:add(gradDiff[1]) -- second part of the gradInput - self.gradInput:add(self.convo:backward(self.input,gradinConvo)) + self.gradInput:add(self.convo:updateGradInput(self.input,gradinConvo)) return self.gradInput end @@ -259,58 +259,3 @@ function SpatialNormalization:type(type) self.diviseMod:type(type) return self end - -function SpatialNormalization:write(file) - parent.write(self,file) - file:writeObject(self.kernel) - file:writeInt(self.nfeatures) - file:writeInt(self.padW) - file:writeInt(self.padH) - file:writeInt(self.kerWisPair) - file:writeInt(self.kerHisPair) - file:writeObject(self.convo) - file:writeObject(self.convostd) - file:writeObject(self.squareMod) - file:writeObject(self.sqrtMod) - file:writeObject(self.subtractMod) - file:writeObject(self.meanDiviseMod) - file:writeObject(self.stdDiviseMod) - file:writeObject(self.thresMod) - file:writeObject(self.diviseMod) - file:writeObject(self.coef) - if type(self.kernel) == 'table' then - file:writeInt(self.pad2W) - file:writeInt(self.pad2H) - file:writeInt(self.ker2WisPair) - file:writeInt(self.ker2HisPair) - end - file:writeInt(self.fixedThres or 0) -end - -function SpatialNormalization:read(file) - parent.read(self,file) - self.kernel = file:readObject() - self.nfeatures = file:readInt() - self.padW = file:readInt() - self.padH = file:readInt() - self.kerWisPair = file:readInt() - self.kerHisPair = file:readInt() - self.convo = file:readObject() - self.convostd = file:readObject() - self.squareMod = file:readObject() - self.sqrtMod = file:readObject() - self.subtractMod = file:readObject() - self.meanDiviseMod = file:readObject() - self.stdDiviseMod = file:readObject() - self.thresMod = file:readObject() - self.diviseMod = file:readObject() - self.coef = file:readObject() - if type(self.kernel) == 'table' then - self.pad2W = file:readInt() - self.pad2H = file:readInt() - self.ker2WisPair = file:readInt() - self.ker2HisPair = file:readInt() - end - self.fixedThres = file:readInt() - if self.fixedThres == 0 then self.fixedThres = nil end -end diff --git a/SpatialPadding.lua b/SpatialPadding.lua index 9c87c2b..ecea707 100644 --- a/SpatialPadding.lua +++ b/SpatialPadding.lua @@ -19,7 +19,7 @@ function SpatialPadding:__init(pad_l, pad_r, pad_t, pad_b) self.pad_b = pad_b or self.pad_l end -function SpatialPadding:forward(input) +function SpatialPadding:updateOutput(input) if input:dim() ~= 3 then error('input must be 3-dimensional') end local h = input:size(2) + self.pad_t + self.pad_b local w = input:size(3) + self.pad_l + self.pad_r @@ -43,7 +43,7 @@ function SpatialPadding:forward(input) return self.output end -function SpatialPadding:backward(input, gradOutput) +function SpatialPadding:updateGradInput(input, gradOutput) if input:dim() ~= 3 then error('input must be 3-dimensional') end self.gradInput:resizeAs(input):zero() -- crop gradInput if necessary @@ -62,20 +62,3 @@ function SpatialPadding:backward(input, gradOutput) cg_input:copy(cg_output) return self.gradInput end - -function SpatialPadding:write(file) - parent.write(self, file) - file:writeInt(self.pad_l) - file:writeInt(self.pad_r) - file:writeInt(self.pad_t) - file:writeInt(self.pad_b) -end - -function SpatialPadding:read(file) - parent.read(self, file) - self.pad_l = file:readInt() - self.pad_r = file:readInt() - self.pad_t = file:readInt() - self.pad_b = file:readInt() -end - diff --git a/SpatialReSampling.lua b/SpatialReSampling.lua index 9aef63c..b738eab 100644 --- a/SpatialReSampling.lua +++ b/SpatialReSampling.lua @@ -30,26 +30,14 @@ function SpatialReSampling:__init(...) ) end -function SpatialReSampling:forward(input) +function SpatialReSampling:updateOutput(input) self.oheight = self.oheight or self.rheight*input:size(2) self.owidth = self.owidth or self.rwidth*input:size(3) - input.nn.SpatialReSampling_forward(self, input) + input.nn.SpatialReSampling_updateOutput(self, input) return self.output end -function SpatialReSampling:backward(input, gradOutput) - input.nn.SpatialReSampling_backward(self, input, gradOutput) +function SpatialReSampling:updateGradInput(input, gradOutput) + input.nn.SpatialReSampling_updateGradInput(self, input, gradOutput) return self.gradInput end - -function SpatialReSampling:write(file) - parent.write(self, file) - file:writeInt(self.owidth) - file:writeInt(self.oheight) -end - -function SpatialReSampling:read(file) - parent.read(self, file) - self.owidth = file:readInt() - self.oheight = file:readInt() -end diff --git a/SpatialRecursiveFovea.lua b/SpatialRecursiveFovea.lua index c9490c2..ef7243b 100644 --- a/SpatialRecursiveFovea.lua +++ b/SpatialRecursiveFovea.lua @@ -275,7 +275,7 @@ function SpatialRecursiveFovea:focus(x,y) self.corners = corners end -function SpatialRecursiveFovea:forward(input,target,x,y) +function SpatialRecursiveFovea:updateOutput(input,target,x,y) -- input must be 3D if input:nDimension() ~= 3 or