ABC2019 Spring

5月26日に開催された「Android Bazaar and Conference 2019 Spring」の発表資料です。

** このプログラムには、一部の成人の方が不愉快に感じられる表現、また未成年には不適切な表現が含まれている可能性があります。聴講には参加者または保護者の判断が必要です。 **

Transcript

1. C-LIS CO., LTD.
   

   View Slide

2. "OESPJE#B[BBSBOE$POGFSFODF4QSJOH
   
   ౦ւେֶߴྠΩϟϯύε
   ͜ͷϓϩάϥϜʹ͸ɺҰ෦ͷ੒ਓͷํ͕ෆշʹײ͡ΒΕΔදݱɺ·ͨະ
   ੒೥ʹ͸ෆద੾ͳදݱؚ͕·Ε͍ͯΔՄೳੑ͕͋Γ·͢ɻௌߨʹ͸ࢀՃऀ
   ·ͨ͸อޢऀͷ൑அ͕ඞཁͰ͢ɻ
   

   View Slide

3. C-LIS CO., LTD.
   
   
   ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ
   $-*4$0 -5%
   Photo : Koji MORIGUCHI (MORIGCHOWDER)
   "OESPJEΞϓϦ։ൃνϣοτσΩϧ
   ػցֶश͸ͪΐͬͱ΍ͬͨ͜ͱ͋Γ·͢
   ΍ͬͯ·ͤΜ
   

   View Slide

4. 
   
   ؟ڸ່ͬͷΠϥετΛ

   Πϯλʔωοτ͔ΒࣗಈͰऩू͍ͨ͠ʂ
   

   View Slide

5. ؟ڸ່ͬ൑ఆ
   
   
   

   View Slide

6. ؟
   ڸ
   ͬ
   ່
   ˜ࠜઇΕ͍
   

   View Slide

7. 5FOTPS'MPXൃදʢ೥݄ʣ
   
   
   

   View Slide

8. ϓϩάϥϛϯάͱػցֶश
   
   
   Programming
   Machine Learning
   Rules
   Data
   Answer
   Rules
   Data
   Answer
   IUUQTXXXZPVUVCFDPNXBUDI WUKT)4*(*U
   

   View Slide

9. ػցֶशΛར༻ͨ͋͠Ε͜Ε
   
   
   IUUQLJWBOUJVNIBUFCMPKQFOUSZ
   TensorFlowͰΞχϝΏΔΏΓͷ੍࡞ձࣾΛࣝผ͢Δ
   IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ
   σΟʔϓϥʔχϯάͰ͓ͦদ͞Μͷ࿡ͭࢠ͸ݟ෼͚ΒΕΔͷ͔ʁ
   IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ
   IUUQDISJTUJOBIBUFOBCMPHDPNFOUSZ
   Deep LearningͰϥϒϥΠϒʂΩϟϥΛࣝผ͢Δ
   

   View Slide

10. 
    
    

    View Slide

11. 
    
    ݱࡏ
    

    View Slide

12. 
    
    

    View Slide

13. https://twitter.com/_meganeco
    .FHBOF$P1MBZHSPVOE
    
    
    https://playground.megane.ai/
    

    View Slide

14. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    

    View Slide

15. 
    
    ը૾σʔλͷऩू
    

    View Slide

16. σʔλͷऔಘݩαʔϏε
    
    
    

    View Slide

17. 
    
    IUUQTCMPHUXJUUFSDPNEFWFMPQFSFO@VTUPQJDTUPPMTOFXEFWFMPQFSSFRVJSFNFOUTUPQSPUFDUPVSQMBUGPSNIUNM
    

    View Slide

18. 
    
    9,572
    1,233
    ը૾ͱΞΧ΢ϯτ͸૝૾Ҏ্ʹফ͑Δ
    

    View Slide

19. .BTUPEPO
    
    
    IUUQTQBXPPOFU
    IUUQTNTUEOKQ
    

    View Slide

20. 
    
    ϞσϧͷΞʔΩςΫνϟ
    

    View Slide

21. 
    
