모바일 얼굴인식 SDK 개발

#FacialRecognition

딥러닝 기반의 얼굴인식 모델을 처리 성능이 빈약한 모바일 및 전용장비에 적용하기 위한 최적화 방법에 대해서 공유하고자 합니다.

발표자 : andrew.js
카카오엔터프라이즈 비전인텔리전스팀의 앤드루와 포입니다. 얼굴인식을 모바일에 적용하는 일을 하고 있습니다.

po.ai
카카오엔터프라이즈 비전인텔리전스팀의 앤드루와 포입니다. 얼굴인식을 모바일에 적용하는 일을 하고 있습니다.

Transcript

1. $PQZSJHIU ,BLBP $PSQ"MMSJHIUTSFTFSWFE3FEJTUSJCVUJPOPSQVCMJDEJTQMBZJTOPUQFSNJUUFEXJUIPVUXSJUUFOQFSNJTTJPOGSPN,BLBP
   ݽ߄ੌ ঴ҷੋध 4%,ѐߊ
   JG LBLBP
   
   ஠஠য় ূఠ೐ۄ੉ૉ
   ন੿ࢳ "OESFXKT ࣠੿Ӕ 1PBJ
   

   View Slide

2. ࠗ
   ঴ҷ ੋध 4%,ࣗѐ +BOVT4%,
   
   

   View Slide

3. ,BLBP&OUFSQSJTFࢎޖप ਕ௼झܖ
   

   View Slide

4. ݽ߄ੌ ঴ҷੋध 4%,੄ ೙ਃࢿ
   ҃۝ച ߂ ࢿמ ӓ؀ച ೙ਃ
   '375,*04,ࠗ࠙ਤী ੸ਊػ ݽ؛਷ ݽ߄ੌী ੸ਊ ೞӝী ցޖ ௼׮
   ޷Ҵ Ҵ݀಴ળӝࣿোҳࣗ /*45
   ੄ ঴ҷ ੋध ߮؊ ߮஖݃௼ పझ౟
   

   View Slide

5. ݽ߄ੌ ঴ҷੋध 4%,੄ ೙ਃࢿ
   ׮নೠ ೒ۖಬ ߂ ٣߄੉झী ੸ਊ оמ
   п ٣߄੉झо о૑Ҋ ੓ח ୭Ҋ੄ ࢿמਵ۽ ز੘ ೧ঠ ೠ׮
   ৈ۞ ٣߄੉झܳ ഒ೤ ਍৔੉ оמೞ׮
   (BMBYZ5BC
   "OESPJE
   4OBQESBHPO
   ઱
   ਬפৡழޭפ౭
   "OESPJE
   *ӝ߈ /16
   "QQMFJ1BE
   J04
   Gӝ߈ /16
   (BMBYZQIPOF
   BOESPJE
   &YZOPT
   

   View Slide

6. ݽ߄ੌ ঴ҷੋध 4%,੄ ೙ਃࢿ
   ׮নೠ ೒ۖಬ ߂ ٣߄੉झী ੸ਊ оמ
   п ٣߄੉झо о૑Ҋ ੓ח ୭Ҋ੄ ࢿמਵ۽ ز੘ ೧ঠ ೠ׮
   ৈ۞ ٣߄੉झܳ ഒ೤ ਍৔੉ оמೞ׮
   (BMBYZ5BC
   "OESPJE
   4OBQESBHPO
   ઱
   ਬפৡழޭפ౭
   "OESPJE
   *ӝ߈ /16
   "QQMFJ1BE
   J04
   Gӝ߈ /16
   (BMBYZQIPOF
   BOESPJE
   &YZOPT
   

   View Slide

7. ઱ਃ ٣߄੉झ߹ प೯ ࣘب
   GQT੿بܳ ࢲ࠺झী ਗഝೠ ࣘب۽ ࠅ ࣻ ੓णפ׮
   ࢖ࢿ੹੗
   (BMBYZ5BC 4
   GQT
   ઱
   ਬפৡழޭפ౭
   6#JP9'BDF1SFNJVN
   GQT
   "QQMF *OD
   J1BE UI HFO
   
