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
pytorchで機械学習しない
Search
Hata Ryosuke
October 21, 2019
Technology
3
1k
pytorchで機械学習しない
pytorchでマクローリン展開とニュートン法を試してみました。
Hata Ryosuke
October 21, 2019
Tweet
Share
More Decks by Hata Ryosuke
See All by Hata Ryosuke
関西Kaggler会 発表スライド
ryosukehata
1
1.1k
Monocular 3D Object Detection Survey
ryosukehata
3
470
量子情報勉強会,量子ゲートについて
ryosukehata
0
220
Other Decks in Technology
See All in Technology
高速なプロダクト開発を実現、創業期から掲げるエンタープライズアーキテクチャ
kawauso
1
130
生成AIで小説を書くためにプロンプトの制約や原則について学ぶ / prompt-engineering-for-ai-fiction
nwiizo
4
3.2k
Liquid Glass革新とSwiftUI/UIKit進化
fumiyasac0921
0
300
監視のこれまでとこれから/sakura monitoring seminar 2025
fujiwara3
11
4k
Fabric + Databricks 2025.6 の最新情報ピックアップ
ryomaru0825
1
150
OPENLOGI Company Profile
hr01
0
67k
LangChain Interrupt & LangChain Ambassadors meetingレポート
os1ma
2
200
GitHub Copilot の概要
tomokusaba
1
140
rubygem開発で鍛える設計力
joker1007
2
260
Understanding_Thread_Tuning_for_Inference_Servers_of_Deep_Models.pdf
lycorptech_jp
PRO
0
150
BrainPadプログラミングコンテスト記念LT会2025_社内イベント&問題解説
brainpadpr
1
180
asken AI勉強会(Android)
tadashi_sato
0
130
Featured
See All Featured
Rails Girls Zürich Keynote
gr2m
94
14k
We Have a Design System, Now What?
morganepeng
53
7.7k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
281
13k
Producing Creativity
orderedlist
PRO
346
40k
Large-scale JavaScript Application Architecture
addyosmani
512
110k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
48
5.4k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
29
1.8k
The Power of CSS Pseudo Elements
geoffreycrofte
77
5.8k
KATA
mclloyd
30
14k
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
3.9k
The Art of Delivering Value - GDevCon NA Keynote
reverentgeek
15
1.5k
A better future with KSS
kneath
239
17k
Transcript
pytorchͰػցֶश͠ͳ͍ ػցֶश/Kaggle͘͘ձ#10ˏ େࡕ(10/21 19:00ʙ) ാɹྒྷհ
͜͡͠ΐ͏͔͍ twitter (@hattan0523) ϝʔΧʔۈ ීஈ$"%ϙνϙνͯ͠ɼ ͦΕͬΆ͍ਤΛग़͢ਓ ɹɹʢ෩ͷͱ͖ʹΨϥεʹςʔϓషΔ͖ͱ͔ܭࢉ͢ΔɻͲΜͳਤͰ࡞Ε·͢ʣ ػցֶशओۀͰ͋Γ·ͤΜ େֶ࣌ʹྔࢠޫֶͷݚڀʹूத͠ɼ
