20190417 畳み込みニューラル ネットワークの基礎と応用

for 実践者向けディープラーニング勉強会 第二回
2019-04-17
https://dl4-practitioners.connpass.com/event/124498/

Transcript

1. 畳み込みニューラル ネットワークの基礎と応⽤ 実践者向けディープラーニング勉強会 第⼆回 https://dl4-practitioners.connpass.com/event/124498/ 2019-04-17 (Wed) @ SB C&S
2. Agenda 19:00 〜 19:05 opening 19:05 〜 19:20 機械学習の導⼊ 19:20 〜 19:30 walkthrough Kerasを⽤いたコーディングサンプル 19:30 〜 19:45 畳み込みニューラルネットワークの基礎 19:45 〜 20:00 walkthrough CNNのフィルタの理解 20:00 〜 20:10 休憩 20:10 〜 20:25 breakout CNNアーキテクチャの紹介 20:25 〜 20:40 Skymind Intelligent Layer (SKIL) の解説 20:40 〜 20:55 walkthrough SKILを⽤いたモデルのデプロイ 20:55 〜 21:00 closing !2
3. !3 15
4. ニューラルネットワークの基礎 !4
5. !5 https://www.coursera.org/learn/introduction-tensorflow
6. Activity Recognition !6 https://www.coursera.org/learn/introduction-tensorflow
7. Activity Recognition !7 https://www.coursera.org/learn/introduction-tensorflow
8. Activity Recognition !8 https://www.coursera.org/learn/introduction-tensorflow
9. Activity Recognition !9 https://www.coursera.org/learn/introduction-tensorflow
10. !10 https://www.coursera.org/learn/introduction-tensorflow
11. Activity Recognition !11 https://www.coursera.org/learn/introduction-tensorflow
12. 機械学習の"Hello World” !12 x = -2, -1, 0, 1, 2, 3, 4 y = -3, -1, 1, 3, 5, 7, 9 y = f(x)
13. 機械学習の"Hello World” !13 x = -2, -1, 0, 1, 2, 3, 4 y = -3, -1, 1, 3, 5, 7, 9 y = f(x) = 2x + 1
14. 機械学習のアプローチ •適当にモデルを初期化 (y_=ax+b) !14
15. 機械学習のアプローチ • 適当にモデルを初期化 (y_=ax+b) •誤差（損失）を計算 (L=1/N Σ(y-y_)2) !15
16. 機械学習のアプローチ • 適当にモデルを初期化 (y_=ax+b) • 誤差（損失）を計算 (L=1/N Σ(y-y_)2) •誤差が⼩さくなるようにパラメータ (a, b) を少し更新 (a ← a - η ∂L/∂a) !16
17. 機械学習のアプローチ !17 • 適当にモデルを初期化 (y_=ax+b) • 誤差（損失）を計算 (L=1/N Σ(y-y_)2) • 誤差が⼩さくなるようにパラメータ (a, b) を少し更新 (a ← a - η ∂L/∂a) •誤差（損失）を計算 (L=1/N Σ(y-y_)2)
18. 機械学習のアプローチ !18 • 適当にモデルを初期化 (y_=ax+b) • 誤差（損失）を計算 (L=1/N Σ(y-y_)2) • 誤差が⼩さくなるようにパラメータ (a, b) を少し更新 (a ← a - η ∂L/∂a) • 誤差（損失）を計算 (L=1/N Σ(y-y_)2) •誤差が⼩さくなるようにパラメータ (a, b) を少し更新 (a ← a - η ∂L/∂a) • …
19. !19
20. !20 Y º
    B
    C Z@
21. !21
22. !22
23. 勾配降下法 !23 「⽬隠しで⾜元の勾配情報のみを使って⼭の頂上を⽬指すようなもの」 学習率 η は歩幅のイメージ（η⼩＝すり⾜、η⼤＝巨⼈の⼀歩） IUUQTUXJUUFSDPNNPNJKJ@GVMMNPPOTUBUVT IUUQTXXXZBNBLFJPOMJOFDPNKPVSOBMEFUBJMQIQ JE
24. Fashion MNIST Dataset • 7万画像 • 10カテゴリ • 28×28 pixels • 実験⽤データセット !24 IUUQTHJUIVCDPN[BMBOEPSFTFBSDIGBTIJPONOJTU
25. !25
26. !26
27. !27
28. !28 ʜ ʜʜʜʜ ʜ ʜ 'MBUUFO
    