    DPOW
    YY
    YY
    SFMV
    Yͷը૾ΛYͷϑΟϧλʔͰ৞ΈࠐΉ
    ͦΜͳ࣌୅΋͋Γ·ͨ͠
    

    View Slide

22. 3FTJEVBM/FUXPSLT
    
    
    DPOW
    YY
    GD
    
    GD
    
    YY
    %SPQPVU
    
    %SPQPVU
    
    DPOW
    YY
    DPOW
    YY
    CPUUMFOFDL
    DPOW
    YY
    TUSJEF
    QPPMJOH
    SFTJEVBM@CMPDL@
    
    SFTJEVBM@CMPDL@
    
    SFTJEVBM@CMPDL@
    
    SFTJEVBM@CMPDL@
    
    SFTJEVBM@CMPDL@
    
    SFTJEVBM@CMPDL@
    
    SFMV
    SFMV
    SFMV
    SFMV
    SFMV
    CO CO
    

    View Slide

23. ֶशɾ܇࿅
    
    
    Ϟσϧ
    [0.0 - 1.0]
    1.0 or 0.0
    ग़ྗ
    ਖ਼ղσʔλ
    ʢڭࢣ৴߸ʣ
    ...
    ϥϕϧ
    4JHNPJE
    

    View Slide

24. 
    
    σʔληοτ
    

    View Slide

25. 
    
    
    ؟ڸ່ͬɹຕ
    ඇ؟ڸ່ͬຕ
    988ຕ
    
    
    

    View Slide

26. 
    
    
    ؟ڸ່ͬɹ ຕ
    ඇ؟ڸ່ͬ ຕ
    4ສ3,432ຕ
    
    
    

    View Slide

27. 
    
    
    ؟ڸ່ͬɹ ຕ
    ඇ؟ڸ່ͬ ຕ
    4ສ4,403ຕ
    
    
    

    View Slide

28. ऩूͨ͠ը૾σʔλ૯਺
    
    
    1,159ສ4,657ຕ
    (2019.05.26)
    

    View Slide

29. ը૾ͷऔಘݩ಺༁
    
    
    twitter.com 10,377,343
    pawoo.net 883,632
    mstdn.jp 311,773
    others 21,909
    (2019.05.26)
    

    View Slide

30. ϥϕϧͷछྨ
    PSJHJOBM@BSU
    OTGX
    GBWPSJUF
    QIPUP
    JMMVTU
    DPNJD
    
    
    GBDF
    GFNBMF
    NFHBOF
    TDISPPM@VOJGPSN
    CMB[FS@VOJGPSN
    TBJMPS@VOJGPSN
    HM
    LFNPOP
    NBMF
    CM
    DBU
    EPH
    GPPE
    EJTMJLF
    

    View Slide

31. 
    
    

    View Slide

32. 
    
    Positive Negative Positive Rate
    megane 18,715 25,689 0.4215
    nsfw 16,720 21,036 0.4428
    favorite 16,004 12,415 0.5631
    photo 16,136 14,010 0.5353
    illust 24,252 14,021 0.6337
    comic 4,052 14,564 0.2177
    cat 5,080 16,090 0.2400
    food 5,795 14,167 0.2903
    (2019.05.26)
    

    View Slide

33. 
    
    ϞσϧͷධՁ
    

    View Slide

34. 
    
    ʰྑ͍Ϟσϧʱͱ͸ͳʹ͔
    

    View Slide

35. 
    
    ϩε͕௿͍Ϟσϧʁ
    

    View Slide

36. 
    
