Check dimension of gram matrix (#87)

* Implement dimension check for prediction gram matrix

* Add tests for gram matrix dimension check

* Disallow providing only support vector related entries

* Adjust error message

* Minor style changes

Resolves #85

src/LIBSVM.jl

@@ -106,7 +106,13 @@ function SVM(smc::SVMModel, y, X, weights, labels, svmtype, kernel)
         unsafe_copyto!(pointer(libsvmweight_label), smc.param.weight_label, nw)
     end
 
-    SVM(svmtype, kernel, weights, size(X,1),
+    if kernel == Kernel.Precomputed
+        nfeatures = size(X, 2)
+    else
+        nfeatures = size(X, 1)
+    end
+
+    SVM(svmtype, kernel, weights, nfeatures,
         smc.nr_class, labels, libsvmlabel, libsvmweight, libsvmweight_label,
         svs, smc.param.coef0, coefs, probA, probB,
         rho, smc.param.degree,
@@ -373,7 +379,9 @@ function svmpredict(model::SVM{T}, X::AbstractMatrix{U}; nt::Integer = 0) where
     set_num_threads(nt)
 
     if model.kernel != Kernel.Precomputed && size(X, 1) != model.nfeatures
-        throw(DimensionMismatch("Model has $(model.nfeatures) but $(size(X, 1)) provided"))
+        throw(DimensionMismatch("Model has $(model.nfeatures) features but $(size(X, 1)) provided"))
+    elseif model.kernel == Kernel.Precomputed && size(X, 1) != model.nfeatures
+            throw(DimensionMismatch("Gram matrix should have $(model.nfeatures) but $(size(X, 1)) provided"))
     end
 
     ninstances = size(X, 2)

test/runtests.jl

@@ -271,6 +271,25 @@ end
         @test model.coefs ≈ model₂.coefs
         @test model.SVs.indices ≈ model₂.SVs.indices
     end
+
+    @testset "Malformed prediction" begin
+        X = [-2 -1 -1 1 1 2;
+            -1 -1 -2 1 2 1]
+        y = [1, 1, 1, 2, 2, 2]
+
+        T = [-1 2 3;
+            -1 2 2]
+
+        K = X' * X
+
+        model = svmtrain(K, y, kernel=Kernel.Precomputed)
+
+        KK = X' * T
+
+        KK_malformed = KK[1:1,:]
+
+        @test_throws DimensionMismatch svmpredict(model, KK_malformed)
+    end
 end
 
 end  # @testset "LIBSVM"