input:size(1) ~= self.nInputPlane then xerror('input must be 3d and have ' .. self.nInputPlane .. ' input planes','nn.SpatialRecursiveFovea') @@ -299,11 +299,11 @@ function SpatialRecursiveFovea:forward(input,target,x,y) -- (1-2) create preprocessed pyramid for idx = 1,nscales do -- (1) generate pyramid - self.pyramid[idx] = self.downsamplers[idx]:forward(input) + self.pyramid[idx] = self.downsamplers[idx]:updateOutput(input) -- (2) preprocess if self.preProcessors[idx] then - self.preProcessed[idx] = self.preProcessors[idx]:forward(self.pyramid[idx]) + self.preProcessed[idx] = self.preProcessors[idx]:updateOutput(self.pyramid[idx]) else self.preProcessed[idx] = self.pyramid[idx] end @@ -312,7 +312,7 @@ function SpatialRecursiveFovea:forward(input,target,x,y) -- (3-7) walk through recursion for idx = 1,nscales do -- (3) pad inputs - self.padded[idx] = self.padders[idx]:forward(self.preProcessed[idx]) + self.padded[idx] = self.padders[idx]:updateOutput(self.preProcessed[idx]) -- (4) is fovea focused ? self.narrowed[idx] @@ -336,15 +336,15 @@ function SpatialRecursiveFovea:forward(input,target,x,y) end -- (6) apply processors to pyramid - self.processed[idx] = self.processors[idx]:forward(self.concatenated[idx]) + self.processed[idx] = self.processors[idx]:updateOutput(self.concatenated[idx]) -- (7) upsample, pad and narrow, for next stage if idx < nscales then -- (7.a) - self.upsampled[idx] = self.upsamplers[idx]:forward(self.processed[idx]) + self.upsampled[idx] = self.upsamplers[idx]:updateOutput(self.processed[idx]) -- (7.b) - self.upsampledPadded[idx] = self.upsampledPadders[idx]:forward(self.upsampled[idx]) + self.upsampledPadded[idx] = self.upsampledPadders[idx]:updateOutput(self.upsampled[idx]) -- (7.c) self.upsampledNarrowed[idx] @@ -355,7 +355,7 @@ function SpatialRecursiveFovea:forward(input,target,x,y) -- (8) optional post processors for idx = 1,nscales do if self.postProcessors[idx] then - self.postProcessed[idx] = self.postProcessors[idx]:forward(self.processed[idx]) + self.postProcessed[idx] = self.postProcessors[idx]:updateOutput(self.processed[idx]) else self.postProcessed[idx] = self.processed[idx] end @@ -410,7 +410,7 @@ function SpatialRecursiveFovea:forward(input,target,x,y) self.targets[idx] = self.targets_scaled[idx] end -- then evaluate the criterion's error - error = error + self.criterions[idx]:forward(self.predicted[idx], self.targets[idx]) + error = error + self.criterions[idx]:updateOutput(self.predicted[idx], self.targets[idx]) end -- normalize error @@ -434,17 +434,17 @@ function SpatialRecursiveFovea:forward(input,target,x,y) return self.output, error end -function SpatialRecursiveFovea:backward(input) +function SpatialRecursiveFovea:updateGradInput(input) -- local params local nscales = #self.ratios local fov = self.fov local sub = self.sub local corners = self.corners - -- (9) backprop through criterions using generated targets (from prev forward call) + -- (9) backprop through criterions using generated targets (from prev updateOutput call) for idx = 1,nscales do -- bprop through criterion - self.gradPredicted[idx] = self.criterions[idx]:backward(self.predicted[idx], self.targets[idx]) + self.gradPredicted[idx] = self.criterions[idx]:updateGradInput(self.predicted[idx], self.targets[idx]) -- then remap partial grad vector self.gradPostProcessed[idx] = self.gradPostProcessed[idx] or torch.Tensor() @@ -460,7 +460,7 @@ function SpatialRecursiveFovea:backward(input) -- (8) backprop through post processors for idx = 1,nscales do if self.postProcessors[idx] then - self.gradProcessed[idx] = self.postProcessors[idx]:backward(self.processed[idx], self.gradPostProcessed[idx]) + self.gradProcessed[idx] = self.postProcessors[idx]:updateGradInput(self.processed[idx], self.gradPostProcessed[idx]) else self.gradProcessed[idx] = self.gradPostProcessed[idx] end @@ -471,7 +471,7 @@ function SpatialRecursiveFovea:backward(input) -- (6) backprop through processors for idx = 1,nscales do - self.gradConcatenated[idx] = self.processors[idx]:backward(self.concatenated[idx], self.gradProcessed[idx]) + self.gradConcatenated[idx] = self.processors[idx]:updateGradInput(self.concatenated[idx], self.gradProcessed[idx]) end -- (5) bprop through concatenators @@ -488,13 +488,13 @@ function SpatialRecursiveFovea:backward(input) -- (3) bprop through padders for idx = 1,nscales do - self.gradPreProcessed[idx] = self.padders[idx]:backward(self.preProcessed[idx], self.gradPadded[idx]) + self.gradPreProcessed[idx] = self.padders[idx]:updateGradInput(self.preProcessed[idx], self.gradPadded[idx]) end -- (2) bprop through preProcessors for idx = 1,nscales do if self.preProcessors[idx] then - self.gradPyramid[idx] = self.preProcessors[idx]:backward(self.pyramid[idx], self.gradPreProcessed[idx]) + self.gradPyramid[idx] = self.preProcessors[idx]:updateGradInput(self.pyramid[idx], self.gradPreProcessed[idx]) else