DevFest Tokyo 2017 TensorFlow

Slide 1

Slide 1 text

C-LIS CO., LTD.

Slide 2

Slide 2 text

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

Slide 3

Slide 3 text

%FW'FTU5PLZP 5FOTPS'MPXͰ  झຯͷը૾ऩूαʔόʔΛ࡞Δ ೥݄߸

Slide 4

Slide 4 text

લճ·Ͱͷ͓࿩

Slide 5

Slide 5 text

5FOTPS'MPXൃදʢ೥݄ʣ

Slide 6

Slide 6 text

ษڧձ΍Ζ͏ͥ

Slide 7

Slide 7 text

(PPHMF%FWFMPQFS(SPVQ

Slide 8

Slide 8 text

IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT

Slide 9

Slide 9 text

Πϯλʔωοτ͔Β ޷Έͷը૾ΛࣗಈͰऩू͍ͨ͠

Slide 10

Slide 10 text

Slide 11

Slide 11 text

γεςϜߏ੒ Downloader Face Detection Megane Detection ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ֶशʢ܇࿅ʣ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ֶशʢ܇࿅ʣ TensorFlow

Slide 12

Slide 12 text

{ "generator": "Region Cropper", "file_name": "haruki_g17.png", "regions": [ { "probability": 1.0, "label": 2, "rect": { "left": 97.0, "top": 251.0, "right": 285.0, "bottom": 383.0 } }, { "probability": 1.0, "label": 2, "rect": { "left": 536.0, "top": 175.0, "right": 730.0, "bottom": 321.0 } } ] } Region Cropper: https://github.com/keiji/region_cropper

Slide 13

Slide 13 text

؟ڸ່ͬ൑ఆ conv 3x3x3 32 conv 3x3x32 64 max_pooling lrn conv 3x3x64  128 conv 3x3x128 128 max_pooling lrn conv 3x3x128  128 conv 3x3x128 256 fc 384 fc 256 output 2

Slide 14

Slide 14 text

ͷ൑ఆਫ਼౓ 1 0

Slide 15

Slide 15 text

ٕज़ܥಉਓࢽΦϯϦʔΠϕϯτ ٕज़ॻయ̍ˍ̎ https://techbookfest.org/tbf01/#A-28 https://techbookfest.org/event/tbf02/circle/5705718560718848

Slide 16

Slide 16 text

੔ཧ଴ͪͷը૾ ຕ ʢ2017.10.08 ࣌఺ʣ

Slide 17

Slide 17 text

σʔληοτͷ੔උʢ੔ཧʣ͕  ௚ۙͷ՝୊

Slide 18

Slide 18 text

Πϥετإݕग़ʢ'BDF%FUFDUJPOʣ

Slide 19

Slide 19 text

γεςϜߏ੒ Downloader Face Detection Megane Detection ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ֶशʢ܇࿅ʣ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ֶशʢ܇࿅ʣ TensorFlow

Slide 20

Slide 20 text

3$// 3FHJPOCBTFE$POWPMVUJPOBM/FVSBM/FUXPSLT

Slide 21

Slide 21 text

4MJEJOH8JOEPX

Slide 22

Slide 22 text

4FMFDUJWF4FBSDI 4FMFDUJWF4FBSDIGPS0CKFDU3FDPHOJUJPO IUUQXXXIVQQFMFOOMQVCMJDBUJPOTTFMFDUJWF4FBSDI%SBGUQEG 848

Slide 23

Slide 23 text

%FW'FTU5PLZP 5FOTPS'MPXͰ Πϥετإݕग़ثΛ࡞Δ

Slide 24

Slide 24 text

ֶशσʔλͷ࡞੒

Slide 25

Slide 25 text

࠲ඪͷਖ਼ଇԽɹਖ਼نԽ 888x613 left : 536 → 0.6036 top : 175 → 0.2854 right : 730 → 0.8220 bottom : 321 → 0.5236 0.1092 ← 97 : left 0.4094 ← 251 : top 0.3209 ← 285 : right 0.6247 ← 383 : bottom

Slide 26

Slide 26 text

ը૾ͷϦαΠζɾάϨʔεέʔϧԽ 256x256 left : 0.6036 top : 0.2854 right : 0.8220 bottom : 0.5236 0.1092 : left 0.4094 : top 0.3209 : right 0.6247 : bottom

