Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
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」
Search
Kiikurage
September 03, 2017
Programming
6
15k
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」
2017/09/03 Deep Learning Acceleration勉強会@DeNAでの発表資料です。
Kiikurage
September 03, 2017
Tweet
Share
Other Decks in Programming
See All in Programming
From Spring Boot 2 to Spring Boot 3 with Java 22 and Jakarta EE
ivargrimstad
0
940
[技育CAMPアカデミア]アイディアを形に!【超入門】スマホアプリ開発〜リリースまでの流れをご紹介
teamlab
PRO
0
350
データアナリストが行うDatabricksを活用したETLの自動化事例
shinoa
0
260
try! Swift Tokyo 初参加報告LT
hinakko2
0
200
Hanami and htmx
bkuhlmann
0
200
チーム力を高めるスクラム実践法:カンバン公開と課題攻略について - ニフティのスクラムトーク Vol. 2 - NIFTY Tech Talk #18
niftycorp
PRO
1
110
Behind VS Code Extensions for JavaScript / TypeScript Linnting and Formatting
unvalley
4
310
脱・初心者!脱・マネコン!AWS CDKを使ってみませんか!?
har1101
0
310
Folding Cheat Sheet #2
philipschwarz
PRO
0
110
ADRを一年運用してみた/adr_after_a_year
hanhan1978
7
2.3k
スキーマ駆動開発による品質とスピードの両立 - 私達は何故、スキーマを書くのか
kentaroutakeda
0
130
見た目から始める生産性向上
ikumatadokoro
4
230
Featured
See All Featured
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
243
20k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
39
4.4k
WebSockets: Embracing the real-time Web
robhawkes
59
7k
The Straight Up "How To Draw Better" Workshop
denniskardys
227
130k
How to Create Impact in a Changing Tech Landscape [PerfNow 2023]
tammyeverts
12
1.5k
Designing for Performance
lara
601
67k
Atom: Resistance is Futile
akmur
258
25k
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
24
2.3k
Fight the Zombie Pattern Library - RWD Summit 2016
marcelosomers
226
16k
Robots, Beer and Maslow
schacon
PRO
155
7.9k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
501
140k
How to name files
jennybc
64
92k
Transcript
Σϒϒϥβ͚ਂֶशϞσϧߴ࣮ߦϑϨʔϜϫʔΫ WebDNN ౦ژେֶ େֶӃใཧֶܥݚڀՊ ใֶઐ߈ ݪాɾڇٱݚڀࣨ म࢜ ༔Ұ !,JJLVSBHF