self.gradPyramid[idx] = self.gradPreProcessed[idx] end @@ -503,7 +503,7 @@ function SpatialRecursiveFovea:backward(input) -- (1) bprop through pyramid self.gradInput:resizeAs(input):zero() for idx = 1,nscales do - local partialGrad = self.downsamplers[idx]:backward(input, self.gradPyramid[idx]) + local partialGrad = self.downsamplers[idx]:updateGradInput(input, self.gradPyramid[idx]) self.gradInput:add(partialGrad) end return self.gradInput @@ -534,80 +534,3 @@ function SpatialRecursiveFovea:decayParameters(decay) end end end - -function SpatialRecursiveFovea:write(file) - parent.write(self, file) - -- params - file:writeInt(self.nInputPlane) - file:writeInt(self.nRecursivePlane) - file:writeInt(self.fov) - file:writeInt(self.sub) - file:writeObject(self.ratios) - -- modules - file:writeObject(self.downsamplers) - file:writeObject(self.padders) - file:writeObject(self.upsamplers) - file:writeObject(self.upsampledPadders) - file:writeObject(self.processors) - file:writeObject(self.preProcessors) - file:writeObject(self.postProcessors) - file:writeObject(self.criterions) - -- states - file:writeObject(self.pyramid) - file:writeObject(self.preProcessed) - file:writeObject(self.padded) - file:writeObject(self.narrowed) - file:writeObject(self.concatenated) - file:writeObject(self.processed) - file:writeObject(self.upsampled) - file:writeObject(self.upsampledPadded) - file:writeObject(self.upsampledNarrowed) - file:writeObject(self.postProcessed) - file:writeObject(self.predicted) -end - -function SpatialRecursiveFovea:read(file) - parent.read(self, file) - -- params - self.nInputPlane = file:readInt() - self.nRecursivePlane = file:readInt() - self.fov = file:readInt() - self.sub = file:readInt() - self.ratios = file:readObject() - self.batchSize = self.fov - -- modules - self.downsamplers = file:readObject() - self.padders = file:readObject() - self.upsamplers = file:readObject() - self.upsampledPadders = file:readObject() - self.processors = file:readObject() - self.preProcessors = file:readObject() - self.postProcessors = file:readObject() - self.criterions = file:readObject() - -- states - self.pyramid = file:readObject() - self.preProcessed = file:readObject() - self.padded = file:readObject() - self.narrowed = file:readObject() - self.concatenated = file:readObject() - self.processed = file:readObject() - self.upsampled = file:readObject() - self.upsampledPadded = file:readObject() - self.upsampledNarrowed = file:readObject() - self.postProcessed = file:readObject() - self.predicted = file:readObject() - -- grad states - self.gradPostProcessed = {} - self.gradUpsampledNarrowed = {} - self.gradUpsampledPadded = {} - self.gradUpsampled = {} - self.gradProcessed = {} - self.gradConcatenated = {} - self.gradNarrowed = {} - self.gradPadded = {} - self.gradPreProcessed = {} - self.gradPyramid = {} - self.gradPredicted = {} - self.targets_scaled = {} - self.targets = {} -end diff --git a/SpatialSparseCriterion.lua b/SpatialSparseCriterion.lua index abb2301..f19475e 100644 --- a/SpatialSparseCriterion.lua +++ b/SpatialSparseCriterion.lua @@ -12,10 +12,10 @@ function SpatialSparseCriterion:__init(...) ) end -function SpatialSparseCriterion:forward(input) +function SpatialSparseCriterion:updateOutput(input) self.fullOutput = self.fullOutput or torch.Tensor() self.fullOutput:resize(input:size(2), input:size(3)) - input.nn.SpatialSparseCriterion_forward(self, input) + input.nn.SpatialSparseCriterion_updateOutput(self, input) if self.sizeAverage then self.output = self.fullOutput:mean() else @@ -24,7 +24,7 @@ function SpatialSparseCriterion:forward(input) return self.output end -function SpatialSparseCriterion:backward(input,target) +function SpatialSparseCriterion:updateGradInput(input,target) -- (1) retrieve adjusted target target = self.target -- (2) resize input gradient map @@ -32,12 +32,12 @@ function SpatialSparseCriterion:backward(input,target) -- (3) compute input gradients, based on the nbGradients param if self.nbGradients == -1 then -- dense gradients - input.nn.SpatialSparseCriterion_backward(self, input, self.gradInput) + input.nn.SpatialSparseCriterion_updateGradInput(self, input, self.gradInput) elseif self.nbGradients == 1 then -- only 1 gradient is computed, sampled in the center self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialSparseCriterion_backward(self, input, self.fullGradInput) + input.nn.SpatialSparseCriterion_updateGradInput(self, input, self.fullGradInput) local y = math.ceil(self.gradInput:size(2)/2) local x = math.ceil(self.gradInput:size(3)/2) self.gradInput:select(3,x):select(2,y):copy(self.fullGradInput:select(3,x):select(2,y)) @@ -45,7 +45,7 @@ function SpatialSparseCriterion:backward(input,target) -- only N gradients are computed, sampled in random locations self.fullGradInput = torch.Tensor() or self.fullGradInput