catch up commit before starting work

quietlight

Mar 1, 2024, 7:47 PM

4BTZNCRM7R6SZPE5W7UENKS2LNXXRWVXGGZAM7QOWDIPEGW3LQ6AC

Dependencies

• [2] E3Y55MPR added perceptual colour maps and changed get_image_from_sample
• [3] NV7FXZ5Q first commit
• [4] RBXUHIO2 made somn small changes and reverted them mostly
• [5] ROFI4OLA catch up, working on colour images
• [6] 3FAESP6N catch up with changes, mainly construct primary dataset
• [7] NMQCXLNG catch up commit, about to do some work on prediction loop
• [8] U46LDPL7 added model, CUDA works again now on ubuntu
• [9] EDYR5C55 fixed ConstructPrimaryDataSet, licence date, Readme, re-arranged Predict.jl
• [10] 2UBDFCJH new files tracked

Change contents

• replacement in src/Train.jl at line 28
  [2.122]→[2.122:174](∅→∅)
  − train("K1-7_R50", 20, glob_pattern, true, 0.90, 32)
  [2.122]

  [3.0]
  + train("K1-8_Colour", 20, glob_pattern, true, 0.90, 64)
• replacement in src/Train.jl at line 233
  [3.342176]→[2.629:773](∅→∅)
  − fst = Metalhead.ResNet(50, pretrain = pretrain).layers

  − lst = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(2048 => classes))
  [3.342176]

  [3.342319]
  + fst = Metalhead.ResNet(18, pretrain = pretrain).layers

  + lst = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(512 => classes))
• replacement in src/Train.jl at line 245
  [3.342795]→[2.774:917](∅→∅)
  − f = Metalhead.ResNet(50, pretrain = false).layers

  − l = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(2048 => model_classes))
  [3.342795]

  [3.342937]
  + f = Metalhead.ResNet(18, pretrain = false).layers

  + l = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(512 => model_classes))
• replacement in src/Train.jl at line 253
  [3.343093]→[2.918:1007](∅→∅)
  − lst = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(2048 => classes))
  [3.343093]

  [3.343181]
  + lst = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(512 => classes))
• replacement in src/Predict.jl at line 33
  [3.22]→[2.1065:1249](∅→∅)
  − model = "/media/david/SSD2/PrimaryDataset/model_K1-7_CPU_epoch-17-0.9126-2024-02-04.jld2"

  − model = "/media/david/SSD2/PrimaryDataset/model_K1-7_R50_CPU_epoch-20-0.9136-2024-02-07.jld2"
  [3.22]

  [3.367291]
  + model = "/media/david/SSD2/PrimaryDataset/model_K1-8_Colour_CPU_epoch-14-0.9129-2024-02-16.jld2"
• replacement in src/Predict.jl at line 61
  [3.462]→[2.1388:1531](∅→∅)
  − f = Metalhead.ResNet(50, pretrain = false).layers

  − l = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(2048 => model_classes))
  [3.462]

  [3.604]
  + f = Metalhead.ResNet(18, pretrain = false).layers

  + l = Flux.Chain(AdaptiveMeanPool((1, 1)), Flux.flatten, Dense(512 => model_classes))
• edit in src/Predict.jl at line 208
  [3.373189]

  [2.1633]
  + #= not needed
• replacement in src/Predict.jl at line 213
  [2.1756]→[2.1756:1807](∅→∅)
  − x -> collect(channelview(float32.(x))) |>
  [2.1756]

  [2.1807]
  + # x -> collect(channelview(float32.(x))) |>
• edit in src/Predict.jl at line 218
  [2.1889]

  [2.1889]
  + =#
• replacement in src/Predict.jl at line 232
  [3.4269]→[3.4269:4349](∅→∅)
  − raw_images = ThreadsX.map(x -> get_image_for_inference(x, f), split_signal)
  [3.4269]

  [3.4349]
  + raw_images = ThreadsX.map(x -> get_image_from_sample(x, f), split_signal)
• replacement in src/Clips.jl at line 209
  [2.2054]→[3.19231:19291](∅→∅),[3.19231]→[3.19231:19291](∅→∅)
  − x -> RGB.(x) |>

  − x -> imresize(x, 224, 224)
  [2.2054]

  [3.19291]
  + #x -> RGB.(x) |>

  + x -> imresize(x, 224, 224) |>

  + x -> Float32.(x)