ウェブブラウザ向け深層学習モデル高速実行フレームワーク「WebDNN」

by Kiikurage

Slide 1

Slide 1 text

΢Σϒϒϥ΢β޲͚ਂ૚ֶशϞσϧߴ଎࣮ߦϑϨʔϜϫʔΫ WebDNN ౦ژେֶ େֶӃ৘ใཧ޻ֶܥݚڀՊ ૑଄৘ใֶઐ߈ ݪాɾڇٱݚڀࣨ म࢜೥ ໦૔ ༔Ұ࿠ !,JJLVSBHF

Slide 2

Slide 2 text

໦૔ ༔Ұ࿠:VJDIJSP,JLVSB ౦ژେֶ େֶӃ৘ใཧ޻ֶܥݚڀՊ ૑଄৘ใֶઐ߈ ݪాɾڇٱݚڀࣨ म࢜೥ ΢Σϒϒϥ΢β޲͚ਂ૚ֶशϞσϧߴ଎࣮ߦϑϨʔϜϫʔΫʮ8FC%//ʯͷ͝঺հ • 8FC%//ͱ͸ • 8FC%//ͷ࢖͍ํ • 8FC%//͕ߦ͍ͬͯΔߴ଎ԽΞϓϩʔν ͸͡Ίʹ

Slide 3

Slide 3 text

8FC%// ΢Σϒϒϥ΢β޲͚ਂ૚ֶशϞσϧߴ଎࣮ߦϑϨʔϜϫʔΫ ֶशࡁΈϞσϧΛਪ࿦ϑΣʔζʹ࠷దԽ͠΢Σϒϒϥ΢β্Ͱ࣮ߦ 5IJTXPSLXBTQBSUJBMMZTVQQPSUFECZ+45 $3&45 (SBOU/VNCFS+1.+$3 +BQBO5IJTXPSLXBTBMTP QBSUJBMMZTVQQPSUFECZUIF.JOJTUSZPG&EVDBUJPO $VMUVSF 4QPSUT 4DJFODFBOE5FDIOPMPHZ .&95 BT l4FNJOBM*TTVFPO1PTU,$PNQVUFSz .BTBUPTIJ)JEBLB :VJDIJSP,JLVSB :PTIJUBLB6TIJLV 5BUTVZB)BSBEB ౦ژେֶ ৘ใཧ޻ֶܥݚڀՊ ݪాɾڇٱݚڀࣨ .BDIJOF*OUFMMJHFODF-BC.*-

Slide 4

Slide 4 text

8FC%// ΢Σϒϒϥ΢β޲͚ਂ૚ֶशϞσϧߴ଎࣮ߦϑϨʔϜϫʔΫ ֶशࡁΈϞσϧΛਪ࿦ϑΣʔζʹ࠷దԽ͠΢Σϒϒϥ΢β্Ͱ࣮ߦ Ϟσϧఆٛ ม ׵ ࠷ ద Խ ੜ ੒ (16#BDLFOE ֶशࡁΈ ύϥϝʔλ ύϥϝʔλ (SBQI5SBOTQJMFS ݚڀ੒Ռͷਝ଎ͳެ։ %//Λ༻͍ͨΞϓϦͷσϞ ར༻ྫ IUUQTNJMUPLZPHJUIVCJPXFCEOO %FTDSJQUPS3VOOFS $16#BDLFOE (SBQI %FTDSJQUPS

Slide 5

Slide 5 text

8FC%//ͷར༻ྫ .BLF(JSMT.PF <:+JOBOE+;IBOH> IUUQNBLFHJSMTNPF ("/ʹΑΔೋ࣍ݩΩϟϥը૾ੜ੒

Slide 6

Slide 6 text

΢Σϒϒϥ΢βͰͷਂ૚ֶश %//Λ࣮ΞϓϦέʔγϣϯ΁Ԡ༻͢Δࡍͷ՝୊ αʔόʔαΠυίϯϐϡʔςΟϯά • ΫϥΠΞϯτ਺ʹର͢ΔεέʔϧΞ΢τ͕༰қͰ͸ͳ͍ • ϓϥΠόγʔʹ఍৮͢ΔσʔλΛѻ͍ͮΒ͍ ΫϥΠΞϯταΠυίϯϐϡʔςΟϯά • ܭࢉ؀ڥηοτΞοϓ͕೉͍͠