Slide 27

Slide 27 text

Ϩίʔυߏ଄ ը૾σʔλʢόΠτྻʣ      

Slide 28

Slide 28 text

XJUIUGQZUIPO@JP5'3FDPSE8SJUFS PVUQVU@QBUI PQUJPOT BTXSJUFS FYBNQMFUGUSBJO&YBNQMF GFBUVSFTUGUSBJO'FBUVSFT GFBUVSF\ GJMF@OBNF@CZUFT@GFBUVSF GJMF@OBNFFODPEF JNBHF@CZUFT@GFBUVSF JNBHF@EBUB SFHJPO@DPVOU@JOU@GFBUVSF SFHJPO@DPVOU SFHJPOT@GMPBU@GFBUVSF GMBUUFO@SFHJPOT MBCFMT@JOU@MJTU@GFBUVSF MBCFMT ^ XSJUFSXSJUF FYBNQMF4FSJBMJ[F5P4USJOH

Slide 29

Slide 29 text

Ϟσϧ conv 5x5x1  128 batch_normalization conv 1x1x128  64 conv 3x3x64  64 batch_normalization conv 1x1x64  128 ReLU conv 1x1x128  128 conv 3x3x128  128 batch_normalization conv 1x1x128  128 ReLU conv 1x1x256  256 conv 3x3x256  256 batch_normalization conv 1x1x256  128 ReLU conv 2x2x128  128 fc 2048 fc 4 256x256x1 Residual block 1 Residual block 2 Residual block 3 × 3 × 3 × 3 avg_pool 2x2 avg_pool 2x2 avg_pool 2x2 ReLU Sigmoid

Slide 30

Slide 30 text

ೖྗͱग़ྗ 256x256x1 Ϟσϧ Sigmoid

Slide 31

Slide 31 text

ReLU Sigmoid ׆ੑԽؔ਺

Slide 32

Slide 32 text

256x256 left : 0.6036 top : 0.2854 right : 0.8220 bottom : 0.5236 0.1092 : left 0.4094 : top 0.3209 : right 0.6247 : bottom ࠲ඪͷ஋Ҭʢʣ

Slide 33

Slide 33 text

ReLU Sigmoid ׆ੑԽؔ਺

Slide 34

Slide 34 text

܇࿅ɾֶश ϛχόον ֶश཰ʢ"EBN0QUJNJ[FSʣ

Slide 35

Slide 35 text

ݕग़݁Ռ

Slide 36

Slide 36 text

̎ਓҎ্͍Δ৔߹͸ʁ

Slide 37

Slide 37 text

Slide 38

Slide 38 text

σʔλ಺༁ 0 (18.57%) 1 (57.64%) 2 (11.93%) إͷ਺ ϑΝΠϧ਺ ׂ߹ 0 3748 18.57% 1 11636 57.64% 2 2408 11.93% 3 844 4.18% 4 596 2.95% 5 325 1.61% 6 214 1.06% 7 108 0.53%

Slide 39

Slide 39 text

ೖྗͱग़ྗ 256x256x1 Ϟσϧ Sigmoid

Slide 40

Slide 40 text

إྖҬͷݕग़

Slide 41

Slide 41 text

إྖҬͷݕग़ dconv 3x3x1 64 dconv 3x3x64 64 dconv 3x3x256 128 dconv 3x3x64 128 dconv 3x3x128 128 conv 3x3x1 64 conv 3x3x64 64 conv 3x3x128 256 conv 3x3x64 128 conv 3x3x128 128

Slide 42

Slide 42 text

إྖҬͷݕग़

Slide 43

Slide 43 text

... ৞ΈࠐΈ૚ʢDPOWEʣͷ໾ׂ ಛ௃Ϛοϓ

Slide 44

Slide 44 text

إྖҬͷݕग़ʢ৞ΈࠐΈΦʔτΤϯίʔμʣ dconv 3x3x3 64 dconv 3x3x64 64 dconv 3x3x256 128 dconv 3x3x64 128 dconv 3x3x128 128 conv 3x3x1 64 conv 3x3x64 64 conv 3x3x128 256 conv 3x3x64 128 conv 3x3x128 128

Slide 45

Slide 45 text

৞ΈࠐΈ૚ͷΈʹͨ͠৔߹ conv 3x3x1 64 conv 3x3x64 64 conv 3x3x128 256 conv 3x3x64 128 conv 3x3x128 128 dconv 3x3x3 64 dconv 3x3x64 64 dconv 3x3x256 128 dconv 3x3x64 128 dconv 3x3x128 128

Slide 46

Slide 46 text

৞ΈࠐΈ૚ͷΈʹͨ͠৔߹ conv 3x3x64 stride 1 conv 3x3x64  stride 1 ReLU ReLU conv 3x3x128  stride 1 conv 3x3x128  stride 1 ReLU conv 3x3x256  stride 1 conv 3x3x256  stride 1 ReLU conv 1x1x1  stride 1 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU Sigmoid 29x29x1

