\uad6c\ud604\uc18c\uc2a4(Julia Notebook \u2013 html \ubc84\uc804) <\/a><\/p>\n\n\n\n\uc774 \uc608\uc81c\uc5d0\uc11c\ub294 GP\ub97c \ud1b5\ud574 f(x1,x2) = 0.9 x1^2 + 0.7 x1<\/em> x2 + 0.3 x2 \uc758 \uadfc\uc0ac\uc2dd\uc744 \ucc3e\uace0 \ud559\uc2b5 \ub418\uc9c0 \uc54a\uc740 \ub370\uc774\ud130 \ub300\ud55c
prediction\uc744 \ud569\ub2c8\ub2e4.<\/p>\n\n\n\nGP2.jl \uc18c\uc2a4<\/strong><\/p>\n\n\n\nmodule GP2\nusing Distributed \nusing Random\nusing ExprOptimization\nusing SharedArrays\n\nmutable struct DGrammar\n grammar::Array{String,1}\n function DGrammar()\n grammar = Array{String,1}(undef,0)\n push!(grammar,\"@grammar begin\")\n push!(grammar,\"end\")\n new(grammar)\n end\nend\n#=\nexamples:\ngrmmar = DGrammar()\ngrmmar(\"R = R + R\")\ngrmmar(\"R = R - R\")\n=#\nfunction (g::DGrammar)(r::String)\n insert!(g.grammar,2,r)\nend\n#=\nReturn Grammar\ngrmmar()\n=#\nfunction (g::DGrammar)()\n eval(Meta.parse(join(g.grammar,\"\\n\")))\nend\n\nmutable struct Vars\n v::Vector{String}\nfunction Vars()\n v = Vector{String}(undef,0)\n new(v)\n end\nend\nfunction (vs::Vars)(s::String)\n push!(vs.v,s)\nend\nfunction (vs::Vars)()\n vars = \"([\"*join(vs.v,\",\")*\"])\"\nend\n\nfunction train(gp::GeneticProgram, grammar::Grammar, \n train_x::Array{Float64,2},train_y::Array{Float64,1},nJob::Int64=1)\n #=\n nJob\uc774 1 \ubcf4\ub2e4 \ud06c\uba74 \ubcd1\ub82c\ucc98\ub9ac\n =#\n if nJob > 1\n pids = addprocs(nJob)\n module_dir = \"\/home\/shpark\/julia_test\/exproptimizatoin\/examples\"\n # everywhere \uc0ac\uc6a9\uc2dc \uac01 process\uc5d0 \uac12\uc744 \ub118\uaca8 \uc8fc\ub824\uba74 interpolation\uc73c\ub85c \ub118\uaca8 \uc918\uc57c \ud55c\ub2e4.\n @everywhere push!(LOAD_PATH, $module_dir) \n eval(macroexpand(Distributed,quote @everywhere using GP2 end))\n nr,nc = size(train_x) \n shared_train_x = SharedArray{Float64}((nr,nc),pids=pids)\n shared_train_y = SharedArray{Float64}((nr,1),pids=pids)\n shared_train_x[:,:] = train_x\n shared_train_y[:,:] = train_y\n \n rps = [@spawnat i GP2._train_parallel(gp,grammar,\n shared_train_x,shared_train_y) \n for i in pids]\n ms = map(x->fetch(rps[x]),1:length(rps))\n # nJob\uac1c\uc758 \ud504\ub85c\uc138\uc2a4\uc5d0\uc11c \uac00\uc838\uc628 \uacb0\uacfc\uc911 loss\uac00 \uac00\uc7a5 \uc791\uc740\n # \uac1c\uccb4\ub97c \uace8\ub77c \ub0b8\ub2e4.\n # m\uc758 \ud0c0\uc785 \ubc0f \ub0b4\ubd80 fields\ub97c \uc54c\ub824\uba74\n # type\ud655\uc778 : typeof(m)\n # fields \ud655\uc778 : fieldnames(typeof(m))\n r = ms[argmin([m.loss for m in ms ])]\n rmprocs(pids) \n else\n r = _train(gp,grammar,train_x,train_y)\n end\n return r\nend\n\n\nfunction _train(gp::GeneticProgram, grammar::Grammar, \n train_x::Array{Float64,2},train_y::Array{Float64,1})\n \n S = SymbolTable(grammar)\n \n function loss(tree::RuleNode, grammar::Grammar)\n ex = get_executable(tree,grammar)\n nr,nc = size(train_x)\n \u03a3\u0394y = 0 \n for r in 1:nr\n for c in 1:nc\n x = train_x[r,c] \n S[Symbol(\"x\",c)] = x \n end\n y_fit = Core.eval(S,ex) \n y = train_y[r]\n \u03a3\u0394y += (y - y_fit)^2\n end \n los = sqrt(\u03a3\u0394y)\/nr\n end \n results_gp = optimize(gp, grammar, :R, loss,verbose=true)\n return results_gp \nend\n \nfunction _train_parallel(gp::GeneticProgram, grammar::Grammar, \n train_x::SharedArray,train_y::SharedArray)\n \n S = SymbolTable(grammar)\n \n function loss(tree::RuleNode, grammar::Grammar)\n ex = get_executable(tree,grammar)\n nr,nc = size(train_x)\n \u03a3\u0394y = 0 \n for r in 1:nr\n for c in 1:nc\n x = train_x[r,c] \n S[Symbol(\"x\",c)] = x \n end\n y_fit = Core.eval(S,ex) \n y = train_y[r]\n \u03a3\u0394y += (y - y_fit)^2\n end \n los = sqrt(\u03a3\u0394y)\/nr\n end \n results_gp = optimize(gp, grammar, :R, loss,verbose=true)\n return results_gp\nend\n\nfunction predict(results_gp::ExprOptResult,grammar::Grammar,test_x::Array{Float64,2})\n #=\n Prediction \n =#\n S = SymbolTable(grammar)\n ex = results_gp.expr\n nr,nc = size(test_x)\n y = Array{Float64,1}(undef,0)\n for r in 1:nr\n for c in 1:nc\n x = test_x[r,c]\n S[Symbol(\"x\",c)] = x\n end\n push!(y, Core.eval(S,ex))\n end\n y\nend \n\nend # module GP2<\/code><\/pre>\n\n\n\nWS.jl \uc18c\uc2a4<\/strong><\/p>\n\n\n\nmodule WS\nusing Random\nusing DataFrames\nusing CSV\n\nmutable struct WindowSlider\n w::Int # window_size - number of time steps to look back\n o::Int # offset between last reading and temperature\n r::Int # response_size - number of time steps to predict\n l::Int # maximum length to slide - (#observation - w)\n p::Int # final predictors - (#predictors * w)\n names::Array{Symbol,1}\n function WindowSlider(;window_size::Int=5)\n w = window_size\n o = 0\n r = 1\n l = 0\n p = 0\n names = Array{Symbol,1}(undef,0)\n new(w,o,r,l,p,names)\n end\nend\n\nfunction re_init(arr::Array{Float64,1})\n cumsum!(arr,arr;dims=1)\n return arr .- arr[1]\nend\n\nfunction collect_windows(ws::WindowSlider,X::DataFrame;window_size::Int=5, offset::Int=0,previous_y = false)\n N, cols = size(X)\n cols -= 1\n ws.o = offset\n ws.w = window_size\n ws.l = N - (ws.w + ws.r) + 1\n ws.p = previous_y ? (cols + 1) * ws.w : cols * ws.w\n \n # create the names of the variables in the window\n # Check first if we need to create that for the response itself\n x = previous_y ? deepcopy(X) : select(X,Not(names(X)[end]))\n \n for col in names(x)\n for i in 1:ws.w\n #= python\uc5d0\uc11c\ub294 \u0394t(1)\uc744 \uceec\ub7fc \uc774\ub984\uc73c\ub85c \uc0ac\uc6a9\ud558\uace0 \uc788\uc73c\ub098\n \uc904\ub9ac\uc544\uc5d0\uc11c\ub294 \uad04\ud638\ub4f1 \ud2b9\uc218 \ubb38\uc790\ub294 Symbol\ub85c \ud5c8\uc6a9\ub418\uc9c0 \uc54a\ub294\ub2e4\n \uc904\ub9ac\uc544\uc758 DataFrame\uc758 \uceec\ub7fc\uc740 symbol\uc774 \uc0ac\uc6a9\ub418\uace0 \ucc38\uc870\uc2dc\uc5d0\ub3c4 \n symbol\uc774 \uc0ac\uc6a9\ub41c\ub2e4.\n \uc608\ub97c \ub4e4\uc5b4 df.\u0394t_1 \uc740 \ud558\uc6a9\ub418\ub098 df.\u0394t(1)\uc740 \ud5c8\uc6a9\ub418\uc9c0 \uc54a\ub294\ub2e4.\n =#\n # t(1),..,t(w1),x1(1),...,x1(w)\n name = Symbol(col,\"_\",i)\n push!(ws.names,name)\n end\n end \n \n # Incorporate the timestamps where we want to predict\n # t(w),..,t(w+r-1)\n for k in 1:ws.r\n name = Symbol(\"\u0394t\",\"_\",ws.w + k)\n push!(ws.names, name)\n end\n push!