NNEFを読めるようになろう

学習済みのニューラルネットワークをエクスポートするためのファイル形式NNEFを解説します
ついでにNNEFの公式にサンプルとして転がっているVGG16を動かします
これは2023年7月22日に行われた Kernel/VM探検隊@東京 No16 での発表資料です

カーネル/VM探検隊について : https://www.kernelvm.org/
カーネル/VM探検隊東京 No16について : https://kernelvm.connpass.com/event/287261/
この発表の動画 : https://www.youtube.com/watch?v=pYNu5e6vMVg
この発表のソースコード : https://github.com/Fadis/gct

Transcript

1. ΛಡΊΔΑ͏ʹͳΖ͏
   ͸ Khronos Group Inc. ͷొ࿥঎ඪͰ͢
   NAOMASA MATSUBAYASHI(@fadis_)
   

   View Slide

2. x0
   x1
   x2
   y0
   y1
   y2
   χϡʔϥϧωοτϫʔΫ
   

   View Slide

3. w01
   w02
   w03
   w04
   ×
   ×
   ×
   ×
   ∑
   ׆
   ੑ
   Խ
   ؔ
   ਺
   ॏΈ ܗࣜχϡʔϩϯ
   ೖྗ0
   ೖྗ1
   ೖྗ2
   ೖྗ3
   ⋯
   ⋮
   ग़ྗ
   

   View Slide

4. x0
   x1
   x2
   

   View Slide

5. x0
   x1
   x2
   ೖ
   ྗ
   ૚
   શ݁߹૚ શ݁߹૚ શ݁߹૚
   

   View Slide

6. x0
   x1
   x2
   

   View Slide

7. Կͱͳ͘૬ؔ͸͋Γͦ͏ͳΜ͚ͩͲ
   Ͳ͏͍͏ؔ܎͔Α͘Θ͔Βͳ͍σʔλ
   ?
   2
   7
   4
   खॻ͖จࣈͷը૾ ॻ͔Ε͍ͯΔ਺ࣈ
   

   View Slide

8. ͱΓ͋͑ͣΑ͘Θ͔Βͳ͍ؔ܎Λ
   χϡʔϥϧωοτϫʔΫʹ͢Δ
   2
   7
   4
   w
   શ
   ݁
   ߹
   ૚
   શ
   ݁
   ߹
   ૚
   શ
   ݁
   ߹
   ૚
   ೖྗ૚
   w w
   

   View Slide

9. ਖ਼͍͠ग़ྗ͕ग़ͯ͘ΔΑ͏ͳ Λݟ͚ͭΔ
   ਺ཧ࠷దԽ໰୊ʹͳΔ
   w
   2
   7
   4
   w
   શ
   ݁
   ߹
   ૚
   શ
   ݁
   ߹
   ૚
   શ
   ݁
   ߹
   ૚
   ೖྗ૚
   w w ֶश
   ͜Ε͕
   ग़ͯ͘ΔΑ͏ʹ
   ͜ΕΛௐ੔
   

   View Slide

10. 4
    w
    ૚ ૚ ૚
    w w
    ͜Ε͕
    ग़ͯ͘ΔΑ͏ʹ
    ͜ΕΛௐ੔
    ֶशΛߦ͏ͨΊͷϑϨʔϜϫʔΫ
    TensorFlow
    PyTorch
    Caffe
    

    View Slide

11. TensorFlow
    ΑΜ 7
    ≠
    ͜ͷঢ়گͰ7͕ग़ΔΑ͏ʹ
    
    
    ॏΈΛमਖ਼
    w w w
    ͜Ε͸ԿͰ͔͢
    ΍Γ௚͠
    ΞϓϦέʔγϣϯ
    ͳͳ
    w w w
    ͜Ε͸ԿͰ͔͢
    ਖ਼͍͠ग़ྗΛు͚ΔΑ͏ʹͳͬͨΒ
    
    
    ΞϓϦέʔγϣϯʹ૊ΈࠐΜͰ࢖͍͍ͨ
    

    View Slide

12. ϑϨʔϜϫʔΫͷ਺͚ͩอଘܗ͕ࣜ͋Δ
    TensorFlow
    PyTorch
    Caffe
    w
    w
    tf.saved_modelܗࣜ
    w
    w
    pickleܗࣜ
    w
    w
    caffemodelܗࣜ
    

    View Slide

13. ΞϓϦέʔγϣϯ͔Βར༻͢Δͷ͕ਏ͍
    w
    w
    tf.saved_modelܗࣜ
    w
    w
    pickleܗࣜ
    w
    w
    caffemodelܗࣜ
    ΞϓϦέʔγϣϯ
    

    View Slide

14. Adobe Photoshop
    PSDܗࣜ
    GIMP
    XCFܗࣜ
    PNGܗࣜ
    ΞϓϦέʔγϣϯ
    ͜͏͍͏ͷແ͍ͷ?
    ม׵
    ม׵
    JPEGܗࣜ
    

    View Slide

15. w
    w
    tf.saved_modelܗࣜ
    TensorFlow
    w
    w
    caffemodelܗࣜ
    Caffe
    ม׵
    ม׵
    w
    w
    ΞϓϦέʔγϣϯ
    ͋Δ
    

    View Slide

16. ೖ
    ྗ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    શ
    ݁
    ߹
    ૚
    શ
    ݁
    ߹
    ૚
    શ
    ݁
    ߹
    ૚
    ֶश݁ՌΛ࠶ݱ͢Δҝʹอଘ͓͔ͯ͠ͳ͚Ε͹ͳΒͳ͍෺
    ͲͷΑ͏ͳ૚͕ͲͷΑ͏ͳॱͰܨ͕͍ͬͯΔ͔
    3x3ͷϑΟϧλͰ
    
    
    3νϟωϧ͔Β64νϟωϧʹม׵
    
    
    ύσΟϯά͸֤ล1ͮͭ
    
    
    dilationͱstride͸ͦΕͧΕ1
    ֤ͦͯ͠૚ͷઃఆ
    

    View Slide

17. ೖ
    ྗ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    ৞
    Έ
    ࠐ
    Έ
    ૚
    maxpool
    ૚
    શ
    ݁
    ߹
    ૚
    શ
    ݁
    ߹
    ૚
    શ
    ݁
    ߹
    ૚
    ֶश݁ՌΛ࠶ݱ͢Δҝʹอଘ͓͔ͯ͠ͳ͚Ε͹ͳΒͳ͍෺
    ֤૚ͷॏΈ
    w w w w w w w w w w w w w w w w
    3x3x3x64ͷ4֊ͷςϯιϧ
    
    
    ֤ཁૉͷ஋͸32bitුಈখ਺఺਺Ͱه࿥
    ͦͯ͠ॏΈ͕ͲͷΑ͏ͳܕͰ
    
    
    ϑΝΠϧʹॻ͔Ε͍ͯΔ͔
    

    View Slide

18. w
    w w
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ੔਺ͷ
    ςϯιϧ
    ੔਺Λ
    ුಈখ਺఺਺ʹ
    ม׵͢Δํ๏
    ࢀর
    ࢀর
    min=-2.0
    max=2.0
    bit=8
    ࢀর
    ωοτϫʔΫͷܗͱઃఆ
    

    View Slide

19. w
    w w
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ੔਺ͷ
    ςϯιϧ
    ੔਺Λ
    ුಈখ਺఺਺ʹ
    ม׵͢Δํ๏
    ࢀর
    ࢀর
    min=-2.0
    max=2.0
    bit=8
    ࢀর
    ωοτϫʔΫͷܗͱઃఆ
    ਓ͕ؒ
    
    
    ಡΈॻ͖Ͱ͖Δ
    
    
    ςΩετܗࣜ
    ૉૣ͘ύʔεͰ͖Δ
    
    
    όΠφϦܗࣜ
    

    View Slide

20. w
    w w
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ੔਺ͷ
    ςϯιϧ
    ੔਺Λ
    ුಈখ਺఺਺ʹ
    ม׵͢Δํ๏
    ࢀর
    ࢀর
    min=-2.0
    max=2.0
    bit=8
    ࢀর
    ωοτϫʔΫͷܗͱઃఆ
    ϔομ
    ਺஋ͷྻ
    128όΠτݻఆ
    ϑΝΠϧΛύʔε͢Δલʹ
    σʔλͷసૹΛ࢝ΊΒΕΔ
    

