AWSの機械学習基盤を使ってみよう

by Takaaki Tanaka

Slide 1

Slide 1 text

AWSͷػցֶशج൫Λ࢖ͬͯΈΑ͏ ౔ ߹ಉษڧձ in େ౎ձԬࢁ -2017 Winter- ాத޹໌

Slide 2

Slide 2 text

"CPVUNF

Slide 3

Slide 3 text

wΫϥεϝιουגࣜձࣾ wϞόΠϧΞϓϦαʔϏε෦ wJ04ΞϓϦΤϯδχΞ wαʔόʔαΠυΞϓϦΤϯδχΞ wαʔόʔϨε։ൃ෦ wΞϓϦέʔγϣϯΤϯδχΞ ాத޹໌ @kongmingtrap

Slide 4

Slide 4 text

ాத޹໌ @kongmingtrap Ԭࢁग़਎

Slide 5

Slide 5 text

ాத޹໌ @kongmingtrap ෱Ԭࡏॅ

Slide 6

Slide 6 text

IUUQTDMBTTNFUIPEKQOFXTOFXP⒏DFGVLVPLB

Slide 7

Slide 7 text

IUUQTDMBTTNFUIPEKQOFXTOFXP⒏DFGVLVPLB ෱ԬҠॅʂʂʂ

Slide 8

Slide 8 text

ؓ࿩ٳ୊

Slide 9

Slide 9 text

ࠓ೥ͷػցֶश ϋΠϥΠτ

Slide 10

Slide 10 text

w5FOTPS'MPX-JUF wߴ଎͔ͭܰྔͳΞϓϦ޲͚ػցֶशϑϨʔϜϫʔΫ w5FOTPS'MPX3FTFBSDI$PVME wτϨʔχϯά͓Αͼਪ࿦ͷ྆ํΛߴ଎Խ͢Δ·ͬͨ ͘৽͍͠(PPHMFͷΫϥ΢υ516 (PPHMF*0

Slide 11

Slide 11 text

w$PSF.- wֶशϞσϧ౳ΛJ04NBD04্Ͱར༻͢Δࡍʹɺ ։ൃऀ͕ઐ໳తͳ஌ࣝΛඞཁͱͤͣʹѻ͑ΔΑ͏ʹ ิॿ͢ΔϑϨʔϜϫʔΫ wDPSFNMUPPMT wػցֶशϑϨʔϜϫʔΫͰ࡞੒ֶͨ͠शϞσϧΛ $PSF.-Ͱར༻Ͱ͖ΔΑ͏ʹม׵ 88%$

Slide 12

Slide 12 text

w"84%FFQ-FOT SF*OWFOU

Slide 13

Slide 13 text

w"84%FFQ-FOT wσΟʔϓϥʔχϯάϞσϧΛػث্Ͱ௚઀࣮ߦͰ͖ ΔɺϓϩάϥϛϯάՄೳͳ৽͍͠ϏσΦΧϝϥ wͲΜͳεΩϧϨϕϧͷ։ൃऀͰ΋෼Ͱ%FFQ -FBSOJOHΛ։࢝Ͱ͖Δ SF*OWFOU

Slide 14

Slide 14 text

w"NB[PO3FLPHOJUJPO7JEFP wը૾෼ੳαʔϏε"NB[PO3FLPHOJUJPO͕ɺಈը Λαϙʔτ w෺ମɺγʔϯɺςΩετɺإͷݕग़ɺ༗໊ਓͷೝࣝ wܞଳి࿩ɺΧϝϥɺ*P5ϏσΦηϯαʔɺ͓ΑͼϦΞ ϧλΠϜϥΠϒετϦʔϜϏσΦॲཧ͔ΒΩϟϓ νϟʔ͞ΕͨಈըΛɺεέʔϥϒϧͰߴਫ਼౓Ͱಈը ෼ੳ͢ΔιϦϡʔγϣϯʹར༻ SF*OWFOU

Slide 15

Slide 15 text

"NB[PO4BHFNBLFS

Slide 16

Slide 16 text

wػցֶशϞσϧͷߏஙͱτϨʔχϯάͷ४ උ͕ΑΓ؆୯ʹ wΞϓϦέʔγϣϯʹ࠷దͳΞϧΰϦζϜͱϑϨʔϜ ϫʔΫΛબ୒ w࠷దԽ͢ΔͨΊʹඞཁͳπʔϧ͕ἧ͍ͬͯΔ "NB[PO4BHF.BLFS

