New element-wise trainer

Project.toml

@@ -1,7 +1,7 @@
 name = "WordCloud"
 uuid = "6385f0a0-cb03-45b6-9089-4e0acc74b26b"
 authors = ["guoyongzhi <[email protected]>"]
-version = "0.2.6"
+version = "0.3.0"
 
 [deps]
 ColorSchemes = "35d6a980-a343-548e-a6ea-1d62b119f2f4"

README.md

@@ -2,16 +2,6 @@
 wordcloud in Julia
 ---
 
-* [x] 排序 & 预放置
-* [x] 基于四叉树碰撞检测
-* [x] 根据局部灰度梯度位置调整（训练迭代）
-* [x] 引入动量加速训练
-* [x] 分代调整以优化性能
-* [x] 控制字体大小和填充密度的策略
-* [x] 重新放置和缩放的策略
-* [x] 文字颜色和方向
-* [ ] 并行计算
-
 # Basic Usage 
 ```julia
 ]add WordCloud
@@ -30,7 +20,7 @@ wc = wordcloud(
     mask = loadmask("res/alice_mask.png", color="#faeef8"),
     colors = (WordCloud.colorschemes[:Set1_5].colors..., ),
     angles = (0, 90),
-    filling_rate = 0.6) |> generate!
+    fillingrate = 0.75) |> generate!
 paint(wc, "alice.png", ratio=0.5)
 ```
 ![alice](res/alice.png)
@@ -41,6 +31,16 @@ paint(wc, "alice.png", ratio=0.5)
 ![compare](res/compare.png)  
 [Comparison of Obama's and Trump's inaugural address](./examples/compare.jl)  
 
+***
+* [x] 排序 & 预放置
+* [x] 基于四叉树碰撞检测
+* [x] 根据局部灰度梯度位置调整（训练迭代）
+* [x] 引入动量加速训练
+* [x] 分代调整以优化性能
+* [x] 控制字体大小和填充密度的策略
+* [x] 重新放置和缩放的策略
+* [x] 文字颜色和方向
+* [ ] 并行计算
 ***
 linux添加中文字体  
 > mv wqy-microhei.ttc ~/.fonts  

examples/animation.jl

@@ -9,7 +9,7 @@ weights = df[!, "Column3"]
 wc = wordcloud(texts, weights, 
 #     colors = ("red", 0.7, "#00ff00"),
     mask=shape(box, 400, 300, 40, color=0.15), 
-    filling_rate=0.65)
+    fillingrate=0.8)
 
 gifdirectory = "guxiang_animation"
 generate_animation!(wc, 100, optimiser=Momentum(η=1/4, ρ=0.5), patient=10, retry=2, outputdir=gifdirectory)

examples/compare.jl

@@ -4,12 +4,12 @@ stwords = ["us", "will"];
 println("==Obama's==")
 cs = WordCloud.randomscheme()
 as = WordCloud.randomangles()
-fr = 0.6 #not too high
+fr = 0.65 #not too high
 wca = wordcloud(
     process(open("res/Barack Obama's First Inaugural Address.txt"), stopwords=WordCloud.stopwords_en ∪ stwords), 
     colors = cs,
     angles = as,
-    filling_rate = fr) |> generate!
+    fillingrate = fr) |> generate!
 println("==Trump's==")
 tb, wb = process(open("res/Donald Trump's Inaugural Address.txt"), stopwords=WordCloud.stopwords_en ∪ stwords)
 samemask = tb .∈ Ref(wca.words)
@@ -29,7 +29,7 @@ wcb = wordcloud(
     mask = wca.mask,
     colors = csb,
     angles = asb,
-    filling_rate = fr)
+    fillingrate = fr)
 for i in 1:length(tb)
     if samemask[i]
         ii = wainds[tb[i]]
@@ -62,4 +62,4 @@ WordCloud.ImageMagick.save("address_compare/compare.png", [ma space mb])
 gif = WordCloud.GIF("address_compare")
 record(wca, "Obama", gif)
 record(wcb, "Trump", gif)
-WordCloud.Render.generate(gif, framerate=0.5)
+WordCloud.Render.generate(gif, framerate=1)