   GQT
   

   View Slide

8. ઱ਃ ٣߄੉झ߹ प೯ ࣘب
   GQT੿بܳ ࢲ࠺झী ਗഝೠ ࣘب۽ ࠅ ࣻ ੓णפ׮
   ࢖ࢿ੹੗
   (BMBYZ5BC 4
   GQT
   ઱
   ਬפৡழޭפ౭
   6#JP9'BDF1SFNJVN
   GQT
   "QQMF *OD
   J1BE UI HFO
   
   GQT
   

   View Slide

9. ࢲߡ "1*
   ױ݈ "1*
   4%,о ૑ਗ ೞח 1MBUGPSN
   $$PSF-JC
   "OESPJE
   ""3-JCSBSZ
   
   J04
   'SBNFXPSL-JCSBSZ
   
   1ZUIPO
   .PEVMF
   
   +BWB ,PUMJO
   +"3-JCSBSZ
   
   ݽ߄ੌҗ ࢲߡജ҃ী ೙ਃೠ -JCSBSZઁҕ
   ݽ߄ੌҗ ࢲߡ জ ࢚ഐ ਍৔੉ ੌ߈੸ ੉׮
   ݽ߄ੌ জ ױة ਍৔ ҳࢿ оמ
   ࢲߡ জ ױة ਍৔ ҳࢿ оמ
   ݽ߄ੌ জ
   .PCJMF"QQ
   
   ࢲߡ জ
   4FSWFS"QQ
   
   

   View Slide

10. ࢲߡ "1*
    ױ݈ "1*
    4%,о ૑ਗ ೞח 1MBUGPSN
    $$PSF-JC
    "OESPJE
    ""3-JCSBSZ
    
    J04
    'SBNFXPSL-JCSBSZ
    
    1ZUIPO
    .PEVMF
    
    +BWB ,PUMJO
    +"3-JCSBSZ
    
    ݽ߄ੌҗ ࢲߡജ҃ী ೙ਃೠ -JCSBSZઁҕ
    ݽ߄ੌҗ ࢲߡ জ ࢚ഐ ਍৔੉ ੌ߈੸ ੉׮
    ݽ߄ੌ জ ױة ਍৔ ҳࢿ оמ
    ࢲߡ জ ױة ਍৔ ҳࢿ оמ
    ݽ߄ੌ জ
    .PCJMF"QQ
    
    ࢲߡ জ
    4FSWFS"QQ
    
    

    View Slide

11. ࢲߡ "1*
    ױ݈ "1*
    4%,о ૑ਗ ೞח 1MBUGPSN
    $$PSF-JC
    "OESPJE
    ""3-JCSBSZ
    
    J04
    'SBNFXPSL-JCSBSZ
    
    1ZUIPO
    .PEVMF
    
    +BWB ,PUMJO
    +"3-JCSBSZ
    
    ݽ߄ੌҗ ࢲߡജ҃ী ೙ਃೠ -JCSBSZઁҕ
    ݽ߄ੌҗ ࢲߡ জ ࢚ഐ ਍৔੉ ੌ߈੸ ੉׮
    ݽ߄ੌ জ ױة ਍৔ ҳࢿ оמ
    ࢲߡ জ ױة ਍৔ ҳࢿ оמ
    ݽ߄ੌ জ
    .PCJMF"QQ
    
    ࢲߡ জ
    4FSWFS"QQ
    
    

    View Slide

12. ࠗ
    
    1MBUGPSN-FWFMীࢲ੄ ബਯച ౲
    

    View Slide

13. ঴ҷ ੋध 4%,੄ ݽٕ 1JQFMJOF
    

    View Slide

14. ঴ҷ ੋध 4%,੄ ݽٕ 1JQFMJOF
    

    View Slide

15. ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    1ZUPSDI
    0//9
    ,FSBT
    5FOTPSGMPX-JUF
    $POWFSUFS
    5FOTPSGMPX-JUF
    *OUFSQSFUFS
    5FOTPSGMPX-JUF
    'PSNBU
    ݽ؛ ѐߊ ױ҅ ߓನ ױ҅
    $PSFNMNPEFM
    $PSF.-
    $PSFNM UPPMT
    

