ABC2019 Spring

C-LIS CO., LTD.

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

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

؟ڸ່ͬͷΠϥετΛ  Πϯλʔωοτ͔ΒࣗಈͰऩू͍ͨ͠ʂ

؟ڸ່ͬ൑ఆ

؟ ڸ ͬ ່ ࠜઇΕ͍

5FOTPS'MPXൃදʢ೥݄ʣ

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

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

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

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

ը૾σʔλͷऩू

σʔλͷऔಘݩαʔϏε

IUUQTCMPHUXJUUFSDPNEFWFMPQFSFO@VTUPQJDTUPPMTOFXEFWFMPQFSSFRVJSFNFOUTUPQSPUFDUPVSQMBUGPSNIUNM

9,572 1,233 ը૾ͱΞΧ΢ϯτ͸૝૾Ҏ্ʹফ͑Δ

.BTUPEPO IUUQTQBXPPOFU IUUQTNTUEOKQ

ϞσϧͷΞʔΩςΫνϟ

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

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

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

σʔληοτ

؟ڸ່ͬɹຕ ඇ؟ڸ່ͬຕ 988ຕ

؟ڸ່ͬɹ ຕ ඇ؟ڸ່ͬ ຕ 4ສ3,432ຕ

؟ڸ່ͬɹ ຕ ඇ؟ڸ່ͬ ຕ 4ສ4,403ຕ

ऩूͨ͠ը૾σʔλ૯਺ 1,159ສ4,657ຕ (2019.05.26)

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

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

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)

ϞσϧͷධՁ

ʰྑ͍Ϟσϧʱͱ͸ͳʹ͔

ϩε͕௿͍Ϟσϧʁ

ग़ྗώετάϥϜ͕ཧ૝తͳϞσϧʁ

σʔληοτ ֶश༻ ධՁ༻ ϞσϧͷධՁ

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

େ͖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)

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

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

4FOTJUJWJUZͱ4QFDJpDJUZͷ૊Έ߹Θͤ

4FOTJUJWJUZͱ4QFDJpDJUZͷฏۉʁ 1SFDJTJPO 4FOTJUJWJUZ 4QFDJpDJUZ

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

(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౓

ػցֶशϫʔΫϑϩʔ

୯ମϞσϧ ୯ମϞσϧ [0.0 - 1.0] ग़ྗ 256x256x3

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

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

ෳ਺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

σʔληοτͷ੔උ ϞσϧͷֶशͱධՁ ϞσϧͷσϓϩΠ طଘσʔληοτͷਪ࿦

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

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

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

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

"SHP8PSLqPXʢ,VCFSOFUFTʣ

"OESPJEΞϓϦͰͷར༻

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

'PPE(BMMFSZ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPX

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

͞·͟·ͳ࣮૷ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXCSBODIFT

"OESPJE//"1*

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

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

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

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

ਫ਼౓͕ۃ୺ʹ௿Լ

ྔࢠԽ

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

ࠓޙͷ՝୊

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

"SHP8PSLqPXͷ҆ఆԽ

σʔληοτ؅ཧը໘ͷ8FCΞϓϦԽ

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͸ɺɺɺäΛׂѪ͍ͯ͠·͢ɻ 5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF

ABC2019 Spring

ABC2019 Spring

More Decks by ARIYAMA Keiji

Other Decks in Technology

Featured

Transcript