    View Slide

21. w
    w w
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ුಈখ਺఺਺ͷ
    ςϯιϧ
    ੔਺ͷ
    ςϯιϧ
    ੔਺Λ
    ුಈখ਺఺਺ʹ
    ม׵͢Δํ๏
    ࢀর
    ࢀর
    min=-2.0
    max=2.0
    bit=8
    ࢀর
    ωοτϫʔΫͷܗͱઃఆ
    version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    }
    graph.nnef
    NNEF 1.0ͷ࢓༷ʹैͬͯॻ͔Ε͍ͯ·͢
    ͜ͷωοτϫʔΫͷ໊લ
    

    View Slide

22. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    }
    graph.nnef
    ωοτϫʔΫʹର͢Δೖྗ஋͸
    
    
    ͜ͷൣғͷதͰఆٛ͞ΕΔ
    
    
    dataͱ͍͏όοϑΝʹॻ͖ࠐΈ·͢
    ͜ͷൣғ
    ωοτϫʔΫͷग़ྗ͸
    
    
    ͜ͷൣғͷதͰఆٛ͞ΕΔ
    
    
    probͱ͍͏όοϑΝ͔ΒಡΈ·͢
    

    View Slide

23. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    VGG16
    SIMONYAN, Karen; ZISSERMAN, Andrew. Very deep convolutional networks for large-
    scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
    https://arxiv.org/abs/1409.1556
    

    View Slide

24. https://github.com/KhronosGroup/NNEF-Tools/tree/main/models#nnef-model-zoo
    ILSVRC༻ʹֶशΛߦͬͨ
    
    
    VGG16ͷcaffemodelΛ
    
    
    ʹม׵ͨ͠΋ͷ
    

    View Slide

25. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    ೖྗ
    ग़ྗ
    

    View Slide

26. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    224x224ͰRGB(3νϟωϧ)ͷը૾͕10ຕ·ͱΊͯ֎͔Βೖͬͯ͘Δ
    external - ֎͔Βೖͬͯ͘ΔσʔλΛఆٛ͢Δ
    external(shape = [10, 3, 224, 224]);
    

    View Slide

27. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    linear - ઢܗม׵
    linear(relu_13, variable_28, variable_29);
    ϕΫτϧrelu_13ʹߦྻvariable_28Λֻ͚ͯϕΫτϧvariable_29Λ଍͢
    

    View Slide

28. x0
    x1
    x2
    x3
    y0
    y1
    y2
    y3
    y4
    ͱ ͕͜͏͍͏ؔ܎ʹ͋Δ࣌
    x y
    શ݁߹૚ - શͯͷ૊Έ߹Θ͕ͤ઀ଓ͞Ε͍ͯΔ
    

    View Slide

29. x0
    x1
    x2
    x3
    y0
    y1
    y2
    y3
    y4
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    w40
    w41
    w42
    w43
    ֤χϡʔϩϯͷॏΈΛฒ΂ͯߦྻʹ͢Δͱ
    

    View Slide

30. ϕ
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    w40
    w41
    w42
    w43
    x0
    x1
    x2
    x3
    =
    y0
    y1
    y2
    y3
    y4
    ׆ੑԽؔ਺
    શ݁߹૚=ߦྻͱϕΫτϧͷੵΛٻΊͯ׆ੑԽؔ਺ʹ௨͢
    

    View Slide

31. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    ReLU - ྲྀߦΓͷ׆ੑԽؔ਺
    relu(linear_1);
    

    View Slide

32. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    ߦྻͱϕΫτϧͷ
    
    
    ੵΛٻΊͯ
    ׆ੑԽؔ਺ʹ௨͢
    ͦͷ݁ՌΛ
    

    View Slide

33. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    શ݁߹૚
    

    View Slide

34. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    softmax - ग़ྗͷ૯࿨͕1ʹͳΔ׆ੑԽؔ਺
    softmax(linear_2, axes = [1]);
    

    View Slide

35. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    ߦྻͱϕΫτϧͷ
    
    
    ੵΛٻΊͯ
    ׆ੑԽؔ਺ʹ௨͢
    ͦͷ݁ՌΛ
    

    View Slide

36. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    શ݁߹૚
    શ݁߹૚
    

    View Slide

37. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    variable_31 = variable(label = 'fc8_blob2', shape = [1, 1000]);
    variable_30 = variable(label = 'fc8_blob1', shape = [1000, 4096]);
    variable_29 = variable(label = 'fc7_blob2', shape = [1, 4096]);
    variable_28 = variable(label = 'fc7_blob1', shape = [4096, 4096]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    variable - ϑΝΠϧ͔Β஋ΛಡΈࠐΉ
    

    View Slide

38. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    variable_31 = variable(label = 'fc8_blob2', shape = [1, 1000]);
    variable_30 = variable(label = 'fc8_blob1', shape = [1000, 4096]);
    variable_29 = variable(label = 'fc7_blob2', shape = [1, 4096]);
    variable_28 = variable(label = 'fc7_blob1', shape = [4096, 4096]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    variable - ϑΝΠϧ͔Β஋ΛಡΈࠐΉ
    fc7_blob2.datͱ͍͏໊લͷϑΝΠϧ͔Β
    4096ཁૉͷϕΫτϧΛಡΜͩ΋ͷΛ
    variable_29ͱݺͿࣄʹ͢Δ
    variable(label = 'fc7_blob2', shape = [1, 4096]);
    

    View Slide

39. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    variable_31 = variable(label = 'fc8_blob2', shape = [1, 1000]);
    variable_30 = variable(label = 'fc8_blob1', shape = [1000, 4096]);
    variable_29 = variable(label = 'fc7_blob2', shape = [1, 4096]);
    variable_28 = variable(label = 'fc7_blob1', shape = [4096, 4096]);
    ...
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    ...
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    graph.nnef
    variable - ϑΝΠϧ͔Β஋ΛಡΈࠐΉ
    fc7_blob1.datͱ͍͏໊લͷϑΝΠϧ͔Β
    4096x4096ͷߦྻΛಡΜͩ΋ͷΛ
    variable_28ͱݺͿࣄʹ͢Δ
    variable(label = 'fc7_blob1', shape = [4096, 4096]);
    

    View Slide

40. ReLU
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    w40
    w41
    w42
    w43
    x0
    x1
    x2
    x3
    +
    aaaaa
    aaaaa
    aaaaa
    aaaaa
    aaaaa
    =
    y0
    y1
    y2
    y3
    y4
    ׆ੑԽؔ਺
    w
    fc7_blob1.dat
    w
    fc7_blob2.dat
    relu_13
    relu_14
    

    View Slide

41. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    VGG16
    SIMONYAN, Karen; ZISSERMAN, Andrew. Very deep convolutional networks for large-
    scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
    https://arxiv.org/abs/1409.1556
    ࠓݟͨ෦෼
    

    View Slide

42. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    ৞ΈࠐΈ૚
    VGG͸ը૾ͷ෼ྨͷҝʹ࡞ΒΕͨχϡʔϥϧωοτϫʔΫͳͷͰ
    
    
    ্ͷํʹ͸ը૾ॲཧ޲͖ͷ૚͕ஔ͔Ε͍ͯΔ
    

    View Slide

43. 0.2
    0.1
    0.1
    0.2 0.7
    0.1
    0.1
    0.2
    0.1
    0.9
    0
    0
    0 0
    0
    0.7
    0.4
    0
    0.3
    0
    0 0
    0.8
    0
    0
    0.31
    ೖྗ
    ग़ྗ
    ೖྗը૾ͷ࿈ଓ͢ΔϐΫηϧʹ
    
    
    ϑΟϧλΛ͔͚ͯ
    
    
    ग़ྗը૾ͷ஋ΛಘΔ
    ϑΟϧλ
    Α͋͘Δը૾ॲཧͷܗ
    

    View Slide

44. 0.2
    0.1
    0.1
    0.2 0.7
    0.1
    0.1
    0.2
    0.1
    0.9
    0
    0
    0 0
    0
    0.7
    0.4
    0
    0.3
    0
    0 0
    0.8
    0
    0
    0.31
    ೖྗ
    ग़ྗ
    ϑΟϧλΛॏΈ ͱͯ͠
    w
    w
    ৞ΈࠐΈ૚
    ֶशͰཉ͍͠৘ใΛऔΓग़ͤΔ
    