Slide 7

Slide 7 text

΢Σϒϒϥ΢βͰͷਂ૚ֶश %//Λ࣮ΞϓϦέʔγϣϯ΁Ԡ༻͢Δࡍͷ՝୊ ΢ΣϒΞϓϦέʔγϣϯ • ΫϥΠΞϯτଆͰܭࢉ͢ΔͨΊεέʔϧΞ΢τ͕༰қ • αʔόʔ΁ͷσʔλΞοϓϩʔυ͕ෆཁͳͨΊಗ໊ੑͷ୲อ͕༰қ • ηοτΞοϓ͸ෆཁ ΢ΣϒαΠτʹΞΫηε͢Δ͚ͩͰར༻Մೳ "JYJMF͞Μ !OBNBOJLV .BLF(JSTU.PFͷ࡞ऀ

Slide 8

Slide 8 text

΢Σϒϒϥ΢βͰͷػցֶश • .*-+4 IUUQNJMUPLZPHJUIVCJPNJMKTIUNM ౰ݚڀࣨͰ։ൃ͍ͯ͠Δ+BWB4DSJQUϕʔεػցֶशϥΠϒϥϦ܈ • 4VTIJ +BWB4DSJQUͰಈ࡞͢ΔߦྻԋࢉϥΠϒϥϦ <,.JVSB ><.)JEBLB > • 4VLJZBLJ 4VTIJΛܭࢉόοΫΤϯυͱ͢ΔػցֶशϥΠϒϥϦ <,.JVSB ><.)JEBLB >

Slide 9

Slide 9 text

8FC%//ͷಛ௃ ΢Σϒ࠷৽࢓༷Λ༻͍ͨߴ଎ͳܭࢉόοΫΤϯυ 8FC(16ɾ8FC(-ɾ8FC"TTFNCMZʹΑΔ$16(16྆ํͰͷߴ଎ͳ࣮ߦ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴ଎Խ ਪ࿦ϑΣʔζʹಛԽͨ͠࠷దԽʹΑΓܭࢉྔɾϞσϧύϥϝʔλαΠζΛ࡟ݮ طଘͷਂ૚ֶशϑϨʔϜϫʔΫͷ޿ൣғͳαϙʔτ ,FSBT $BGGF $IBJOFS 5FOTPS'MPXʹରԠ

Slide 10

Slide 10 text

8FC%//ͷಛ௃ ΢Σϒ࠷৽࢓༷Λ༻͍ͨߴ଎ͳܭࢉόοΫΤϯυ • ༷ʑͳ؀ڥʹରԠՄೳͳछྨͷܭࢉόοΫΤϯυΛ࣮૷ • ϒϥ΢βʹ࣮૷͞Ε͍ͯΔ"1*ʹԠͯࣗ͡ಈతʹ࠷దͳόοΫΤϯυΛબ୒ • ΄΅શͯͷ؀ڥͰ(16ʹΑΔϋʔυ΢ΣΞΞΫηϥϨʔγϣϯ͕ར༻Մೳ 8FC(16 #BDLFOE 8FC(- #BDLFOE 8FC"TTFNCMZ #BDLFOE 'BMMCBDL #BDLFOE

Slide 11

Slide 11 text

8FC%//ͷಛ௃ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴ଎Խ • ਪ࿦ϑΣʔζʹಛԽ͢Δ͜ͱͰɺΑΓੵۃతͳܭࢉάϥϑͷ࠷దԽΛ࣮ࢪ

Slide 12

Slide 12 text

8FC%//ͷಛ௃ طଘͷਂ૚ֶशϑϨʔϜϫʔΫͷ޿ൣғͳαϙʔτ • ,FSBT $BGGF $IBJOFS 5FOTPS'MPXͷֶशࡁΈϞσϧ͕ม׵Մೳ • ಠ࣮ࣗ૷ͨؔ͠਺΁ͷରԠʢτϥϯεύΠϥͷ֦ுʣ΋༰қ $IBJOFS chainer.Chain ,FSBT keras.Model $BGGF .proto 5FOTPS'MPX tf.Session