༔Ұ:VJDIJSP,JLVSB ౦ژେֶ େֶӃใཧֶܥݚڀՊ ใֶઐ߈ ݪాɾڇٱݚڀࣨ म࢜ Σϒϒϥβ͚ਂֶशϞσϧߴ࣮ߦϑϨʔϜϫʔΫʮ8FC%//ʯͷ͝հ • 8FC%//ͱ
• 8FC%//ͷ͍ํ • 8FC%//͕ߦ͍ͬͯΔߴԽΞϓϩʔν ͡Ίʹ
8FC%// Σϒϒϥβ͚ਂֶशϞσϧߴ࣮ߦϑϨʔϜϫʔΫ ֶशࡁΈϞσϧΛਪϑΣʔζʹ࠷దԽ͠Σϒϒϥβ্Ͱ࣮ߦ 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 ใཧֶܥݚڀՊ ݪాɾڇٱݚڀࣨ .BDIJOF*OUFMMJHFODF-BC.*-
8FC%// Σϒϒϥβ͚ਂֶशϞσϧߴ࣮ߦϑϨʔϜϫʔΫ ֶशࡁΈϞσϧΛਪϑΣʔζʹ࠷దԽ͠Σϒϒϥβ্Ͱ࣮ߦ Ϟσϧఆٛ ม ࠷ ద Խ ੜ
(16#BDLFOE ֶशࡁΈ ύϥϝʔλ ύϥϝʔλ (SBQI5SBOTQJMFS ݚڀՌͷਝͳެ։ %//Λ༻͍ͨΞϓϦͷσϞ ར༻ྫ IUUQTNJMUPLZPHJUIVCJPXFCEOO %FTDSJQUPS3VOOFS $16#BDLFOE (SBQI %FTDSJQUPS
8FC%//ͷར༻ྫ .BLF(JSMT.PF <:+JOBOE+;IBOH> IUUQNBLFHJSMTNPF ("/ʹΑΔೋ࣍ݩΩϟϥը૾ੜ
ΣϒϒϥβͰͷਂֶश %//Λ࣮ΞϓϦέʔγϣϯԠ༻͢Δࡍͷ՝ αʔόʔαΠυίϯϐϡʔςΟϯά • ΫϥΠΞϯτʹର͢ΔεέʔϧΞτ͕༰қͰͳ͍ • ϓϥΠόγʔʹ৮͢ΔσʔλΛѻ͍ͮΒ͍ ΫϥΠΞϯταΠυίϯϐϡʔςΟϯά • ܭࢉڥηοτΞοϓ͕͍͠
ΣϒϒϥβͰͷਂֶश %//Λ࣮ΞϓϦέʔγϣϯԠ༻͢Δࡍͷ՝ ΣϒΞϓϦέʔγϣϯ • ΫϥΠΞϯτଆͰܭࢉ͢ΔͨΊεέʔϧΞτ͕༰қ • αʔόʔͷσʔλΞοϓϩʔυ͕ෆཁͳͨΊಗ໊ੑͷ୲อ͕༰қ • ηοτΞοϓෆཁ ΣϒαΠτʹΞΫηε͢Δ͚ͩͰར༻Մೳ
"JYJMF͞Μ !OBNBOJLV .BLF(JSTU.PFͷ࡞ऀ
ΣϒϒϥβͰͷػցֶश • .*-+4 IUUQNJMUPLZPHJUIVCJPNJMKTIUNM ݚڀࣨͰ։ൃ͍ͯ͠Δ+BWB4DSJQUϕʔεػցֶशϥΠϒϥϦ܈ • 4VTIJ +BWB4DSJQUͰಈ࡞͢ΔߦྻԋࢉϥΠϒϥϦ <,.JVSB ><.)JEBLB
> • 4VLJZBLJ 4VTIJΛܭࢉόοΫΤϯυͱ͢ΔػցֶशϥΠϒϥϦ <,.JVSB ><.)JEBLB >
8FC%//ͷಛ Σϒ࠷৽༷Λ༻͍ͨߴͳܭࢉόοΫΤϯυ 8FC(16ɾ8FC(-ɾ8FC"TTFNCMZʹΑΔ$16(16྆ํͰͷߴͳ࣮ߦ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴԽ ਪϑΣʔζʹಛԽͨ͠࠷దԽʹΑΓܭࢉྔɾϞσϧύϥϝʔλαΠζΛݮ طଘͷਂֶशϑϨʔϜϫʔΫͷൣғͳαϙʔτ ,FSBT
$BGGF $IBJOFS 5FOTPS'MPXʹରԠ
8FC%//ͷಛ Σϒ࠷৽༷Λ༻͍ͨߴͳܭࢉόοΫΤϯυ • ༷ʑͳڥʹରԠՄೳͳछྨͷܭࢉόοΫΤϯυΛ࣮ • ϒϥβʹ࣮͞Ε͍ͯΔ"1*ʹԠͯࣗ͡ಈతʹ࠷దͳόοΫΤϯυΛબ • ΄΅શͯͷڥͰ(16ʹΑΔϋʔυΣΞΞΫηϥϨʔγϣϯ͕ར༻Մೳ 8FC(16
#BDLFOE 8FC(- #BDLFOE 8FC"TTFNCMZ #BDLFOE 'BMMCBDL #BDLFOE
8FC%//ͷಛ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴԽ • ਪϑΣʔζʹಛԽ͢Δ͜ͱͰɺΑΓੵۃతͳܭࢉάϥϑͷ࠷దԽΛ࣮ࢪ
8FC%//ͷಛ طଘͷਂֶशϑϨʔϜϫʔΫͷൣғͳαϙʔτ • ,FSBT $BGGF $IBJOFS 5FOTPS'MPXͷֶशࡁΈϞσϧ͕มՄೳ • ಠ࣮ࣗͨؔ͠ͷରԠʢτϥϯεύΠϥͷ֦ுʣ༰қ
$IBJOFS chainer.Chain ,FSBT keras.Model $BGGF .proto 5FOTPS'MPX tf.Session
ϕϯνϚʔΫ 3FT/FU<)F > ը૾ຕͷਪʹཁ͢Δ࣌ؒ • ೖྗαΠζ • ൺֱର • ,FSBTKT
IUUQTHJUIVCDPNUSBOTDSBOJBMLFSBTKT LFSBTͷֶशࡁΈϞσϧΛ࣮ߦՄೳͳϥΠϒϥϦɻ8FC(-ʹΑΔ(16ར༻ɻ (1,224,224,3)