    
    ϑΟϧλΛ֫ಘ͢Δ
    

    View Slide

45. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [1, 1],
    groups = 1,
    padding = [(1, 1), (1, 1)],
    stride = [1, 1]
    );
    conv - ৞ΈࠐΈ
    

    View Slide

46. conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [1, 1],
    groups = 1,
    padding = [(1, 1), (1, 1)],
    stride = [1, 1]
    );
    conv - ৞ΈࠐΈ
    ೖྗը૾
    ϑΟϧλ
    νϟωϧຖʹՃ͑Δఆ਺
    

    View Slide

47. conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [1, 1],
    groups = 1,
    padding = [(1, 1), (1, 1)],
    stride = [1, 1]
    );
    dilation=1 dilation=2
    ͜Ε
    

    View Slide

48. conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [1, 1],
    groups = 1,
    padding = [(1, 1), (1, 1)],
    stride = [1, 1]
    );
    padding=0 padding=1
    ͜Ε
    

    View Slide

49. conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [1, 1],
    groups = 1,
    padding = [(1, 1), (1, 1)],
    stride = [1, 1]
    );
    stride=1 stride=2
    ͜Ε
    

    View Slide

50. ೖྗ3νϟωϧ ग़ྗ4νϟωϧ
    ͷϑΟϧλ͕ ݸඞཁ
    3 × 3 3 × 4
    

    View Slide

51. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    RGBͷ3νϟωϧ
    3νϟωϧΛ64νϟωϧʹ͢Δ
    64νϟωϧΛ64νϟωϧʹ͢Δ
    

    View Slide

52. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    ϑΝΠϧconv1_1_blob1.datʹॻ͔Εͨ
    
    
    ͷ4֊ͷςϯιϧΛϑΟϧλͱͯ͠࢖͏
    3 × 3 × 3 × 64
    

    View Slide

53. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    3νϟωϧ͔Β64νϟωϧ΁ͷ৞ΈࠐΈ૚
    

    View Slide

54. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    ϑΝΠϧconv1_2_blob1.datʹॻ͔Εͨ
    
    
    ͷ4֊ͷςϯιϧΛϑΟϧλͱͯ͠࢖͏
    3 × 3 × 64 × 64
    

    View Slide

55. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    64νϟωϧ͔Β64νϟωϧ΁ͷ৞ΈࠐΈ૚
    

    View Slide

56. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    max_pool(
    relu_1,
    border = 'ignore',
    padding = [(0, 0), (0, 0), (0, 0), (0, 0)],
    size = [1, 1, 2, 2],
    stride = [1, 1, 2, 2]
    );
    max_pool - ϑΟϧλൣғ಺ͷ࠷େ஋Λग़ྗ͢Δ
    

    View Slide

57. 4
    3
    6
    5 3
    0
    1
    0
    8
    6
    1
    9 1
    3
    7
    0
    8
    ೖྗ
    ग़ྗ
    Max Pooling૚
    ϑΟϧλͷൣғ಺Ͱ࠷େͷ஋Λ
    
    
    ग़ྗը૾ͷରԠ͢ΔҐஔʹు͘
    ͜ͷൣғ಺Ͱ
    
    
    ࠷େ
    ͜ͷ૚͸ॏΈ Λ࣋ͨͳ͍
    w
    

    View Slide

58. max_pool(
    relu_1,
    border = 'ignore',
    padding = [(0, 0), (0, 0), (0, 0), (0, 0)],
    size = [1, 1, 2, 2],
    stride = [1, 1, 2, 2]
    );
    padding=0 padding=1
    ͜Ε
    

    View Slide

59. max_pool(
    relu_1,
    border = 'ignore',
    padding = [(0, 0), (0, 0), (0, 0), (0, 0)],
    size = [1, 1, 2, 2],
    stride = [1, 1, 2, 2]
    );
    size=2 size=3 size=4
    ͜Ε
    

    View Slide

60. max_pool(
    relu_1,
    border = 'ignore',
    padding = [(0, 0), (0, 0), (0, 0), (0, 0)],
    size = [1, 1, 2, 2],
    stride = [1, 1, 2, 2]
    );
    stride=1 stride=2
    ͜Ε
    

    View Slide

61. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    ...
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
    groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0),
    (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    ...
    }
    graph.nnef
    Max PoolingͰը૾Λ224x224͔Β112x112ʹॖখ
    

    View Slide

62. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    VGG16
    SIMONYAN, Karen; ZISSERMAN, Andrew. Very deep convolutional networks for large-
    scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
    https://arxiv.org/abs/1409.1556
    ࠓݟͨ෦෼
    

    View Slide

63. input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    ReLU
    linear
    softmax
    ͷը૾͕10ຕग़ΔΑ
    7 × 7 × 512
    ཁૉͷϕΫτϧ͕10ݸཉ͍͠Α
    25088
    σʔλͱͯ͠͸ಉ͕ͩ͡
    
    
    ܕͷม׵͕͍Δ
    

    View Slide

64. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    ...
    relu_12 = relu(conv_12);
    max_pool_4 = max_pool(relu_12, border = 'ignore', padding = [(0, 0), (0, 0), (0,
    0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    reshape = reshape(max_pool_4, shape = [10, -1]);
    linear = linear(reshape, variable_26, variable_27);
    relu_13 = relu(linear);
    ...
    }
    graph.nnef
    reshape(max_pool_4, shape = [10, -1]);
    ͜ͷόοϑΝͷ಺༰Λ
    10౳෼ʹͯ͠ɺ10ຊͷϕΫτϧʹ͢Δ
    reshape - σʔλͷղऍͷมߋ
    

    View Slide

65. version 1.0;
    graph VGG_ILSVRC_16_layers(data) -> (prob)
    {
    variable_15 = variable(label = 'conv4_1_blob2', shape = [1, 512]);
    variable_14 = variable(label = 'conv4_1_blob1', shape = [512, 256, 3, 3]);
    variable_13 = variable(label = 'conv3_3_blob2', shape = [1, 256]);
    variable_31 = variable(label = 'fc8_blob2', shape = [1, 1000]);
    variable_30 = variable(label = 'fc8_blob1', shape = [1000, 4096]);
    variable_29 = variable(label = 'fc7_blob2', shape = [1, 4096]);
    variable_28 = variable(label = 'fc7_blob1', shape = [4096, 4096]);
    variable_27 = variable(label = 'fc6_blob2', shape = [1, 4096]);
    variable_26 = variable(label = 'fc6_blob1', shape = [4096, 25088]);
    variable_25 = variable(label = 'conv5_3_blob2', shape = [1, 512]);
    variable_24 = variable(label = 'conv5_3_blob1', shape = [512, 512, 3, 3]);
    variable_23 = variable(label = 'conv5_2_blob2', shape = [1, 512]);
    variable_22 = variable(label = 'conv5_2_blob1', shape = [512, 512, 3, 3]);
    variable_21 = variable(label = 'conv5_1_blob2', shape = [1, 512]);
    variable_20 = variable(label = 'conv5_1_blob1', shape = [512, 512, 3, 3]);
    variable_19 = variable(label = 'conv4_3_blob2', shape = [1, 512]);
    variable_18 = variable(label = 'conv4_3_blob1', shape = [512, 512, 3, 3]);
    variable_17 = variable(label = 'conv4_2_blob2', shape = [1, 512]);
    variable_16 = variable(label = 'conv4_2_blob1', shape = [512, 512, 3, 3]);
    variable_12 = variable(label = 'conv3_3_blob1', shape = [256, 256, 3, 3]);
    variable_10 = variable(label = 'conv3_2_blob1', shape = [256, 256, 3, 3]);
    variable_9 = variable(label = 'conv3_1_blob2', shape = [1, 256]);
    variable_8 = variable(label = 'conv3_1_blob1', shape = [256, 128, 3, 3]);
    variable_6 = variable(label = 'conv2_2_blob1', shape = [128, 128, 3, 3]);
    variable_11 = variable(label = 'conv3_2_blob2', shape = [1, 256]);
    variable_5 = variable(label = 'conv2_1_blob2', shape = [1, 128]);
    variable_4 = variable(label = 'conv2_1_blob1', shape = [128, 64, 3, 3]);
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    variable_7 = variable(label = 'conv2_2_blob2', shape = [1, 128]);
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_2 = conv(max_pool, variable_4, variable_5, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    graph.nnef
    ϑΝΠϧ͔Β஋ΛಡΉ
    