Slide 13

Slide 13 text

ϕϯνϚʔΫ 3FT/FU<)F > ը૾ຕͷਪ࿦ʹཁ͢Δ࣌ؒ • ೖྗαΠζ • ൺֱର৅ • ,FSBTKT IUUQTHJUIVCDPNUSBOTDSBOJBMLFSBTKT LFSBTͷֶशࡁΈϞσϧΛ࣮ߦՄೳͳϥΠϒϥϦɻ8FC(-ʹΑΔ(16ར༻ɻ (1,224,224,3)

Slide 14

Slide 14 text

ϕϯνϚʔΫ 3FT/FUը૾ຕͷਪ࿦ʹཁ͢Δ࣌ؒ ,FSBTKT 8FC%// ॴཁ࣌ؒ $ISPNF 'JSF'PY 4BGBSJ51 ,FSBTKT 8FC%// ॴཁ࣌ؒ $ISPNF 'JSF'PY 4BGBSJ51 8FC%// 8FC%// .BD#PPL1SP&BSMZ()[*OUFM$PSFJ$16*OUFM*SJT(SBQIJDT(16 $16 (16 1$

Slide 15

Slide 15 text

ϕϯνϚʔΫ 3FT/FUը૾ຕͷਪ࿦ʹཁ͢Δ࣌ؒ ,FSBTKT 8FC%// ॴཁ࣌ؒ $ISPNF 'JSF'PY 4BGBSJ ,FSBTKT 8FC%// ॴཁ࣌ؒ $ISPNF 'JSF'PY 4BGBSJ 8FC%// 8FC%// ˞ܽଛՕॴ͸࣮ߦෆՄೳɾαϙʔτ֎ 91&3*"9;"OESPJEY4OBQESBHPO()[Y$16"ESFOP(16 $ISPNF 'JSF'PY J1IPOFJ04"QQMF" 4BGBSJ εϚʔτϑΥϯ $16 (16

Slide 16

Slide 16 text

࢖༻ํ๏ ,FSBTɾ$IBJOFS෇ଐͷ3FT/FUֶशࡁΈϞσϧͷม׵ IUUQTNJMUPLZPHJUIVCJPXFCEOOEPDTUVUPSJBMLFSBTIUNM IUUQTNJMUPLZPHJUIVCJPXFCEOOEPDTUVUPSJBMDIBJOFSIUNM

Slide 17

Slide 17 text

࢖༻ํ๏ ֶशࡁΈϞσϧΛม׵͢Δ,FSBT from keras.applications import resnet50 from webdnn.frontend.keras import KerasConverter from webdnn.backend import generate_descriptor graph = KerasConverter().convert(resnet50.ResNet50()) # (1) generate_descriptor('webgl', graph).save('model') # (2) ,FSBT keras.Model (SBQI%FTDSJQUPS GPS8FC(- #BDLFOE 8FC%//*3 1ZUIPO

Slide 18

Slide 18 text

࢖༻ํ๏ ֶशࡁΈϞσϧΛม׵͢Δ$IBJOFS import numpy as np from chainer import Variable, links as L model = L.ResNet50Layers() ܭࢉάϥϑΛ࡞ΔͨΊʹμϛʔσʔλΛྲྀ͢ x = Variable(np.zeros((1, 3, 224, 224), dtype=np.float32)) y = model(x, layers=['prob'])['prob'] $IBJOFS chainer.Chain (SBQI%FTDSJQUPS GPS8FC(- #BDLFOE 8FC%//*3 1ZUIPO

Slide 19

Slide 19 text

࢖༻ํ๏ ֶशࡁΈϞσϧΛม׵͢Δ$IBJOFS from webdnn.frontend.chainer import ChainerConverter from webdnn.backend import generate_descriptor graph = ChainerConverter().convert([x], [y]) # (1) generate_descriptor('webgl', graph).save('model') # (2) $IBJOFS chainer.Chain (SBQI%FTDSJQUPS GPS8FC(- #BDLFOE 8FC%//*3 1ZUIPO

Slide 20

Slide 20 text

࢖༻ํ๏ ม׵݁ՌΛϒϥ΢β͔ΒಡΈࠐΈ࣮ߦ͢Δ let runner = await WebDNN.load('model'); let x = runner.getInputViews()[0]; let y = runner.getOutputViews()[0]; جຊతͳલɾޙॲཧʹؔ͢ΔαϙʔτϥΠϒϥϦ͕෇ଐ x.set(await WebDNN.Image.getImageArray('cat.jpg', { dstW: 224, dstH: 224 })); await runner.run(); ࣮ߦ console.log(y.toActual()); JavaScript +BWB4DSJQU