examples/trainer_metric.jl

@@ -0,0 +1,28 @@
+using WordCloud
+using Random
+
+println("This test may take several minutes")
+
+words = [Random.randstring(rand(1:8)) for i in 1:200]
+weights = randexp(length(words)) .* 2000 .+ rand(20:100, length(words));
+wc1 = wordcloud(words, weights, mask=shape(ellipse, 500, 500, color=0.15), angles=(0,90,45), fillingrate=0.7)
+
+words = [Random.randstring(rand(1:8)) for i in 1:500]
+weights = randexp(length(words)) .* 2000 .+ rand(20:100, length(words));
+wc2 = wordcloud(words, weights, mask=shape(ellipse, 500, 500, color=0.15), angles=(0,90,45))
+
+words = [Random.randstring(rand(1:8)) for i in 1:5000]
+weights = randexp(length(words)) .* 2000 .+ rand(20:100, length(words));
+wc3 = wordcloud(words, weights, mask=shape(box, 2000, 2000, 100, color=0.15), angles=(0,90,45))
+
+wcs = [wc1, wc1, wc2, wc3] #repeat wc1 to trigger compiling
+ts = [WordCloud.trainepoch_E!,WordCloud.trainepoch_EM!,WordCloud.trainepoch_EM2!,
+        WordCloud.trainepoch_P!,WordCloud.trainepoch_P2!,WordCloud.trainepoch_level!]
+for (i,wc) in enumerate(wcs)
+    println("\n\n", "*"^10, "wordcloud - $(length(wc.words)) words on mask$(size(wc.mask))", "*"^10)
+    for (j,t) in enumerate(ts)
+        println("\n", i-1, "==== ", j, "/", length(ts), " ", nameof(t))
+        placement!(wc)
+        @time generate!(wc, trainer=t, retry=1)
+    end
+end

src/interface.jl

@@ -21,7 +21,7 @@ function randommask(color, sz=800)
 end
 function randomangles()
     a = rand((0, (0,90),(0,90,45),(0,-90),(0,-45,-90),-90:90))
-    println("angles: ", a)
+    println("angles = ", a)
     a
 end
 
@@ -80,7 +80,7 @@ end
 * angles = (0, 90, 45) #choose entries randomly  
 * angles = 0:180 #choose entries randomly  
 * angles = [0, 22, 4, 1, 100, 10, ......] #use entries sequentially in cycle  
-* filling_rate = 0.5  
+* fillingrate = 0.5  
 * border = 1  
 ### mask kargs
 * mask = loadmask("res/heart.jpg", 256, 256) #see doc of `loadmask`  
@@ -94,7 +94,7 @@ wordcloud(counter::AbstractVector{<:Pair}; kargs...) = wordcloud(first.(counter)
 
 function wordcloud(words::AbstractVector{<:AbstractString}, weights::AbstractVector{<:Real}; 
                 colors=randomscheme(), angles=randomangles(), font="",
-                filling_rate=0.5, border=1, minfontsize=:auto, kargs...)
+                fillingrate=0.65, border=1, minfontsize=:auto, kargs...)
 
     @assert length(words) == length(weights) > 0
 #     @show words,weights
@@ -147,11 +147,11 @@ function wordcloud(words::AbstractVector{<:AbstractString}, weights::AbstractVec
     end
     weights = weights ./ √(sum(weights.^2 .* length.(words)) / length(weights))
     params[:weights] = weights
-    scale = find_weight_scale(words, weights, groundoccupied, border=border, initial_scale=0, 
-    filling_rate=filling_rate, max_iter=5, error=0.03, minfontsize=minfontsize)
+    scale = find_weight_scale(words, weights, groundoccupied, border=border, initialscale=0, 
+    fillingrate=fillingrate, maxiter=5, error=0.03, minfontsize=minfontsize)
     params[:scale] = scale
-    params[:filling_rate] = filling_rate
-    println("set filling_rate to $filling_rate, with scale=$scale")
+    params[:fillingrate] = fillingrate
+    println("set fillingrate to $fillingrate, with scale=$scale")
     imgs, mimgs, qtrees = prepareforeground(words, weights * scale, colors, angles, groundsize, 
     bgcolor=(0, 0, 0, 0), border=border, font=font, minfontsize=minfontsize);
     params[:mimgs] = mimgs
@@ -195,8 +195,16 @@ function record(wc::wordcloud, label::AbstractString, gif_callback=x->x)
     gif_callback(resultpic)
 end
 