    View Slide

16. ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    1ZUPSDI
    0//9
    ,FSBT
    5FOTPSGMPX-JUF
    $POWFSUFS
    5FOTPSGMPX-JUF
    *OUFSQSFUFS
    5FOTPSGMPX-JUF
    'PSNBU
    ݽ؛ ѐߊ ױ҅ ߓನ ױ҅
    $PSFNMNPEFM
    $PSF.-
    $PSFNM UPPMT
    

    View Slide

17. ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    1ZUPSDI
    0//9
    ,FSBT
    5FOTPSGMPX-JUF
    $POWFSUFS
    5FOTPSGMPX-JUF
    *OUFSQSFUFS
    5FOTPSGMPX-JUF
    'PSNBU
    ݽ؛ ѐߊ ױ҅ ߓನ ױ҅
    $PSFNMNPEFM
    $PSF.-
    $PSFNM UPPMT
    

    View Slide

18. 5FOTPSGMPX-JUFܳ ݽ߄ੌ ೐ۨ੐ਕ௼۽ ࢶఖೠ ੉ਬ
    -BUFODZ ࢿמ੉ જ׮
    ׮নೠ ߔূ٘ #BDLFOE
    ૑ਗ
    DQV HQV OOBQJ YOOQBDL DPSFNM FUD
    "SNӝ߈ Gীࢲח YOOQBDLӝ߈ CBDLFOEо જ਷ ࢿמ
    ୭Ӕ੄ %FWJDFীࢲ HQV ӝ߈ ࢿמ੉ ੼੼ જই ૑Ҋ੓׮
    //"1* /FVSBM/FUXPSL"1*
    ח উ٘۽੉٘੄ ಴ળ ݠन۞׬ झಖ੉׮
    ࢚؀੸ਵ۽ ߊৌী ஂডೞ૑ ঋ׮
    উ੿੸ੋ ন੗ച 2VBOUJ[BUJPO
    ૑ਗ
    ׮নೠ ೒ۖಬ J04 "OESPJE 4FSWFS
    ૑ਗ
    ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    

    View Slide

19. 5FOTPSGMPX-JUFܳ ݽ߄ੌ ೐ۨ੐ਕ௼۽ ࢶఖೠ ੉ਬ
    -BUFODZ ࢿמ੉ જ׮
    ׮নೠ ߔূ٘ #BDLFOE
    ૑ਗ
    DQV HQV OOBQJ YOOQBDL DPSFNM FUD
    "SNӝ߈ Gীࢲח YOOQBDLӝ߈ CBDLFOEо જ਷ ࢿמ
    ୭Ӕ੄ %FWJDFীࢲ HQV ӝ߈ ࢿמ੉ ੼੼ જই ૑Ҋ੓׮
    //"1* /FVSBM/FUXPSL"1*
    ח উ٘۽੉٘੄ ಴ળ ݠन۞׬ झಖ੉׮
    ࢚؀੸ਵ۽ ߊৌী ஂডೞ૑ ঋ׮
    উ੿੸ੋ ন੗ച 2VBOUJ[BUJPO
    ૑ਗ
    ׮নೠ ೒ۖಬ J04 "OESPJE 4FSWFS
    ૑ਗ
    ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    