ϕϯνϚʔΫ 3FT/FUը૾ຕͷਪʹཁ͢Δ࣌ؒ ,FSBTKT 8FC%//
ॴཁ࣌ؒ <NTຕ> $ISPNF 'JSF'PY 4BGBSJ51 ,FSBTKT 8FC%// ॴཁ࣌ؒ<NTຕ> $ISPNF 'JSF'PY 4BGBSJ51 8FC%// 8FC%// .BD#PPL1SP&BSMZ()[*OUFM$PSFJ$16*OUFM*SJT(SBQIJDT(16 $16 (16 1$
ϕϯνϚʔΫ 3FT/FUը૾ຕͷਪʹཁ͢Δ࣌ؒ
,FSBTKT 8FC%// ॴཁ࣌ؒ <NTຕ> $ISPNF 'JSF'PY 4BGBSJ ,FSBTKT 8FC%// ॴཁ࣌ؒ <NTຕ> $ISPNF 'JSF'PY 4BGBSJ 8FC%// 8FC%// ˞ܽଛՕॴ࣮ߦෆՄೳɾαϙʔτ֎ 91&3*"9;"OESPJEY4OBQESBHPO()[Y$16"ESFOP(16 $ISPNF 'JSF'PY J1IPOFJ04"QQMF" 4BGBSJ εϚʔτϑΥϯ $16 (16
༻ํ๏ ,FSBTɾ$IBJOFSଐͷ3FT/FUֶशࡁΈϞσϧͷม IUUQTNJMUPLZPHJUIVCJPXFCEOOEPDTUVUPSJBMLFSBTIUNM IUUQTNJMUPLZPHJUIVCJPXFCEOOEPDTUVUPSJBMDIBJOFSIUNM
༻ํ๏ ֶशࡁΈϞσϧΛม͢Δ,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
༻ํ๏ ֶशࡁΈϞσϧΛม͢Δ$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'] $IBJOFS chainer.Chain (SBQI%FTDSJQUPS GPS8FC(- #BDLFOE 8FC%//*3 1ZUIPO
༻ํ๏ ֶशࡁΈϞσϧΛม͢Δ$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
༻ํ๏ ม݁ՌΛϒϥβ͔ΒಡΈࠐΈ࣮ߦ͢Δ let runner = await WebDNN.load('model'); let x =
runner.getInputViews()[0]; let y = runner.getOutputViews()[0]; جຊతͳલɾޙॲཧʹؔ͢ΔαϙʔτϥΠϒϥϦ͕ଐ x.set(await WebDNN.Image.getImageArray('cat.jpg', { dstW: 224, dstH: 224 })); await runner.run(); ࣮ߦ console.log(y.toActual()); JavaScript +BWB4DSJQU
ߴԽ ΣϒϒϥβίϯϐϡʔςΟϯάʹ͓͍ͯͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ ϝϞϦଳҬෆ
ߴԽ ΣϒϒϥβίϯϐϡʔςΟϯάʹ͓͍ͯͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ ϝϞϦଳҬෆ
+BWB4DSJQUͷΦʔόʔϔου • +BWB4DSJQUͦͷͷͷ࣮ߦ͕͍ • ϨΠϠΛ৮ΕΔ"1*͕૿͖͑ͯͨ • 7BOJMMB+BWB4DSJQUˠ8FC"TTFNCMZ • 8FC(-ˠ 8FC(-ˠ
8FC(16
8FC"TTFNCMZ Σϒϒϥβ͔ΒΞηϯϒϥϨϕϧͰͷ໋ྩΛѻ͑Δ༷ • +BWB4DSJQUͷΠϯλϓϦλ࣮ߦʹΑΔΦʔόʔϔου͕ͳͤ͘Δ • 7BOJMMB+BWB4DSJQUͰͰ͖ͳ͔༷ͬͨʑͳ࠷దԽ͕Մೳ • 4*.%ϚϧνεϨουՄೳʹ $ISPNF'JSF'PYͰळʙౙTIJQ͞Εͦ͏ C
C++ Rust .wasm ίϯύΠϧ ϒϥβ͔ΒόΠφϦΛ ಡΈࠐΈ࣮ߦ ϒϥβ
8FC(- • Σϒϒϥβʹ(1(16༻"1*͕ଘࡏ͠ͳ͍ • 8FC(- • ը૾ॲཧ༻"1*Ͱ͋Γ$PNQVUJOH4IBEFSαϙʔτ͞Εͯͳ͍ • ؤுΕ(1(16Ͳ͖Մೳ •
༷ʑͳ੍ͷ͍ͤͰΦʔόʔϔου͕େ͖͍ • 8FC(- • (1(16͘͢͠ͳΔػೳ͕Ճ • Φʔόʔϔου͕͔ͳΓݮΒͤΔ