Slide 17

Slide 17 text

wτϨʔχϯάσʔλΛ؆୯ʹ෼ੳ͠ՄࢹԽ wϗετܕͷ+VQZUFS/PUFCPPLΛඋ͍͑ͯΔ w4ͷσʔλʹ௚઀઀ଓͰ͖Δ w"NB[PO%ZOBNP%#ɺ"NB[PO3FETIJGU͔Β ͷσʔλΛ4ʹҠಈͯͦ͠ΕΒͷσʔλΛ /PUFCPPLͰ෼ੳͰ͖Δ "NB[PO4BHF.BLFS

Slide 18

Slide 18 text

Ͳ͏มΘΔͷ͔ʁ

Slide 19

Slide 19 text

ैདྷͰ͸ʜ w(16͕ࡌ͍ͬͯΔΠϯελϯεΛ༻ҙ͢Δ

Slide 20

Slide 20 text

ैདྷͰ͸ʜ wػցֶशϑϨʔϜϫʔΫΛ౥ࡌ͍ͯ͠Δ".* Λىಈ͢Δ

Slide 21

Slide 21 text

ैདྷͰ͸ʜ wֶश༻ͷϓϩάϥϜΛ४උ͢Δ model = Sequential() model.add(Conv2D(32, (3, 3), padding='same', input_shape=X_train.shape[1:])) model.add(Activation('relu')) model.add(Conv2D(32, (3, 3))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Conv2D(64, (3, 3), padding='same')) model.add(Activation('relu')) model.add(Conv2D(64, (3, 3))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(512)) model.add(Activation('relu')) model.add(Dropout(0.5)) model.add(Dense(nb_classes)) model.add(Activation('softmax')) model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

Slide 22

Slide 22 text

ैདྷͰ͸ʜ w࣮ߦ

Slide 23

Slide 23 text

ैདྷͰ͸ʜ w(16Πϯελϯεͷ༻ҙɺֶश༻ͷϓϩά ϥϜͷ࡞੒ͳͲɺ৭ʑͱϋʔυϧ͕͋ͬͨ w5FOTPS'MPX,FSBTͱ͍ͬͨɺػցֶ श༻ͷϑϨʔϜϫʔΫΛ࢖Θͳ͍ͱݫ͍͠ w্هͷϑϨʔϜϫʔΫͷ஌͕ࣝෆՄܽ wͦ΋ͦ΋ΞϧΰϦζϜͷ஌ࣝ΋ඞཁ

Slide 24

Slide 24 text

$*'"3 IUUQXXXDTUPSPOUPFEVdLSJ[DJGBSIUNM

Slide 25

Slide 25 text

$*'"3 wYͷը૾ຕͷσʔληοτ wτϨʔχϯάσʔλ͕ຕ wςετσʔλ͕ຕ wτϨʔχϯάσʔλͰֶशͨ͠ͷͪɺςε τσʔλΛ࢖ͬͯݕূ͢Δ

Slide 26

Slide 26 text

৞ΈࠐΈχϡʔϥϧωοτϫʔΫ IUUQTXXXZPVUVCFDPNXBUDI UJNF@DPOUJOVFW2;)$1OXX

Slide 27

Slide 27 text

৞ΈࠐΈχϡʔϥϧωοτϫʔΫ IUUQTLFSBTJPKB wରԠ͍ͯ͠ΔϥΠϒϥϦͷબఆ

Slide 28

Slide 28 text

wϥΠϒϥϦʹैͬͯίʔσΟϯά model = Sequential() model.add(Conv2D(32, (3, 3), padding='same', input_shape=X_train.shape[1:])) model.add(Activation('relu')) model.add(Conv2D(32, (3, 3))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Conv2D(64, (3, 3), padding='same')) model.add(Activation('relu')) model.add(Conv2D(64, (3, 3))) model.add(Activation('relu')) model.add(MaxPooling2D(pool_size=(2, 2))) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(512)) model.add(Activation('relu')) model.add(Dropout(0.5)) model.add(Dense(nb_classes)) model.add(Activation('softmax')) model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy']) ৞ΈࠐΈχϡʔϥϧωοτϫʔΫ