    View Slide

66. variable_6 = variable(label = 'conv2_2_blob1', shape = [128, 128, 3, 3]);
    variable_11 = variable(label = 'conv3_2_blob2', shape = [1, 256]);
    variable_5 = variable(label = 'conv2_1_blob2', shape = [1, 128]);
    variable_4 = variable(label = 'conv2_1_blob1', shape = [128, 64, 3, 3]);
    variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    variable_7 = variable(label = 'conv2_2_blob2', shape = [1, 128]);
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_2 = conv(max_pool, variable_4, variable_5, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_2 = relu(conv_2);
    conv_3 = conv(relu_2, variable_6, variable_7, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_3 = relu(conv_3);
    max_pool_1 = max_pool(relu_3, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_4 = conv(max_pool_1, variable_8, variable_9, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_4 = relu(conv_4);
    conv_5 = conv(relu_4, variable_10, variable_11, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_5 = relu(conv_5);
    conv_6 = conv(relu_5, variable_12, variable_13, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_6 = relu(conv_6);
    max_pool_2 = max_pool(relu_6, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_7 = conv(max_pool_2, variable_14, variable_15, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_7 = relu(conv_7);
    conv_8 = conv(relu_7, variable_16, variable_17, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_8 = relu(conv_8);
    conv_9 = conv(relu_8, variable_18, variable_19, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_9 = relu(conv_9);
    max_pool_3 = max_pool(relu_9, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_10 = conv(max_pool_3, variable_20, variable_21, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_10 = relu(conv_10);
    conv_11 = conv(relu_10, variable_22, variable_23, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_11 = relu(conv_11);
    conv_12 = conv(relu_11, variable_24, variable_25, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_12 = relu(conv_12);
    max_pool_4 = max_pool(relu_12, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    reshape = reshape(max_pool_4, shape = [10, -1]);
    linear = linear(reshape, variable_26, variable_27);
    relu_13 = relu(linear);
    linear_1 = linear(relu_13, variable_28, variable_29);
    input
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    

    View Slide

67. variable_2 = variable(label = 'conv1_2_blob1', shape = [64, 64, 3, 3]);
    variable_1 = variable(label = 'conv1_1_blob2', shape = [1, 64]);
    variable_7 = variable(label = 'conv2_2_blob2', shape = [1, 128]);
    variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
    variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
    data = external(shape = [10, 3, 224, 224]);
    conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu = relu(conv);
    conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_1 = relu(conv_1);
    max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_2 = conv(max_pool, variable_4, variable_5, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_2 = relu(conv_2);
    conv_3 = conv(relu_2, variable_6, variable_7, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_3 = relu(conv_3);
    max_pool_1 = max_pool(relu_3, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_4 = conv(max_pool_1, variable_8, variable_9, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_4 = relu(conv_4);
    conv_5 = conv(relu_4, variable_10, variable_11, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_5 = relu(conv_5);
    conv_6 = conv(relu_5, variable_12, variable_13, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_6 = relu(conv_6);
    max_pool_2 = max_pool(relu_6, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_7 = conv(max_pool_2, variable_14, variable_15, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_7 = relu(conv_7);
    conv_8 = conv(relu_7, variable_16, variable_17, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_8 = relu(conv_8);
    conv_9 = conv(relu_8, variable_18, variable_19, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_9 = relu(conv_9);
    max_pool_3 = max_pool(relu_9, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_10 = conv(max_pool_3, variable_20, variable_21, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_10 = relu(conv_10);
    conv_11 = conv(relu_10, variable_22, variable_23, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_11 = relu(conv_11);
    conv_12 = conv(relu_11, variable_24, variable_25, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_12 = relu(conv_12);
    max_pool_4 = max_pool(relu_12, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    reshape = reshape(max_pool_4, shape = [10, -1]);
    linear = linear(reshape, variable_26, variable_27);
    relu_13 = relu(linear);
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    

    View Slide

68. max_pool = max_pool(relu_1, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_2 = conv(max_pool, variable_4, variable_5, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_2 = relu(conv_2);
    conv_3 = conv(relu_2, variable_6, variable_7, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_3 = relu(conv_3);
    max_pool_1 = max_pool(relu_3, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_4 = conv(max_pool_1, variable_8, variable_9, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_4 = relu(conv_4);
    conv_5 = conv(relu_4, variable_10, variable_11, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_5 = relu(conv_5);
    conv_6 = conv(relu_5, variable_12, variable_13, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_6 = relu(conv_6);
    max_pool_2 = max_pool(relu_6, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_7 = conv(max_pool_2, variable_14, variable_15, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_7 = relu(conv_7);
    conv_8 = conv(relu_7, variable_16, variable_17, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_8 = relu(conv_8);
    conv_9 = conv(relu_8, variable_18, variable_19, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_9 = relu(conv_9);
    max_pool_3 = max_pool(relu_9, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    conv_10 = conv(max_pool_3, variable_20, variable_21, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_10 = relu(conv_10);
    conv_11 = conv(relu_10, variable_22, variable_23, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_11 = relu(conv_11);
    conv_12 = conv(relu_11, variable_24, variable_25, border = 'constant', dilation = [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
    relu_12 = relu(conv_12);
    max_pool_4 = max_pool(relu_12, border = 'ignore', padding = [(0, 0), (0, 0), (0, 0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
    reshape = reshape(max_pool_4, shape = [10, -1]);
    linear = linear(reshape, variable_26, variable_27);
    relu_13 = relu(linear);
    linear_1 = linear(relu_13, variable_28, variable_29);
    relu_14 = relu(linear_1);
    linear_2 = linear(relu_14, variable_30, variable_31);
    prob = softmax(linear_2, axes = [1]);
    }
    convolution
    ReLU
    convolution
    ReLU
    convolution
    ReLU
    maxpool
    linear
    ReLU
    linear
    softmax
    ReLU
    convolution
    ReLU
    maxpool
    

    View Slide

69. constant(
    shape = [ 3, 4 ],
    value = [42.0]
    );
    constant - શͯͷཁૉ͕ಛఆͷ஋ʹͳ͍ͬͯΔόοϑΝΛ࡞Δ
    [
    42.0 42.0 42.0 42.0
    42.0 42.0 42.0 42.0
    42.0 42.0 42.0 42.0
    ]
    

    View Slide

70. conv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [2, 2],
    groups = 1,
    padding = [(0, 0), (0, 0)],
    stride = [1, 1]
    );
    deconv - ٯ৞ΈࠐΈ
    deconv(
    data, variable, variable_1,
    border = 'constant',
    dilation = [2, 2],
    groups = 1,
    padding = [(0, 0), (0, 0)],
    stride = [1, 1]
    );
    

    View Slide

71. box - ϑΟϧλͷൣғ಺ͷ૯࿨Λฦ͢
    box(
    data,
    size = [ 1, 3, 3 ],
    border = 'constant',
    dilation = [2, 2],
    groups = 1,
    padding = [(0, 0), (0, 0)],
    stride = [1, 1]
    );
    debox(
    data,
    size = [ 1, 3, 3 ],
    border = 'constant',
    dilation = [2, 2],
    groups = 1,
    padding = [(0, 0), (0, 0)],
    stride = [1, 1]
    );
    9
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    1
    4
    1
    1
    1
    1
    1
    1
    1
    1
    1
    

    View Slide

72. 6
    9
    8
    2
    7
    6
    8
    7
    8
    5
    3
    6
    6
    1
    0
    6 4
    7
    3
    4
    7
    argmax_pool - ࠷େ஋Λ࣋ͭཁૉͷΠϯσοΫεΛฦ͢
    sample - ΠϯσοΫεͰଞͷςϯιϧ͔Β஋Λरͬͯ͘Δ
    

    View Slide

73. 22
    9
    2
    4
    7
    9
    9
    2
    4
    7
    2
    9
    2
    4
    7
    0
    9
    2
    4
    7
    1
    9
    2
    4
    7
    0
    1
    0
    1
    1
    1
    1
    0
    1
    1
    reduce - ಛఆͷ࣠ํ޲ͷશͯͷ஋Λ1ͭͷ஋ʹू໿͢Δ
    sum_reduce max_reduce min_reduce argmax_reduce
    argmin_reduce all_reduce any_reduce
    