self.fullGradInput:resizeAs(input):zero() - input.nn.SpatialSparseCriterion_backward(self, input, self.fullGradInput) + input.nn.SpatialSparseCriterion_updateGradInput(self, input, self.fullGradInput) for i = 1,self.nbGradients do local x = math.random(1,self.gradInput:size(1)) local y = math.random(1,self.gradInput:size(2)) @@ -54,15 +54,3 @@ function SpatialSparseCriterion:backward(input,target) end return self.gradInput end - -function SpatialSparseCriterion:write(file) - parent.write(self, file) - file:writeDouble(self.resampleTarget) - file:writeInt(self.nbGradients) -end - -function SpatialSparseCriterion:read(file) - parent.read(self, file) - self.resampleTarget= file:readDouble() - self.nbGradients = file:readInt() -end diff --git a/SpatialUpSampling.lua b/SpatialUpSampling.lua index 90d0126..0d26bcd 100644 --- a/SpatialUpSampling.lua +++ b/SpatialUpSampling.lua @@ -25,26 +25,14 @@ function SpatialUpSampling:__init(...) {arg='dH', type='number', help='stride height', req=true}) end -function SpatialUpSampling:forward(input) +function SpatialUpSampling:updateOutput(input) self.output:resize(input:size(1), input:size(2) * self.dH, input:size(3) * self.dW) - input.nn.SpatialUpSampling_forward(self, input) + input.nn.SpatialUpSampling_updateOutput(self, input) return self.output end -function SpatialUpSampling:backward(input, gradOutput) +function SpatialUpSampling:updateGradInput(input, gradOutput) self.gradInput:resizeAs(input) - input.nn.SpatialUpSampling_backward(self, input, gradOutput) + input.nn.SpatialUpSampling_updateGradInput(self, input, gradOutput) return self.gradInput end - -function SpatialUpSampling:write(file) - parent.write(self, file) - file:writeInt(self.dW) - file:writeInt(self.dH) -end - -function SpatialUpSampling:read(file) - parent.read(self, file) - self.dW = file:readInt() - self.dH = file:readInt() -end diff --git a/Sqrt.lua b/Sqrt.lua deleted file mode 100644 index 4c0427b..0000000 --- a/Sqrt.lua +++ /dev/null @@ -1,34 +0,0 @@ - -local Sqrt, parent = torch.class('nn.Sqrt','nn.Module') - -function Sqrt:__init(args) - parent.__init(self) - if args then - error(xlua.usage('nn.Sqrt', - 'a simple component-wise mapping: sqrt()', - 'sq = nn.Sqrt()\n'.. - 'sqrt = sq:forward(sometensor)', - {type='nil', help='no arg required'})) - end -end - -function Sqrt:forward(input) - self.output:resizeAs(input):copy(input) - self.output:sqrt() - return self.output -end - -function Sqrt:backward(input, gradOutput) - self.gradInput:resizeAs(input):copy(gradOutput) - self.gradInput:cdiv(self.output):mul(0.5) - return self.gradInput -end - - -function Sqrt:write(file) - parent.write(self,file) -end - -function Sqrt:read(file) - parent.read(self,file) -end diff --git a/Square.lua b/Square.lua deleted file mode 100644 index 30e8cb9..0000000 --- a/Square.lua +++ /dev/null @@ -1,33 +0,0 @@ -local Square, parent = torch.class('nn.Square','nn.Module') - -function Square:__init(args) - parent.__init(self) - if args then - error(xlua.usage('nn.Square', - 'a simple component-wise mapping: square()', - 'sq = nn.Square()\n'.. - 'squared = sq:forward(sometensor)', - {type='nil', help='no arg required'})) - end -end - -function Square:forward(input) - self.output:resizeAs(input):copy(input) - self.output:cmul(input) - return self.output -end - -function Square:backward(input, gradOutput) - self.gradInput:resizeAs(input):copy(gradOutput) - self.gradInput:cmul(input):mul(2) - return self.gradInput -end - - -function Square:write(file) - parent.write(self,file) -end - -function Square:read(file) - parent.read(self,file) -end diff --git a/SuperCriterion.lua b/SuperCriterion.lua index c73d716..983639a 100644 --- a/SuperCriterion.lua +++ b/SuperCriterion.lua @@ -13,32 +13,32 @@ function SuperCriterion:add(criterion, weight) table.insert(self.weights, weight) end -function SuperCriterion:forward(input, target) +function SuperCriterion:updateOutput(input, target) self.output = 0 if type(target) == 'table' then for i,criterion in ipairs(self.criterions) do - self.output = self.output + self.weights[i]*criterion:forward(input[i],target[i]) + self.output = self.output + self.weights[i]*criterion:updateOutput(input[i],target[i]) end else for i,criterion in ipairs(self.criterions) do - self.output = self.output + self.weights[i]*criterion:forward(input[i],target) + self.output = self.output + self.weights[i]*criterion:updateOutput(input[i],target) end end return self.output end -function SuperCriterion:backward(input, target) +function SuperCriterion:updateGradInput(input, target) if type(target) == 'table' then for i,criterion in ipairs(self.criterions) do self.gradInput[i] = torch.Tensor() or self.gradInput[i] self.gradInput[i]:resizeAs(input[i]):zero() - self.gradInput[i]:add(self.weights[i], criterion:backward(input[i],target[i])) + self.gradInput[i]:add(self.weights[i], criterion:updateGradInput(input[i],target[i])) end else for i,criterion in ipairs(self.criterions) do self.gradInput[i] = torch.Tensor() or self.gradInput[i] self.gradInput[i]:resizeAs(input[i]):zero() - self.gradInput[i]:add(self.weights[i], criterion:backward(input[i],target)) + self.gradInput[i]:add(self.weights[i], criterion:updateGradInput(input[i],target)) end end return self.gradInput diff --git a/Threshold.lua b/Threshold.lua deleted file mode 100644 index fbd5c54..0000000 --- a/Threshold.lua +++ /dev/null @@ -1,36 +0,0 @@ -local Threshold, parent = torch.class('nn.Threshold','nn.Module') - -function Threshold:__init(th,v) - parent.