|
| 1 | +import Random |
| 2 | +getRNG(seed::Integer = -1) = seed >= 0 ? Random.MersenneTwister(seed) : Random.GLOBAL_RNG |
| 3 | +getRNG(seed::Union{Random.MersenneTwister,Random._GLOBAL_RNG}) = seed |
| 4 | + |
| 5 | +import LightGraphs |
| 6 | +const LG = LightGraphs |
| 7 | + |
| 8 | +import SimpleWeightedGraphs |
| 9 | +const SWG = SimpleWeightedGraphs |
| 10 | + |
| 11 | +using GraphPlot, Colors |
| 12 | + |
| 13 | +function ising_dependency_graph( |
| 14 | + dims::Tuple{T,T}; |
| 15 | + periodic::Bool = false, |
| 16 | + rng = -1, |
| 17 | +)::SWG.SimpleWeightedGraph{Int64,Float64} where {T} |
| 18 | + rng = getRNG(rng) |
| 19 | + g = SWG.SimpleWeightedGraph(LG.grid(dims, periodic = periodic)) |
| 20 | + for e in LG.edges(g) |
| 21 | + i, j = Tuple(e) |
| 22 | + @inbounds g.weights[i, j] = g.weights[j, i] = rand(rng) |
| 23 | + end |
| 24 | + return g |
| 25 | +end |
| 26 | + |
| 27 | +""" |
| 28 | +Find bad |
| 29 | +U_ij > exp(- |J| (1 - sign(J) x_i x_j)) |
| 30 | +log(U_ij) > - |J| (1 - sign(J) x_i x_j) |
| 31 | +sign(J) x_i x_j < 0 & log(U_ij) > - 2 |J| |
| 32 | +""" |
| 33 | +function ising_find_bad_states( |
| 34 | + g::SWG.SimpleWeightedGraph{T, U}, |
| 35 | + states::Vector{T}, |
| 36 | + J::U; |
| 37 | + rng = -1, |
| 38 | +)::Set{T} where {T, U} |
| 39 | + |
| 40 | + @assert LG.nv(g) == length(states) |
| 41 | + |
| 42 | + rng = getRNG(rng) |
| 43 | + |
| 44 | + _2absJ, sign_J = -2.0 * abs(J), sign(J) |
| 45 | + |
| 46 | + bad = Set{T}() |
| 47 | + for e ∈ LG.edges(g) |
| 48 | + i, j = Tuple(e) |
| 49 | + # Break constraint: |
| 50 | + # log(U_ij) > - |J| (1 - sign(J) x_i x_j) |
| 51 | + # <=> sign(J) x_i x_j < 0 & log(U_ij) > - 2 |J| |
| 52 | + if ((sign_J * states[i] * states[j]) < 0.0) & (log(e.weight) > _2absJ) |
| 53 | + union!(bad, i, j) |
| 54 | + end |
| 55 | + end |
| 56 | + return bad |
| 57 | +end |
| 58 | + |
| 59 | +""" |
| 60 | +Find Res |
| 61 | +""" |
| 62 | +function ising_find_states_to_resample!( |
| 63 | + g::SWG.SimpleWeightedGraph{T, U}, |
| 64 | + states::Vector{T}, |
| 65 | + J::U; |
| 66 | + rng = -1 |
| 67 | +)::Set{T} where {T, U} |
| 68 | + |
| 69 | + rng = getRNG(rng) |
| 70 | + sign_J = sign(J) |
| 71 | + R = ising_find_bad_states(g, states, J, rng = rng) |
| 72 | + ∂R, ∂R_tmp = copy(R), Set{T}() |
| 73 | + while !isempty(∂R) |
| 74 | + for i ∈ ∂R |
| 75 | + for j ∈ LG.neighbors(g, i) |
| 76 | + # Break constraint: |
| 77 | + # log(U_ij) > - |J| (1 - sign(J) x_i x_j) |
| 78 | + # <=> sign(J) x_i x_j < 0 & log(U_ij) > - 2 |J| |
| 79 | + if j ∈ R |
| 80 | + if (sign_J * states[i] * states[j]) < 0.0 |
| 81 | + # U_ij can be increased |
| 82 | + @inbounds g.weights[i, j] = g.weights[j, i] = rand(rng) |
| 83 | + end |
| 84 | + else # if j ∉ R |
| 85 | + # x_j can be flipped to make sign(J) x_i x_j < 0 |
| 86 | + union!(R, j) |
| 87 | + union!(∂R_tmp, j) |