ໟΛࣦ͍ത࢜Λऔಘɻ http://www.breault.com/software/asap-nextgen https://www.muratasoftware.com/products/ examples/watgal004/
ಈػ pytorchͱ͔ਂֶशͷϥΠϒϥϦɼ ίϯϖʹ͍ͬͺ͍ΘΕͯΔ͠ ը૾ͱ͔NLPͷਂֶशΛΔͱ͖ ͱΓ͋͑ͣͬͱ͖Ό͑͑Ζɻ https://twitter.com/nino_pira/status/1181913845507354626
ಈػ pytorchԿΛͬͯΔͷʁ ͡Ͳ͏ͼͿΜʁ ͠Μͦ͏͕͘͠ΎʔͷϥΠϒϥϦͰʁ ͦ͏ͩʂ ඍΛ͠Α͏ʂ
ࠓճɿࣗಈඍΛͬͯܭࢉ ϚΫϩʔϦϯల։ sinؔΛ10࣍·Ͱ sinؔΛ20࣍·Ͱ χϡʔτϯ๏ɹɹɹɹɹɹɹɹͷղΛग़͢
ࣗಈඍ is Կʁ https://github.com/pytorch/pytorch/blob/edb88b5f3af03718b443d015f195faa1832ce95b/caffe2/operators/sin_op.cu ɾඍ ࣮ࡍʹಋؔͷఆٛʹैͬͯతʹܭࢉ͢Δɻ ޡ͕ࠩͰ͖Δɻ ɾࣗಈඍ ܭࢉ͢ΔؔͷಋؔΛ༧Ίఆ͓ٛͯ͘͠ɻ ؔʹೖ͢Δ͚ͩͳͷͰޡࠩ΄΅ͳ͍ɻ
pytorchͷsinͷಋؔͷఆٛ ֻ͚ͷͱ͜ΖΛݟΔͱɼ ಋؔ(cos)͕ఆٛ͞Ε͍ͯΔ ͜ͱ͕Θ͔Δɻ ࢀߟɿࣗಈඍΛ࣮ͯ͠ཧղ͢Δ https://qiita.com/lotz/items/39c52f08cc9b5d8439ca https://qiita.com/lotz/items/f1d4ab1d83dc13a5d81a
ͬͯΈΔ ී௨ʹΈͳ͞Μ͕࣮ߦ͍ͯ͠Δ͜ͱɻ requires_grad=Trueͱ͢Δ͜ͱͰɼඍ͢ΔΑʔͬͯએݴ͢Δɻ backward()ͰܭࢉάϥϑʹԊͬͯࣗಈඍ͕ݺͼग़͞Εɼ x.gradͰඍ͕ಘΒΕ·͢ɻ ࢀߟɿPyTorchͰߴ֊ภඍ (https://qiita.com/tmasada/items/ 9dee38e5bc1482217493)
ೋ֊ඍ ‘torch.autograd.grad(f, x, create_graph=True)’Ͱɼfʹ͍ͭͯxͰඍ ͢ΔΑʔͱએݴ͢Δɻ ܭࢉάϥϑΛ࡞Βͳ͍ͱඍͯ͘͠Εͳ͍ɻ ࡉ͔͍͜ͱ্ͷQiitaͷهࣄͰɻ ࢀߟɿPyTorchͰߴ֊ภඍ (https://qiita.com/tmasada/items/ 9dee38e5bc1482217493)
ԿճͰඍͰ͖·͢ʂ ࢀߟɿPyTorchͰߴ֊ภඍ (https://qiita.com/tmasada/items/ 9dee38e5bc1482217493) ඍ͕Ͱ͖ΔʂͰ͖Δͧʂʂ
ϚΫϩʔϦϯల։ Β͔ͳؔΛ্ͷΑ͏ͳܗʹల։ͯ͠ɼ ۙࣅతͳؔͱͯ͠ද͢͜ͱ͕Ͱ͖Δɻ ֶಘҙ͡Όͳ͍ͷͰɼ ݫີͳఆٛΑ͘Θ͔͍ͬͯ·ͤΜɻ ͜ΕΛpytorchͰ࣮ͯ͠ΈΑ͏ʂ (kaggleͷnotebookࢀর)
sinؔͰಘΒΕͨάϥϑ ͬͨ͜ͱ 1.ҙͷ࣍·ͰܭࢉάϥϑΛ࡞ͯ͠ɼඍΛܭࢉ͢Δ 2.֤ԣ࣠ͷʹରͯ͠ܭࢉͨ͠ඍͰॎ࣠ͷΛग़͢ɻ େมͳͷͰ1,2ΛߦྻͰܭࢉ͢Δ sinؔΛ10࣍·Ͱ sinؔΛ20࣍·Ͱ https://www.kaggle.com/hattan0523/pytorch-maclaurin-series? scriptVersionId=22206309
χϡʔτϯ๏ ඍΛͬͯํఔࣜΛղ͘ํ๏ ʹ͍ͭͯࣗಈඍΛͬͯղ͍ͯΈΔɻ ղ x=ln2≒0.69314718056 C++ - ඇઢܗํఔࣜͷղ๏ʢχϡʔ τϯ๏ʣʂ https://www.mk-mode.com/blog/
2012/11/21/21002047/# ΑΓҾ༻
σϞ̎ɿ݁Ռ ࣮ίʔυ: https://www.kaggle.com/hattan0523/newton-raphson-method-by- pytorch ղ x=ln2≒0.69314718056 ॳظͱؔΛ༩͑ΕɼҙͷճܭࢉΛߦͬͯ͘ΕΔɻ ֓ͶऩଋͰ͖ͨʂ
·ͱΊ ɾpytorchࣗಈඍͷϥΠϒϥϦ Ͱ͌ʔΒʔʹΜ͙Ͱͳͦ͞͏ɻ ɾܭࢉάϥϑΛ໌ࣔతʹ࡞ͬͯ͋͛Δ͜ͱͰ ɹඍՄೳͳؔΛ͍ͬͯΖΜͳ༡ͼ͕Ͱ͖ͦ͏ɻ ɾྫͱͯ͠ɼ sinؔͷϚΫϩʔϦϯల։ χϡʔτϯ๏ΛͬͨํఔࣜͷղΛಋग़ͨ͠ɻ ɾඍָ͍͠ΊΔͬ