    View Slide

20. 5FOTPSGMPX-JUFܳ ݽ߄ੌ ೐ۨ੐ਕ௼۽ ࢶఖೠ ੉ਬ
    -BUFODZ ࢿמ੉ જ׮
    ׮নೠ ߔূ٘ #BDLFOE
    ૑ਗ
    DQV HQV OOBQJ YOOQBDL DPSFNM FUD
    "SNӝ߈ Gীࢲח YOOQBDLӝ߈ CBDLFOEо જ਷ ࢿמ
    ୭Ӕ੄ %FWJDFীࢲ HQV ӝ߈ ࢿמ੉ ੼੼ જই ૑Ҋ੓׮
    //"1* /FVSBM/FUXPSL"1*
    ח উ٘۽੉٘੄ ಴ળ ݠन۞׬ झಖ੉׮
    ࢚؀੸ਵ۽ ߊৌী ஂডೞ૑ ঋ׮
    উ੿੸ੋ ন੗ച 2VBOUJ[BUJPO
    ૑ਗ
    ׮নೠ ೒ۖಬ J04 "OESPJE 4FSWFS
    ૑ਗ
    ٩۞׬ ೐ۨ੐ਕ௼ %- 'SBNFXPSL
    
    

    View Slide

21. %- ݽ؛੄ ѐࣻܳ ઴ৈۄ
    ੋध /FUXPSLҗ ݃झ௼ Ѩ୹ /FUXPSLਸ ా೤
    Ѣ੄ ݃झ௼ Ѩ୹ /FUXPSL ݅ఀ੄ -BUFODZਸ ੺ড ೡ ࣻ ੓঻׮
    ୶о ௏झ౟о Ѣ੄ হ੉ ׮ܲ ࣘࢿٜب ୶о ೡ ࣻ ੓঻׮
    

    View Slide

22. %- ݽ؛੄ ѐࣻܳ ઴ৈۄ
    ੋध /FUXPSLҗ ݃झ௼ Ѩ୹ /FUXPSLਸ ా೤
    Ѣ੄ ݃झ௼ Ѩ୹ /FUXPSL ݅ఀ੄ -BUFODZਸ ੺ড ೡ ࣻ ੓঻׮
    ୶о ௏झ౟о Ѣ੄ হ੉ ׮ܲ ࣘࢿٜب ୶о ೡ ࣻ ੓঻׮
    

    View Slide

23. /&0/ਸ ੸ӓ੸ਵ۽ ഝਊೞۄ
    /&0/਷ "SN੉ ૑ਗೞח 4*.% 4JOHMF*OTUSVDUJPO.VMUJQMF%BUB
    
    2
    2
    2
    
    4PVSDF
    3FHJTUFS
    %FTUJOBUJPO
    3FHJTUFS
    7"%%2G
    

    View Slide

24. /&0/ਸ ੉޷૑ ୊ܻ *NBHF1SPDFTTJOH
    ী ੸ਊ
    Image(YUV) Image(RGB)
    (BMBYZ4 6CJP9
    /&0/੹ N4 N4
    /&0/റ N4 N4
    $PMPSNPEFMDPOWFSTJPO
    3PUBUF'MJQ
    (BMBYZ4 6CJP9
    ঴ҷ ৔৉ Ѩ୹ӝ
    ੹ N4 N4
    റ N4 N4
    ঴ҷ ے٘݃௼ Ѩ୹ӝ
    ੹ N4 N4
    റ N4 N4
    ঴ҷ ਤઑ х૑ӝ
    ੹ N4 N4
    റ N4 N4
    ౠ૚ ࢤࢿӝ
    ੹ N4 N4
    റ N4 N4
    1JYFM/PSNBMJ[BUJPO
    
    à
    PS
    
    N4ب ࣗ઺೤פ׮
    

    View Slide

25. /&0/ਸ ੉޷૑ ୊ܻ *NBHF1SPDFTTJOH
    ী ੸ਊ
    Image(YUV) Image(RGB)
    (BMBYZ4 6CJP9
    /&0/੹ N4 N4
    /&0/റ N4 N4
    $PMPSNPEFMDPOWFSTJPO
    3PUBUF'MJQ
    (BMBYZ4 6CJP9
    ঴ҷ ৔৉ Ѩ୹ӝ
    ੹ N4 N4
    റ N4 N4
    ঴ҷ ے٘݃௼ Ѩ୹ӝ
    ੹ N4 N4
    റ N4 N4
    ঴ҷ ਤઑ х૑ӝ
    ੹ N4 N4
    റ N4 N4
    ౠ૚ ࢤࢿӝ
    ੹ N4 N4
    റ N4 N4
    1JYFM/PSNBMJ[BUJPO
    