| 88 | + # followed by an increase of U_ij |
| 89 | + @inbounds g.weights[i, j] = g.weights[j, i] = rand(rng) |
| 90 | + end |
| 91 | + end |
| 92 | + end |
| 93 | + ∂R, ∂R_tmp = ∂R_tmp, Set{T}() |
| 94 | + end |
| 95 | + return R |
| 96 | +end |
| 97 | + |
| 98 | +""" |
| 99 | +Resample |
| 100 | +""" |
| 101 | +function ising_sample_states!( |
| 102 | + states::Vector{T}, |
| 103 | + res_ind::Set{T}, |
| 104 | + probas::Vector{Float64}; |
| 105 | + rng = -1, |
| 106 | +) where {T} |
| 107 | + n = length(states) |
| 108 | + @assert n == length(probas) |
| 109 | + rng = getRNG(rng) |
| 110 | + for i in res_ind |
| 111 | + states[i] = rand(rng) < probas[i] ? 1 : -1 |
| 112 | + end |
| 113 | +end |
| 114 | + |
| 115 | +sigmoid(x) = @. 1 / (1 + exp(-x)) |
| 116 | + |
| 117 | +function ising_prs( |
| 118 | + dims::Tuple{T, T}, |
| 119 | + h::Vector{U}, |
| 120 | + J::U; |
| 121 | + periodic::Bool=false, |
| 122 | + rng=-1, |
| 123 | +) where {T, U} |
| 124 | + rng = getRNG(rng) |
| 125 | + |
| 126 | + g = ising_dependency_graph(dims, periodic=periodic, rng=rng) |
| 127 | + |
| 128 | + n = LG.nv(g) |
| 129 | + states = Vector{Int64}(undef, n) |
| 130 | + probas = sigmoid.(2.0 .* h) |
| 131 | + res = Set{T}(1:LG.nv(g)) |
| 132 | + |
| 133 | + cnt = 0 |
| 134 | + while !isempty(res) |
| 135 | + ising_sample_states!(states, res, probas, rng=rng) |
| 136 | + res = ising_find_states_to_resample!(g, states, J, rng = rng) |
| 137 | + cnt += 1 |
| 138 | + end |
| 139 | + |
| 140 | + return g, states, cnt |
| 141 | +end |
| 142 | + |
| 143 | +function ising_prs( |
| 144 | + dims::Tuple{T, T}, |
| 145 | + h::U, |
| 146 | + J::U; |
| 147 | + periodic::Bool = false, |
| 148 | + rng = -1, |
| 149 | +) where {T, U} |
| 150 | + h_vec = fill(h, prod(dims)) |
| 151 | + return ising_prs(dims, h_vec, J, periodic=periodic, rng=rng) |
| 152 | +end |
| 153 | + |
| 154 | +function plot_ising(g, dims, state) |
| 155 | + pos = collect(Iterators.product(1:dims[1], 1:dims[2]))[:] |
| 156 | + locs_x, locs_y = map(x->x[1], pos), map(x->x[2], pos) |
| 157 | + |
| 158 | + col_nodes = ifelse.(state .== 1, colorant"gray", colorant"white") |
| 159 | + |
| 160 | + p = gplot(g, |
| 161 | + locs_x, reverse(locs_y), |
| 162 | + nodefillc=col_nodes, |
| 163 | + # nodelabel=LG.vertices(g), |
| 164 | + # arrowlengthfrac=0.05 |
| 165 | + # edgestrokec=col_edges |
| 166 | + ) |
| 167 | + display(p) |
| 168 | +end |
| 169 | + |
| 170 | + |
| 171 | +dims = (14, 14) # if > (14, 14) the display becomes all black, don't know why ! |
| 172 | +H, J = 0.0, -0.02 # Use Float |
| 173 | + |
| 174 | +periodic = false |
| 175 | +seed = -1 |
| 176 | +g, config, cnt = ising_prs(dims, H, J; periodic=periodic, rng=seed) |
| 177 | + |
| 178 | +plot_ising(g, dims, config) |
| 179 | + |
| 180 | +println("Number of resampling steps") |
| 181 | +println(cnt) |
0 commit comments