Slide 21

Slide 21 text

ߴ଎Խ ΢Σϒϒϥ΢βίϯϐϡʔςΟϯάʹ͓͍ͯ཯଎ͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ଍ ϝϞϦଳҬෆ଍

Slide 22

Slide 22 text

ߴ଎Խ ΢Σϒϒϥ΢βίϯϐϡʔςΟϯάʹ͓͍ͯ཯଎ͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ଍ ϝϞϦଳҬෆ଍

Slide 23

Slide 23 text

+BWB4DSJQUͷΦʔόʔϔου • +BWB4DSJQUͦͷ΋ͷͷ࣮ߦ଎౓͕஗͍ • ௿ϨΠϠΛ৮ΕΔ"1*͕૿͖͑ͯͨ • 7BOJMMB+BWB4DSJQUˠ8FC"TTFNCMZ • 8FC(-ˠ 8FC(-ˠ 8FC(16

Slide 24

Slide 24 text

8FC"TTFNCMZ ΢Σϒϒϥ΢β͔ΒΞηϯϒϥϨϕϧͰͷ໋ྩΛѻ͑Δ࢓༷ • +BWB4DSJQUͷΠϯλϓϦλ࣮ߦʹΑΔΦʔόʔϔου͕ͳͤ͘Δ • 7BOJMMB+BWB4DSJQUͰ͸Ͱ͖ͳ͔༷ͬͨʑͳ࠷దԽ͕Մೳ • 4*.%΍ϚϧνεϨου΋Մೳʹ $ISPNF'JSF'PYͰळʙౙTIJQ͞Εͦ͏ C C++ Rust .wasm ίϯύΠϧ ϒϥ΢β͔ΒόΠφϦΛ ௚઀ಡΈࠐΈ࣮ߦ ϒϥ΢β

Slide 25

Slide 25 text

8FC(- • ΢Σϒϒϥ΢βʹ͸(1(16༻"1*͕ଘࡏ͠ͳ͍ • 8FC(- • ը૾ॲཧ༻"1*Ͱ͋Γ$PNQVUJOH4IBEFS͸αϙʔτ͞Εͯͳ͍ • ؤுΕ͹(1(16΋Ͳ͖΋Մೳ • ༷ʑͳ੍໿ͷ͍ͤͰΦʔόʔϔου͕େ͖͍ • 8FC(- • (1(16͠΍͘͢ͳΔػೳ͕௥Ճ • Φʔόʔϔου͕͔ͳΓݮΒͤΔ

Slide 26

Slide 26 text

8FC(16 • IUUQTXFCLJUPSHXQDPOUFOUVQMPBETXFCHQVBQJQSPQPTBMIUNM • "QQMF͕த৺ͱͳͬͯٞ࿦தͷ࣍ظ"1* • $PNQVUJOH4IBEFSΛαϙʔτɺඈ༂తͳύϑΥʔϚϯε޲্ • 8FC,JUʹ࣮૷ͨ͠ • NBD04)JHI4JFSSBJ04ͷTBGBSJ͔Β࢖༻Մೳ 8(ͷٞࣄ࿥Λݟ͍ͯΔͱ ʮ΍ͬͺ$PNQVUJOH4IBEFSෆཁͰ͸ʁʯ ͱ͍͏ٞ࿦͕ͳ͞Ε͓ͯΓɺएׯӢߦ͖͕ո͍͠

Slide 27

Slide 27 text

ߴ଎Խ ΢Σϒϒϥ΢βίϯϐϡʔςΟϯάʹ͓͍ͯ཯଎ͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ଍ ϝϞϦଳҬෆ଍

Slide 28

Slide 28 text

ϝϞϦଳҬෆ଍ͷରࡦ ,FSOFM.FSHJOH • ແବͳಡΈॻ͖ΛݮΒ͠ԋࢉີ౓ΛߴΊ • ෳ਺ͷΧʔωϧΛ̍ͭʹϚʔδ͢Δ • ൃߦͷΦʔόʔϔου΋ݮΒͤΔ float x1 = X1[i]; float x2 = X2[i]; H[i] = x1 + x2; float h = H[i]; Y[i] = h>0?h:0; float x1 = X1[i]; float x2 = X2[i]; float h = x1 + x2; Y[i] = h>0?h:0; 3FBE 3FBE 3FBE 8SJUF 8SJUF 3FBE 3FBE 8SJUF