    ग़ྗώετάϥϜ͕ཧ૝తͳϞσϧʁ
    

    View Slide

37. σʔληοτ
    ֶश༻
    
    ධՁ༻
    
    ϞσϧͷධՁ
    
    
    

    View Slide

38. ֶश༻
    
    ධՁ༻
    
    ϞσϧͷධՁ
    
    
    ධՁ༻
    Positive Samples
    O
    Negative Samples
    O
    
    

    View Slide

39. େ͖ͳภΓͷ͋Δσʔλ
    
    
    Positive (>= 0.5) Negative (< 0.5) Positive Rate
    megane 164,872 5,905,699 0.02715
    nsfw 827,529 5,136,725 0.1387
    (2019.05.26)
    

    View Slide

40. ධՁσʔλ
    Positive Samples
    True PositiveʢTPʣ False PositiveʢFPʣ
    Negative Samples
    True NegativeʢTNʣ False NegativeʢFNʣ
    
    
    1PTJUJWF/FHBUJWFαϯϓϧ͸ಠཱͯ͠ධՁ͢Δ
    

    View Slide

41. 4FOTJUJWJUZͱ4QFDJpDJUZ
    
    
    ˠ4FOTJUJWJUZʢײ౓ʣ
    ˠ4QFDJpDJUZʢಛҟ౓ʣ
    51'1
    51
    5/'/
    5/
    

    View Slide

42. 4FOTJUJWJUZͱ4QFDJpDJUZͷ૊Έ߹Θͤ
    
    
    

    View Slide

43. 4FOTJUJWJUZͱ4QFDJpDJUZͷ૊Έ߹Θͤ
    
    
    

    View Slide

44. 4FOTJUJWJUZͱ4QFDJpDJUZͷฏۉʁ
    
    
    1SFDJTJPO 4FOTJUJWJUZ4QFDJpDJUZ
    
    

    View Slide

45. ݱࡏͷධՁؔ਺
    
    
    WBSJBODFWBSJBODF 4FOTJUJWJUZ 4QFDJpDJUZ
    
    QSFDJTJPO 4FOTJUJWJUZ4QFDJpDJUZ
    
    QSFDJTJPOQSFDJTJPO WBSJBODF
    
    

    View Slide

46. 
    
    (2019.05.26)
    Positive Negative Positive Rate Sensitivity Specificity
    megane 18,715 25,689 0.4215 0.8917 0.9469
    nsfw 16,720 21,036 0.4428 0.8242 0.8786
    favorite 16,004 12,415 0.5631 0.8729 0.7648
    photo 16,136 14,010 0.5353 0.9764 0.9339
    illust 24,252 14,021 0.6337 0.9616 0.9539
    comic 4,052 14,564 0.2177 0.9516 0.9794
    cat 5,080 16,090 0.2400 0.9156 0.9376
    food 5,795 14,167 0.2903 0.8898 0.9691
    ֤Ϟσϧͷਫ਼౓
    

    View Slide

47. 
    
    ػցֶशϫʔΫϑϩʔ
    

    View Slide

48. ୯ମϞσϧ
    
    
    ୯ମϞσϧ
    [0.0 - 1.0]
    ग़ྗ
    256x256x3
    

    View Slide

49. ϥϕϧͷछྨ
    PSJHJOBM@BSU
    OTGX
    GBWPSJUF
    QIPUP
    JMMVTU
    DPNJD
    
    
    GBDF
    GFNBMF
    NFHBOF
    TDISPPM@VOJGPSN
    CMB[FS@VOJGPSN
    TBJMPS@VOJGPSN
    HM
    LFNPOP
    NBMF
    CM
    DBU
    EPH
    GPPE
    EJTMJLF
    

    View Slide

50. ౷߹Ϟσϧ
    
    
    ग़ྗ
    [0.0 - 1.0]
    [0.0 - 1.0]
    :
    ౷߹Ϟσϧ
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    256x256x3
    

    View Slide

51. 
    
    ෳ਺ͷ୯ମϞσϧ͔Β౷߹ϞσϧΛֶशɾ܇࿅͢Δ

    ,OPXMFEHF$PODFOUSBUJPO
    ౷߹Ϟσϧ
    ֶशࡁΈ୯ମϞσϧ
    [0.0 - 1.0]
    ਖ਼ղσʔλ
    ʢڭࢣ৴߸ʣ
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    : :
    ग़ྗ
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    [0.0 - 1.0]
    :
    256x256x3
    

    View Slide

52. 
    
