Code optimization and verification

1 files changed, 42 insertions, 27 deletions

diff --git a/QDRiemaNNLinear.lua b/QDRiemaNNLinear.lua
index 76befe6..8dc4793 100644
--- a/QDRiemaNNLinear.lua
+++ b/QDRiemaNNLinear.lua
@@ -1,48 +1,63 @@
 --
--- Author: Gaetan Marceau Caron (gaetan.marceau-caron@inria.fr)
+-- Author: Gaetan Marceau Caron (gaetan.marceau-caron@inria.fr) and Yann Ollivier
 -- Description: Implementation of the quasi-diagonal reduction
--- based on the Practical Riemannian Neural Networks (Yann Ollivier and Gaetan Marceau Caron) paper (http://arxiv.org/abs/1602.08007)
+-- based on the Practical Riemannian Neural Networks paper (http://arxiv.org/abs/1602.08007)
 -- 
 local QDRiemaNNLinear, parent = torch.class('nnx.QDRiemaNNLinear', 'nn.Linear')
 
 function QDRiemaNNLinear:__init(inputSize, outputSize, gamma, qdFlag)
    parent.__init(self,inputSize, outputSize)
-   self.qdFlag = qdFlag or true -- Flag for choosing between diagonal or quasi-diagonal reductions
+   if qdFlag == nil then -- Flag for choosing between diagonal or quasi-diagonal reductions
+	  self.qdFlag = true 
+   else
+	  self.qdFlag = qdFlag
+   end
    self.gamma = gamma  or 0.01 -- update rate of the metric
    self.matReg = 1e-12 -- numerical regularization 
    self.initMetric = true -- flag for first update
    self.Mii = torch.Tensor(outputSize, inputSize)
    if self.qdFlag then self.M0i = torch.Tensor(outputSize, inputSize) end
    self.M00 = torch.Tensor(outputSize)
+   self.accGradientFlag = true
+   self.accMetricFlag = true
+end
+
+function QDRiemaNNLinear:setAccFlag(accGradientFlag,accMetricFlag)
+   self.accGradientFlag = accGradientFlag
+   self.accMetricFlag = accMetricFlag
 end
 
 function QDRiemaNNLinear:accGradParameters(input, gradOutput)
-   parent.accGradParameters(self,input,gradOutput)
-   
-   gradOutputSqT = torch.pow(gradOutput,2):t()
-   
-   if self.initMetric then
-	  self.Mii:mm(gradOutputSqT,torch.pow(input,2))
-	  self.M00:mv(gradOutputSqT,self.addBuffer)
-	  if self.qdFlag then self.M0i:mm(gradOutputSqT,input) end
-	  self.initMetric = false
-   else
-	  self.Mii:mul(1.-self.gamma):addmm(self.gamma,gradOutputSqT,torch.pow(input,2))
-	  if self.qdFlag then self.M0i:mul(1.-self.gamma):addmm(self.gamma,gradOutputSqT,input) end
-	  self.M00:mul(1.-self.gamma):addmv(self.gamma,gradOutputSqT,self.addBuffer)
+   if self.accGradientFlag then
+	  parent.accGradParameters(self,input,gradOutput)
    end
-   
-   if self.qdFlag then
-	  local numerator = torch.add(torch.cmul(self.gradWeight,self.M00:view(-1,1):expandAs(self.gradWeight)), -1.0, torch.cmul(self.M0i,self.gradBias:view(-1,1):expandAs(self.M0i)))
-	  local denominator = torch.add(torch.cmul(self.Mii,self.M00:view(-1,1):expandAs(self.Mii)),-1.0,torch.pow(self.M0i,2)):clamp(self.matReg,1e25)
-	  self.gradWeight:copy(torch.cdiv(numerator,denominator))
+
+   if self.accMetricFlag then
+	  local gradOutputSqT = torch.pow(gradOutput,2):t()
 	  
-	  local temp = torch.cmul(torch.cdiv(self.M0i,self.M00:view(-1,1):expandAs(self.M0i)),self.gradWeight)
-	  self.gradBias:copy(torch.add(torch.cdiv(self.gradBias,self.M00),-1.0,torch.sum(temp,2)))
-   
-   else
-	  self.gradWeight:cdiv(self.Mii:add(self.matReg))
-	  self.gradBias:cdiv(self.M00:add(self.matReg))
+	  if self.initMetric then
+		 self.Mii:mm(gradOutputSqT,torch.pow(input,2))
+		 self.M00:mv(gradOutputSqT,self.addBuffer)
+		 if self.qdFlag then self.M0i:mm(gradOutputSqT,input) end
+		 self.initMetric = false
+	  else
+		 self.Mii:mul(1.-self.gamma):addmm(self.gamma,gradOutputSqT,torch.pow(input,2))
+		 if self.qdFlag then self.M0i:mul(1.-self.gamma):addmm(self.gamma,gradOutputSqT,input) end
+		 self.M00:mul(1.-self.gamma):addmv(self.gamma,gradOutputSqT,self.addBuffer)
+	  end
+	  
+	  if self.qdFlag then
+		 local numerator = torch.add(torch.cmul(self.gradWeight,self.M00:view(-1,1):expandAs(self.gradWeight)), -1.0, torch.cmul(self.M0i,self.gradBias:view(-1,1):expandAs(self.M0i)))
+		 local denominator = torch.add(torch.cmul(self.Mii,self.M00:view(-1,1):expandAs(self.Mii)),-1.0,torch.pow(self.M0i,2)):clamp(self.matReg,1e25)
+		 self.gradWeight:copy(numerator:cdiv(denominator))
+
+		 local temp = torch.cmul(self.M0i,self.gradWeight):sum(2)
+		 self.gradBias:add(-1.,temp):cdiv(torch.add(self.M00,self.matReg))
+
+	  else
+		 self.gradWeight:cdiv(self.Mii:add(self.matReg))
+		 self.gradBias:cdiv(self.M00:add(self.matReg))
+	  end
    end
 end