8FC(16 • IUUQTXFCLJUPSHXQDPOUFOUVQMPBETXFCHQVBQJQSPQPTBMIUNM • "QQMF͕த৺ͱͳͬͯٞதͷ࣍ظ"1* • $PNQVUJOH4IBEFSΛαϙʔτɺඈ༂తͳύϑΥʔϚϯε্ • 8FC,JUʹ࣮ͨ͠ •
NBD04)JHI4JFSSBJ04ͷTBGBSJ͔Β༻Մೳ 8(ͷٞࣄΛݟ͍ͯΔͱ ʮͬͺ$PNQVUJOH4IBEFSෆཁͰʁʯ ͱ͍͏͕ٞͳ͞Ε͓ͯΓɺएׯӢߦ͖͕ո͍͠
ߴԽ ΣϒϒϥβίϯϐϡʔςΟϯάʹ͓͍ͯͱͳΔཁҼ +BWB4DSJQUͷΦʔόʔϔου ԋࢉίΞෆ ϝϞϦଳҬෆ
ϝϞϦଳҬෆͷରࡦ ,FSOFM.FSHJOH • ແବͳಡΈॻ͖ΛݮΒ͠ԋࢉີΛߴΊ • ෳͷΧʔωϧΛ̍ͭʹϚʔδ͢Δ • ൃߦͷΦʔόʔϔουݮΒͤΔ 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
άϥϑߏ࠷దԽ ࣄલʹొ͞ΕͨύλʔϯʹԊͬͯάϥϑΛม $POTUBOU'PMEJOH Add Var Const1 Const2 Const1 + Const2
ύλʔϯ ม ఆ ม
άϥϑߏ࠷దԽ ࣄલʹొ͞ΕͨύλʔϯʹԊͬͯάϥϑΛม Add x1 Add x3 x2 c1 c2
άϥϑߏ࠷దԽ ࣄલʹొ͞ΕͨύλʔϯʹԊͬͯάϥϑΛม Add x1 Add x3 x2 c1 c2 $POTUBOU'PMEJOH
Add x3 x2 c1 + c2
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 c1 c2 Add Add x2 Add (x1
+ c1) + (x2 + c2)
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 x2 c2 Add Add c1 Add (x1
+ c1) + (x2 + c2) = (x1 + x2) + (c1 + c2)
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 x2 c2 Add Add c1 Add $POTUBOU'PMEJOH
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Add Add c1 + c2 x2
/PSNBMJ[BUJPO &YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 w Conv2D x2 s Mul x3
b Add
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 w Conv2D x2 s Mul x3 b
Add &YQSFTTJPO4JNQMJGJDBUJPO
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add x4 s Mul
w
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add x4 s Mul
w $POTUBOU'PMEJOH
&YQSFTTJPO4JNQMJGJDBUJPO ఆ߲Λ·ͱΊɺࣜΛ୯७Խ͠ଞͷ࠷దԽΛ༰қʹ͢Δ x1 Conv2D x3 b Add w * s
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ 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
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ 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
&MFNFOUXJTF ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ 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
&YQSFTTJPO4JNQMJGJDBUJPO ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ 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
