Kaggle ã¨åºä¼ã£ã¦ãã Master ã«ãªãã¾ã§ãããã¦ããããã
ã¯ããã«
以ä¸ã® Tweet ã§è¨ã£ã¦ãã"é·ãæ è·¯"ã®è©³ç´°ã§ããå®å ¨ã«ãã¨ã ãã¤é·æ*1ããããèªå·±é¶é ãå¤åã«å«ãã èªåèªãã§ãã
æãã¤å¿ã«ä½è£ããã£ã¦ä½ã§ãå¯å¤§ã«è¨±ããããªå ´åã«ãèªã¿ãã ããã
å¥è·¡ã® solo gold ãç²å¾ããéã« Kaggle Master ã«ãªãã¾ããï¼
— 俵 (@tawatawara) 2020å¹´3æ18æ¥
Kaggle ã¨åºä¼ã£ã¦ãã4å¹´åãä¸å¿µçºèµ·ãã¦åæã³ã³ãã«æ¬æ°ã§åãçµã¿å§ãã¦ãã1å¹´åãé·ãæ è·¯ã§ããã
ä»ã¾ã§ãã¼ã ãçµãã ãè°è«ããããã¦ä¸ãã£ãçãããæ¬å½ã«ãããã¨ããããã¾ããï¼ããããããããããé¡ãè´ãã¾ãï¼ pic.twitter.com/IzWin2rioa
話ãããããããã¯ï¼å¹´... ããï¼å¹´ãããåã®è©±ã ã£ãã...
ã¨ãã£ãæãã®ãããã¼ããããããã®è©±*2ãKaggle ã¨åºä¼ã£ã¦ããã®äººç(ã®ä¸é¨)ã®ç´¹ä»ã¿ãããªãã®ã«ãªã£ã¦ãã¾ãã
â»æ³¨æï¼ãã®è¨äºã§ã¯ä¸»æ¨ã¨ãã¦éè¦ãªã®ããã£ã¦è²ã ãªæ¹ã ãã¼ãããå½¢ããã㯠by name (Twitterã»Kaggleä¸ã«åå¨ããæ¹ã ã®ã¿) ã§ç´¹ä»ãã¦ãã¾ãã ãã£ã¨ã¼ããã¦æ¬²ããã¨ãããã¨ãããã°è¨ã£ã¦ãã ãããé常ã«å 輪ãã¿æãããã¾ãããã¾ãå人ã®ããã°ãªã®ã§ã誰ãããã¤!?ãã¦ãªã£ã¦ãã許ããã ããðââï¸
ç®æ¬¡
- ã¯ããã«
- Kaggle ã¨ã®åºä¼ã
- åæã³ã³ãæ¬æ ¼åæ¦åã®è©±
- å 輩ã®èªãããå§ã¾ãã³ã³ãçæ´»ï¼ACM RecSys Challenge 2017
- åãã¦ã®ã½ãã³ã³ãï¼SIGNATE ã¢ããã«å¼è¶ãéè¦äºæ¸¬
- ã¢ããã¼ã·ã§ã³ââï¼ã¹ãã¢ã©ã人工ç¥è½ã»ããã¼è´è¬
- å¾è¼©ã¨ã®ææ¦ï¼PAN Author Profiling Task 2018
- ã¨ããã§ï¼Kaggle ã¯ãã£ã¦ãªãã£ãã®ï¼
- åæã³ã³ãæ¬æ ¼åæ¦ã決æãã¦ããã®è©±
- èæ°´ã®é£ãæ·ãï¼ã¾ãã¯è²·ã£ãããªã³ãã¬ãã·ã³
- é¾å¤ã§ã¤ã±ã¤ã±ðãã³ãã³ð : SIGNATE å°é¢¨ã³ã³ã
- éããåå ð å ¨ç¶ããããð : Kaggle Tokyo Meetup #5
- åãã¦ã®ãã«ãã©ãã«åé¡ : Human Proteins Atras Image Classification
- çæåæ¦ã§ç¡äºæ»äº¡ð: é»ç· & Pet
- ã½ãã§ã®ç»åã³ã³ãåã³ðï¼iMetCollection2019 - FGVC6
- ç¥ããå±ãð : Instant Gratification (1 week)
- é å¼µã£ã¦LTã«ææ¦ : Kaggle Tokyo Meetup #6
- ã¡ãã£ã¨ä¸ä¼ã¿ : 4ãæããã
- 復活ã®ãã£ãã : Kaggle Days Tokyo