__init(self) - self.threshold = th or 1e-6 - self.val = v or 0 - if (th and type(th) ~= 'number') or (v and type(v) ~= 'number') then - error(xlua.usage('nn.Threshold', - 'a threhold module, if input < threshold, then output = value', - nil, - {type='number', help='threshold'}, - {type='number', help='value'})) - end -end - -function Threshold:forward(input) - input.nn.Threshold_forward(self, input) - return self.output -end - -function Threshold:backward(input, gradOutput) - input.nn.Threshold_backward(self, input, gradOutput) - return self.gradInput -end - -function Threshold:write(file) - parent.write(self,file) - file:writeDouble(self.threshold) - file:writeDouble(self.val) -end - -function Threshold:read(file) - parent.read(self,file) - self.threshold = file:readDouble() - self.val = file:readDouble() -end diff --git a/Trainer.lua b/Trainer.lua index b7da770..34d1d33 100644 --- a/Trainer.lua +++ b/Trainer.lua @@ -9,18 +9,6 @@ end function Trainer:train(dataset) end -function Trainer:write(file) - file:writeDouble(self.learningRate) - file:writeDouble(self.learningRateDecay) - file:writeInt(self.maxIteration) -end - -function Trainer:read(file) - self.learningRate = file:readDouble() - self.learningRateDecay = file:readDouble() - self.maxIteration = file:readInt() -end - function Trainer:share(mlp, ...) for i,v in ipairs(arg) do if self[v] ~=nil then self[v]:set(mlp[v]) end @@ -18,24 +18,24 @@ function Type:add(module) return self end -function Type:forward(input) - input = self.convert_input:forward(input) - local output = parent.forward(self, input) - self.output = self.convert_output:forward(output) +function Type:updateOutput(input) + input = self.convert_input:updateOutput(input) + local output = parent.updateOutput(self, input) + self.output = self.convert_output:updateOutput(output) return self.output end -function Type:backward(input, gradOutput) - input = self.convert_input:forward(input) - gradOutput = self.convert_gradOutput:forward(gradOutput) - local gradInput = parent.backward(self, input, gradOutput) - self.gradInput = self.convert_gradInput:forward(gradInput) +function Type:updateGradInput(input, gradOutput) + input = self.convert_input:updateOutput(input) + gradOutput = self.convert_gradOutput:updateOutput(gradOutput) + local gradInput = parent.updateGradInput(self, input, gradOutput) + self.gradInput = self.convert_gradInput:updateOutput(gradInput) return self.gradInput end function Type:accGradParameters(input, gradOutput) - input = self.convert_input:forward(input) - gradOutput = self.convert_gradOutput:forward(gradOutput) + input = self.convert_input:updateOutput(input) + gradOutput = self.convert_gradOutput:updateOutput(gradOutput) parent.accGradParameters(self, input, gradOutput) end diff --git a/generic/Abs.c b/generic/Abs.c index c45c0ab..8c65813 100644 --- a/generic/Abs.c +++ b/generic/Abs.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/Abs.c" #else -static int nn_(Abs_forward)(lua_State *L) +static int nn_(Abs_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id)); @@ -14,7 +14,7 @@ static int nn_(Abs_forward)(lua_State *L) return 1; } -static int nn_(Abs_backward)(lua_State *L) +static int nn_(Abs_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -28,8 +28,8 @@ static int nn_(Abs_backward)(lua_State *L) } static const struct luaL_Reg nn_(Abs__) [] = { - {"Abs_forward", nn_(Abs_forward)}, - {"Abs_backward", nn_(Abs_backward)}, + {"Abs_updateOutput", nn_(Abs_updateOutput)}, + {"Abs_updateGradInput", nn_(Abs_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/DistMarginCriterion.c b/generic/DistMarginCriterion.c index 6e94c9d..8859bdd 100644 --- a/generic/DistMarginCriterion.c +++ b/generic/DistMarginCriterion.