    
    
                         
29. Dense Layer の⽋点 • ⼊⼒のベクトルの全要素の相関をみている > 住宅価格予測みたいな話ならまだいい > もう作ってる特徴量と特徴量の組み合わせ - 例）東京墨⽥区 & 床⾯積 30m2 & 1K & ⾵呂トイレ別 & 新築 => 家賃⽉10万円 • 「画像の特徴量」を抽出してからDense Layerに渡せば効率的 !29 ৞ࠐΈχϡʔϥϧωοτϫʔΫ
    $POWPMVUJPOBM
    /FVSBM
    /FUXPSL
    $//
30. 畳み込みニューラルネットワークの基礎 !30
31. 畳み込み (Convolution) CURRENT_PIXEL_VALUE = 82 NEW_PIXEL_VALUE = (-1 * 144) + (0 * 60) + (-2 * 19) + (0.5 * 188) + (4.5 * 82) + (-1.5 * 32) + (1.5 * 156) + (2 * 55) + (-3 * 27) !31                   u ijm = K 1 X k=0 W 1 X p=0 H 1 X q=0 z(l 1) i+p,j+q,k h pqkm + b ijm <latexit sha1_base64="bp48ep/Dp5jA49Knv3mGnitdbqk=">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</latexit> https://news.yahoo.co.jp/byline/yuasamakoto/20190403-00120722/
32. !32                            https://www.bbkong.net/fs/alleyoop/molten_BGL7
33. stride=(1, 1), padding=ʻvalidʼ !33
34. stride=(1, 1), padding=ʻvalidʼ !34
35. stride=(1, 1), padding=ʻvalidʼ !35
36. stride=(1, 1), padding=ʻvalidʼ !36
37. stride=(1, 1), padding=ʻvalidʼ !37
38. stride=(1, 1), padding=ʻvalidʼ !38
39. stride=(1, 1), padding=ʻvalidʼ !39
40. stride=(2, 2), padding=ʻvalidʼ !40
41. stride=(2, 2), padding=ʻvalidʼ !41
42. stride=(2, 2), padding=ʻvalidʼ !42
43. stride=(1, 1), padding=ʻsameʼ !43
44. stride=(1, 1), padding=ʻsameʼ !44
45. stride=(1, 1), padding=ʻsameʼ !45
46. stride=(1, 1), padding=ʻsameʼ !46
47. Max Pooling !47 https://www.coursera.org/learn/introduction-tensorflow
48. AlexNet !48
49. !49 TMJEF
    GSPN
    ,FZOPUF
    4QFFDI
    CZ
    -BVSFODF
    .PSPOFZ
    BU
    %FFQ
    -FBSOJOH
    %BZ
    