    σʔληοτͷ੔උ
    ϞσϧͷֶशͱධՁ
    ϞσϧͷσϓϩΠ
    طଘσʔληοτͷਪ࿦
    

    View Slide

53. train_merge
    train_single: megane
    train_single: nsfw
    train_single: illust
    train_single: photo
    train_single: favorite
    train_single: cat
    train_single: food
    ֶशͱධՁͷϫʔΫϑϩʔ
    
    
    create_dataset
    

    View Slide

54. 
    
    ධՁ༻αʔόʔ
    ܇࿅ɾֶश༻αʔόʔ܈
    σʔληοτసૹ
    ʢTFRecordʣ ֶशࡁϞσϧऔಘ
    ը૾औಘ
    ը૾औಘ
    ϥϕϧ
    ෇͚
    σʔληοτ؅ཧ
    αʔόʔ
    ը૾ऩूݩ
    ը૾ऩू
    ϥϕϧ
    ෇͚
    σʔληοτ

    ؅ཧΞϓϦ
    playground.megane.ai
    ֶशࡁΈϞσϧ഑ஔ
    

    View Slide

55. ߴՐྗίϯϐϡʔςΟϯάαʔόʔʢݕূ࣮ݧػʣ
    $16ɹɹɹɹ9FPO&WίΞʷ
    .FNPSZɹɹ(#
    ετϨʔδɹ44%(#ʷʢ3"*%ʣ
    (16ɹɹɹ/7*%*"5*5"/9(#
    ɹɹɹɹɹ/7*%*"5*5"/9(#
    ɹɹɹɹɹ/7*%*"(F'PSDF(595J(#
    ɹɹɹɹɹ/7*%*"(F'PSDF(595J(#
    
    
    ʢDriver Version: 410.48ʣ
    

    View Slide

56. ߴՐྗίϯϐϡʔςΟϯάαʔόʔʢݕূ࣮ݧػʣ
    $16ɹɹɹɹ9FPO&WίΞʷ
    .FNPSZɹɹ(#
    ετϨʔδɹ44%(#ʷʢ3"*%ʣ
    (16ɹɹɹ".%3BEFPO7FHB(#
    ".%3BEFPO7FHB(#
    
    
    

    View Slide

57. "SHP8PSLqPXʢ,VCFSOFUFTʣ
    
    
    

    View Slide

58. 
    
    %FNP
    

    View Slide

59. 
    
    "OESPJEΞϓϦͰͷར༻
    

    View Slide

60. ͞·͟·ͳख๏
    5FOTPS'MPXGPS.PCJMF
    5FOTPS'MPX-JUF
    5FOTPS'MPX-JUFྔࢠԽϞσϧ
    5FOTPS'MPX-JUFྔࢠԽϞσϧ//ʢ/FVSBM/FUXPSLTʣ"1*
    
    
    

    View Slide

61. 
    
    

    View Slide

62. 'PPE(BMMFSZ
    
    
    IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPX
    

    View Slide

63. ৯΂෺ʢ'PPEʣ൑ผϞσϧ
    
    
    FoodϞσϧ
    [0.0 - 1.0]
    ग़ྗ
    128x128x3
    

    View Slide

64. ͞·͟·ͳ࣮૷
    
    
    IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXCSBODIFT
    

    View Slide

65. 
    
    "OESPJE//"1*
    

    View Slide

66. "OESPJE/FVSBM/FUXPSLT"1*
    
    
    IUUQTEFWFMPQFSBOESPJEDPNOELHVJEFTOFVSBMOFUXPSLT
    ϞόΠϧ୺຤্ͰػցֶशͷܭࢉॲཧΛ࣮ߦ͢ΔͨΊʹઃܭ͞Εͨ

    "OESPJE$"1*ɻ"OESPJEʢ"1*ϨϕϧʣҎ߱ͰରԠɻ
    

    View Slide

67. 
    
