Merge pull request #8 from Evizero/master

Add squared hinge loss for L2-SVM

README.md

@@ -33,6 +33,7 @@ This package provides basic components for implementing regularized empirical ri
   - [x] quantile loss
   - [x] huber loss
   - [x] hinge loss
+  - [x] squared hinge loss
   - [x] smoothed hinge loss
   - [x] logistic loss
   - [x] sum squared loss (for multivariate prediction)

src/EmpiricalRisks.jl

@@ -23,6 +23,7 @@ export
     QuantileLoss,
     HuberLoss,
     HingeLoss,
+    SqrHingeLoss,
     SmoothedHingeLoss,
     LogisticLoss,
     SumLoss,

src/loss.jl

@@ -102,7 +102,7 @@ end
 
 
 
-## Hinge loss (for SVM)
+## Hinge loss (for L1-SVM)
 #
 #   loss(p, y) := max(1 - y * p, 0)
 #
@@ -118,6 +118,26 @@ function value_and_deriv{T<:BlasReal}(::HingeLoss, p::T, y::T)
 end
 
 
+## Squared Hinge loss (for L2-SVM)
+#
+#   loss(p, y) := max(1 - y * p, 0)^2
+#
+immutable SqrHingeLoss <: UnivariateLoss
+end
+
+function value{T<:BlasReal}(::SqrHingeLoss, p::T, y::T)
+    yp = y * p
+    yp >= one(T) ? zero(T) : half(abs2(nonneg(one(T) - yp)))
+end
+
+deriv{T<:BlasReal}(::SqrHingeLoss, p::T, y::T) = y * p < one(T) ? (p - y) : zero(T)
+
+function value_and_deriv{T<:BlasReal}(::SqrHingeLoss, p::T, y::T)
+    yp = y * p
+    yp >= one(T) ? (zero(T), zero(T)) : (half(abs2(one(T) - yp)), (p - y))
+end
+
+
 ## Smoothed HingeLoss
 #
 #   loss(p, y) := 0              ... y * p > 1 + h

test/uniloss.jl

@@ -75,6 +75,12 @@ _hingef(u::Dual, y) = y * real(u) < 1.0 ? 1.0 - y * u : dual(0.0, 0.0)
 verify_uniloss(HingeLoss(), _hingef, -2.0:0.5:2.0, [-1.0, 1.0])
 
 
+# SquaredHingeLoss
+
+_sqrhingef(u::Dual, y) = y * real(u) < 1.0 ? .5(1.0 - y * u).^2 : dual(0.0, 0.0)
+verify_uniloss(SqrHingeLoss(), _sqrhingef, -2.0:0.5:2.0, [-1.0, 1.0])
+
+
 # SmoothedHingeLoss
 
 function _sm_hingef(h::Float64, u::Dual, y)