uses LIBSVM instead of LIBLINEAR

Author: sadit

Project.toml

@@ -6,7 +6,7 @@ version = "0.6.0"
 [deps]
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 InvertedFiles = "b20bd276-2c70-11ec-161a-3d1e1109a1c3"
-LIBLINEAR = "2d691ee1-e668-5016-a719-b2531b85e0f5"
+LIBSVM = "b1bec4e5-fd48-53fe-b0cb-9723c09d164b"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MLUtils = "f1d291b0-491e-4a28-83b9-f70985020b54"
 Parameters = "d96e819e-fc66-5662-9728-84c9c7592b0a"
@@ -20,7 +20,7 @@ TextSearch = "7f6f6c8a-3b03-11e9-223d-e7d88259bd6c"
 
 [compat]
 InvertedFiles = "0.8"
-LIBLINEAR = "0.7"
+LIBSVM = "0.8"
 MLUtils = "0.4"
 Parameters = "0.12"
 SearchModels = "0.4"

src/TextClassification.jl

@@ -9,7 +9,7 @@ using Parameters, InvertedFiles
 include("scores.jl")
 include("textconfigspace.jl")
 include("textmodelspace.jl")
-include("liblinearconfig.jl")
+include("libsvmconfig.jl")
 include("microtcconfig.jl")
 include("microtc.jl")
 include("utils.jl")

src/liblinearconfig.jl

@@ -0,0 +1,61 @@
+# This file is part of TextClassification.jl
+
+export LIBSVMConfig, LIBSVMConfigSpace
+
+using LIBSVM
+
+@with_kw struct LIBSVMConfig
+    C::Float64 = 1.0
+    weights = :balance
+end
+
+struct LIBSVMWrapper{LIBSVMModel}
+    dim::Int
+    cls::LIBSVMModel
+end
+
+function balanced_weights(y)
+    C = countmap(y)
+    s = sum(values(C))
+    nc = length(C)
+    Dict{Any,Float64}(label => (s / (nc * count)) for (label, count) in C)
+end
+
+function create(config::LIBSVMConfig, train_X, train_y, dim)
+    train_X_ = sparse(train_X, dim)
+    nt = Threads.nthreads()
+    verbose = true
+    kernel = Kernel.Linear
+    weights = balanced_weights(train_y)
+    cls = svmtrain(train_X_, train_y; nt, weights, verbose, kernel, cost=config.C, )
+    LIBSVMWrapper(dim, cls)
+end
+
+@with_kw struct LIBSVMConfigSpace <: AbstractSolutionSpace
+    C = [1.0]
+    eps = [0.1]
+    weights = [:balance, nothing]
+    scale_C = (lower=0.001, s=3.0, upper=1000.0)
+    scale_eps = (lower=0.0001, s=3.0, upper=0.99)
+end
+
+Base.eltype(::LIBSVMConfigSpace) = LIBSVMConfig
+
+function Base.rand(space::LIBSVMConfigSpace)
+    LIBSVMConfig(rand(space.C), rand(space.weights))
+end
+
+function combine(a::LIBSVMConfig, b::LIBSVMConfig)
+    LIBSVMConfig(a.C, rand([a.weights, b.weights]))
+end
+
+function mutate(space::LIBSVMConfigSpace, a::LIBSVMConfig, iter)
+    C = space.scale_C === nothing ? a.C : SearchModels.scale(a.C; space.scale_C...)
+    weights = rand([a.weights, rand(space.weights)])
+    LIBSVMConfig(C, weights)
+end
+
+function predict(w::LIBSVMWrapper, vec::SVEC)
+    ypred = svmpredict(w.cls, sparse([vec], w.dim); nt=Threads.nthreads())
+    ypred[1][1]
+end

src/libsvmconfig.jl

@@ -10,7 +10,7 @@ end
 function MicroTC_Config(;
         textconfig=TextConfig(),
         textmodel=EntModelConfig(),
-        cls=LiblinearConfig(1.0, 0.1)
+        cls=LIBSVMConfig(1.0, 0.1)
     )
 
     MicroTC_Config(textconfig, textmodel, cls)
@@ -38,7 +38,7 @@ Base.isequal(a::MicroTC_Config, b::MicroTC_Config) = repr(a) == repr(b)
 
 function MicroTC_ConfigSpace(;
         textmodel=[VectorModelConfigSpace(), EntModelConfigSpace()],
-        cls=[LiblinearConfigSpace()],
+        cls=[LIBSVMConfigSpace()],
         textconfig::TextConfigSpace = TextConfigSpace()
     )