TensorFlowと機械学習の今

%FWFMPQFST*0 & ฏాܓ "84ίϯαϧςΟϯά෦ Ϋϥεϝιου ⡥$MBTTNFUIPE *OD ೥݄೔ 5FOTPS'MPXͱػցֶशͷࠓ

⡥$MBTTNFUIPE *OD ࣗݾ঺հ ໊લ : ฏా ܓ (@masuwo3) ॴଐ :
σʔλ෼ੳνʔϜ ݱࡏ : 2015೥5݄ΑΓΫϥεϝιουೖࣾ աڈ : ಛ೚ॿڭɺSEɺITߨࢣɺetc... ML : ݚڀࣨͰ͔ͬͨ͡ఔ౓

⡥$MBTTNFUIPE *OD

⡥$MBTTNFUIPE *OD ࠓ೔͓࿩͢͠Δ͜ͱ ਂ૚ֶशͱ5FOTPS'MPXʹ͍ͭͯɺ ࣗ෼͕͜Ε·ͰʹֶΜͰ͖ͨ͜ͱΛɺ ͬ͘͟Γͱ࿩͠·͢ɻ

⡥$MBTTNFUIPE *OD ΞδΣϯμ w ਂ૚ֶशʹ͍ͭͯͷ֓ཁ w 5FOTPS'MPXͷ࠷ۙͷಈ޲ w ਂ૚ֶशͷࣄྫʹ͍ͭͯ %FFQ2/FUXPSL

ਂ૚ֶश

⡥$MBTTNFUIPE *OD ਂ૚ֶशͷ֓ཁ w ਂ૚ֶशͱ͸ͳʹ͔ʁ w ଟ૚͔ΒͳΔχϡʔϥϧωοτϫʔΫʹΑΔֶश

⡥$MBTTNFUIPE *OD w χϡʔϥϧωοτϫʔΫͱ͸ʁ w ਆܦࡉ๔ͷಈ͖Λ࠶ݱ͢ΔͨΊʹ࡞ΒΕͨϞσϧ w ݱࡏ͸ύʔηϓτϩϯϞσϧΛ༻͍ͨܭࢉωοτϫʔΫશൠ χϡʔϥϧωοτϫʔΫͷ͓͞Β͍
x1 x2 x3 w3 w2 w1 u y y = u ( w1x1 + w2x2 + ... )

⡥$MBTTNFUIPE *OD χϡʔϥϧωοτϫʔΫͷ͓͞Β͍ w χϡʔϥϧωοτϫʔΫʹ͓͚Δֶश w ֶश͸ύʔηϓτϩϯͷॏΈΛߋ৽ͯ͠ߦ͏ w
ޡࠩؔ਺Λ࠷খͱ͢ΔॏΈΛ୳ࡧతʹٻΊΔ x1 x2 x3 u y w1 w2 w3

⡥$MBTTNFUIPE *OD χϡʔϥϧωοτϫʔΫͷ͓͞Β͍ w ύʔηϓτϩϯͷू߹ֶशϞσϧ

⡥$MBTTNFUIPE *OD w ଟ૚ωοτϫʔΫͷݶք w ૚Ҏ্ͷωοτϫʔΫͰ͸ֶश͕҆ఆ͠ͳ͍ w $//ͳͲɺҰ෦ͷख๏Ͱ͸੒ޭ w ͳͥࠓ·Ͱଟ૚ԽͰ͖ͳ͔ͬͨͷ͔ʁ
w ޯ഑ফࣦ໰୊ w ଟ૚ͷ৔߹ɺೖྗ૚ʹۙͮ͘΄Ͳޡ͕ࠩগͳ͘ͳΔ w ݁Ռͱͯ͠ɺֶश͕҆ఆͤͣɺਫ਼౓͕޲্͠ͳ͍ ଟ૚ωοτϫʔΫͷ໰୊఺