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D
Add x7 x8 x10 Relu b1 b2 x1 w1 * s1 x5 w2 * s2
&YQSFTTJPO4JNQMJGJDBUJPO ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9
Conv2D Add x7 x8 x10 Relu b1 b2 x1 w1 * s1 x5 w2 * s2
$POTUBOU'PMEJOH ,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9
Conv2D Add x8 x10 Relu b2 x1 w1 * s1 x5 w2 * s2 x7 b1
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D
x10 Relu x1 w1 * s1 x5 w2 * s2 x7 b1 + b2
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D Add Add x3 x4 x9 Conv2D
x10 Relu x1 w1 * s1 x5 w2 * s2 x7 b1 + b2 &MFNFOUXJTF,FSOFM.FSHJOH
,FSOFM.FSHJOH &MFNFOUXJTFͳΧʔωϧΛ̍ͭʹ·ͱΊɺϝϞϦͷॻ͖ࠐΈ ΧʔωϧͷσΟεύονʹΑΔΦʔόʔϔουΛݮΒ͢ Conv2D MergedElementwise x3 Conv2D x10 x1 w1
* s1 x5 w2 * s2 x7 b1 + b2 ճ ˠճ 3FBEɾ8SJUF
,FSOFM.FSHJOH ,FSOFM.FSHJOHʹΑΔ࠷దԽͷޮՌ 4BGBSJ51.BD#PPL1SP&BSMZ 3FT/FU࣮ߦ࣌ؒ <NTJNBHF> WebGPU backend 最適化なし WebGPU
backend ,FSOFM.FSHJOH
ϝϞϦଳҬෆͷରࡦ %BUB'PSNBU0QUJNJ[BUJPO ܭࢉάϥϑதͷมΛɺϝϞϦ্ʹͲͷΑ͏ʹ֨ೲ͢Δ͔Λ࠷దԽ • %BUB0SEFS 0QFSBUPSͷछྨɾೖग़ྗมͷܗঢ়ͳͲʹΑΓదͨ͠σʔλΦʔμʔҟͳ Δ • ͋Δ0QFSBUPS/)8$Ͱɺผͷ0QFSBUPS/$)8ͰσʔλΛѻ͍͍ͨɺ ͱ͍ͬͨ͜ͱ͕͋Δ
• લޙͷ0QFSBUPSՃຯͯ͠ೖग़ྗมͷ%BUB 0SEFSΛܾΊΔ
ϝϞϦଳҬෆͷରࡦ %BUB'PSNBU0QUJNJ[BUJPO ܭࢉάϥϑதͷมΛɺϝϞϦ্ʹͲͷΑ͏ʹ֨ೲ͢Δ͔Λ࠷దԽ • "MJHONFOUBOE1BEEJOH 4*.%໋ྩ(16ΧʔωϧதͰͷॲཧൣғͷׂׂΓͯʢ5JMJOHʣΛ༰қʹ ͢ΔͨΊʹߦྻͷαΠζɾΦϑηοτΛἧ͑Δ
ϝϞϦଳҬෆͷରࡦ ྫ (&..$"# ߦྻͷαΠζ͕λΠϧαΠζͰ ៉ྷʹׂΓΕΔΑ͏ ༧ΊύσΟϯά͓ͯ͘͠ " # $ 5JMF
ϝϞϦଳҬෆͷରࡦ ྫ (&..$"# ߦྻͷαΠζ͕λΠϧαΠζͰ ៉ྷʹׂΓΕΔΑ͏ ༧ΊύσΟϯά͓ͯ͘͠ " 1BEEJOH # $
%BUB'PSNBU 0QUJNJ[BUJPO %BUB'PSNBU 0QUJNJ[BUJPOʹΑΔ࠷దԽͷޮՌ 4BGBSJ51.BD#PPL1SP&BSMZ 3FT/FU࣮ߦ࣌ؒ <NTJNBHF> WebGPU backend 最適化なし
WebGPU backend + Elementwise Kernel Merging WebGPU backend + Elementwise Kernel Merging %BUB'PSNBU 0QUJNJ[BUJPO
·ͱΊ 8FC%//Σϒϒϥβ͚ਂֶशϞσϧߴ࣮ߦϑϨʔϜϫʔΫ IUUQTNJMUPLZPHJUIVCJPXFCEOO Σϒ࠷৽༷Λ༻͍ͨߴͳܭࢉόοΫΤϯυ 8FC(16ɾ8FC(-ɾ8FC"TTFNCMZʹΑΔ$16(16྆ํͰͷߴͳ࣮ߦ ܭࢉάϥϑ࠷దԽʹΑΔܭࢉͷߴԽ ਪϑΣʔζʹಛԽͨ͠࠷దԽʹΑΓܭࢉྔɾϞσϧύϥϝʔλαΠζΛݮ
طଘͷਂֶशϑϨʔϜϫʔΫͷൣғͳαϙʔτ ,FSBT $BGGF $IBJOFS 5FOTPS'MPXʹରԠ