(ws.names, :Y)\n \n df = DataFrame(zeros(ws.l,ws.p + ws.r + 1), ws.names) \n \n # Populate by rows in the new dataframe\n for i in 1:ws.l \n # 2\ucc28\uc6d0 \ubc30\uc5f4 0 \ud06c\uae30\uc758 cols\uac1c\uc758 \uceec\ub7fc\uc744 \uac00\uc9c4 \ubc30\uc5f4\ub85c \ucd08\uae30\ud654\n slices = Array{Float64,1}(undef,0)\n # Flatten the lags of predictions\n _, cols = size(x)\n # Julia\uc5d0\uc11c a[1:3] \uc774\uba74 1 \ubd80\ud130 3 \uae4c\uc9c0 index\uc758 3\uac1c\uc758 \uac12\uc744 \uac00\uc838\uc628\ub2e4.\n # Python\uc5d0\uc11c\ub294 1 \ubd80\ud130 2\uae4c\uc9c0 2\uac1c\uc758 \uac12\uc744 \uac00\uc838\uc628\ub2e4.\n for p in 1:cols\n line = X[i:ws.w + i - 1,p]\n # Reinitialization at every window for \u0394t\n if p == 1 \n line = re_init(line) \n end\n \n # Concatenate the lines in one slice\n push!(slices, line...)\n end\n \n # Incorporate the timestamps where we want to predict\n # predict \ub77c\uc778\uc758 \uc989 \uc708\ub3c4\uc6b0\uc5d0\uc11c w+r \ubc88\uc9f8\uc758 \n line = re_init(X[i:ws.w + ws.r + i - 1, 1])[end]\n y = X[ws.w + ws.r + i - 1, end]\n push!(slices,line..., y...)\n \n # Incorporate the slice to the cake(df)\n df[i,:] .= slices\n end\n return df\nend\n\n\nfunction _data_set(;N::Int=600,window_size::Int=5, seed::Int=123)\n Random.seed!(seed)\n N = N\n #t = Float64[0:N-1...]\n t = Float64[0:N-1;]\n t = t .+ rand(N) \/ 4\n t = t .- rand(N) \/ 7\n t = round.(t; digits=2)\n x1 = round.(rand(N)*5;digits=2)\n x2 = round.(rand(N)*5;digits=2)\n x3 = round.(rand(N)*5;digits=2)\n n = round.(rand(N)*2;digits=2)\n\n x1_1 = circshift(x1,1) # t-1\n x2_1 = circshift(x2,1) # t-1\n x3_3 = circshift(x3,3) # t-3 \n\n y = @. log(abs(2 + x1))-x2_1^2 + 0.02*x3_3*exp(x1_1)\n y = @. round(y+n,digits=2) \n \u0394t = t[2:end] .- t[1:end-1]\n insert!(\u0394t,1,0)\n df = DataFrame([t \u0394t x1 x2 x3 y],[:t,:\u0394t,:x1,:x2,:x3,:y])\n return df\nend\n\nstruct Dataset\n t::Array{Float64,1}\n n_train::UnitRange{Int64}\n n_test::UnitRange{Int64}\n train::DataFrame\n test::DataFrame\n train_windows::DataFrame\n test_windows::DataFrame\n train_windows_y_inc::DataFrame\n test_windows_y_inc::DataFrame\nend\n\nfunction data_set(;N::Int=600,window_size::Int=5, seed::Int=2018,\n n_train=1:160,n_test=161:200)\n \n df = _data_set(N=N,window_size=window_size,seed=seed)\n t = df.t\n train = df[n_train,2:end]\n test = df[n_test,2:end] \n train_windows = collect_windows(WindowSlider(),train,previous_y=false) \n test_windows = collect_windows(WindowSlider(),test,previous_y=false)\n train_windows_y_inc = collect_windows(WindowSlider(),train,previous_y=true) \n test_windows_y_inc = collect_windows(WindowSlider(),test,previous_y=true)\n return Dataset(t,n_train, n_test,\n train,test,\n train_windows,test_windows,\n train_windows_y_inc,test_windows_y_inc)\nend\n\nfunction data_kospi(;N::Int=600,window_size::Int=5, seed::Int=2018,\n n_train=1:160,n_test=161:200)\n df = CSV.read(\"kospi.csv\") \n df2 = df[n_train[1]:n_test[end],:]\n select!(df2, Not([:close,:date]))\n # \ub2e4\uc74c\ub0a0 \uc885\uac00\n# next_day_close = df[n_train[2]:n_test[end]+1,end]\n# _,nc = size(df2)\n# insertcols!(df2,nc+1,:next_day_close=>next_day_close)\n #=\n \uac01 \uceec\ub7fc \ub370\uc774\ud130\ub97c \uc815\uaddc\ud654 \ud55c\ub2e4.\n \ub2e8 train data\uc758 max\uac12\uc73c\ub85c train \ubc0f test \ub370\uc774\ud130\ub97c \uc815\uaddc\ud654 \ud55c\ub2e4.