-function generate!(wc::wordcloud, nepoch::Number=100, args...; retry=3,
-    trainer=trainepoch_gen!, optimiser=Momentum(η=1/4, ρ=0.5), patient=10, krags...)
+"""
+# Positional Args
+* wc: the wordcloud to train
+* nepoch: training epoch nums
+# Keyword Args
+* retry: shrink & retrain times, default 3
+* patient: number of epochs before teleporting & number of identical teleportation before giving up
+* trainer: appoint a training engine
+"""
+function generate!(wc::wordcloud, args...; retry=3, krags...)
     ep, nc = -1, -1
     for r in 1:retry
         # fr = feelingoccupied(wc.params[:mimgs])/wc.params[:groundoccupied]
@@ -205,8 +213,7 @@ function generate!(wc::wordcloud, nepoch::Number=100, args...; retry=3,
             rescale!(wc, sc)
         end
         println("#$r. scale = $(wc.params[:scale])")
-        ep, nc = train_with_teleport!(wc.qtrees, wc.maskqtree, nepoch, args...; 
-            trainer=trainer, optimiser=optimiser, patient=patient, krags...)
+        ep, nc = train!(wc.qtrees, wc.maskqtree, args...; krags...)
         wc.params[:epoch] += ep
         if nc == 0
             break
@@ -215,15 +222,17 @@ function generate!(wc::wordcloud, nepoch::Number=100, args...; retry=3,
     @show ep, nc
     if nc == 0
         wc.params[:completed] = true
-    else
-        wc.params[:completed] = false
+    else #check
         colllist = first.(listcollision(wc.qtrees, wc.maskqtree))
         get_text(i) = i>0 ? wc.words[i] : "#MASK#"
         collwords = [(get_text(i), get_text(j)) for (i,j) in colllist]
         if length(colllist) > 0
+            wc.params[:completed] = false
             println("have $(length(colllist)) collision.",
-            " try setting a larger `nepoch` and `retry`, or lower `filling_rate` in `wordcloud` to fix that")
+            " try setting a larger `nepoch` and `retry`, or lower `fillingrate` in `wordcloud` to fix that")
             println("$collwords")
+        else
+            wc.params[:completed] = true
         end
     end
     wc

src/qtree.jl

@@ -223,7 +223,7 @@ callefttop(t::ShiftedQtree, center) = center .- kernelsize(t) .÷  2
 setcenter!(t::ShiftedQtree, center) = setshift!(t, 1, callefttop(t, center))
 
 function inbounds(bgqt::ShiftedQtree, qt::ShiftedQtree)
-    inbounds(bgqt[1], (getshift(qt[1]) .+ kernelsize(qt[1]) .÷ 2)...)
+    inbounds(bgqt[1], center(qt)...)
 end
 function outofbounds(bgqt::ShiftedQtree, qts)
     [i for (i,t) in enumerate(qts) if !inbounds(bgqt, t)]

src/qtreetools.jl

@@ -313,7 +313,8 @@ function locate!(qts::AbstractVector, loctree::QtreeNode=LocQtreeInt((levelnum(q
     end
     loctree
 end
-function locate!(qts::AbstractVector, inds::AbstractVector{Int}, loctree::QtreeNode=LocQtreeInt((levelnum(qts[1]), 1, 1))) #must have same levelnum
+function locate!(qts::AbstractVector, inds::Union{AbstractVector{Int}, AbstractSet{Int}}, 
+        loctree::QtreeNode=LocQtreeInt((levelnum(qts[1]), 1, 1))) #must have same levelnum
     for i in inds
         locate!(qts[i], loctree, label=i, newnode=LocQtreeInt)
     end
@@ -353,7 +354,7 @@ function listcollision_qtree(qtrees::AbstractVector, mask::AbstractStackedQtree;
     loctree = locate!(mask, loctree, label=0, newnode=LocQtreeInt)
     listcollision_qtree(qtrees, mask, loctree; kargs...)
 end
-function listcollision_qtree(qtrees::AbstractVector, mask::AbstractStackedQtree, inds::AbstractVector{Int}; kargs...)
+function listcollision_qtree(qtrees::AbstractVector, mask::AbstractStackedQtree, inds::Union{AbstractVector{Int}, AbstractSet{Int}}; kargs...)
     loctree = locate!(qtrees, inds)
     loctree = locate!(mask, loctree, label=0, newnode=LocQtreeInt)
     listcollision_qtree(qtrees, mask, loctree; kargs...)