    View Slide

74. split - ಛఆͷ࣠ํ޲ʹςϯιϧΛ෼ׂ͢Δ
    split concat pad
    

    View Slide

75. 3
    9
    5
    4
    gather - ಛఆͷ࣠ํ޲ͷཁૉ͔Β1ͭΛબͿ
    1
    0
    3
    2
    3
    9
    5
    4 9.0
    2.0
    7.0
    7.0
    3.0
    6.0
    0.0
    8.0
    8.0
    5.0
    7.0
    6.0
    2.0
    0.0
    7.0
    9.0
    9
    2
    7
    7
    3
    6
    0
    8
    8
    5
    7
    6
    2
    0
    7
    9
    cast - ςϯιϧͷཁૉͷܕΛม͑Δ
    

    View Slide

76. matmul - ߦྻੵ
    result = matmul( input1, input2 );
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    =
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    input1 input2
    result
    result = transpose( input, [ 1, 0 ] );
    transpose - సஔ
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    =
    w00
    w01
    w02
    w03
    w10
    w11
    w12
    w13
    w20
    w21
    w22
    w23
    w30
    w31
    w32
    w33
    T
    result input
    

    View Slide

77. ׆ੑԽؔ਺
    sigmoid relu prelu leaky_relu elu gelu silu softmax softplus
    ϓʔϦϯά
    max_pool_with_index max_pool avg_pool rms_pool
    ਖ਼نԽ
    local_response_normalization local_mean_normalization
    local_variance_normalization local_contrast_normalization
    l1_normalization l2_normalization batch_normalization
    ྔࢠԽ
    min_max_linear_quantize zero_point_linear_quantize
    logarithmetic_quantize
    

    View Slide

78. ؔ৺ྖҬϓʔϦϯά
    avg_roi_pool max_roi_pool roi_resample avg_roi_align max_roi_align
    ϦαΠζ
    nearest_downsample area_downsample nearest_upsample
    multilinear_upsample
    ୯߲ԋࢉࢠ
    copy neg rcp exp log sin cos tan sinh cosh tanh asin acos atan
    asinh acosh atanh abs sign not floor ceil round
    ೋ߲ԋࢉࢠ
    add sub mul div pow lt gt le ge eq ne and or
    ࡾ߲ԋࢉࢠ
    select
    ͦͷଞͷؔ਺
    sqr sqrt rsqr rsqrt log2 min max clamp
    

    View Slide

79. https://github.com/Fadis/gct
    GPU
    
    
    Computing
    
    
    Toolkit
    VulkanΛ࢖ͬͯ
    
    
    GPUΛ࢖͏ΞϓϦέʔγϣϯ͕
    
    
    Α͘ߦ͏ॲཧΛ
    
    
    ؆ܿʹॻ͚ΔΑ͏ʹ͢Δ
    

    View Slide

80. https://github.com/Fadis/gct
    GPU
    
    
    Computing
    
    
    Toolkit
    VulkanΛ࢖ͬͯ
    
    
    GPUΛ࢖͏ΞϓϦέʔγϣϯ͕
    
    
    Α͘ߦ͏ॲཧΛ
    
    
    ؆ܿʹॻ͚ΔΑ͏ʹ͢Δ
    Λ౉ͨ͠Β
    
    
    χϡʔϥϧωοτϫʔΫͷධՁΛ
    
    
    GPUͰγϡοͱ࣮ߦͰ͖ΔΑ͏ʹ͍ͨ͠
    

    View Slide

81. ͷΦϖϨʔλͷ͏ͪ
    
    
    VGGΛಈ͔ͨ͢Ίʹ࣮૷͞Ε͍ͯͳ͚Ε͹ͳΒͳ͍෺
    external
    variable
    conv
    relu
    max_pool
    reshape
    linear
    softmax
    

    View Slide

82. ͷΦϖϨʔλͷ͏ͪ
    
    
    VGGΛಈ͔ͨ͢Ίʹ࣮૷͞Ε͍ͯͳ͚Ε͹ͳΒͳ͍෺
    external
    variable
    conv
    relu
    max_pool
    reshape
    linear
    softmax
    GPUͷϝϞϦ֬อ͢Δ͚ͩ
    GPUͷϝϞϦ֬อͯ͠σʔλΛసૹ͢Δ͚ͩ
    ςϯιϧͷαΠζΛม͑Δ͚ͩ
    

    View Slide

83. #version 450
    #extension GL_ARB_separate_shader_objects : enable
    #extension GL_ARB_shading_language_420pack : enable
    #extension GL_KHR_shader_subgroup_basic : enable
    #extension GL_KHR_shader_subgroup_arithmetic : enable
    #extension GL_KHR_memory_scope_semantics : enable
    layout(local_size_x = 32, local_size_y = 32, local_size_z = 1 ) in;
    layout(std430, binding = 0) buffer input_vector {
    float input_data[];
    };
    layout(std430, binding = 1) buffer weight {
    float weight_data[];
    };
    layout(std430, binding = 2) buffer output_vector {
    float output_data[];
    };
    layout(std430, binding = 3) buffer bias {
    float bias_data[];
    };
    layout(constant_id = 1) const int filter_size_x = 0;
    layout(constant_id = 2) const int filter_size_y = 0;
    ೖྗ
    όΠΞε
    ग़ྗ
    ϑΟϧλ
    conv
    

    View Slide

84. layout(constant_id = 1) const int filter_size_x = 0;
    layout(constant_id = 2) const int filter_size_y = 0;
    layout(constant_id = 3) const int filter_size_z = 0;
    layout(constant_id = 4) const int lpadding = 0;
    layout(constant_id = 5) const int rpadding = 0;
    layout(constant_id = 6) const int tpadding = 0;
    layout(constant_id = 7) const int bpadding = 0;
    layout(constant_id = 8) const int stride_x = 0;
    layout(constant_id = 9) const int stride_y = 0;
    layout(constant_id = 10) const int dilation_x = 0;
    layout(constant_id = 11) const int dilation_y = 0;
    layout(constant_id = 12) const int input_dim_x = 0;
    layout(constant_id = 13) const int input_dim_y = 0;
    layout(constant_id = 14) const int input_dim_z = 0;
    layout(constant_id = 15) const int output_dim_x = 0;
    layout(constant_id = 16) const int output_dim_y = 0;
    layout(constant_id = 17) const int output_dim_z = 0;
    layout(constant_id = 18) const float border_value = 0.0;
    layout(constant_id = 19) const int bias_mode = 0;
    layout(constant_id = 20) const float bias_value = 0;
    int get_filter_length() {
    return filter_size_x * filter_size_y * filter_size_z;
    }
    ϑΟϧλαΠζ
    όΠΞεͷ৐ͤํ
    ൣғ֎ͷ஋
    padding
    stride
    dilation
    ೖྗͷαΠζ
    ग़ྗͷαΠζ
    

    View Slide

85. shared float[filter_size_x*filter_size_y*filter_size_z] filter_cache;
    void load_filter() {
    const uint local_id =
    gl_LocalInvocationID.x +
    gl_LocalInvocationID.y * gl_WorkGroupSize.x +
    gl_LocalInvocationID.z * gl_WorkGroupSize.y * gl_WorkGroupSize.x;
    const uint local_size = gl_WorkGroupSize.x * gl_WorkGroupSize.y * gl_WorkGroupSize.z;
    const uint filter_size = uint( get_filter_length() );
    const uint cycles = filter_size / local_size + ( bool( filter_size % local_size ) ?
    1 : 0 );
    for( uint c = 0u; c != cycles; c++ ) {
    uint i = c * local_size + local_id;
    const int filter_offset = get_filter_offset( int( i ) );
    if( i < filter_size ) {
    filter_cache[ i ] = weight_data[ filter_offset ];
    }
    }
    }
    float get_input( int i ) {
    const int input_offset = get_input_offset( i );
    return
    ( input_offset < 0 ) ?
    1εϨου͕ग़ྗͷ1ཁૉΛ୲౰͢Δ
    ෳ਺ͷεϨου͕ಉ͡ϑΟϧλΛ࢖͏ͷͰ
    