50. !50 https://medium.com/@lmoroney_40129/codelabs-from-googlemlsummit-f9d53cac8d24
51. !51 https://medium.com/@lmoroney_40129/codelabs-from-googlemlsummit-f9d53cac8d24
52. 応⽤的なCNNアーキテクチャの紹介 !52
53. ImageNet コンペの優勝モデル !53 https://www.slideshare.net/ren4yu/ss-84282514
54. ResNet • 2015年のImageNetコンペ (ILSVRC) 優勝モデル • Residualモジュール（ショートカット機構）の導⼊ !54 http://image-net.org/challenges/talks/ilsvrc2015_deep_residual_learning_kaiminghe.pdf Revolution of Depth 3.57 6.7 7.3 11.7 16.4 25.8 28.2 ILSVRC'15 ResNet ILSVRC'14 GoogleNet ILSVRC'14 VGG ILSVRC'13 ILSVRC'12 AlexNet ILSVRC'11 ILSVRC'10 ImageNet Classification top-5 error (%) shallow 8 layers 19 layers 22 layers 152 layers Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. 8 layers
55. !55 Revolution of Depth ResNet, 152 layers 1x1 conv, 64 3x3 conv, 64 1x1 conv, 256 1x1 conv, 64 3x3 conv, 64 1x1 conv, 256 1x1 conv, 64 3x3 conv, 64 1x1 conv, 256 1x2 conv, 128, /2 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 256, /2 3x3 conv, 256 7x7 conv, 64, /2, pool/2 Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. (there was an animation here)
56. !56 Revolution of Depth ResNet, 152 layers 1x1 conv, 512 1x1 conv, 128 3x3 conv, 128 1x1 conv, 512 1x1 conv, 256, /2 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. (there was an animation here)
57. !57 Revolution of Depth ResNet, 152 layers 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. (there was an animation here)
58. !58 Revolution of Depth ResNet, 152 layers 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 256 3x3 conv, 256 1x1 conv, 1024 1x1 conv, 512, /2 3x3 conv, 512 1x1 conv, 2048 1x1 conv, 512 3x3 conv, 512 1x1 conv, 2048 1x1 conv, 512 3x3 conv, 512 1x1 conv, 2048 ave pool, fc 1000 Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. (there was an animation here)
59. ResNet !59 Deep Residual Learning • Plaint net Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015. any two stacked layers () weight layer weight layer relu relu is any desired mapping, hope the 2 weight layers fit () Deep Residual Learning • Residual net Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian hop hop weight layer weight layer relu relu = + identity () http://image-net.org/challenges/talks/ilsvrc2015_deep_residual_learning_kaiminghe.pdf
60. !60 CIFAR-10 experiments 0 1 2 3 4 5 6 0 5 10 20 iter. (1e4) error (%) plain-20 plain-32 plain-44 plain-56 20-layer 32-layer 44-layer 56-layer CIFAR-10 plain nets 0 1 2 3 4 5 6 0 5 10 20 iter. (1e4) error (%) ResNet-20 ResNet-32 ResNet-44 ResNet-56 ResNet-110 CIFAR-10 ResNets 56-layer 44-layer 32-layer 20-layer 110-layer • Deep ResNets can be trained without difficulties • Deeper ResNets have lower training error, and also lower test error solid: test dashed: train Kaiming He, Xiangyu Zhang, Shaoqing Ren, & Jian Sun. “Deep Residual Learning for Image Recognition”. arXiv 2015.
61. Further Reading • ⼀年半前 (2017-12) の資料ですが素晴らしいスライドです • 畳み込みニューラルネットワークの研究動向 by DeNA 内⽥さん • https://www.slideshare.net/ren4yu/ss-84282514 !61
62. YOLO (You Only Look Once) !62 https://www.youtube.com/watch?v=MPU2HistivI
63. YOLOのアーキテクチャ !63
64. Labeling / Annotation !64 https://github.com/Microsoft/VoTT
65. AIモデルの運⽤・デプロイ !65
66. !66 https://www.programming-fun.net/article/article_40.html
67. !67 https://www.youtube.com/watch?v=-UYyyeYJAoQ
68. モデルの利⽤ !68 O݄ؒϥϜμϊʔτ7PM
    /PΑΓ
69. モデルの利⽤ !69 O݄ؒϥϜμϊʔτ7PM
    /PΑΓ
70. !70
71. !71
72. !72
73. !73
74. !74
75. CNNモデルのデプロイ !75
76. !76
77. !77
78. !78 curl --request POST \ --url http://<SKILhostIP>:9008/login \ --header 'accept: application/json' \ --header 'content-type: application/json' \ --data '{"userId":"<userId>","password":"<password>"}'
79. Links • SKIL Download Page • https://skymind.ai/download • SKIL Documentation • https://docs.skymind.ai/docs • SKIL REST API reference • https://docs.skymind.ai/reference?newApiExplorer=true • SKIL Clients • https://github.com/SkymindIO/skil-clients !79
80. まとめ !80