    IUUQTCMPHLFJKJJPUFOTPSqPXBEWFOU@DBMFOEBSIUNM
    ͍·//"1*ʢ5FOTPS'MPX-JUFʣ͸࢖͑Δͷ͔
    

    View Slide

68. ͞·͟·ͳ੍໿
    εΧϥʔͰͷԋࢉ͕Ͱ͖ͳ͍ɻʷOPSNBMJ[FE@JNBHFJNBHF@QI
    εϥΠε͕࢖͑ͳ͍ɻʷJNBHFJNBHF< >
    )ZQFSCPMJD5BOHFOU͕࢖͑ͳ͍ɻʷPVUQVUUGUBOI PVU@P⒎TFU
    
    
    
    IUUQTXXXUFOTPSqPXPSHMJUFHVJEFPQT@DPNQBUJCJMJUZVOTVQQPSUFE@PQFSBUJPOT
    

    View Slide

69. 
    
    ػछ໊ NN API͋Γ NN APIͳ͠
    Essential PH-1 556,323ns 185,372,624ns
    Pixel 2 450,807ns 187,395,464ns
    Pixel 3 477,489ns 129,994,563ns
    ਪ࿦଎౓ͷൺֱ
    IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXSFMFBTFTUBHUqJUF@OOBQJ
    

    View Slide

70. 
    
    ਫ਼౓͕ۃ୺ʹ௿Լ
    

    View Slide

71. ྔࢠԽ
    
    
    

    View Slide

72. 
    
    max: 14.586626
    min: -3.8083103
    ϞσϧΛߏ੒͍ͯ͠Δύϥϝʔλʔͷ࠷େɾ࠷খ஋Λ֬ೝ
    ྔࢠԽͷࡍʹߟྀ͢Δඞཁ͕͋Δʁ
    

    View Slide

73. 
    
    ࠓޙͷ՝୊
    

    View Slide

74. WARNING:tensorflow:From /Users/keiji_ariyama/PycharmProjects/dataset_manager/server/picture_single_discriminator/tf_model/model_res5.py:
    28: conv2d (from tensorflow.python.layers.convolutional) is deprecated and will be removed in a future version.
    Instructions for updating:
    Use keras.layers.conv2d instead.
    WARNING:tensorflow:From /Users/keiji_ariyama/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py:263:
    colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
    Instructions for updating:
    Colocations handled automatically by placer.
    WARNING:tensorflow:From /Users/keiji_ariyama/PycharmProjects/dataset_manager/server/picture_single_discriminator/tf_model/model_res5.py:
    35: batch_normalization (from tensorflow.python.layers.normalization) is deprecated and will be removed in a future version.
    Instructions for updating:
    Use keras.layers.batch_normalization instead.
    WARNING:tensorflow:From /Users/keiji_ariyama/PycharmProjects/dataset_manager/server/picture_single_discriminator/tf_model/model_res5.py:
    68: flatten (from tensorflow.python.layers.core) is deprecated and will be removed in a future version.
    Instructions for updating:
    Use keras.layers.flatten instead.
    WARNING:tensorflow:From /Users/keiji_ariyama/PycharmProjects/dataset_manager/server/picture_single_discriminator/tf_model/model_res5.py:
    74: dense (from tensorflow.python.layers.core) is deprecated and will be removed in a future version.
    Instructions for updating:
    Use keras.layers.dense instead.
    WARNING:tensorflow:From /Users/keiji_ariyama/PycharmProjects/dataset_manager/server/picture_single_discriminator/tf_model/model_res5.py:
    75: dropout (from tensorflow.python.layers.core) is deprecated and will be removed in a future version.
    
    
    5FOTPS'MPXରԠ
    

    View Slide

75. "SHP8PSLqPXͷ҆ఆԽ
    
    
    

    View Slide

76. σʔληοτ؅ཧը໘ͷ8FCΞϓϦԽ
    
    
    

    View Slide

77. 
    
    

    View Slide

78. C-LIS CO., LTD.
    ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ
    5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF
    ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺ˜ɺšɺäΛׂѪ͍ͯ͠·͢ɻ
    5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF
    

    View Slide