Slide 47

Slide 47 text

άϦουʢYʣ

Slide 48

Slide 48 text

άϦουͷͲ͜ʹإ͕͋Δ͔

Slide 49

Slide 49 text

࠲ඪͷதԝͷηϧΛબ୒

Slide 50

Slide 50 text

֤ηϧʹσʔλΛઃఆ confidence : 1.0 center_x : 0.7128 center_y : 0.4045 width : 0.2184 height : 0.2382 1.0 : confidence 0.2150 : center_x 0.5169 : center_y 0.2117 : width 0.2151 : height confidence : 0.0 center_x : 0.0 center_y : 0.0 width : 0.0 height : 0.0

Slide 51

Slide 51 text

Ϩίʔυߏ଄ ը૾σʔλʢόΠτྻʣ      

Slide 52

Slide 52 text

Ϟσϧ conv 3x3x64 stride 1 conv 3x3x64  stride 1 ReLU ReLU conv 3x3x128  stride 1 conv 3x3x128  stride 1 ReLU conv 3x3x256  stride 1 conv 3x3x256  stride 1 ReLU conv 1x1x5  stride 1 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU Sigmoid 29x29x5

Slide 53

Slide 53 text

ग़ྗ conﬁdence center_x center_y width height 29 × 29 × (4 + 1) = 4,205

Slide 54

Slide 54 text

܇࿅ɾֶश ࠷దԽΞϧΰϦζϜ"EBN0QUJNJ[FS ֶश཰

Slide 55

Slide 55 text

Yͷݶք إதԝҎ֎ͷηϧͷDPOpEFODF΋ߴ͍ DPOpEFODFͷߴ͍ηϧ͔Β  ਖ਼͍͠࠲ඪ͕ಘΒΕͳ͍

Slide 56

Slide 56 text

৞ΈࠐΈ૚ͷ໾ׂ conv 3x3x64 stride 1 conv 3x3x64  stride 1 ReLU ReLU conv 3x3x128  stride 1 conv 3x3x128  stride 1 ReLU conv 3x3x256  stride 1 conv 3x3x256  stride 1 ReLU conv 1x1x1  stride 1 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU Sigmoid 29x29x1

Slide 57

Slide 57 text

... લ૚͔ΒͷೖྗΛεΩϟϯͯ͠ಛ௃ϚοϓΛग़ྗ ಛ௃Ϛοϓ

Slide 58

Slide 58 text

ಛ௃Ϛοϓ͸ɺ খ͘͞ͳΔ΄ͲେҬͷಛ௃Λද͢

Slide 59

Slide 59 text

ಛ௃Ϛοϓ͸ɺ খ͘͞ͳΔ΄ͲେҬͷಛ௃Λද͢

Slide 60

Slide 60 text

దͨ͠αΠζͷηϧʹׂΓ౰ͯΔ

Slide 61

Slide 61 text

৞ΈࠐΈΛ܁Γฦ͢աఔͰग़ྗ͢Δ Single Shot MultiBox Detector https://arxiv.org/abs/1512.02325

Slide 62

Slide 62 text

Ϟσϧ

Slide 63

Slide 63 text

conv 3x3x64 stride 1 conv 3x3x64  stride 1 ReLU ReLU conv 3x3x128  stride 1 conv 3x3x128  stride 1 ReLU conv 3x3x256  stride 1 conv 3x3x256  stride 1 ReLU 256x256x1 max_pool 2x2 stride 2 max_pool 2x2 stride 2 ReLU ReLU 29x29x256

Slide 64

Slide 64 text

conv 3x3x256  stride 2 conv 3x3x256  stride 2 conv 3x3x256  stride 2 conv 1x1x5  stride 1 Sigmoid conv 1x1x5  stride 1 conv 1x1x5  stride 1 Sigmoid conv 1x1x5  stride 1 Sigmoid conv 1x1x5  stride 1 Sigmoid 1x1x5 15x15x5 8x8x5 2x2x5 4x4x5 29x29x256 conv 1x1x5  stride 1 Sigmoid conv 3x3x256  stride 2 conv 3x3x256  stride 2 Sigmoid 29x29x5

Slide 65

Slide 65 text

ֶशσʔλ

Slide 66

Slide 66 text

ֶशσʔλ confidence : 1.0 center_x : 0.7128 center_y : 0.4045 width : 0.2184 height : 0.2382 1.0 : confidence 0.2150 : center_x 0.5169 : center_y 0.2117 : width 0.2151 : height confidence : 0.0 center_x : 0.0 center_y : 0.0 width : 0.0 height : 0.0

