binary Search in find_weight_scale

examples/animation.jl

@@ -6,7 +6,7 @@ df = CSV.File(pkgdir(WordCloud)*"/res/guxiang_frequency.txt", header=false)|> Da
 words = df[!, "Column2"]
 weights = df[!, "Column3"]
 
-wc = wordcloud(words, weights, density=0.7)
+wc = wordcloud(words, weights, density=0.65)
 gifdirectory = "guxiang_animation"
 #eval# try rm("guxiang_animation", force=true, recursive=true) catch end 
 generate_animation!(wc, 100, outputdir=gifdirectory)

examples/benchmark.jl

@@ -27,7 +27,7 @@ for (i,wc) in enumerate(wcs)
         println("\n", i-1, "==== ", j, "/", length(ts), " ", nameof(t))
         placement!(wc)
         @time e = @elapsed generate!(wc, trainer=t, retry=1)
-        push!(es[i],nameof(t)=>e)
+        push!(es[i], string(nameof(t)) * (getstate(wc)==:generate! ? " ✔" : " ✘")=>e)
     end
 end
 println("SUMMARY")

examples/highdensity.jl

@@ -6,7 +6,7 @@ using WordCloud
 #md#     mask = shape(box, 400, 300, 10),
 #md#     colors = :Dark2_3,
 #md#     angles = (0, 90),
-#md#     density = 0.75) |> generate!
+#md#     density = 0.7) |> generate!
 #md# paint(wc, "highdensity.png")
 #md# ```
 #md# But you may find that doesn't work. That is because there should be at least 1 pixel gap between two words, which is controlled by the `border` parameter (default 1) in `wordcloud`. While, when the picture is small, 1 pixel is expensive. So, that can be done as follows:
@@ -15,7 +15,7 @@ wc = wordcloud(
     mask = shape(box, 400*2, 300*2, 10*2),
     colors = :Dark2_3,
     angles = (0, 90),
-    density = 0.75) |> generate!
+    density = 0.7) |> generate!
 paint(wc, "highdensity.png", ratio=0.5)
 #md# 
 println("results are saved to highdensity.png")

src/strategy.jl

@@ -18,7 +18,7 @@ end
 function feelingoccupied(imgs, border=0, bgvalue=imgs[1][1])
     bs = boxoccupied.(imgs, border)
     os = occupied.(imgs, bgvalue)
-    s = (0.8 * sum(bs) .+ 0.2 * sum(os)) / 0.93 #兼顾饱满字体（华文琥珀）和清瘦字体（仿宋）
+    s = (0.8 * sum(bs) + 0.2 * sum(os)) / 0.93 #兼顾饱满字体（华文琥珀）和清瘦字体（仿宋）
     # sum(os) ≈ 2/3 sum(bs), 故除以0.93还原到sum(bs)的大小
     th = 10quantile(bs, 0.1)
     bigind = findall(x->x>th, bs)
@@ -53,7 +53,7 @@ function prepareword(word, fontsize, color, angle; backgroundcolor=(0,0,0,0), fo
         angle=angle, border=border, font=font, type=:both)
 end
 
-wordmask(img, bgcolor, border) = dilate(img.!=img[1], border) 
+wordmask(img, bgcolor, border) = dilate(img.!=img[1], border)
 #use `img[1]` instead of `convert(eltype(img), parsecolor(bgcolor))`
 #https://github.com/JuliaGraphics/Luxor.jl/issues/107
 
@@ -75,12 +75,17 @@ function find_weight_scale!(wc::WC; initialscale=0, density=0.3, maxiter=5, erro
     ground_size = wc.params[:groundoccupied]
     words = wc.words
     if initialscale <= 0
-        initialscale = √(ground_size/length(words)/0.4*density) #初始值假设字符的字面框面积占正方格比率为0.4（低估了汉字）
+        initialscale = √(ground_size/length(words)/0.45*density) #初始值假设字符的字面框面积占正方格比率为0.45（低估了汉字）
     end
     @assert sum(wc.weights.^2 .* length.(words)) / length(wc.weights) ≈ 1.0
+    target = density * ground_size
     target_lower = (density - error) * ground_size
     target_upper = (density + error) * ground_size
     target = density * ground_size
+    best_tar_H = Inf
+    best_tar_L = -Inf
+    best_scale_H = Inf
+    best_scale_L = -Inf
     step = 0
     sc1 = initialscale
     fonts = getfonts(wc, words)
@@ -89,24 +94,47 @@ function find_weight_scale!(wc::WC; initialscale=0, density=0.3, maxiter=5, erro
     while true
         step = step + 1
         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`."
+            @warn "find_weight_scale reach maxiter. This may be caused by too small background or too many words or too big `minfontsize`."
             break
         end
+        # cal tg1
+        @assert sc1 < 50initialscale #防止全空白words的输入，计算出sc过大渲染字体耗尽内存
         wc.params[:scale] = sc1
         tg1 = textoccupied(words, getfontsizes(wc, words), fonts)
         dens = tg1 / ground_size
         println("scale=$(wc.params[:scale]), density=$dens\t", dens>density ? "↑" : "↓")
+        if tg1 > target
+            if best_tar_H > tg1
+                best_tar_H = tg1
+                best_scale_H = sc1
+            end
+        else
+            if best_tar_L <= tg1
+                best_tar_L = tg1
+                best_scale_L = sc1
+            end
+        end
+        # cal sc2
         sc2 = scalestep(sc0, tg0, sc1, tg1, target)
+#         @show sc0, tg0, sc1, tg1, sc2
+        if !(best_scale_L < sc2 < best_scale_H)
+            if isfinite(best_tar_H + best_tar_L)
+                sc2_ = √((best_scale_H^2 + best_scale_L^2)/2.)
+                println("bisection takes effect: scale $sc2 -> $sc2_")
+                sc2 = sc2_
+#                 @show best_scale_L best_scale_H
+            else
+                Base.error("scalestep failed!")
+            end
+        end
+        # next iter init
         tg0 = tg1
         sc0 = sc1
         sc1 = sc2
-        @assert sc1 < 50initialscale #防止全空白words的输入，计算出sc过大渲染字体耗尽内存
         if target_lower <= tg1 <= target_upper
             break
         end
-
     end
-#     @show textoccupied(words, weights, sc, radius=border)
     wc.params[:scale] = sc1
     return sc1
 end