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/DistMarginCriterion.c" #else -static int nn_(DistMarginCriterion_forward)(lua_State *L) +static int nn_(DistMarginCriterion_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage"); @@ -82,7 +82,7 @@ static int nn_(DistMarginCriterion_forward)(lua_State *L) return 1; } -static int nn_(DistMarginCriterion_backward)(lua_State *L) +static int nn_(DistMarginCriterion_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage"); @@ -172,8 +172,8 @@ static int nn_(DistMarginCriterion_backward)(lua_State *L) } static const struct luaL_Reg nn_(DistMarginCriterion__) [] = { - {"DistMarginCriterion_forward", nn_(DistMarginCriterion_forward)}, - {"DistMarginCriterion_backward", nn_(DistMarginCriterion_backward)}, + {"DistMarginCriterion_updateOutput", nn_(DistMarginCriterion_updateOutput)}, + {"DistMarginCriterion_updateGradInput", nn_(DistMarginCriterion_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/HardShrink.c b/generic/HardShrink.c index d3bcfcd..d75e071 100644 --- a/generic/HardShrink.c +++ b/generic/HardShrink.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/HardShrink.c" #else -static int nn_(HardShrink_forward)(lua_State *L) +static int nn_(HardShrink_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id)); @@ -16,7 +16,7 @@ static int nn_(HardShrink_forward)(lua_State *L) return 1; } -static int nn_(HardShrink_backward)(lua_State *L) +static int nn_(HardShrink_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -32,8 +32,8 @@ static int nn_(HardShrink_backward)(lua_State *L) } static const struct luaL_Reg nn_(HardShrink__) [] = { - {"HardShrink_forward", nn_(HardShrink_forward)}, - {"HardShrink_backward", nn_(HardShrink_backward)}, + {"HardShrink_updateOutput", nn_(HardShrink_updateOutput)}, + {"HardShrink_updateGradInput", nn_(HardShrink_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SparseCriterion.c b/generic/SparseCriterion.c index 76da569..f4ebb8e 100644 --- a/generic/SparseCriterion.c +++ b/generic/SparseCriterion.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SparseCriterion.c" #else -static int nn_(SparseCriterion_forward)(lua_State *L) +static int nn_(SparseCriterion_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage"); @@ -19,7 +19,7 @@ static int nn_(SparseCriterion_forward)(lua_State *L) return 1; } -static int nn_(SparseCriterion_backward)(lua_State *L) +static int nn_(SparseCriterion_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage"); @@ -34,8 +34,8 @@ static int nn_(SparseCriterion_backward)(lua_State *L) } static const struct luaL_Reg nn_(SparseCriterion__) [] = { - {"SparseCriterion_forward", nn_(SparseCriterion_forward)}, - {"SparseCriterion_backward", nn_(SparseCriterion_backward)}, + {"SparseCriterion_updateOutput", nn_(SparseCriterion_updateOutput)}, + {"SparseCriterion_updateGradInput", nn_(SparseCriterion_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialClassNLLCriterion.c b/generic/SpatialClassNLLCriterion.c index 517af80..ddbbe1f 100644 --- a/generic/SpatialClassNLLCriterion.c +++ b/generic/SpatialClassNLLCriterion.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SpatialClassNLLCriterion.c" #else -static int nn_(SpatialClassNLLCriterion_forward)(lua_State *L) +static int nn_(SpatialClassNLLCriterion_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *target = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -20,7 +20,7 @@ static int nn_(SpatialClassNLLCriterion_forward)(lua_State *L) return 1; } -static int nn_(SpatialClassNLLCriterion_backward)(lua_State *L) +static int nn_(SpatialClassNLLCriterion_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *target = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -39,8 +39,8 @@ static int nn_(SpatialClassNLLCriterion_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialClassNLLCriterion__) [] = { - {"SpatialClassNLLCriterion_forward", nn_(SpatialClassNLLCriterion_forward)}, - {"SpatialClassNLLCriterion_backward", nn_(SpatialClassNLLCriterion_backward)}, + {"SpatialClassNLLCriterion_updateOutput", nn_(SpatialClassNLLCriterion_updateOutput)}, + {"SpatialClassNLLCriterion_updateGradInput", nn_(SpatialClassNLLCriterion_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialGraph.c b/generic/SpatialGraph.c index 29d8bbf..a1db79a 100644 --- a/generic/SpatialGraph.c +++ b/generic/SpatialGraph.c @@ -7,7 +7,7 @@ #endif #define square(x) ((x)*(x)) -static int nn_(SpatialGraph_forward)(lua_State *L) +static int nn_(SpatialGraph_updateOutput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -109,7 +109,7 @@ static int nn_(SpatialGraph_forward)(lua_State *L) return 1; } -static int nn_(SpatialGraph_backward)(lua_State *L) +static int nn_(SpatialGraph_updateGradInput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -245,8 +245,8 @@ static int nn_(SpatialGraph_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialGraph__) [] = { - {"SpatialGraph_forward", nn_(SpatialGraph_forward)}, - {"SpatialGraph_backward", nn_(SpatialGraph_backward)}, + {"SpatialGraph_updateOutput", nn_(SpatialGraph_updateOutput)}, + {"SpatialGraph_updateGradInput", nn_(SpatialGraph_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialLinear.c b/generic/SpatialLinear.c index 903c3b8..524d313 100644 --- a/generic/SpatialLinear.c +++ b/generic/SpatialLinear.