Slide 67

Slide 67 text

ηϧͷׂΓ౰ͯج४ ÷ = Jaccard Overlap ҰఆҎ্ͷ஋ʢॏͳΓʣͷηϧʹׂΓ౰ͯ ↓ + - ) (

Slide 68

Slide 68 text

EFG@MBOE SFDU XJEUISFDU<*/%&9@3*()5>SFDU<*/%&9@-&'5> IFJHIUSFDU<*/%&9@#0550.>SFDU<*/%&9@501> JGXJEUIPSIFJHIU SFUVSO SFUVSOXJEUIIFJHIU EFGKBDDBSE@PWFSMBQ SFDU SFDU PWFSMBQ@MFGUNBY SFDU<*/%&9@-&'5> SFDU<*/%&9@-&'5> PWFSMBQ@SJHIUNJO SFDU<*/%&9@3*()5> SFDU<*/%&9@3*()5> PWFSMBQ@UPQNBY SFDU<*/%&9@501> SFDU<*/%&9@501> PWFSMBQ@CPUUPNNJO SFDU<*/%&9@#0550.> SFDU<*/%&9@#0550.> PWFSMBQ@MBOE VOJPO@MBOE SFDU @MBOE SFDU PWFSMBQ SFUVSOPWFSMBQVOJPO

Slide 69

Slide 69 text

ޡࠩؔ਺ ֬৴౓ʢ$POpEFODFʣˠ.4&ʢ.FBO4RVBSFE&SSPSʣ ࠲ඪʢ3FHJPOʣˠ4NPPUI--PTT ʢ'BTU3$//IUUQTBSYJWPSHBCTʣ

Slide 70

Slide 70 text

def _smooth_l1_loss(x): with tf.name_scope("smooth_l1"): abs_x = tf.abs(x) less_mask = tf.cast(abs_x < 1.0, tf.float32) return less_mask * (0.5 * tf.square(x)) + (1.0 - less_mask) * (abs_x - 0.5) def _loc_loss(logits, groundtruth): offset = logits - groundtruth return _smooth_l1_loss(offset)

Slide 71

Slide 71 text

MSEʢMean Squared Errorʣ Smooth L1 Loss

Slide 72

Slide 72 text

ֶशʹࣦഊʜʜ

Slide 73

Slide 73 text

1PTJUJWF4BNQMFͷׂ߹ 841 29x29 225 15x15 64 8x8 1 4 2x2 16 4x4 1,151

Slide 74

Slide 74 text

)BSE/FHBUJWF.JOJOH

Slide 75

Slide 75 text

1PTJUJWFαϯϓϧʹՃ͑ͯɺޡࠩͷେ͖ͳαϯϓϧΛબ୒

Slide 76

Slide 76 text

/&("5*7&@$06/5@3"5& EFG@DBMD@MPTT@XJUI@IBSE@OFHBUJWF@NJOJOH MPTTFT QPTJUJWF@NBTL QPTJUJWF@DPVOU OFHBUJWF@DPVOUQPTJUJWF@DPVOU/&("5*7&@$06/5@3"5& CBUDI@TJ[FMPTTFTHFU@TIBQF <>WBMVF QPTJUJWF@MPTTFTMPTTFTQPTJUJWF@NBTL OFHBUJWF@MPTTFTMPTTFTQPTJUJWF@MPTTFT UPQ@OFHBUJWF@MPTTFT @UGOOUPQ@L UGSFTIBQF OFHBUJWF@MPTTFT LUGDBTU OFHBUJWF@DPVOU UGJOU MPTT UGSFEVDF@TVN QPTJUJWF@MPTTFTQPTJUJWF@DPVOU UGSFEVDF@TVN UPQ@OFHBUJWF@MPTTFTOFHBUJWF@DPVOU SFUVSOMPTT

Slide 77

Slide 77 text

ֶशʹࣦഊʜʜ

Slide 78

Slide 78 text

إ͸্ख͘ݕग़Ͱ͖͍ͯΔ͕ ࠲ඪ͕ਖ਼͘͠ͳ͍

Slide 79

Slide 79 text

EFGBVMU#PY ʢ"ODIPSʣ

Slide 80

Slide 80 text

EFGBVMUCPYͷಋೖ left : 0.5000 top : 0.2500 right : 0.7500 bottom : 0.5000 0.0000 : left 0.5000 : top 0.2500 : right 0.7500 : bottom

Slide 81

Slide 81 text

EFGBVMUCPYͱ࠲ඪ left : 0.5000 top : 0.2500 right : 0.7500 bottom : 0.5000 0.0000 : left 0.5000 : top 0.2500 : right 0.7500 : bottom left : 0.6036 top : 0.2854 right : 0.8220 bottom : 0.5236 0.1092 : left 0.4094 : top 0.3209 : right 0.6247 : bottom