⡥$MBTTNFUIPE *OD ਂ૚ֶशͷొ৔ w ଟ૚ωοτϫʔΫʹΑΔֶश͕࠶஫໨ w ͖͔͚ͬ͸%FFQ#FMJFG/FUXPSL w 3#.Λ༻͍ͯஈ֊తʹ//Λߏங w
ͳͥͰ͖ΔΑ͏ʹͳͬͨͷ͔ʁ w ஈ֊తʹ//Λߏங͢Δ͜ͱͰɺॳظ஋Λ҆ఆԽ w ଟ૚//ʹ஫໨͕ू·Δ w ݱࡏ͸༷ʑͳख๏Ͱফࣦ໰୊Λճආ͍ͯ͠Δ

⡥$MBTTNFUIPE *OD ਂ૚ֶशͷಛ௃ w ෳࡶԽ͢ΔϞσϧ w HPPH-F/FU w ֶशʹ͔͔Δܭࢉίετ͕രൃ
w ख๏ͷݕূ͕ࠔ೉

5FOTPS'MPX

⡥$MBTTNFUIPE *OD 5FOTPS'MPXʹ͍ͭͯ w (PPHMF੡ͷػցֶशϥΠϒϥϦ w %FFQ-FBSOJOHҎ߱ w 1ZUIPO w
$6%" 2015/6/9 2015/11/20 2013/10/20 caffe v0.1 chainer ެ։ TensorFlow ެ։ …..

⡥$MBTTNFUIPE *OD w ॊೈੑ w ػցֶशͷϞσϧΛॊೈʹهड़Ͱ͖Δ w /FVSBM/FUXPSLʹݶఆ͠ͳ͍ w ϙʔλϏϦςΟ
w ؀ڥʹ߹ΘͤͯܭࢉॲཧΛߦ͏ w $16(16 ϥοϓτοϓαʔόͳͲ w ݚڀ੒Ռͱ੡඼ͷ࿈݁ w ݚڀ੒ՌͷݕূΛߦ͍΍͘͢ w ϓϩμΫτʹస༻͠΍͘͢ 5FOTPS'MPXͷಛ௕ ެ͔ࣜΒ

⡥$MBTTNFUIPE *OD 5FOTPS'MPXʹ͓͚Δܭࢉॲཧ w 5FOTPS'MPX͸ܭࢉΛάϥϑߏ଄Ͱදݱ͢Δ Y ˎ ʴ Z
C 8

⡥$MBTTNFUIPE *OD 5FOTPSͱ0QFSBUJPO σʔλ͸Tensor ܭࢉॲཧ͸Operation 5FOTPS 5FOTPS op 5FOTPS
5FOTPS

⡥$MBTTNFUIPE *OD 5FOTPSʹ͍ͭͯ ֊਺0ͷςϯιϧ (εΧϥ) ֊਺1ͷςϯιϧ (ϕΫτϧ) ֊਺2ͷςϯιϧ (ߦྻ)
֊਺3ͷςϯιϧ < ʜ> << ʜ> < ʜ> < ʜ> ʜ> <<< ʜ> ʜ> << ʜ> ʜ> << ʜ> ʜ> ʜ> B B<> B<><> B<><><>

⡥$MBTTNFUIPE *OD 4FTTJPOʹ͍ͭͯ w ܭࢉॲཧ͸ϑΣΠζ %FpOF3VO Define Run •
ܭࢉͷάϥϑϞσϧΛߏங͢ΔϑΣΠζ • ͜ͷ࣌఺Ͱ͸ܭࢉ݁Ռ͸֬ఆ͠ͳ͍ • άϥϑϞσϧ͔Βܭࢉ݁ՌΛ֬ఆ͢ΔϑΣΠζ • SessionʹϞσϧΛ౤ೖ͠ɺܭࢉ݁ՌΛಘΔ