\n \uc989 train\uc2dc \uc801\uc6a9\ub418\uc5c8\ub358 \uc870\uac74\uc744 test data\uc5d0 \uc801\uc6a9\ud55c\ub2e4.\n =# \n for i in 1:size(df2)[2]\n c_max = max(df2[n_train,i]...)\n df2[!,i] = df2[!,i]\/c_max \n end \n t = Float64[(n_train[1]:n_test[end])...]\n \u0394t=ones(n_test[end])\n insertcols!(df2,1,:t=>t)\n insertcols!(df2,2,:\u0394t=>\u0394t)\n\n train= df2[n_train,2:end]\n test= df2[n_test,2:end] \n train_windows= collect_windows(WindowSlider() ,train,previous_y=false,window_size=window_size) \n test_windows= collect_windows(WindowSlider(),test,previous_y=false,window_size=window_size)\n train_windows_y_inc= collect_windows(WindowSlider(),train,previous_y=true,window_size=window_size) \n test_windows_y_inc= collect_windows(WindowSlider(),test,previous_y=true,window_size=window_size)\n return Dataset(t,n_train, n_test,\n train,test,\n train_windows,test_windows,\n train_windows_y_inc,test_windows_y_inc)\nend\n\nend # module WS<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"\ud574\ub2f9 \uc81c\ubaa9\uc5d0 \ub300\ud55c \ub0b4\uc6a9\uc740 \uae30\uc874\uc5d0 \uc62c\ub9b0 “Julia\uc5d0\uc11c \uc2dc\uacc4\uc5f4 \ub370\uc774\ud130 \ucc98\ub9ac \uc608\uc81c”\ub4e4\uc744 \ucc38\uc870 \ud558\uba74\ub429\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 \ub2e4\uc911 \ud504\ub85c\uc138\uc2a4\ub85c Genetic Programming(GP)\uc744 \ubcd1\ub82c\ucc98\ub9ac \ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec \uc90d\ub2c8\ub2e4. \ub2e4\uc911 \ud504\ub85c\uc138\uc2a4\uac00 train, test\ud560 \ub370\uc774\ud130\ub294 SharedArray\ub85c \uacf5\uc720 \ud569\ub2c8\ub2e4. \uc544\ub798 \ubaa8\ub4c8\uc740 “\/home\/shpark\/julia_test\/exproptimizatoin\/examples” \ud3f4\ub354 \uc788\ub2e4\uace0 \uac00\uc815WS.jl : \uc5f0\uc2b5\uc6a9 \ub370\uc774\ud130\ub97c \ub9cc\ub4dc\ub294 \ubaa8\ub4c8 \/ \uc774 \uac8c\uc2dc\uae00\uc5d0\uc11c\ub294 WindowSlider \uad6c\uc870\uccb4\ub9cc \uc0ac\uc6a9\ud568GP2.jl: Genetic Programming \uc801\uc6a9 \ubaa8\ub4c8\uc544\ub798\uc5d0 \uc18c\uc2a4\ub97c \uac8c\uc81c \ud569\ub2c8\ub2e4.\uadf8\ub9ac\uace0 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"default","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"default","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[18,19,20,29,22],"tags":[],"_links":{"self":[{"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/posts\/2350"}],"collection":[{"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/julialang.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2350"}],"version-history":[{"count":5,"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/posts\/2350\/revisions"}],"predecessor-version":[{"id":2357,"href":"https:\/\/julialang.kr\/index.php?rest_route=\/wp\/v2\/posts\/2350\/revisions\/2357"}],"wp:attachment":[{"href":"https:\/\/julialang.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2350"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/julialang.kr\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2350"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/julialang.kr\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2350"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}