    
    ڞ༗ϝϞϦʹϑΟϧλΛϩʔυ͢Δ
    

    View Slide

86. const int output_channel = int( gl_GlobalInvocationID.z % output_dim_z );
    return
    ( bias_mode == 0 ) ?
    bias_value :
    bias_data[ output_channel ];
    }
    void set_output( float v ) {
    const int output_offset = get_output_offset();
    if( output_offset >= 0 ) {
    output_data[ output_offset ] = v;
    }
    }
    void main() {
    load_filter();
    barrier();
    const int filter_length = get_filter_length();
    float sum = 0.0;
    for( int i = 0; i != filter_length; i++ ) {
    sum += get_input( i ) * get_filter( i );
    }
    sum += get_bias();
    set_output( sum );
    }
    ϑΟϧλͷൣғ಺ͷೖྗ஋ʹϑΟϧλΛ͔͚ͯ
    
    
    ૯࿨ΛऔΓɺόΠΞεΛՃ͑ͯग़ྗʹॻ͘
    

    View Slide

87. #version 450
    #extension GL_ARB_separate_shader_objects : enable
    #extension GL_ARB_shading_language_420pack : enable
    #extension GL_KHR_shader_subgroup_basic : enable
    #extension GL_KHR_shader_subgroup_arithmetic : enable
    #extension GL_KHR_memory_scope_semantics : enable
    layout(local_size_x = 1024, local_size_y = 1, local_size_z = 1 ) in;
    layout(std430, binding = 0) buffer input_vector {
    float input_data[];
    };
    layout(std430, binding = 1) buffer weight {
    float weight_data[];
    };
    layout(std430, binding = 2) buffer output_vector {
    float output_data[];
    };
    layout(std430, binding = 3) buffer bias {
    float bias_data[];
    };
    layout(constant_id = 1) const uint input_length = 32;
    layout(constant_id = 2) const uint bias_mode = 0;
    layout(constant_id = 3) const float bias_value = 0.0;
    linear
    ೖྗ
    όΠΞε
    ग़ྗ
    ϑΟϧλ
    

    View Slide

88. };
    layout(constant_id = 1) const uint input_length = 32;
    layout(constant_id = 2) const uint bias_mode = 0;
    layout(constant_id = 3) const float bias_value = 0.0;
    shared float[32] temp;
    float get_bias() {
    const uint output_index = gl_GlobalInvocationID.y;
    return
    ( bias_mode == 0 ) ?
    bias_value :
    bias_data[ output_index ];
    }
    void main() {
    const uint x = gl_GlobalInvocationID.x;
    const uint y = gl_GlobalInvocationID.y;
    const uint batch = gl_GlobalInvocationID.z;
    float sum = float( 0 );
    const uint x_blocks =
    input_length / gl_WorkGroupSize.x +
    ( bool( input_length % gl_WorkGroupSize.x ) ? 1 : 0 );
    for( uint x_index = 0; x_index != x_blocks; x_index++ ) {
    const uint x_global = x_index * gl_WorkGroupSize.x + x;
    ೖྗͷαΠζ
    όΠΞεͷ৐ͤํ
    ਫฏՃࢉʹ࢖͏ڞ༗ϝϞϦ
    

    View Slide

89. const uint batch = gl_GlobalInvocationID.z;
    float sum = float( 0 );
    const uint x_blocks =
    input_length / gl_WorkGroupSize.x +
    ( bool( input_length % gl_WorkGroupSize.x ) ? 1 : 0 );
    for( uint x_index = 0; x_index != x_blocks; x_index++ ) {
    const uint x_global = x_index * gl_WorkGroupSize.x + x;
    const uint batch_offset = input_length * batch;
    bool mask = ( x_global < input_length );
    float result = float( 0 );
    result = mask ?
    ( weight_data[ x_global + y * input_length ] * input_data[ batch_offset + x_global ] ) :
    float( 0 );
    result = subgroupAdd( result );
    if( gl_SubgroupInvocationID == 0 ) {
    temp[ gl_SubgroupID ] = result;
    }
    barrier();
    mask = ( x < gl_NumSubgroups );
    result = subgroupAdd( mask ? temp[ x ] : float( 0 ) );
    sum += result;
    }
    if( x == 0 ) {
    const uint output_length = gl_WorkGroupSize.y * gl_NumWorkGroups.y;
    const uint batch_offset = output_length * batch;
    output_data[ batch_offset + y ] = sum + get_bias();
    }
    ߦྻͷ1ཁૉʹ͖ͭ1εϨου
    ਫฏՃࢉ໋ྩͰ
    
    
    ೖྗͱॏΈ1ྻͷ಺ੵΛٻΊΔ
    1ߦ໨Λ୲౰͢ΔεϨου͕
    
    
    ܭࢉ݁ՌΛॻ͘
    

    View Slide

90. #version 450
    #extension GL_ARB_separate_shader_objects : enable
    #extension GL_ARB_shading_language_420pack : enable
    #extension GL_KHR_shader_subgroup_basic : enable
    #extension GL_KHR_shader_subgroup_arithmetic : enable
    #extension GL_KHR_memory_scope_semantics : enable
    layout(local_size_x = 32, local_size_y = 32, local_size_z = 1 ) in;
    layout(std430, binding = 0) buffer input_vector {
    float input_data[];
    };
    layout(std430, binding = 1) buffer output_vector {
    float output_data[];
    };
    layout(constant_id = 1) const int filter_size_x = 0;
    layout(constant_id = 2) const int filter_size_y = 0;
    layout(constant_id = 3) const int filter_size_z = 0;
    layout(constant_id = 4) const int filter_size_w = 0;
    layout(constant_id = 5) const int lpadding = 0;
    layout(constant_id = 6) const int rpadding = 0;
    layout(constant_id = 7) const int tpadding = 0;
    layout(constant_id = 8) const int bpadding = 0;
    layout(constant_id = 9) const int stride_x = 0;
    layout(constant_id = 10) const int stride_y = 0;
    layout(constant_id = 11) const int stride_z = 0;
    layout(constant_id = 12) const int stride_w = 0;
    max_pool
    ೖྗ
    ग़ྗ
    

    View Slide

91. };
    layout(constant_id = 1) const int filter_size_x = 0;
    layout(constant_id = 2) const int filter_size_y = 0;
    layout(constant_id = 3) const int filter_size_z = 0;
    layout(constant_id = 4) const int filter_size_w = 0;
    layout(constant_id = 5) const int lpadding = 0;
    layout(constant_id = 6) const int rpadding = 0;
    layout(constant_id = 7) const int tpadding = 0;
    layout(constant_id = 8) const int bpadding = 0;
    layout(constant_id = 9) const int stride_x = 0;
    layout(constant_id = 10) const int stride_y = 0;
    layout(constant_id = 11) const int stride_z = 0;
    layout(constant_id = 12) const int stride_w = 0;
    layout(constant_id = 13) const int input_dim_x = 0;
    layout(constant_id = 14) const int input_dim_y = 0;
    layout(constant_id = 15) const int input_dim_z = 0;
    layout(constant_id = 16) const int input_dim_w = 0;
    layout(constant_id = 17) const int output_dim_x = 0;
    layout(constant_id = 18) const int output_dim_y = 0;
    layout(constant_id = 19) const int output_dim_z = 0;
    layout(constant_id = 20) const int output_dim_w = 0;
    layout(constant_id = 21) const float border_value = 0.0;
    int get_filter_length() {
    return filter_size_x * filter_size_y * filter_size_z * filter_size_w;
    }
    int get_input_offset( int i ) {
    ϑΟϧλαΠζ
    ൣғ֎ͷ஋
    padding
    stride
    ೖྗͷαΠζ
    ग़ྗͷαΠζ
    

    View Slide

92. }
    float get_input( int i ) {
    const int input_offset = get_input_offset( i );
    return
    ( input_offset < 0 ) ?
    border_value :
    input_data[ input_offset ];
    }
    void set_output( float v ) {
    const int output_offset = get_output_offset();
    if( output_offset >= 0 ) {
    output_data[ output_offset ] = v;
    }
    }
    void main() {
    const int filter_length = get_filter_length();
    float v = -10000.0;
    for( int i = 0; i != filter_length; i++ ) {
    const int input_offset = get_input_offset( i );
    if( input_offset != -1 ) {
    v = max( input_data[ input_offset ], v );
    }
    }
    set_output( v );
    }
    ϑΟϧλͷൣғ಺Ͱ
    