    à
    PS
    
    N4ب ࣗ઺೤פ׮
    

    View Slide

26. 1JYFM/PSNBMJ[BUJPO୭੸ച
    ా೤ оמೠ ো࢑ਸ ޷ܻ ࣻ೯ 0GGMJOF
    
    #BUDI /PSNBMJ[BUJPO-BZFS੄ GVTJOHҗ ਬࢎ
    2VBOUJ[BUJPOਸ Ҋ۰ೞৈ ࣗࣻ੼੉ ߊࢤೞ૑ ঋח NFBO
    TUE
    ਸ ࢎਊೣ
    ੉
    ޷
    ૑
    QJYFM@NFBO
    ࡐӝ
    QJYFM@TUE
    ա־ӝ
    QJYFM@NFBO
    ࡐӝ
    88TUE
    ੉
    ޷
    ૑
    ೟ण җ੿ 5SBJOJOH5JNF
    
    ѐࢶ ੹ *OGFSFODF5JNF
    ѐࢶ റ *OGFSFODF5JNF
    

    View Slide

27. 1JYFM/PSNBMJ[BUJPO୭੸ച
    ా೤ оמೠ ো࢑ਸ ޷ܻ ࣻ೯ 0GGMJOF
    
    #BUDI /PSNBMJ[BUJPO-BZFS੄ GVTJOHҗ ਬࢎ
    2VBOUJ[BUJPOਸ Ҋ۰ೞৈ ࣗࣻ੼੉ ߊࢤೞ૑ ঋח NFBO
    TUE
    ਸ ࢎਊೣ
    ੉
    ޷
    ૑
    QJYFM@NFBO
    ࡐӝ
    QJYFM@TUE
    ա־ӝ
    QJYFM@NFBO
    ࡐӝ
    88TUE
    ੉
    ޷
    ૑
    ೟ण җ੿ 5SBJOJOH5JNF
    
    ѐࢶ ੹ *OGFSFODF5JNF
    ѐࢶ റ *OGFSFODF5JNF
    

    View Slide

28. ࠗ
    ݽ؛ ন੗ച 2VBOUJ[BUJPO
    ੸ਊद੄ ౲
    

    View Slide

29. ݽ؛ ন੗ച 2VBOUJ[BUJPO
    ী ੄ೠ -BUFODZ ѐࢶ
    ֙ ੉റ੄ ٣߄੉झীࢲח JOUҗ G੄ ࢿמ ର੉о ഛোೞ׮
    (16ࠁ׮ب જ׮
    model precision cpu xnnpack gpu nnapi
    Galaxy S7
    f32 32.27 19.50 25.42 32.52
    i8 24.89 24.68 17.38 24.63
    Galaxy S10
    f32 22.36 11.97 12.25 29.41
    i8 14.26 15.27 13.19 16.15
    Galaxy Note20
    f32 12.11 9.31 7.50 14.50
    i8 5.55 5.20 5.81 6.17
    Galaxy S21
    f32 10.36 6.95 5.43 18.51
    i8 4.1 4.4 5.63 8.09
    .PCJMFOFU 7੄ -BUFODZஏ੿ NJMMJ4FDPOET
    
    

    View Slide

30. ݽ؛ ন੗ച 2VBOUJ[BUJPO
    ী ੄ೠ -BUFODZ ѐࢶ
    ֙ ੉റ੄ ٣߄੉झীࢲח JOUҗ G੄ ࢿמ ର੉о ഛোೞ׮
    (16ࠁ׮ب જ׮
    model precision cpu xnnpack gpu nnapi
    Galaxy S7
    f32 32.27 19.50 25.42 32.52
    i8 24.89 24.68 17.38 24.63
    Galaxy S10
    f32 22.36 11.97 12.25 29.41
    i8 14.26 15.27 13.19 16.15
    Galaxy Note20
    f32 12.11 9.31 7.50 14.50
    i8 5.55 5.20 5.81 6.17
    Galaxy S21
    f32 10.36 6.95 5.43 18.51
    i8 4.1 4.4 5.63 8.09
    .PCJMFOFU 7੄ -BUFODZஏ੿ NJMMJ4FDPOET
    
