swtich to GLMakie

benchmarking/Project.toml

@@ -1,23 +1,25 @@
 [deps]
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+Colors = "5ae59095-9a9b-59fe-a467-6f913c188581"
 DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
 Flux = "587475ba-b771-5e3f-ad9e-33799f191a9c"
+GLMakie = "e9467ef8-e4e7-5192-8a1a-b1aee30e663a"
 MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
 Metalhead = "dbeba491-748d-5e0e-a39e-b530a07fa0cc"
 Optimisers = "3bd65402-5787-11e9-1adc-39752487f4e2"
 ProgressMeter = "92933f4c-e287-5a05-a399-4b506db050ca"
 TimerOutputs = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
-UnicodePlots = "b8865327-cd53-5732-bb35-84acbb429228"
 cuDNN = "02a925ec-e4fe-4b08-9a7e-0d78e3d38ccd"
 
 [compat]
+Colors = "0.12"
 CUDA = "5"
 DataFrames = "1"
 Flux = "0.14"
+GLMakie = "0.20"
 MLDatasets = "0.7"
 Metalhead = "0.9"
 Optimisers = "0.3"
 ProgressMeter = "1.9"
 TimerOutputs = "0.5"
-UnicodePlots = "3.6"
 cuDNN = "1.2"

benchmarking/benchmark.jl

@@ -1,14 +1,15 @@
 
+using Colors
 using CUDA, cuDNN
 using DataFrames
 using Flux
 using Flux: logitcrossentropy, onecold, onehotbatch
+using GLMakie
 using Metalhead
 using MLDatasets
 using Optimisers
 using ProgressMeter
 using TimerOutputs
-using UnicodePlots
 
 include("tooling.jl")
 
@@ -64,7 +65,10 @@ modelstrings = (
     # "UNet(; $common)" # doesn't support kwargs "inchannels", "nclasses"
     )
 df = DataFrame(; model=String[], train_loss=Float64[], train_acc=Float64[], test_loss=Float64[], test_acc=Float64[])
-plt = lineplot([0], [0], title="test accuracy vs. time (s)", ylim = (0,1), xlim=(0,60))
+cols = distinguishable_colors(length(modelstrings), [RGB(1,1,1), RGB(0,0,0)], dropseed=true)
+f = Figure()
+ax = Axis(f[1, 1], title="CIFAR-10 Training on a Nvidia 3090, batch 1000\nTest accuracy vs. time over 45 epochs", xlabel="Time (s)", ylabel="Testset Accuracy")
+display(f)
 max_x = 0
 for (i, modstring) in enumerate(modelstrings)
     @timeit to "$modstring" begin
@@ -86,14 +90,15 @@ for (i, modstring) in enumerate(modelstrings)
         else
             elapsed, train_loss_hist, train_acc_hist, test_loss_hist, test_acc_hist = ret
             max_x = max(maximum(elapsed), max_x)
-            lineplot!(plt, elapsed, test_acc_hist, name=modstring)
+            lines!(ax, elapsed, test_acc_hist, label=modstring, color=cols[i])
             elapsed, train_loss_hist[end], train_acc_hist[end], test_loss_hist[end], test_acc_hist[end]
         end
         push!(df, (modstring, train_loss, train_acc, test_loss, test_acc), promote=true)
     end
     GC.gc(true)
 end
-display(plt)
+f[1, 2] = Legend(f, ax, "Models", framevisible = false)
+display(f)
 display(df)
 print_timer(to; sortby = :firstexec)
 end

benchmarking/tooling.jl

@@ -49,7 +49,6 @@ function _train(model, train_loader, test_loader; epochs = 45,
     train_loss_hist, train_acc_hist = Float64[], Float64[]
     test_loss_hist, test_acc_hist = Float64[], Float64[]
     elapsed = Float64[]
-    start = time()
     @showprogress "training" for epoch in 1:epochs
         i = 0
         for (x, y) in train_loader
@@ -72,7 +71,7 @@ function _train(model, train_loader, test_loader; epochs = 45,
         @timeit to "testing" test_loss, test_acc = loss_and_accuracy(test_loader, model, device; limit)
         push!(train_loss_hist, train_loss); push!(train_acc_hist, train_acc);
         push!(test_loss_hist, test_loss); push!(test_acc_hist, test_acc);
-        push!(elapsed, time() - start)
+        push!(elapsed, time())
         if show_plots
             plt2 = lineplot(1:epoch, train_loss_hist, name = "train_loss", xlabel="epoch", ylabel="loss")
             lineplot!(plt2, 1:epoch, test_loss_hist, name = "test_loss")
@@ -85,6 +84,7 @@ function _train(model, train_loader, test_loader; epochs = 45,
             GC.gc() # GPU will OOM without this
         end
     end
+    elapsed = elapsed .- elapsed[1]
     train_loss, train_acc, test_loss, test_acc = train_loss_hist[end], train_acc_hist[end], test_loss_hist[end], test_acc_hist[end]
     @info "results after $epochs epochs $(repr(map(x->round(x, digits=3), (; train_loss, train_acc, test_loss, test_acc))))"
     return elapsed, train_loss_hist, train_acc_hist, test_loss_hist, test_acc_hist