Slide 29

Slide 29 text

άϥϑߏ଄࠷దԽ ࣄલʹొ࿥͞ΕͨύλʔϯʹԊͬͯάϥϑΛม׵ $POTUBOU'PMEJOH Add Var Const1 Const2 Const1 + Const2 ύλʔϯ ม׵ ఆ਺ ม਺

Slide 30

Slide 30 text

άϥϑߏ଄࠷దԽ ࣄલʹొ࿥͞ΕͨύλʔϯʹԊͬͯάϥϑΛม׵ Add x1 Add x3 x2 c1 c2

Slide 31

Slide 31 text

άϥϑߏ଄࠷దԽ ࣄલʹొ࿥͞ΕͨύλʔϯʹԊͬͯάϥϑΛม׵ Add x1 Add x3 x2 c1 c2 $POTUBOU'PMEJOH Add x3 x2 c1 + c2

Slide 32

Slide 32 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 c1 c2 Add Add x2 Add (x1 + c1) + (x2 + c2)

Slide 33

Slide 33 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 x2 c2 Add Add c1 Add (x1 + c1) + (x2 + c2) = (x1 + x2) + (c1 + c2)

Slide 34

Slide 34 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 x2 c2 Add Add c1 Add $POTUBOU'PMEJOH

Slide 35

Slide 35 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Add Add c1 + c2 x2

Slide 36

Slide 36 text

/PSNBMJ[BUJPO &YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 w Conv2D x2 s Mul x3 b Add

Slide 37

Slide 37 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 w Conv2D x2 s Mul x3 b Add &YQSFTTJPO4JNQMJGJDBUJPO

Slide 38

Slide 38 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add x4 s Mul w

Slide 39

Slide 39 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add x4 s Mul w $POTUBOU'PMEJOH

Slide 40

Slide 40 text

&YQSFTTJPO4JNQMJGJDBUJPO ఆ਺߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add w * s

Slide 41

Slide 41 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D x2 Add Mul Add x3 x4 x9 Conv2D x6 Mul Add x7 x8 x10 Relu s1 b1 b2 s2 x1 w1 x5 w2

Slide 42

Slide 42 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D x2 Add Mul Add x3 x4 x9 Conv2D x6 Mul Add x7 x8 x10 Relu s1 b1 b2 s2 x1 w1 x5 w2 &MFNFOUXJTF

Slide 43

Slide 43 text

&MFNFOUXJTF ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D x2 Add Mul Add x3 x4 x9 Conv2D x6 Mul Add x7 x8 x10 Relu s1 b1 b2 s2 x1 w1 x5 w2 ճ 3FBEɾ8SJUF

Slide 44

Slide 44 text

&YQSFTTJPO4JNQMJGJDBUJPO ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D x2 Add Mul Add x3 x4 x9 Conv2D x6 Mul Add x7 x8 x10 Relu s1 b1 b2 s2 x1 w1 x5 w2 &YQSFTTJPO4JNQMJGJDBUJPO

Slide 45

Slide 45 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D Add x7 x8 x10 Relu b1 b2 x1 w1 * s1 x5 w2 * s2

Slide 46

Slide 46 text

&YQSFTTJPO4JNQMJGJDBUJPO ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D Add x7 x8 x10 Relu b1 b2 x1 w1 * s1 x5 w2 * s2

Slide 47

Slide 47 text

$POTUBOU'PMEJOH ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D Add x8 x10 Relu b2 x1 w1 * s1 x5 w2 * s2 x7 b1

Slide 48

Slide 48 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D x10 Relu x1 w1 * s1 x5 w2 * s2 x7 b1 + b2

Slide 49

Slide 49 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D x10 Relu x1 w1 * s1 x5 w2 * s2 x7 b1 + b2 &MFNFOUXJTF,FSOFM.FSHJOH

Slide 50

Slide 50 text

,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦ΁ͷॻ͖ࠐΈ΍ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D MergedElementwise x3 Conv2D x10 x1 w1 * s1 x5 w2 * s2 x7 b1 + b2 ճ ˠճ 3FBEɾ8SJUF