- ãã¿ããå¼ãã¦éã ããð : ASHRAE - Great Energy Predictor III (2 Hours)
- 念é¡ã® Solo Medal ð : 2019 Data Science Bowl
- Shakeup åãé ããï¼ : AtmaCup#3
- è¦è³ãç¶ãã 4 Daysð / Solo Gold ã¸ã® 4 Daysð : Bengali.AI Handwritten Grapheme Classification
- ã¾ã¨ã
- ããããã Kaggle ããã®ï¼
Kaggle ã¨ã®åºä¼ã
ãåé ã® Tweet ã Kaggle acount ã®ç»é²æ¥ã®éããåã Kaggle ã«åºä¼ã£ãã®ã¯å¤§ä½4å¹´ååãå¦çæ代(M2)ã¾ã§é¡ãã¾ããå½æããã¼ã¿ãã¤ãã³ã°ã¨ããè¨èã¯é »ç¹ã«ä½¿ããã®ã®æ©æ¢°å¦ç¿ã«é¢ãã¦ã¯å ¨ã馴æã¿ãç¡ããã¾ããã®æç¹ã ã¨ç 究室ã®å çæ¹ããã¾ãè¯å®çã§ã¯ãªãæãã§ãã*3ã
ãããªä¸ã ãã®åå¹´ã«ç 究ç§ã«èµ´ä»»ãããæ©æ¢°å¦ç¿æ¿ã¤ãææ*4ã®éããå ¬éã»ããã¼ã§ä»¥ä¸ã®è¬æ¼ãè¡ããã¾ãã
ãããKaggler ã®çãããåç¥ãæ¥æ¬ã代表ãã Grand Master ã®ä¸äººã§ãã smly ãã ã®è¬æ¼ã§ããã ã¨ãã£ã¦ããã®æç¹ã§ã¯åã¯ãã¾ããã®ããããèªèãã¦ããããå½æã® Tweet ãæ¼ã£ã¦ã¿ãã¨ãããªæãã
KDD cup åªåè ã®äººã®è©±çµæ§é¢ç½ãã£ã
— 俵 (@tawatawara) October 16, 2015
(ãªãã§ããã¤ä¸ããç®ç·ãªãã...) ããããã£ããã¨ãªããå人ã®èªãããã£ã¦ Kaggle ã«ç»é²ããã®ã§ããã
ããã ãçµå±ãã®æã¯ã¡ããã£ã¨ã ã触ã£ã¦ããé¢ãã¦ãã¾ãã¾ããM2 ã§ä¿®è«ãå·®ãè¿«ã£ã¦ãããã¨*5ãããã¦å½æã¯ç 究室ã®å 輩ã®å½±é¿ã§ç«¶æããã°ã©ãã³ã°ã«ç±å¿ã«åãçµãã§ãããã¨ã大ããªçç±ã§ãããåã® Python (㨠C++) ã®æ¨æºã©ã¤ãã©ãªã¸ã®ç解åã³ã¢ã«ã´ãªãºã ã®å®è£ è½åã¯ã»ã¼å ¨ã¦ç«¶ããããã¼ã¹ã¨ãªã£ã¦ãã¾ãã
ãã£ããããããããï¼ï¼ï¼ï¼ï¼éã ããããããï¼ï¼ï¼ï¼ pic.twitter.com/y53TBJKZqu
— 俵 (@tawatawara) July 19, 2016
M1ã®å¬ããå§ããM2 ã®ç§ã«ç®æ¨ã¨ãã¦ãã CODE FESTIVAL æ¬æ¦ã«ç¡äºåå ( + ã©ãã«ãããã«ããã¼ã«ã¼ç²å¾)ã社ä¼äººã«ãªã£ã¦ããããã¾ããã£ã¨ç®æ¨ã«ãã¦ãã Top Coder Div1 (é) ææ ¼ãæããã¾ãããã®ãã¨ãç¶ãããããããã£ãã®ã§ãããå¾ã
ã«é¢ãã¦ãã¾ã£ã¦ãã£ãããããªããªã£ã¦ãã¾ãã¾ãã . . . ð
å ã¿ã«åãä¸çªç«¶ããã«ç±ä¸ãã¦ãã¨ãã£ã¦ AtCoder ã®ã³ã³ãã¹ãåå è
ã 500 ãè¡ã£ã¦ãªãã£ãæ代ã§ãä»ã ã¨ãã®10ï½20åã®åå è
ããã¦æ¬å½ã«ããã¯ãªãã¾ããä»ãã£ããéè²ãªãã®çµ¶å¯¾ç¡çã ãããª...