    

    View Slide

31. ন੗ച 2VBOUJ[BUJPO
    ী ஘ച੸ ݽ؛ ࢸ҅ о੉٘
    ֎౟ਕ௼੄ %FQUIо ծਸ ࣻ۾ ਬܻ
    8JEUIо ևਸ ࣻ۾ ਬܻ
    -BZFS
    -BZFS
    -BZFS
    

    View Slide

32. ন੗ച 2VBOUJ[BUJPO
    ী ஘ച੸ ݽ؛ ࢸ҅ о੉٘
    ֎౟ਕ௼੄ %FQUIо ծਸ ࣻ۾ ਬܻ
    8JEUIо ևਸ ࣻ۾ ਬܻ
    -BZFS
    -BZFS
    -BZFS
    

    View Slide

33. ন੗ച 2VBOUJ[BUJPO
    ী ஘ച੸ ݽ؛ ࢸ҅ о੉٘
    ਋ഥ҃۽ 4IPSU$POOFDUJPO
    ੉ ੓ਸ ࣻ۾ ਬܻೞ׮
    -BZFS
    -BZFS
    -BZFS
    
    
    
    
    

    View Slide

34. ۨ੉য ࠙ೡ۽ ন੗ച ࢿמ ೞۅ ѐࢶ
    ୹۱ ౠࢿ੉ ׮ܲ ۨ੉য੄ ޛܻ੸ ࠙ೡ੉ ন੗ചী ਬܻೡ ࣻ ੓׮ 1PTU5SBJOJOH4QMJU
    
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<7x128>
    Bias<7>
    4MJDF
    begin<0>
    end<6>
    4MJDF
    begin<6>
    end<7>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<6x128>
    Bias<6>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<1x128>
    Bias<1>
    변경
    -128 127
    -128 127
    -128 127
    f32
    f32 f32
    

    View Slide

35. ۨ੉য ࠙ೡ۽ ন੗ച ࢿמ ೞۅ ѐࢶ
    ୹۱ ౠࢿ੉ ׮ܲ ۨ੉য੄ ޛܻ੸ ࠙ೡ੉ ন੗ചী ਬܻೡ ࣻ ੓׮ 1PTU5SBJOJOH4QMJU
    
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<7x128>
    Bias<7>
    4MJDF
    begin<0>
    end<6>
    4MJDF
    begin<6>
    end<7>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<6x128>
    Bias<6>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<1x128>
    Bias<1>
    변경
    -128 127
    -128 127
    -128 127
    f32
    f32 f32
    

    View Slide

36. ۨ੉য ࠙ೡ۽ ন੗ച ࢿמ ೞۅ ѐࢶ
    ୹۱ ౠࢿ੉ ׮ܲ ۨ੉য੄ ޛܻ੸ ࠙ೡ੉ ন੗ചী ਬܻೡ ࣻ ੓׮ 1PTU5SBJOJOH4QMJU
    
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<7x128>
    Bias<7>
    4MJDF
    begin<0>
    end<6>
    4MJDF
    begin<6>
    end<7>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<128x256>
    bias<128>
    'VMMZ$POOFDUFE
    weights<6x128>
    Bias<6>
    -PHJTUJD
    $PODBU
    'VMMZ$POOFDUFE
    weights<1x128>
    Bias<1>
    변경
    -128 127
    -128 127
    -128 127
    f32
    f32 f32
    