Slide 82

Slide 82 text

EFGBVMUCPYͱ࠲ඪͷ ࠩʢP⒎TFUʣΛઃఆ left : 0.6036 top : 0.2854 right : 0.8220 bottom : 0.5236 left : 0.1036 top : 0.0354 right : 0.0720 bottom : 0.0236 0.1092 : left -0.0906 : top 0.0709 : right -0.1253 : bottom 0.1092 : left 0.4094 : top 0.3209 : right 0.6247 : bottom

Slide 83

Slide 83 text

ग़ྗ ʹҙຯ͕ੜ·ΕΔ

Slide 84

Slide 84 text

׆ੑԽؔ਺ͷมߋʢग़ྗ૚ʣ left : 0.1036 top : 0.0354 right : 0.0720 bottom : 0.0236 0.1092 : left -0.0906 : top 0.0709 : right -0.1253 : bottom

Slide 85

Slide 85 text

׆ੑԽؔ਺ͷมߋʢग़ྗ૚ʣ Sigmoid Hyperbolic Tangent

Slide 86

Slide 86 text

ֶशɾ܇࿅ εςοϓʢόοναΠζ Ϛϧν(16ʣ  ࠷దԽΞϧΰϦζϜ"EBN0QUJNJ[FS ֶश཰ ݮਰ TUFQ

Slide 87

Slide 87 text

؀ڥʢ͘͞ΒͷߴՐྗίϯϐϡʔςΟϯάʣ $169FPO$PSFʷ .FNPSZ(# 44%(# (F'PSDF(595*5"/9ʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ (F'PSDF(595Jʢ1BTDBMΞʔΩςΫνϟʣ(#ʷ

Slide 88

Slide 88 text

ޡࠩͷมԽ

Slide 89

Slide 89 text

ݕূ

Slide 90

Slide 90 text

ݕূ

Slide 91

Slide 91 text

ࠓޙͷ՝୊

Slide 92

Slide 92 text

ΞεϖΫτൺ͕େ͖͍ը૾΁ͷରԠ

Slide 93

Slide 93 text

ΞεϖΫτൺͷҟͳΔEFGBVMUCPYͷಋೖ

Slide 94

Slide 94 text

ٕज़ॻయ̏ https://techbookfest.org/event/tbf03/ circle/5686003050217472 ೔࣌ɿ2017೥10݄22೔ʢ೔ʣ  ɹɹɹ11࣌ʙ17࣌ ৔ॴɿΞΩόɾεΫΤΞ ओ࠵ɿTechbooster / ୡਓग़൛ձ Illustration: ࠜઇΕ͍

Slide 95

Slide 95 text

ࢀߟ 4JOHMF4IPU.VMUJ#PY%FUFDUPS  IUUQTBSYJWPSHBCT 44%4JOHMF4IPU.VMUJ#PY%FUFDUPSʢ೔ຊޠ༁ʣ2JJUB  IUUQTRJJUBDPNEFUBJUFNTBFDFGDBB 5FOTPS'MPXͰإݕग़ثΛࣗ࡞͢Δ͗͢ΌʔΜϝϞ  IUUQNFNPTVHZBODPNFOUSZ 5FOTPS'MPXͰ෺ମྖҬ༧ଌʢ3FHJPO1SPQPTBMʣΛࢼͯ͠ΈΔ  IUUQXPSLQJMFTDPNUFOTPSqPXTTESFHJPOQSPQPTBM

Slide 96

Slide 96 text

ࢀߟ 'BTU3$//  IUUQTBSYJWPSHBCT 'BTUFS3$//5PXBSET3FBM5*NF0CKFDU%FUFDUJPOXJUI3FHJPO1SPQPTBM/FUXPSLT  IUUQTBSYJWPSHBCT :PV0OMZ-PPL0ODF6OJpFE 3FBM5JNF0CKFDU%FUFDUJPO  IUUQTBSYJWPSHBCT +BDDBSE*OEFY  IUUQTFOXJLJQFEJBPSHXJLJ+BDDBSE@JOEFY

Slide 97

Slide 97 text

C-LIS CO., LTD. ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻܝࡌ͞Ε͍ͯΔΠϥετ͸ɺಛʹهࡌ͕ͳ͍৔߹͸ࠜઇΕ͍ͷஶ࡞෺Ͱ͢ɻ ຊࢿྉͷશ෦ɺ·ͨ͸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 https://speakerdeck.com/keiji/devfest-tokyo-2017-tensorﬂow