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SpatialLinear.c" #else -static int nn_(SpatialLinear_forward)(lua_State *L) +static int nn_(SpatialLinear_updateOutput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -43,7 +43,7 @@ static int nn_(SpatialLinear_forward)(lua_State *L) return 1; } -static int nn_(SpatialLinear_backward)(lua_State *L) +static int nn_(SpatialLinear_updateGradInput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -118,8 +118,8 @@ static int nn_(SpatialLinear_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialLinear__) [] = { - {"SpatialLinear_forward", nn_(SpatialLinear_forward)}, - {"SpatialLinear_backward", nn_(SpatialLinear_backward)}, + {"SpatialLinear_updateOutput", nn_(SpatialLinear_updateOutput)}, + {"SpatialLinear_updateGradInput", nn_(SpatialLinear_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialMSECriterion.c b/generic/SpatialMSECriterion.c index 8f59497..33fc2e3 100644 --- a/generic/SpatialMSECriterion.c +++ b/generic/SpatialMSECriterion.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SpatialMSECriterion.c" #else -static int nn_(SpatialMSECriterion_forward)(lua_State *L) +static int nn_(SpatialMSECriterion_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *target = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -14,7 +14,7 @@ static int nn_(SpatialMSECriterion_forward)(lua_State *L) return 1; } -static int nn_(SpatialMSECriterion_backward)(lua_State *L) +static int nn_(SpatialMSECriterion_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *target = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -40,8 +40,8 @@ static int nn_(SpatialMSECriterion_retarget)(lua_State *L) } static const struct luaL_Reg nn_(SpatialMSECriterion__) [] = { - {"SpatialMSECriterion_forward", nn_(SpatialMSECriterion_forward)}, - {"SpatialMSECriterion_backward", nn_(SpatialMSECriterion_backward)}, + {"SpatialMSECriterion_updateOutput", nn_(SpatialMSECriterion_updateOutput)}, + {"SpatialMSECriterion_updateGradInput", nn_(SpatialMSECriterion_updateGradInput)}, {"SpatialMSECriterion_retarget", nn_(SpatialMSECriterion_retarget)}, {NULL, NULL} }; diff --git a/generic/SpatialMaxSampling.c b/generic/SpatialMaxSampling.c index 5b795b3..fcdd54d 100644 --- a/generic/SpatialMaxSampling.c +++ b/generic/SpatialMaxSampling.c @@ -9,7 +9,7 @@ #define MIN(a,b) ( ((a)<(b)) ? (a) : (b) ) #endif -static int nn_(SpatialMaxSampling_forward)(lua_State *L) +static int nn_(SpatialMaxSampling_updateOutput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -99,7 +99,7 @@ static int nn_(SpatialMaxSampling_forward)(lua_State *L) return 1; } -static int nn_(SpatialMaxSampling_backward)(lua_State *L) +static int nn_(SpatialMaxSampling_updateGradInput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -165,8 +165,8 @@ static int nn_(SpatialMaxSampling_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialMaxSampling__) [] = { - {"SpatialMaxSampling_forward", nn_(SpatialMaxSampling_forward)}, - {"SpatialMaxSampling_backward", nn_(SpatialMaxSampling_backward)}, + {"SpatialMaxSampling_updateOutput", nn_(SpatialMaxSampling_updateOutput)}, + {"SpatialMaxSampling_updateGradInput", nn_(SpatialMaxSampling_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialReSampling.c b/generic/SpatialReSampling.c index 4312615..927add2 100644 --- a/generic/SpatialReSampling.c +++ b/generic/SpatialReSampling.c @@ -9,7 +9,7 @@ #define MIN(a,b) ( ((a)<(b)) ? (a) : (b) ) #endif -static int nn_(SpatialReSampling_forward)(lua_State *L) +static int nn_(SpatialReSampling_updateOutput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -85,7 +85,7 @@ static int nn_(SpatialReSampling_forward)(lua_State *L) return 1; } -static int nn_(SpatialReSampling_backward)(lua_State *L) +static int nn_(SpatialReSampling_updateGradInput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -166,8 +166,8 @@ static int nn_(SpatialReSampling_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialReSampling__) [] = { - {"SpatialReSampling_forward", nn_(SpatialReSampling_forward)}, - {"SpatialReSampling_backward", nn_(SpatialReSampling_backward)}, + {"SpatialReSampling_updateOutput", nn_(SpatialReSampling_updateOutput)}, + {"SpatialReSampling_updateGradInput", nn_(SpatialReSampling_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialSparseCriterion.c b/generic/SpatialSparseCriterion.c index 6b83476..3a3c540 100644 --- a/generic/SpatialSparseCriterion.c +++ b/generic/SpatialSparseCriterion.