Slide 51

Slide 51 text

,FSOFM.FSHJOH ,FSOFM.FSHJOHʹΑΔ࠷దԽͷޮՌ 4BGBSJ51.BD#PPL1SP&BSMZ 3FT/FU࣮ߦ࣌ؒ WebGPU backend 最適化なし WebGPU backend ,FSOFM.FSHJOH

Slide 52

Slide 52 text

ϝϞϦଳҬෆ଍΁ͷରࡦ %BUB'PSNBU0QUJNJ[BUJPO ܭࢉάϥϑதͷม਺ΛɺϝϞϦ্ʹͲͷΑ͏ʹ֨ೲ͢Δ͔Λ࠷దԽ • %BUB0SEFS 0QFSBUPSͷछྨɾೖग़ྗม਺ͷܗঢ়ͳͲʹΑΓదͨ͠σʔλΦʔμʔ͸ҟͳ Δ • ͋Δ0QFSBUPS͸/)8$Ͱɺผͷ0QFSBUPS͸/$)8ͰσʔλΛѻ͍͍ͨɺ ͱ͍ͬͨ͜ͱ͕͋Δ • લޙͷ0QFSBUPS΋Ճຯͯ͠ೖग़ྗม਺ͷ%BUB 0SEFSΛܾΊΔ

Slide 53

Slide 53 text

ϝϞϦଳҬෆ଍΁ͷରࡦ %BUB'PSNBU0QUJNJ[BUJPO ܭࢉάϥϑதͷม਺ΛɺϝϞϦ্ʹͲͷΑ͏ʹ֨ೲ͢Δ͔Λ࠷దԽ • "MJHONFOUBOE1BEEJOH 4*.%໋ྩ΍(16ΧʔωϧதͰͷॲཧൣғͷ෼ׂׂΓ౰ͯʢ5JMJOHʣΛ༰қʹ ͢ΔͨΊʹߦྻͷαΠζɾΦϑηοτΛἧ͑Δ

Slide 54

Slide 54 text

ϝϞϦଳҬෆ଍΁ͷରࡦ ྫ (&..$"# ߦྻͷαΠζ͕λΠϧαΠζͰ ៉ྷʹׂΓ੾ΕΔΑ͏ ༧ΊύσΟϯά͓ͯ͘͠ " # $ 5JMF

Slide 55

Slide 55 text

ϝϞϦଳҬෆ଍΁ͷରࡦ ྫ (&..$"# ߦྻͷαΠζ͕λΠϧαΠζͰ ៉ྷʹׂΓ੾ΕΔΑ͏ ༧ΊύσΟϯά͓ͯ͘͠ " 1BEEJOH # $

Slide 56

Slide 56 text

%BUB'PSNBU 0QUJNJ[BUJPO %BUB'PSNBU 0QUJNJ[BUJPOʹΑΔ࠷దԽͷޮՌ 4BGBSJ51.BD#PPL1SP&BSMZ 3FT/FU࣮ߦ࣌ؒ WebGPU backend 最適化なし WebGPU backend + Elementwise Kernel Merging WebGPU backend + Elementwise Kernel Merging %BUB'PSNBU 0QUJNJ[BUJPO

Slide 57

Slide 57 text

·ͱΊ 8FC%//΢Σϒϒϥ΢β޲͚ਂ૚ֶशϞσϧߴ଎࣮ߦϑϨʔϜϫʔΫ IUUQTNJMUPLZPHJUIVCJPXFCEOO ΢Σϒ࠷৽࢓༷Λ༻͍ͨߴ଎ͳܭࢉόοΫΤϯυ 8FC(16ɾ8FC(-ɾ8FC"TTFNCMZʹΑΔ$16(16྆ํͰͷߴ଎ͳ࣮ߦ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴ଎Խ ਪ࿦ϑΣʔζʹಛԽͨ͠࠷దԽʹΑΓܭࢉྔɾϞσϧύϥϝʔλαΠζΛ࡟ݮ طଘͷਂ૚ֶशϑϨʔϜϫʔΫͷ޿ൣғͳαϙʔτ ,FSBT $BGGF $IBJOFS 5FOTPS'MPXʹରԠ