    
    ࠷େͷ஋Λग़ྗʹॻ͘
    

    View Slide

93. #version 450
    #extension GL_ARB_separate_shader_objects : enable
    #extension GL_ARB_shading_language_420pack : enable
    #extension GL_KHR_shader_subgroup_basic : enable
    #extension GL_KHR_shader_subgroup_arithmetic : enable
    #extension GL_KHR_memory_scope_semantics : enable
    layout(local_size_x = 1024, local_size_y = 1, local_size_z = 1 ) in;
    layout(std430, binding = 0) buffer input_vector {
    float input_data[];
    };
    layout(std430, binding = 1) buffer output_vector {
    float output_data[];
    };
    layout(constant_id = 1) const uint input_length = 32;
    void main() {
    const uint index =
    gl_GlobalInvocationID.x +
    gl_NumWorkGroups.x * gl_WorkGroupSize.x * gl_GlobalInvocationID.y +
    gl_NumWorkGroups.x * gl_WorkGroupSize.x * gl_NumWorkGroups.y * gl_WorkGroupSize.y * gl_GlobalInvocationID.z;
    if( index >= input_length ) return;
    float v = input_data[ index ];
    output_data[ index ] = ( v >= 0.0 ) ? v : 0.0;
    }
    ೖྗ
    ग़ྗ
    ग़ྗ = max( ೖྗ, 0 )
    relu
    

    View Slide

94. #version 450
    #extension GL_ARB_separate_shader_objects : enable
    #extension GL_ARB_shading_language_420pack : enable
    #extension GL_KHR_shader_subgroup_basic : enable
    #extension GL_KHR_shader_subgroup_arithmetic : enable
    #extension GL_KHR_memory_scope_semantics : enable
    layout(local_size_x = 1024, local_size_y = 1, local_size_z = 1 ) in;
    layout(std430, binding = 0) buffer input_vector {
    float input_data[];
    };
    layout(std430, binding = 1) buffer output_vector {
    float output_data[];
    };
    layout(constant_id = 1) const uint input_length = 32;
    shared float[32] temp;
    void main() {
    const uint offset = gl_LocalInvocationID.x;
    const uint batch = gl_GlobalInvocationID.z;
    float max = 0.0;
    for(
    uint i = offset;
    i < input_length;
    i += 1024
    ) {
    softmax
    ೖྗ
    ग़ྗ
    

    View Slide

95. const uint offset = gl_LocalInvocationID.x;
    const uint batch = gl_GlobalInvocationID.z;
    float max = 0.0;
    for(
    uint i = offset;
    i < input_length;
    i += 1024
    ) {
    float v = input_data[ i + input_length * batch ];
    if( v > max ) {
    max = v;
    }
    }
    const float smax = subgroupMax( max );
    if( gl_SubgroupInvocationID == 0 ) {
    temp[ gl_SubgroupID ] = smax;
    }
    barrier();
    const float gmax = subgroupMax( temp[ gl_SubgroupInvocationID ] );
    float sum = 0.0;
    for(
    uint i = offset;
    i < input_length;
    i += 1024
    ) {
    sum += exp( input_data[ i + input_length * batch ] - gmax );
    }
    const float ssum = subgroupAdd( sum );
    softmaxʹ͸
    
    
    ೖྗ஋ͷexpͷ૯࿨͕ཁΔͷͰ
    
    
    શͯͷ஋Λڞ༗ϝϞϦ͕ಧ͘
    
    
    1024εϨουͷதͰย෇͚Δ
    ೖྗͷ࠷େ஋ΛٻΊΔ
    

    View Slide

96. const float smax = subgroupMax( max );
    if( gl_SubgroupInvocationID == 0 ) {
    temp[ gl_SubgroupID ] = smax;
    }
    barrier();
    const float gmax = subgroupMax( temp[ gl_SubgroupInvocationID ] );
    float sum = 0.0;
    for(
    uint i = offset;
    i < input_length;
    i += 1024
    ) {
    sum += exp( input_data[ i + input_length * batch ] - gmax );
    }
    const float ssum = subgroupAdd( sum );
    if( gl_SubgroupInvocationID == 0 ) {
    temp[ gl_SubgroupID ] = ssum;
    }
    barrier();
    const float gsum = subgroupAdd( temp[ gl_SubgroupInvocationID ] );
    for(
    uint i = offset;
    i < input_length;
    i += 1024
    ) {
    output_data[ i + input_length * batch ] =
    exp( input_data[ i + input_length * batch ] - gmax ) / gsum;
    }
    }
    ೖྗͷexpͷ૯࿨ΛٻΊΔ
    ೖྗͷexpΛೖྗͷexpͷ૯࿨ͰׂΔ
    

    View Slide

97. https://github.com/KhronosGroup/NNEF-Tools
    ͜Ε
    NNEFͷެࣜϥΠϒϥϦ
    
    
    NNEF-Toolsͷதʹ
    
    
    NNEFͷύʔαؚ͕·Ε͍ͯΔ
    Pythonͷ࣮૷ͱ
    
    
    C++ͷ࣮૷͕༻ҙ͞Ε͍ͯΔ
    

    View Slide

98. graph::graph(
    const std::shared_ptr< device_t > &device,
    const std::shared_ptr< allocator_t > &allocator,
    const std::shared_ptr< descriptor_pool_t > &descriptor_pool,
    const std::shared_ptr< pipeline_cache_t > &pipeline_cache,
    const std::filesystem::path &dir,
    const std::filesystem::path &shader_dir,
    command_buffer_recorder_t &rec
    ) {
    nnef::Graph parsed;
    std::string error;
    if( !nnef::load_graph( dir.string(), parsed, error, "" ) ) {
    std::cerr << error << std::endl;
    throw -1;
    }
    if( !nnef::infer_shapes( parsed, error ) ) {
    std::cerr << error << std::endl;
    throw -1;
    }
    if( !nnef::allocate_buffers( parsed, error ) ) {
    std::cerr << error << std::endl;
    throw -1;
    }
    gct/src/gct/dnn/graph.cpp
    NNEF-Toolsͷ
    
    
    ύʔαͰ
    
    
    NNEFΛಡΈࠐΉ
    

    View Slide

99. for( const auto &o: parsed.operations ) {
    if( o.name == "variable" ) {
    const std::string name = get_output_name( o );
    const auto label = std::find_if(
    o.attribs.begin(), o.attribs.end(),
    []( const auto &v ) {
    return v.first == "label";
    }
    );
    if( label == o.attribs.end() ) {
    throw -1;
    }
    const auto data_name = label->second.string();
    const auto data_filename = dir / ( data_name + ".dat" );
    auto nnef_data = rec.load_nnef_data(
    allocator,
    std::filesystem::absolute( data_filename ),
    vk::BufferUsageFlagBits::eStorageBuffer|
    vk::BufferUsageFlagBits::eTransferDst
    );
    bufs.insert(
    std::make_pair(
    name,
    nnef_data
    )
    );
    }
    gct/src/gct/dnn/graph.cpp
    ඞཁͳϑΝΠϧͷ಺༰Λ
    
    
    GPUͷϝϞϦʹૹΔ
    

    View Slide

100. for( const auto &o: parsed.operations ) {
     if( o.name == "conv" ) {
     const std::string name = get_output_name( o );
     auto op = std::make_shared< operation::convolution >(
     allocator,
     descriptor_pool,
     pipeline_cache,
     o,
     shaders,
     bufs
     );
     bufs.insert(
     std::make_pair(
     name,
     op->get_output()
     )
     );
     ops.push_back( op );
     }
     else if( o.name == "linear" ) {
     const std::string name = get_output_name( o );
     auto op = std::make_shared< operation::linear >(
     allocator,
     descriptor_pool,
     pipeline_cache,
     o,
     shaders,
     bufs
     gct/src/gct/dnn/graph.cpp
     ֤૚ʹରԠ͢Δ
     