ãããªãããªã§ãããããæ«ãã³ã³ãã¹ãã»ã³ã³ããã£ã·ã§ã³ã«é¢ããã®ãªãæéãç¶ãã®ã§ãããã¨ãããã£ããããåæã³ã³ãã«å¾ã ã«ã®ããè¾¼ãã§ããã¾ãã
åæã³ã³ãæ¬æ ¼åæ¦åã®è©±
å 輩ã®èªãããå§ã¾ãã³ã³ãçæ´»ï¼ACM RecSys Challenge 2017
ãã¨ã¦ãæ°ã®åãè·å ´ã®ä¸ã¤ä¸ã®å 輩 ã¨å°æ¬ãã大å 輩*6 ã®èªããåãã¦ä½ãããããããã¾ã¾åå ãã¾ãããã¹ã³ã¢ãä¸ãã£ã¦ããã®ã楽ããã¦åèªèº«ãã£ã¡ãé å¼µããå 輩æ¹ã¨åãçµéããçµæ 3ä½ã¨ããåã«ã¨ã£ã¦æå¤ã®ç´ æ´ãããçµæãå¾ããã¾ããã
XGBoost ã®ä½¿ãæ¹ããããç¨ãã(content-base ã®) cold start recommendation ã«ã¤ãã¦å¦ã¹ãã®ã¯å¤§ããªçµé¨ã¨ãªã£ã¦ãã¾ããããªã«ãããã®æããããã³ã³ãã§è¯ãæ績ãåããã¨ãã快楽ã«å¾ã ã«æã¾ã£ã¦ããã¾ããéå»ã«åãçµãã§ãã競ããã«ã¯ãé£ãã(ã¨åãæãã)åé¡ã解ãã¦ACãããã¨ããã¬ã¼ããä¸ãã£ãã¨ãã«åã³ãè¦ãã¦ç±ä¸ãã¦ããããã§ãç´ é¤ã¯åäºåã«ãã£ãã®ã§ãããã
åãã¦ã®ã½ãã³ã³ãï¼SIGNATE ã¢ããã«å¼è¶ãéè¦äºæ¸¬
ãå人åå ã¨ããæå³ã§ã¯åãã¡ããã¨åãçµããã¨ãåºæ¥ãæåã®ã³ã³ãã§ãããã·ã³ãã«ãªãã¼ã¿ãã¤ã«ã¬ã³ãã¼çã¨ãã¬ã³ããèæ ®ãã¦äºæ¸¬ã¨ãããå½æã®åã®ä»äºã«è¿ãã®ããã£ã¦é å¼µã£ã¦ã¾ããã 32 ä½ã¨ããå¾®å¦ãªçµæã«çµãã£ã¦ãã¾ãæ®å¿µããã ãã·ã³ãã«ãªãã¼ã¿ãªã ãã«è²ã 試ããããã£ãã®ã§æ¥½ããã§ããããã«æãã¾ãã
å ã¿ã«ãã¨ããè¦è¿ãã¦ã¿ãã¨ãå®ã¯ãã®ã³ã³ãã¯1ä½ããã® mamas å çã ã£ãããé ä½è¡¨ã«ã¡ãã»ãç¥ã£ã¦ãã kaggler ãå± ããã¨ã¡ãã£ã¨é¢ç½ãã§ãã
ã¢ããã¼ã·ã§ã³ââï¼ã¹ãã¢ã©ã人工ç¥è½ã»ããã¼è´è¬
第14å: ãã¼ã¿åæã³ã³ãã¹ãã«ãããæè¡ã¨æè¿ã®é²å±
ã¹ãã¢ã©ã人工ç¥è½ã·ã³ãã¸ã¦ã 2018: ãã¼ã¿åæã³ã³ãã¹ãã«æ³¨ç®ãã¦
åè 㯠smly ããããå¾è 㯠smly ãã㨠iwiwi ãã ãå«ã speaker ããè¬æ¼ããããåã«ã¨ã£ã¦æé«ã®ã¤ãã³ãã§ããã