    View Slide

37. .PEFM0QUJNJ[BUJPO
    ࣘب৬ ಌನݢझ ࢎ੉੄ ୭੸੼ ଺ӝ
    

    View Slide

38. -PX-BUFODZ )JHI1FSGPSNBODF
    Computation
    Performance
    

    View Slide

39. ࠗ
    0WFSBMM.PEFM0QUJNJ[BUJPO1SPDFTT
    

    View Slide

40. 0WFSBMM1SPDFTT
    4FMFDU5BSHFU%FWJDF 4FBSDI"SDIJUFDUVSF 1FSGPSNBODF0QUJNJ[BUJPO
    

    View Slide

41. 0WFSBMM1SPDFTT
    4FMFDU5BSHFU%FWJDF 4FBSDI"SDIJUFDUVSF 1FSGPSNBODF0QUJNJ[BUJPO
    

    View Slide

42. ࠗ
    $PNNPO.PEFM0QUJNJ[BUJPO)FVSJTUJDT
    

    View Slide

43. 4JNQMF$PNQVUBUJPOBM(SBQI 3FEVDF$PNQVUBUJPO 6TF0QUJNJ[FE0QT
    

    View Slide

44. 4JNQMF$PNQVUBUJPOBM(SBQI 3FEVDF$PNQVUBUJPO 6TF0QUJNJ[FE0QT
    

    View Slide

45. ࠗ
    %FWJDFBXBSF.PEFM 0QUJNJ[BUJPO)FVSJTUJDT
    

    View Slide

46. $16*OGFSFODF
    10
    15
    20
    25
    30
    120 180 240 300
    Latency(ms)
    FLOPS(M)
    Lower FLOPS, Lower Latency
    

    View Slide

47. (16/16*OGFSFODF
    Lower Complexity, Lower Latency
    Better
    Worse
    

    View Slide

48. ࠗ
    %FTJHO4USVDUVSF
    

    View Slide

49. "SDIJUFDUVSF੄ ҳࢿ ਃࣗ
    

    View Slide

50. %FTJHO
    input
    output
    depth
    width
    resolution
    

    View Slide

51. .JDSPBSDIJUFDUVSF
    

    View Slide

52. %FTJHOPG"SDIJUFDUVSF
    

    View Slide

53. 4MJN'BDF %FTJHO
    'BDF3FDPHOJUJPO਷ ҕѐػ QSFUSBJOFECBDLCPOFٜ੄ ೟ण ജ҃җ ௼ѱ ׮ܰ׮
    ୭੸੄ EFTJHOਸ ਤ೧ࢲח → Re-design
    

    View Slide

54. 4MJN'BDF %FTJHO
    most effective receptive field
    most effective receptive field
    Total Capacity
    input output
    1x 2x 4x 8x
    1x 2x 4x 8x
    Receptive Field의 디자인이 전체 디자인의 performance를 결정
    

    View Slide

55. ࠗ
    %FTJHO.JDSPBSDIJUFDUVSF
    

    View Slide

56. .JDSPBSDIJUFDUVSF%FTJHO
    

    View Slide

57. 5XP5SBDLT
    

    View Slide

58. SimpleFace
    NAS
    

    View Slide

59. 'PS$16/"4 /FVSBM"SDIJUFDUVSF4FBSDI
    
    /"4ۆ
    ઱য૓ ܻࣗझ ؘ੉ఠ కझ௼ী ୭੸ചػ
    PQFSBUJPOҗ TUSVDUVSFܳ ঌҊ્ܻਵ۽ ఐ࢝
    DPNQVUBUJPOBMHSBQI੄ п OPEFীࢲ যڃ
    PQFSBUJPOਸ ࢎਊ೧ ׮਺ OPEF۽ ૓೯ೡ૑ Ѿ੿
    

    View Slide

60. 'PS$16/"4 /FVSBM"SDIJUFDUVSF4FBSDI
    
    1/" 1SPYZMFTT /FVSBM"SDIJUFDUVSF
    "EBQUBUJPO
    
    
    TFBSDITQBDFо ևয૑ݶ /"4੄ ബਯ਷ хࣗ
    ӝઓ IVNBOEFTJHOFENPEVMFীࢲ
    DPOOFDUJPO੄ ਬޖ݅ਸ TFBSDIೞৈ ࢿמ ೱ࢚
    1) D.G. Kim and H.C. Lee, "Proxyless Neural Architecture Adaptation at Once," in IEEE Access, 2022
    

