1 files changed, 43 insertions, 0 deletions
diff --git a/contrib/lua-torch/nn/HingeEmbeddingCriterion.lua b/contrib/lua-torch/nn/HingeEmbeddingCriterion.lua
new file mode 100644
index 000000000..13ad00f19
--- /dev/null
+++ b/contrib/lua-torch/nn/HingeEmbeddingCriterion.lua
@@ -0,0 +1,43 @@
+local HingeEmbeddingCriterion, parent = torch.class('nn.HingeEmbeddingCriterion', 'nn.Criterion')
+
+function HingeEmbeddingCriterion:__init(margin)
+   parent.__init(self)
+   self.margin = margin or 1
+   self.sizeAverage = true
+end
+
+function HingeEmbeddingCriterion:updateOutput(input,y)
+   self.buffer = self.buffer or input.new()
+   if not torch.isTensor(y) then
+      self.ty = self.ty or input.new():resize(1)
+      self.ty[1]=y
+      y=self.ty
+   end
+
+   self.buffer:resizeAs(input):copy(input)
+   self.buffer[torch.eq(y, -1)] = 0
+   self.output = self.buffer:sum()
+
+   self.buffer:fill(self.margin):add(-1, input)
+   self.buffer:cmax(0)
+   self.buffer[torch.eq(y, 1)] = 0
+   self.output = self.output + self.buffer:sum()
+
+   if (self.sizeAverage == nil or self.sizeAverage == true) then
+      self.output = self.output / input:nElement()
+   end
+
+   return self.output
+end
+
+function HingeEmbeddingCriterion:updateGradInput(input, y)
+   if not torch.isTensor(y) then self.ty[1]=y; y=self.ty end
+   self.gradInput:resizeAs(input):copy(y)
+   self.gradInput[torch.cmul(torch.eq(y, -1), torch.gt(input, self.margin))] = 0
+
+   if (self.sizeAverage == nil or self.sizeAverage == true) then
+      self.gradInput:mul(1 / input:nElement())
+   end
+
+   return self.gradInput
+end