MultiModalMarginRankingCriterion.lua

MultiModalMarginRankingCriterion.lua

@@ -0,0 +1,132 @@
+local MultiModalMarginRankingCriterion, parent = 
+    torch.class('nn.MultiModalMarginRankingCriterion', 'nn.Criterion')
+
+-- loss(x, y) = max(0, -y * (x[1] - x[2]) + margin) + 
+--              max(0, -y * (x[1] - x[3]) + margin)
+
+function MultiModalMarginRankingCriterion:__init(margin)
+   parent.__init(self)
+   margin=margin or 1
+   self.margin = margin
+   self.gradInput = {torch.Tensor(1), torch.Tensor(1), torch.Tensor(1)}
+   self.sizeAverage = true
+end
+
+function MultiModalMarginRankingCriterion:updateOutput(input, y)
+    if torch.type(y) == 'number' then -- non-batch mode
+      --self.output = math.max(0, -y * (input[1][1] - input[2][1]) + self.margin)
+      self.output = math.max(0, -y * (input[1][1] - input[2][1]) + self.margin)
+                + math.max(0, -y * (input[1][1] - input[3][1]) + self.margin)
+   else
+
+      self._output1 = self._output1 or input[1]:clone()
+      self._output1:resizeAs(input[1])
+      self._output1:copy(input[1])
+      self._output2 = self._output1:clone()
+
+-- max(0, -y * (x[1] - x[2]) + margin) + 
+      self._output1:add(-1, input[2])
+      self._output1:mul(-1):cmul(y)
+      self._output1:add(self.margin)
+
+      self._output1:cmax(0)
+
+-- max(0, -y * (x[1] - x[3]) + margin)
+      self._output2:add(-1, input[3])
+      self._output2:mul(-1):cmul(y)
+      self._output2:add(self.margin)
+
+      self._output2:cmax(0)
+
+
+      self.output = self._output1:sum() + self._output2:sum()
+
+      if self.sizeAverage then
+         self.output = self.output/y:size(1)
+      end
+   end
+
+   return self.output
+end
+
+
+
+
+-- loss(x, y) = max(0, -y * (x[1] - x[2]) + margin) + 
+--              max(0, -y * (x[1] - x[3]) + margin)
+
+function MultiModalMarginRankingCriterion:updateGradInput(input, y)
+    if torch.type(y) == 'number' then -- non-batch mode
+      --local dist = -y * (input[1][1] - input[2][1]) + self.margin
+      local dist1 = -y * (input[1][1] - input[2][1]) + self.margin
+      local dist2 = -y * (input[1][1] - input[3][1]) + self.margin
+
+      if dist1 < 0 then
+         self.gradInput[1][1] = 0;
+         self.gradInput[2][1] = 0;
+      else
+         self.gradInput[1][1] = -y
+         self.gradInput[2][1] = y
+      end
+
+      if dist2 < 0 then
+         self.gradInput[1][1] = 0 + self.gradInput[1][1];
+         self.gradInput[3][1] = 0;
+      else
+         self.gradInput[1][1] = -y + self.gradInput[1][1];
+         self.gradInput[3][1] = y
+      end
+
+   else
+      self.dist1 = self.dist1 or input[1].new()
+      self.dist1 = self.dist1:resizeAs(input[1]):copy(input[1])
+      local dist1 = self.dist1
+      self.dist2 = self.dist1:clone()
+      local dist2 = self.dist2
+
+      dist1:add(-1, input[2])
+      dist1:mul(-1):cmul(y)
+      dist1:add(self.margin)
+
+      dist2:add(-1, input[3])
+      dist2:mul(-1):cmul(y)
+      dist2:add(self.margin)
+
+      self.mask1 = self.mask1 or input[1].new()
+      self.mask1 = self.mask1:resizeAs(input[1]):copy(dist1)
+      local mask1 = self.mask1
+
+      self.mask2 = self.mask2 or input[1].new()
+      self.mask2 = self.mask2:resizeAs(input[1]):copy(dist2)
+      local mask2 = self.mask2
+
+      mask1:ge(dist1, 0)
+      mask2:ge(dist2, 0)
+
+      self.gradInput[1]:resize(dist1:size())
+      self.gradInput[2]:resize(dist1:size())
+      self.gradInput[3]:resize(dist2:size())
+
+      self.gradInput[1]:copy(mask1)
+      self.gradInput[1]:mul(-1):cmul(y)
+      self.gradInput[2]:copy(mask1)
+      self.gradInput[2]:cmul(y)
+
+      local gradInput_ = self.gradInput[1]:clone()
+      gradInput_:copy(mask2)
+      gradInput_:mul(-1):cmul(y)
+      self.gradInput[1]:add(gradInput_)
+      self.gradInput[3]:copy(mask2)
+      self.gradInput[3]:cmul(y)
+
+
+      if self.sizeAverage then
+         self.gradInput[1]:div(y:size(1))
+         self.gradInput[2]:div(y:size(1))
+         self.gradInput[3]:div(y:size(1))
+      end
+
+   end
+   return self.gradInput
+end
+