Merge pull request #157 from szagoruyko/multimargin

L2 MultiMarginCriterion

4 files changed, 20 insertions, 6 deletions

diff --git a/MultiMarginCriterion.lua b/MultiMarginCriterion.lua
index e8de9d9..e308870 100644
--- a/MultiMarginCriterion.lua
+++ b/MultiMarginCriterion.lua
@@ -1,11 +1,15 @@
 local MultiMarginCriterion, parent = torch.class('nn.MultiMarginCriterion', 'nn.Criterion')
 
-function MultiMarginCriterion:__init()
+function MultiMarginCriterion:__init(p)
+   assert(p == nil or p == 1 or p == 2, 'only p=1 and p=2 supported')
+   self.p = p or 1
    parent.__init(self)
    self.sizeAverage = true
 end
 
 function MultiMarginCriterion:updateOutput(input, target)
+   -- backward compatibility
+   self.p = self.p or 1
    return input.nn.MultiMarginCriterion_updateOutput(self, input, target)
 end
 
diff --git a/doc/criterion.md b/doc/criterion.md
index 586dba1..2fb7477 100755
--- a/doc/criterion.md
+++ b/doc/criterion.md
@@ -220,13 +220,13 @@ i.e. the mlp successfully separates the two data points such that they both have
 ## MultiMarginCriterion ##
 
 ```lua
-criterion = MultiMarginCriterion()
+criterion = MultiMarginCriterion(p)
 ```
 
 Creates a criterion that optimizes a multi-class classification hinge loss (margin-based loss) between input `x`  (a Tensor of dimension 1) and output `y` (which is a target class index, 1 <= y <= x:size(1)) :
 
 ```lua
-loss(x,y) = forward(x,y) = sum_i(max(0, 1 - (x[y] - x[i]))) / x:size(1)
+loss(x,y) = forward(x,y) = sum_i(max(0, 1 - (x[y] - x[i]))^p) / x:size(1)
 ```
 where i = 1 to x:size(1) and i ~= y
 
diff --git a/generic/MultiMarginCriterion.c b/generic/MultiMarginCriterion.c
index 33f6e94..df7fc25 100644
--- a/generic/MultiMarginCriterion.c
+++ b/generic/MultiMarginCriterion.c
@@ -6,6 +6,7 @@ static int nn_(MultiMarginCriterion_updateOutput)(lua_State *L)
 {
   THTensor *input = luaT_checkudata(L, 2, torch_Tensor);  
   int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage");
+  int p = luaT_getfieldchecknumber(L, 1, "p");
   real *input_data, *target_data;
   long nframe, dim;
   long t, d;
@@ -54,7 +55,7 @@ static int nn_(MultiMarginCriterion_updateOutput)(lua_State *L)
         continue;
     
       if(z > 0)
-        sum += z;
+        sum += (p==1) ? z : z*z;
     }
     input_data += dim;
   }
@@ -75,6 +76,7 @@ static int nn_(MultiMarginCriterion_updateGradInput)(lua_State *L)
 {
   THTensor *input = luaT_checkudata(L, 2, torch_Tensor);
   int sizeAverage = luaT_getfieldcheckboolean(L, 1, "sizeAverage");
+  int p = luaT_getfieldchecknumber(L, 1, "p");
   THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_Tensor);
   real *input_data;
   real *gradInput_data;
@@ -127,8 +129,9 @@ static int nn_(MultiMarginCriterion_updateGradInput)(lua_State *L)
     
       if(z > 0)
       {
-        gradInput_target -= g;
-        gradInput_data[d] = g;
+        real h = (p == 1) ? g : 2*g*z;
+        gradInput_target -= h;
+        gradInput_data[d] = h;
       }
       else
         gradInput_data[d] = 0;
diff --git a/test.lua b/test.lua
index f555b50..a8f81f5 100644
--- a/test.lua
+++ b/test.lua
@@ -693,6 +693,13 @@ function nntest.MarginCriterion()
    criterionJacobianTest1D(cri, input, target)
 end
 
+function nntest.MultiMarginCriterion()
+   local input = torch.rand(100)
+   local target = math.random(1,100)
+   local cri = nn.MultiMarginCriterion(math.random(1,2))
+   criterionJacobianTest1D(cri, input, target)
+end
+
 function nntest.WeightedMSECriterion()
    local input = torch.rand(10)
    local target = input:clone():add(torch.rand(10))