Slide 29

Slide 29 text

Ͳ͔͜ΒखΛ͚ͭΕ͹͍͍ͷ͔ʜ

Slide 30

Slide 30 text

"84SF*OWFOU/&8 -"6/$)*OUSPEVDJOH"NB[PO 4BHF.BLFS .$-

Slide 31

Slide 31 text

IUUQTXXXZPVUVCFDPNXBUDI WQC9ETK;Y@L

Slide 32

Slide 32 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 33

Slide 33 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 34

Slide 34 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 35

Slide 35 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 36

Slide 36 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 37

Slide 37 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 38

Slide 38 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 39

Slide 39 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε 㲔

Slide 40

Slide 40 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε 㲔

Slide 41

Slide 41 text

ֶशϞσϧ࡞੒·Ͱͷγʔέϯε

Slide 42

Slide 42 text

OPUFCPPLͷ࡞੒

Slide 43

Slide 43 text

OPUFCPPLͷ࡞੒

Slide 44

Slide 44 text

OPUFCPPLͷ࡞੒

Slide 45

Slide 45 text

OPUFCPPLͷىಈ w*O4FSWJDFʹͳͬͨΒɺ0QFOΛΫϦοΫ͢ Δͱىಈ͢Δ

Slide 46

Slide 46 text

OPUFCPPLͷىಈ IUUQTHJUIVCDPNBXTMBCTBNB[POTBHFNBLFSFYBNQMFTUSFFNBTUFS TBHFNBLFSQZUIPOTELNYOFU@HMVPO@DJGBS wαϯϓϧͷʮTBNQMFOPUFCPPLʯ ʮTBHFNBLFSQZSIPOTELʯ ʮNYOFU@DJGBSʯΛࢼ͠ʹ࣮ߦ͢Δ

Slide 47

Slide 47 text

δϣϒͷ࡞੒ͷ४උ

Slide 48

Slide 48 text

δϣϒͷ࡞੒ͷ४උ wඞཁͳϥΠϒϥϦͷΠϯετʔϧ

Slide 49

Slide 49 text

δϣϒͷ࡞੒ͷ४උ

Slide 50

Slide 50 text

δϣϒͷ࡞੒ w*O<>ͷۭཝΛΫϦοΫ͢Δ

Slide 51

Slide 51 text

δϣϒͷ࡞੒

Slide 52

Slide 52 text

δϣϒͷ࡞੒ wδϣϒ͕࡞੒͞ΕΔͱֶश͕࣮ߦ͞ΕΔ

Slide 53

Slide 53 text

δϣϒͷ࣮ߦ

Slide 54

Slide 54 text

δϣϒͷ࣮ߦ wOPUFCPPL͔ΒֶशͷਐḿΛ֬ೝͰ͖Δ

Slide 55

Slide 55 text

ֶशϞσϧͷ࡞੒ wδϣϒ͕੒ޭ͢ΔͱɺֶशϞσϧ͕࡞੒͞ ΕΔ

Slide 56

Slide 56 text

ֶशϞσϧͷ࡞੒ wδϣϒ͕੒ޭ͢ΔͱɺֶशϞσϧ͕࡞੒͞ ΕΔ

Slide 57

Slide 57 text

ֶशϞσϧͷ࡞੒ wֶशϞσϧͷৄࡉΛ֬ೝ

Slide 58

Slide 58 text

ֶशϞσϧͷ࡞੒

Slide 59

Slide 59 text

ֶशϞσϧͷ࡞੒ wֶशϞσϧ͸μ΢ϯϩʔυ͢Δ͜ͱ͕Մೳ

Slide 60

Slide 60 text

wֶश؀ڥΛ४උ͢Δ·Ͱʹ͔͔͍ͬͯͨί ετΛ௿ݮ wֶशͷਐḿΛՄࢹԽ wαϯϓϧΛར༻͢Δ͜ͱͰʮࣗ਎ͷֶशʯ ͷͱ͔͔ͬΓʹ͢Δ͜ͱ͕Մೳ ·ͱΊ

Slide 61

Slide 61 text

ࠓޙ΍Γ͍ͨ͜ͱ IUUQTBXTBNB[PODPNKQCMPHTOFXTBNB[POTBHFNBLFS

Slide 62

Slide 62 text

ࠓޙ΍Γ͍ͨ͜ͱ IUUQTBXTBNB[PODPNKQCMPHTOFXTCSJOHNBDIJOFMFBSOJOHUPJPTBQQT VTJOHBQBDIFNYOFUBOEBQQMFDPSFNM