     
     ύΠϓϥΠϯΛ࡞Γ
     
     
     σεΫϦϓληοτʹ
     
     
     όοϑΝΛ݁ͼ͚ͭΔ
     

     View Slide

101. void convolution::operator()( command_buffer_recorder_t &rec ) {
     rec.compute_barrier( { input.buffer }, {} );
     rec.bind_descriptor_set(
     vk::PipelineBindPoint::eCompute,
     pipeline_layout,
     descriptor_set
     );
     rec.bind_pipeline( pipeline );
     rec.dispatch_threads( exec_dim[ 0 ], exec_dim[ 1 ], exec_dim[ 2 ] );
     }
     gct/src/gct/dnn/convolution.cpp
     ίϚϯυόοϑΝʹ
     
     
     ඞཁͳϝϞϦόϦΞͱίϯϐϡʔτύΠϓϥΠϯͷ࣮ߦΛੵΉ
     

     View Slide

102. std::vector< std::uint8_t > temp( dest.buffer->get_props().get_basic().size, 0u );
     std::unordered_map< std::string, int > channel_order{
     { "R", 2 },
     { "G", 1 },
     { "B", 0 }
     };
     for( int c = 0; c != spec.nchannels; ++c ) {
     const auto order = channel_order.find( spec.channelnames[ c ] );
     if( order != channel_order.end() ) {
     file->read_image(
     c,
     c + 1u,
     type,
     reinterpret_cast< std::uint8_t* >(
     std::next( temp.data(), spec.width * spec.height * dest.type.depth/8u * order->second ) )
     );
     }
     }
     constexpr std::array< float, 3u > mean{ 123.68f, 116.779f, 103.939f };
     for( int c = 0; c != spec.nchannels; ++c ) {
     for( unsigned int y = 0; y != spec.height; ++y ) {
     for( unsigned int x = 0; x != spec.width; ++x ) {
     const auto index = x + y * spec.width + c * spec.width * spec.height;
     reinterpret_cast< float* >( temp.data() )[ index ] =
     reinterpret_cast< float* >( temp.data() )[ index ] * 255.0f - mean[ c ];
     }
     }
     }
     gct/src/gct/dnn/load_image.cpp
     AlexNetޓ׵ͷೖྗը૾ͷલॲཧ
     νϟωϧΛBGRͷॱʹฒ΂ସ͑
     
     
     ImageNetͷֶशσʔλͷνϟωϧຖͷฏۉ஋ΛҾ͘
     VGGؚΉଟ͘ͷը૾ॲཧܥͷϞσϧ͕͜ͷલॲཧΛ࠾༻͍ͯ͠Δ
     

     View Slide

103. --- /home/fadis/vgg16.orig/graph.nnef 2019-05-21 21:24:49.000000000 +0900
     +++ /home/fadis/vgg16/graph.nnef 2023-07-17 20:32:06.938232809 +0900
     @@ -34,7 +34,7 @@
     variable_7 = variable(label = 'conv2_2_blob2', shape = [1, 128]);
     variable = variable(label = 'conv1_1_blob1', shape = [64, 3, 3, 3]);
     variable_3 = variable(label = 'conv1_2_blob2', shape = [1, 64]);
     - data = external(shape = [10, 3, 224, 224]);
     + data = external(shape = [1, 3, 224, 224]);
     conv = conv(data, variable, variable_1, border = 'constant', dilation = [1, 1],
     groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
     relu = relu(conv);
     conv_1 = conv(relu, variable_2, variable_3, border = 'constant', dilation = [1, 1],
     groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
     @@ -66,7 +66,7 @@
     conv_12 = conv(relu_11, variable_24, variable_25, border = 'constant', dilation =
     [1, 1], groups = 1, padding = [(1, 1), (1, 1)], stride = [1, 1]);
     relu_12 = relu(conv_12);
     max_pool_4 = max_pool(relu_12, border = 'ignore', padding = [(0, 0), (0, 0), (0,
     0), (0, 0)], size = [1, 1, 2, 2], stride = [1, 1, 2, 2]);
     - reshape = reshape(max_pool_4, shape = [10, -1]);
     + reshape = reshape(max_pool_4, shape = [1, -1]);
     linear = linear(reshape, variable_26, variable_27);
     relu_13 = relu(linear);
     linear_1 = linear(relu_13, variable_28, variable_29);
     ධՁ͸ը૾1ຕͰߦ͍͍ͨͷͰόοναΠζΛ10͔Β1ʹมߋ
     

     View Slide

104. https://github.com/KhronosGroup/NNEF-Tools/tree/main/models#nnef-model-zoo
     ͜ͷϞσϧ͸
     
     
     ೖྗͱͯ͠ը૾Λड͚औΓ
     
     
     ͦΕ͕ԿͰ͋Δ͔Λද͢
     
     
     IDΛฦ͢Α͏ʹ
     
     
     ֶश͕ͳ͞Ε͍ͯΔ
     

     View Slide

105. https://www.kaggle.com/c/imagenet-object-localization-challenge/data?select=LOC_synset_mapping.txt
     ImageNetͷ഑෍ݩ͔Β
     
     
     ͲͷID͕Կͳͷ͔ͷ
     
     
     ରԠදΛरͬͯ͘Δ
     

     View Slide

106. Ϙ΢ϧʹཛԫɺคνʔζɺࠇމᑦΛೖΕͯͨ·͕͝ۉҰʹͳΔ·ͰࠞͥΔ
     ುʹਫͱԘΛೖΕͯ෸ಅͤ͞ɺύελΛାʹॻ͔Εͨ࣌ؒ௨ΓʹᣐͰΔ
     ϑϥΠύϯͰܰ͘ম͍ͨϕʔίϯΛϘ΢ϧʹՃ͑Δ
     ುͷத਎Λ͟Δʹ͋͛ɺ͟Δͷத਎ΛϘ΢ϧʹҠ͢
     ༨೤Ͱคνʔζ͕ͱ͚Δ·ͰΑ͋͑͘Δ
     ΧϧϘφʔϥͷϏϧυखॱ
     ϐΫηϧͷΧϧϘφʔϥͷը૾ͷ׬੒
     224 × 224
     

     View Slide

107. $ dnn_eval -m ~/vgg16 -i ~/002.jpg -l ~/LOC_synset_mapping.txt
     959 carbonara 0.997535
     923 plate 0.00213662
     940 spaghetti squash 0.000186059
     937 broccoli 2.2913e-05
     762 restaurant, eating house, eating place, eatery 1.6662e-05
     809 soup bowl 1.57591e-05
     935 mashed potato 1.37334e-05
     962 meat loaf, meatloaf 1.27348e-05
     934 hotdog, hot dog, red hot 1.23476e-05
     925 consomme 6.84554e-06
     ͜ͷ෺ମ͕
     
     
     ΧϧϘφʔϥͰ͋ΔՄೳੑ
     99.75%
     

     View Slide

108. $ dnn_eval -m ~/vgg16 -i ~/001.jpg -l ~/LOC_synset_mapping.txt
     951 lemon 0.986053
     950 orange 0.0096886
     961 dough 0.001014
     954 banana 0.00058848
     928 ice cream, icecream 0.000531914
     953 pineapple, ananas 0.000395049
     949 strawberry 0.000154118
     952 fig 0.000151808
     940 spaghetti squash 0.000140983
     948 Granny Smith 0.000132113
     ͜ͷ෺ମ͕
     
     
     ϨϞϯͰ͋ΔՄೳੑ98.61%
     

     View Slide

109. $ dnn_eval -m ~/vgg16 -i ~/003.jpg -l ~/LOC_synset_mapping.txt
     784 screwdriver 0.937934
     845 syringe 0.0574703
     696 paintbrush 0.000626332
     418 ballpoint, ballpoint pen, ballpen, Biro 0.000506295
     840 swab, swob, mop 0.000496733
     629 lipstick, lip rouge 0.000402968
     749 quill, quill pen 0.000344308
     731 plunger, plumber's helper 0.000300347
     813 spatula 0.000254442
     623 letter opener, paper knife, paperknife 0.000160844
     ͜ͷ෺ମ͕
     
     
     υϥΠόʔͰ͋ΔՄೳੑ93.79%
     

     View Slide

110. ·ͱΊ
     ֶशࡁΈͷχϡʔϥϧωοτϫʔΫΛ
     
     
     ΤΫεϙʔτ͢ΔϑΝΠϧܗࣜ
     άϥϑఆٛ෦෼͸ςΩετܗࣜͳͷͰ
     
     
     ਓ͕ؒ௚઀ಡΊΔ
     

     View Slide