⡥$MBTTNFUIPE *OD 5FOTPSͱ0QFSBUJPO w 4FTTJPO w όοΫΤϯυͷ$ ϞδϡʔϧͱͷίωΫγϣϯ w ࣮ࡍͷԋࢉ͸͜ͷ$
Ϟδϡʔϧ্ͰߦΘΕΔ w ܭࢉϦιʔε͕ࣗಈతʹׂΓ౰ͯΒΕΔ Y ˎ ʴ Z C 8 cpu:0 gpu:0

⡥$MBTTNFUIPE *OD 0QUJNJ[FSʹ͍ͭͯ w ଛࣦؔ਺ͷ࠷খԽ w ػցֶशͷʮֶशʯ͸ɺଛࣦؔ਺ͷ࠷খԽʹஔ͖׵͑ΒΕΔ w ࠷খ఺ͷ୳ࡧʹ͸ɺޯ഑ܭࢉ͕ඞཁ w
5FOTPS'MPXͰ͸0QUJNJ[FSͰֶशΛҰׅͯ͠ߦ͏

⡥$MBTTNFUIPE *OD 0QUJNJ[FSʹ͍ͭͯ w 0QUJNJTFS w ޯ഑ܭࢉ͔Β୳ࡧ·ͰΛҰׅͰߦ͏ w ޯ഑ͷܭࢉ͸ɺάϥϑϞσϧΛࣗಈతʹม׵ w
ֶशख๏ʹ߹Θͤͯ಺෦ͷॏΈΛߋ৽͍ͯ͘͠ w άϥϑͰදݱ͞Εͨଛࣦؔ਺͔ΒɺࣗಈతʹֶशΛߦ͏

⡥$MBTTNFUIPE *OD 5FOTPS'MPXͷػೳ w 5FOTPS#PBSE ֶशաఔ΍Ϟσϧͷߏ଄ΛՄࢹԽ

⡥$MBTTNFUIPE *OD ͳͥ.-ϥΠϒϥϦ͕ٻΊΒΕ͍ͯΔ͔ w ੒ՌΛڞ༗͢Δ্Ͱͷڞ௨ج൫ͱͯ͠ w ͓ޓ͍ͷ੒ՌΛڞ༗͢Δ্Ͱͷڞ௨ݴޠͱͯ͠ w ϞσϧΛͲͷΑ͏ʹߏங͢Δ͔͸ϊ΢ϋ΢Ϩϕϧ w
ࢼߦࡨޡʹΑͬͯ৽ͨͳൃݟΛظ଴ w ίετ໰୊ͷղܾ w ΑΓޮ཰తͳܭࢉॲཧͷͨΊͷج൫͕ඞཁ w (16Λ༻͍ͨฒྻͳԋࢉ w Ϛϧν(16ʹΑΔฒྻॲཧ w େن໛ΫϥελʹΑΔεέʔϧΞ΢τ

⡥$MBTTNFUIPE *OD %JTUSJCVUFE5FOTPS'MPX w ʮϚγϯΫϥελΛͭͷֶशػցͱͯ͠ѻ͏ʯ w 5FOTPS'MPX͸ɺʮ͍͔ʹͯ͠ܭࢉΛ෼ࢄͤ͞Δ͔ʯΛॏࢹ w ෳ਺Ϛγϯʹద༻ͨ͠ϥΠϒϥϦΛۙ೔ެ։༧ఆ
w ؾʹͳΔ఺ w ֶशύϥϝʔλͷಉظ͕ϘτϧωοΫ w (PPHMF͸௿஗ԆωοτϫʔΫʹΑͬͯߴ଎ԽΛ࣮ݱ w ಈ͘؀ڥ͸͔ͳΓݶΒΕͦ͏

%FFQ2-FBSOJOH

⡥$MBTTNFUIPE *OD %FFQ2/FUXPSL w %FFQ-FBSOJOH ڧԽֶश w Ϟχλը૾͔ΒήʔϜͷઓུΛֶश w ಉ͡ϞσϧͰ༷ʑͳήʔϜΛ߈ུ
w ήʔϜʹΑͬͯ͸ϓϩҎ্ͷ੒Ռ XXXOBUVSFDPNOBUVSFKPVSOBMWOGVMMOBUVSFIUNM