    View Slide

61. 'PS$16/"4 /FVSBM"SDIJUFDUVSF4FBSDI
    
    1/" 1SPYZMFTT /FVSBM"SDIJUFDUVSF
    "EBQUBUJPO
    
    
    TFBSDITQBDFо ևয૑ݶ /"4੄ ബਯ਷ хࣗ
    ӝઓ IVNBOEFTJHOFENPEVMFীࢲ
    DPOOFDUJPO੄ ਬޖ݅ਸ TFBSDIೞৈ ࢿמ ೱ࢚
    1) D.G. Kim and H.C. Lee, "Proxyless Neural Architecture Adaptation at Once," in IEEE Access, 2022
    

    View Slide

62. 'PS$16/"4 /FVSBM"SDIJUFDUVSF4FBSDI
    
    1/" 1SPYZMFTT /FVSBM"EBQUBUJPO
    
    CBTFMJOF ؀࠺ TLJQDPOOFDUJPOѐ ୶о
    زੌ ো࢑۝ীࢲ ӝઓ 4MJN'BDF ؀࠺ ੋध पಁਯ хࣗ
    

    View Slide

63. 'PS(16/164JNQMF'BDF
    Slower
    Better
    Performance
    Faster
    Moderate
    Performance
    

    View Slide

64. 'PS(16/164JNQMF'BDF
    Training Inference
    Better
    Performance
    Faster
    

    View Slide

65. 'PS(16/164JNQMF'BDF
    4MJN'BDF ؀࠺ زੌ ۨ੉ఢदীࢲ ੋध पಁਯ хࣗ
    4MJN'BDF ؀࠺ زੌ ࢿמীࢲ JOGFSFODFद ࣘب ߓ ૐо
    

    View Slide

66. ࠗ
    ,OPXMFEHF%JTUJMMBUJPOUP*NQSPWF1FSGPSNBODF
    

    View Slide

67. ,OPXMFEHF%JTUJMMBUJPO
    Teacher(Pre-trained)
    Student(From scratch)
    Knowledge Distillation
    teacher에서 student로 지식(knowledge)을 증류(distill)하여 전달
    

    View Slide

68. ,OPXMFEHF%JTUJMMBUJPO
    feature extractor(=model) embedding classifier
    logit
    얼굴에 대한 정보를 추상적으로 함축
    학습 데이터의 identity를 확률 분포로 표현
    knowledge to distill
    ঴ҷ ੋध੄ ੹୓ ౵੉೐ۄੋীࢲ ୶୹ೡ LOPXMFEHFܳ Ѿ੿
    

    View Slide

69. ,OPXMFEHF%JTUJMMBUJPO
    hard soft
    student의 학습을 방해하지 않기 위해 smoothing을 통해 soft한 지식을 전달
    

    View Slide

70. ,OPXMFEHF%JTUJMMBUJPO
    UFBDIFS৬ TUVEFOU੄ ର੉о ௼ݶ TUVEFOUо ઁ؀۽ ߓ਋૑ ޅೞӝ ٸޙী TUVEFOUо ୽࠙൤ ழঠ ೞח
    ޙઁ
    ؊ ֫਷ ࢿמਸ о૓ UFBDIFSࠁ׮ زੌೠ ࢿמਸ о૓ UFBDIFSܳ ࢎਊೞח Ѫ੉ ؊ જ਷ ബҗ
    EJTUJMMBUJPOਸ ૓೯ೠ TUVEFOUܳ ੤ର TUVEFOU۽ ॄࢲ ৈ۞ ߣ EJTUJMMBUJPOೡ ҃਋ ࢿמ ೱ࢚
    small gap similar performance distill multiple times
    

    View Slide

71. хࢎ೤פ׮
    

    View Slide