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SpatialSparseCriterion.c" #else -static int nn_(SpatialSparseCriterion_forward)(lua_State *L) +static int nn_(SpatialSparseCriterion_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage"); @@ -23,7 +23,7 @@ static int nn_(SpatialSparseCriterion_forward)(lua_State *L) return 0; } -static int nn_(SpatialSparseCriterion_backward)(lua_State *L) +static int nn_(SpatialSparseCriterion_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *gradInput = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -45,8 +45,8 @@ static int nn_(SpatialSparseCriterion_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialSparseCriterion__) [] = { - {"SpatialSparseCriterion_forward", nn_(SpatialSparseCriterion_forward)}, - {"SpatialSparseCriterion_backward", nn_(SpatialSparseCriterion_backward)}, + {"SpatialSparseCriterion_updateOutput", nn_(SpatialSparseCriterion_updateOutput)}, + {"SpatialSparseCriterion_updateGradInput", nn_(SpatialSparseCriterion_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/SpatialUpSampling.c b/generic/SpatialUpSampling.c index 8a21efe..ea9828c 100644 --- a/generic/SpatialUpSampling.c +++ b/generic/SpatialUpSampling.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/SpatialUpSampling.c" #else -static int nn_(SpatialUpSampling_forward)(lua_State *L) +static int nn_(SpatialUpSampling_updateOutput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -44,7 +44,7 @@ static int nn_(SpatialUpSampling_forward)(lua_State *L) return 1; } -static int nn_(SpatialUpSampling_backward)(lua_State *L) +static int nn_(SpatialUpSampling_updateGradInput)(lua_State *L) { // get all params THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); @@ -92,8 +92,8 @@ static int nn_(SpatialUpSampling_backward)(lua_State *L) } static const struct luaL_Reg nn_(SpatialUpSampling__) [] = { - {"SpatialUpSampling_forward", nn_(SpatialUpSampling_forward)}, - {"SpatialUpSampling_backward", nn_(SpatialUpSampling_backward)}, + {"SpatialUpSampling_updateOutput", nn_(SpatialUpSampling_updateOutput)}, + {"SpatialUpSampling_updateGradInput", nn_(SpatialUpSampling_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/Template.c b/generic/Template.c index 9d1f444..6a75b88 100644 --- a/generic/Template.c +++ b/generic/Template.c @@ -2,19 +2,19 @@ #define TH_GENERIC_FILE "generic/Template.c" #else -static int nn_(Template_forward)(lua_State *L) +static int nn_(Template_updateOutput)(lua_State *L) { } -static int nn_(Template_backward)(lua_State *L) +static int nn_(Template_updateGradInput)(lua_State *L) { } static const struct luaL_Reg nn_(Template__) [] = { - {"Template_forward", nn_(Template_forward)}, - {"Template_backward", nn_(Template_backward)}, + {"Template_updateOutput", nn_(Template_updateOutput)}, + {"Template_updateGradInput", nn_(Template_updateGradInput)}, {NULL, NULL} }; diff --git a/generic/Threshold.c b/generic/Threshold.c index f25fddd..760e842 100644 --- a/generic/Threshold.c +++ b/generic/Threshold.c @@ -2,7 +2,7 @@ #define TH_GENERIC_FILE "generic/Threshold.c" #else -static int nn_(Threshold_forward)(lua_State *L) +static int nn_(Threshold_updateOutput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); real val = luaT_getfieldchecknumber(L, 1, "val"); @@ -16,7 +16,7 @@ static int nn_(Threshold_forward)(lua_State *L) return 1; } -static int nn_(Threshold_backward)(lua_State *L) +static int nn_(Threshold_updateGradInput)(lua_State *L) { THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id)); THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id)); @@ -32,8 +32,8 @@ static int nn_(Threshold_backward)(lua_State *L) } static const struct luaL_Reg nn_(Threshold__) [] = { - {"Threshold_forward", nn_(Threshold_forward)}, - {"Threshold_backward", nn_(Threshold_backward)}, + {"Threshold_updateOutput", nn_(Threshold_updateOutput)}, + {"Threshold_updateGradInput", nn_(Threshold_updateGradInput)}, {NULL, NULL} }; @@ -54,28 +54,6 @@ torch.include('nnx', 'Probe.lua') -- OpenMP module: torch.include('nnx', 'OmpModule.lua') --- those packages are available in a beta branch of Torch7, --- and will soon disapear from here -if not nn.Abs then - -- pointwise modules: - torch.include('nnx', 'Abs.lua') - torch.include('nnx', 'Power.lua') - torch.include('nnx', 'Square.lua') - torch.include('nnx', 'Sqrt.lua') - torch.include('nnx', 'HardShrink.lua') - torch.include('nnx', 'Threshold.lua') - - -- table-based modules: - torch.include('nnx', 'CMulTable.lua') - torch.include('nnx', 'CAddTable.lua') - torch.include('nnx', 'CDivTable.lua') - torch.include('nnx', 'CSubTable.lua') - - -- reshapers: - torch.include('nnx', 'Narrow.lua') - torch.include('nnx', 'Replicate.lua') -end - -- pointwise modules: torch.include('nnx', 'Minus.lua') @@ -128,24 +106,6 @@ torch.include('nnx', 'DataSet.lua') torch.include('nnx', 'DataList.lua') torch.include('nnx', 'DataSetLabelMe.lua') --- torch helpers (should not be here): -function torch.save(filename, object, mode) - mode = mode or 'binary' - local file = torch.DiskFile(filename, 'w') - file[mode](file) - file:writeObject(object) - file:close() -end - -function torch.load(filename, mode) - mode = mode or 'binary' - local file = torch.DiskFile(filename, 'r') - file[mode](file) - local object = file:readObject() - file:close() - return object -end - -- nn helpers: function nnx.empty(module) if module.modules then |