⡥$MBTTNFUIPE *OD ڧԽֶश w ؀ڥͱΤʔδΣϯτ w ΤʔδΣϯτ͸؀ڥΛ஌֮͠ɺͦΕʹର͢ΔߦಈΛܾఆ w ߦಈʹΑͬͯརಘΛಘΒΕΔ ಘΒΕͳ͍৔߹΋͋Δ
w աڈͷܦݧ͔Βɺ؀ڥʹରͯ͠࠷దͳߦಈΛܾఆ͢Δ ΤʔδΣϯτ ؀ڥ

⡥$MBTTNFUIPE *OD 2-FBSOJOH w ߦಈՁ஋ؔ਺ʹΑͬͯߦಈΛܾఆ͢Δ w ঢ়ଶɺߦಈΛೖྗͱ͢ΔߦಈՁ஋ؔ਺ΛٻΊΔ w աڈͷߦಈཤྺ͔ΒߦಈՁ஋ؔ਺Λߋ৽͠ɺ࠷దͳߦಈΛܾ ఆ͢Δ
ߦಈঢ়ଶ T T T ʜ B B B ʜ

⡥$MBTTNFUIPE *OD 2/FUXPSL w ߦಈՁ஋ؔ਺Λ//Ͱදݱ w ݹయతͳςʔϒϧؔ਺ͩͱঢ়ଶ͕രൃ͠΍͍͢ w ϝϞϦͷ໰୊ɺαϯϓϦϯάͷ໰୊ w
//ʹΑͬͯߦಈՁ஋ؔ਺Λۙࣅͤ͞Δ w ߦಈՁ஋ؔ਺ͷਫ਼౓͸//ͷදݱྗʹґଘ

⡥$MBTTNFUIPE *OD %FFQ2/FUXPSL w ৞ΈࠐΈ// $// ΛධՁؔ਺ʹར༻ w ը໘৘ใΛ௚઀ঢ়ଶͱͯ͠ѻ͍ɺߦಈΛܾఆ͢Δ

⡥$MBTTNFUIPE *OD ࣮ݧ݁Ռ w "UBSJ7JEFP0MZNQJDT

⡥$MBTTNFUIPE *OD %2/ͷ໘ന͍ͱ͜Ζ w ΞΠσΞ͸ͱͯ΋෼͔Γ΍͍͢ w 2/FUXPSLʹରͯ͠$//Λద༻ͯ͠Έͨ w ਂ૚ֶश͕ଞ෼໺ʹରͯ͠΋Өڹ͍ͯ͠Δ w
ಛ௃ྔͷࣗಈநग़ͱ͍͏ੑ࣭Λ͏·͘ར༻ w ಉҰͷֶशϞσϧ͔ΒɺϧʔϧͷҟͳΔήʔϜΛ߈ུ͢Δ w $//Λར༻͢Δ͜ͱͰɺΑΓ޿൚ͳڧԽֶशϞσϧ

⡥$MBTTNFUIPE *OD ·ͱΊ w ػցֶशͷࠓਂ૚ֶश w ͨͩ͠ɺະͩʹख๏ʹ͍ͭͯ͸ࢼߦࡨޡ w .-ϥΠϒϥϦ 5FOTPS'MPX
ͷొ৔ w ૿େ͢Δܭࢉίετ໰୊ͷղܾ w ੒ՌΛڞ༗͢ΔͨΊͷڞ௨ج൫ w ਂ૚ֶशͷ੒Ռ͸ଞ෼໺ͷػցֶशʹ΋Өڹ

%FWFMPQFST*0 " ⡥$MBTTNFUIPE *OD #cmdevio2016 ͝੩ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠ɻ ϒϩά͸ޙ΄Ͳެ։͠·͢ɻ

TensorFlowと機械学習の今

TensorFlowと機械学習の今

More Decks by Kei Hirata

Other Decks in Technology

Featured

Transcript