Tweet
Tweet

Slide 1

Slide 1 text

C-LIS CO., LTD.

Slide 2

Slide 2 text

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

Slide 3

Slide 3 text

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

Slide 4

Slide 4 text

؟ڸ່ͬͷΠϥετΛ
 Πϯλʔωοτ͔ΒࣗಈͰऩू͍ͨ͠ʂ

Slide 5

Slide 5 text

؟ڸ່ͬ൑ఆ

Slide 6

Slide 6 text

؟ ڸ ͬ ່ ˜ࠜઇΕ͍

Slide 7

Slide 7 text

5FOTPS'MPXൃදʢ೥݄ʣ

Slide 8

Slide 8 text

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

Slide 9

Slide 9 text

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

Slide 10

Slide 10 text

Slide 11

Slide 11 text

ݱࡏ

Slide 12

Slide 12 text

Slide 13

Slide 13 text

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

Slide 14

Slide 14 text

ධՁ༻αʔόʔ ܇࿅ɾֶश༻αʔόʔ܈ σʔληοτసૹ ʢTFRecordʣ ֶशࡁϞσϧऔಘ ը૾औಘ ը૾औಘ ϥϕϧ ෇͚ σʔληοτ؅ཧ αʔόʔ ը૾ऩूݩ ը૾ऩू ϥϕϧ ෇͚ σʔληοτ
 ؅ཧΞϓϦ playground.megane.ai ֶशࡁΈϞσϧ഑ஔ

Slide 15

Slide 15 text

ը૾σʔλͷऩू

Slide 16

Slide 16 text

σʔλͷऔಘݩαʔϏε

Slide 17

Slide 17 text

IUUQTCMPHUXJUUFSDPNEFWFMPQFSFO@VTUPQJDTUPPMTOFXEFWFMPQFSSFRVJSFNFOUTUPQSPUFDUPVSQMBUGPSNIUNM

Slide 18

Slide 18 text

9,572 1,233 ը૾ͱΞΧ΢ϯτ͸૝૾Ҏ্ʹফ͑Δ

Slide 19

Slide 19 text

.BTUPEPO IUUQTQBXPPOFU IUUQTNTUEOKQ

Slide 20

Slide 20 text

ϞσϧͷΞʔΩςΫνϟ

Slide 21

Slide 21 text

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

Slide 22

Slide 22 text

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

Slide 23

Slide 23 text

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

Slide 24

Slide 24 text

σʔληοτ

Slide 25

Slide 25 text

؟ڸ່ͬɹຕ ඇ؟ڸ່ͬຕ 988ຕ

Slide 26

Slide 26 text

؟ڸ່ͬɹ ຕ ඇ؟ڸ່ͬ ຕ 4ສ3,432ຕ

Slide 27

Slide 27 text

؟ڸ່ͬɹ ຕ ඇ؟ڸ່ͬ ຕ 4ສ4,403ຕ

Slide 28

Slide 28 text

ऩूͨ͠ը૾σʔλ૯਺ 1,159ສ4,657ຕ (2019.05.26)

Slide 29

Slide 29 text

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

Slide 30

Slide 30 text

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

Slide 31

Slide 31 text

Slide 32

Slide 32 text

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)

Slide 33

Slide 33 text

ϞσϧͷධՁ

Slide 34

Slide 34 text

ʰྑ͍Ϟσϧʱͱ͸ͳʹ͔

Slide 35

Slide 35 text

ϩε͕௿͍Ϟσϧʁ

Slide 36

Slide 36 text

ग़ྗώετάϥϜ͕ཧ૝తͳϞσϧʁ

Slide 37

Slide 37 text

σʔληοτ ֶश༻ ධՁ༻ ϞσϧͷධՁ

Slide 38

Slide 38 text

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

Slide 39

Slide 39 text

େ͖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)

Slide 40

Slide 40 text

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

Slide 41

Slide 41 text

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

Slide 42

Slide 42 text

4FOTJUJWJUZͱ4QFDJpDJUZͷ૊Έ߹Θͤ

Slide 43

Slide 43 text

4FOTJUJWJUZͱ4QFDJpDJUZͷ૊Έ߹Θͤ

Slide 44

Slide 44 text

4FOTJUJWJUZͱ4QFDJpDJUZͷฏۉʁ 1SFDJTJPO 4FOTJUJWJUZ4QFDJpDJUZ

Slide 45

Slide 45 text

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

Slide 46

Slide 46 text

(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౓

Slide 47

Slide 47 text

ػցֶशϫʔΫϑϩʔ

Slide 48

Slide 48 text

୯ମϞσϧ ୯ମϞσϧ [0.0 - 1.0] ग़ྗ 256x256x3

Slide 49

Slide 49 text

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

Slide 50

Slide 50 text

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

Slide 51

Slide 51 text

ෳ਺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

Slide 52

Slide 52 text

σʔληοτͷ੔උ ϞσϧͷֶशͱධՁ ϞσϧͷσϓϩΠ طଘσʔληοτͷਪ࿦

Slide 53

Slide 53 text

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

Slide 54

Slide 54 text

ධՁ༻αʔόʔ ܇࿅ɾֶश༻αʔόʔ܈ σʔληοτసૹ ʢTFRecordʣ ֶशࡁϞσϧऔಘ ը૾औಘ ը૾औಘ ϥϕϧ ෇͚ σʔληοτ؅ཧ αʔόʔ ը૾ऩूݩ ը૾ऩू ϥϕϧ ෇͚ σʔληοτ
 ؅ཧΞϓϦ playground.megane.ai ֶशࡁΈϞσϧ഑ஔ

Slide 55

Slide 55 text

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

Slide 56

Slide 56 text

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

Slide 57

Slide 57 text

"SHP8PSLqPXʢ,VCFSOFUFTʣ

Slide 58

Slide 58 text

%FNP

Slide 59

Slide 59 text

"OESPJEΞϓϦͰͷར༻

Slide 60

Slide 60 text

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

Slide 61

Slide 61 text

Slide 62

Slide 62 text

'PPE(BMMFSZ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPX

Slide 63

Slide 63 text

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

Slide 64

Slide 64 text

͞·͟·ͳ࣮૷ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXCSBODIFT

Slide 65

Slide 65 text

"OESPJE//"1*

Slide 66

Slide 66 text

"OESPJE/FVSBM/FUXPSLT"1* IUUQTEFWFMPQFSBOESPJEDPNOELHVJEFTOFVSBMOFUXPSLT ϞόΠϧ୺຤্ͰػցֶशͷܭࢉॲཧΛ࣮ߦ͢ΔͨΊʹઃܭ͞Εͨ
 "OESPJE$"1*ɻ"OESPJEʢ"1*ϨϕϧʣҎ߱ͰରԠɻ

Slide 67

Slide 67 text

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

Slide 68

Slide 68 text

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

Slide 69

Slide 69 text

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

Slide 70

Slide 70 text

ਫ਼౓͕ۃ୺ʹ௿Լ

Slide 71

Slide 71 text

ྔࢠԽ

Slide 72

Slide 72 text

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

Slide 73

Slide 73 text

ࠓޙͷ՝୊

Slide 74

Slide 74 text

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

Slide 75

Slide 75 text

"SHP8PSLqPXͷ҆ఆԽ

Slide 76

Slide 76 text

σʔληοτ؅ཧը໘ͷ8FCΞϓϦԽ

Slide 77

Slide 77 text

Slide 78

Slide 78 text

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