src/strategy.jl

@@ -15,8 +15,8 @@ function box_occupied(imgs::AbstractVector)
     return sum(box_occupied, imgs)
 end
 function feelingoccupied(imgs)
-    m = length(imgs) ÷ 4
-    occupied(imgs[1:m]) + box_occupied(imgs[m+1:end]) #兼顾大字的内隙和小字的占据
+    m = length(imgs) ÷ 100
+    occupied(imgs[1:m])/4 + 3box_occupied(imgs[1:m])/4 + box_occupied(imgs[m+1:end]) #兼顾大字的内隙和小字的占据
 end
 
 function text_occupied(words, weights, scale; font="", border=0, minfontsize=0)
@@ -65,29 +65,29 @@ function prepareforeground(words, weights, colors, angles, groundsize; bgcolor=(
 end
 
 ## weight_scale
-function cal_weight_scale(words, weights, target; border=0, initial_scale=64, kargs...)
-    input = initial_scale
+function cal_weight_scale(words, weights, target; border=0, initialscale=64, kargs...)
+    input = initialscale
     output = text_occupied(words, weights, input; border=border, kargs...)
 #     @show input,output
     return output, sqrt(target/output) * (input+2border) - 2border# 假设output=k*(input+2border)^2
 end
 
-function find_weight_scale(words, weights, ground_size; initial_scale=0, filling_rate=0.3, max_iter=5, error=0.05, kargs...)
-    if initial_scale <= 0
-        initial_scale = √(ground_size/length(words))
+function find_weight_scale(words, weights, ground_size; initialscale=0, fillingrate=0.3, maxiter=5, error=0.05, kargs...)
+    if initialscale <= 0
+        initialscale = √(ground_size/length(words))
     end
     @assert sum(weights.^2 .* length.(words)) / length(weights) ≈ 1.0
-    target_lower = (filling_rate - error) * ground_size
-    target_upper = (filling_rate + error) * ground_size
+    target_lower = (fillingrate - error) * ground_size
+    target_upper = (fillingrate + error) * ground_size
     step = 0
-    sc = initial_scale
+    sc = initialscale
     while true
         step = step + 1
-        if step > max_iter
-            @warn "find_weight_scale reach max_iter. This may be caused by too small background image or too many words or too big `minfontsize`."
+        if step > maxiter
+            @warn "find_weight_scale reach maxiter. This may be caused by too small background image or too many words or too big `minfontsize`."
             break
         end
-        tg, sc = cal_weight_scale(words, weights, filling_rate * ground_size, initial_scale=sc; kargs...)
+        tg, sc = cal_weight_scale(words, weights, fillingrate * ground_size, initialscale=sc; kargs...)
         @show sc, tg, tg/ground_size
         if target_lower <= tg <= target_upper
             break
@@ -112,29 +112,6 @@ function max_collisional_index(qtrees, mask)
     nothing
 end
 
-# function max_collisional_index_rand(qtrees, mask)
-#     l = length(qtrees)
-#     b = l - floor(Int, l / 8 * randexp()) #从末尾1/8起
-#     getqtree(i) = i==0 ? mask : qtrees[i]
-#     for i in b:-1:1
-#         for j in 0:i-1
-#             cp = collision(getqtree(i), getqtree(j))
-#             if cp[1] >= 0
-#                 return i
-#             end
-#         end
-#     end
-#     for i in l:-1:b+1
-#         for j in 0:i-1
-#             cp = collision(getqtree(i), getqtree(j))
-#             if cp[1] >= 0
-#                 return i
-#             end
-#         end
-#     end
-#     nothing
-# end
-
 function max_collisional_index_rand(qtrees, mask; collpool)
     l = length(collpool)
     b = l - floor(Int, l / 8 * randexp()) #从末尾1/8起
@@ -156,7 +133,8 @@ function max_collisional_index_rand(qtrees, mask; collpool)
     end
     return nothing
 end
-function collisional_indexes_rand(qtrees, mask; collpool)
+
+function collisional_indexes_rand(qtrees, mask, collpool::Vector{Tuple{Int,Int}})
     cinds = Vector{Int}()
     l = length(collpool)
     if l == 0
@@ -180,7 +158,9 @@ function collisional_indexes_rand(qtrees, mask; collpool)
     end
     return cinds
 end
-
+function collisional_indexes_rand(qtrees, mask, collpool::Vector{QTree.ColItemType})
+    collisional_indexes_rand(qtrees, mask, first.(collpool))
+end
 function rescale!(wc::wordcloud, scale::Real)
     qts = wc.qtrees
     centers = QTree.center.(qts)