競ããããã£ã¦ããåãããã㨠iwiwi ããã¯ç¥ã«çããåå¨ãªããã§ãã*7ãåãã¦åå ããã³ã³ãããããåãã¦æãã ç»åã³ã³ã㧠solo gold ãåããã¦ãã¦ããã®éç¨ã失æããé¨åããã£ããããããã¡ããã¡ããããããçºè¡¨ã§ããã
smly ããã®çºè¡¨ãèããã¨ãããã®å
容ã©ããã§... ããï¼å¦çæ代ã«è´ãããã¨ãããªï¼ãã¨è¨æ¶ãèãã¾ããå
容ãå
¨é¨ç解ã§ããªãã¦ã smly ãããããã« Kaggle ã好ãããä¼ãã£ã¦ããçºè¡¨ã§ãä»æ¹ãã¦è¦è¿ãã¦ã¿ãã¨æè¡çãªå
容ãã ãã¶ç解ã§ããã®ã§ããã¯æé·ãæãããã¤ã³ãã
ã¾ãã(確ãåè
ã®è¬æ¼ã®ã¨ãã ã£ãã¨æãã®ã§ããã) åããçµæ§é·æéã®åå ã«ãªãã¨æããã§ããã©ãã©ããã£ã¦ã¢ããã¼ã·ã§ã³ãä¿ããã¦ãããã§ããï¼ãã¨è³ªåããã¨ãã«ã
ãé ä½è¡¨ãä¸æ°ã«ä¸ãã£ãç¬éãè³å ç©è³ªãããã¡ã¨åºãæè¦ããããã§ãããã¢ã¬ãä¸åº¦ç¥ã£ã¦ãã¾ã£ãããã¿ã¤ãã«ãªã£ã¦å¿ããããªããã§ããããã
ã¨ãã£ããã¨*8ããã£ããã£ã¦ãããã¾ãã«ä»ã®åããã®ç¶æ
ã§ãð
ããéãããã¨ã¯ã§ããªã...
ç¥ããªãã£ãåæç¥è¦ãã¢ããã¼ãã®ä»æ¹ãå¾ãããã®ã Kaggle ã®éè¦ãªãã¡ã¯ã¿ã¼ã§ããåãçµãçç±ã§ãããã³ã³ãã¨ãã¦ã¯é ä½ãåçã«ä¸ãã£ãã¨ãã®è奮ã¯å¿«æ¥½ã§ãã麻è¬ã§ãã
ããããã£ãçºè¡¨ãä½åº¦ããèãããæ©ä¼ã¯ããã¾ãããããç»åã³ã³ã...ãã£ã¦ã¿ãããã©æ·±å±¤å¦ç¿ããããããä¸ã«GPUãè¦ãã使ãæ¹ãããããããã...ãã¨å½æã¯å®å
¨ã«åã³è
°ã§ããããã¤ã ã£ãã Kaggle ã®ã³ã³ãä¸è¦§ãè¦ã¦ç»åã³ã³ãã°ã£ãã ã£ãã¨ãã«ãããã£... ãã¼ãã«ç¡ãã®ï½ï¼ãã£ã¦ãªã£ãè¨æ¶ãããã¾ãã
éã«ä»ã¯ç»åã³ã³ãã°ã£ãã ã¨ããªã©ãã¯ã¯ã¯ã¯ãã£ãï½ï¼ãã¨ãªãã®ã§ãããæé·ãã¤ã³ãã§ãããã
å¾è¼©ã¨ã®ææ¦ï¼PAN Author Profiling Task 2018
ãä¸è¨ã®å
輩*9ã¨ä¸ã¤ä¸ã®å¾è¼©ã¨çµãã§åå ãããã¨ããã¨ã¼ãããã®ä¼è°ã® WS ã§ã®ã³ã³ãã§ãããé¡ã¯ Text x Image ã®ãã«ãã¢ã¼ãã«ã¿ã¹ã¯ã§ãããå¾è¼©ã NLP ãå°éã¨ãã¦ããã®ã§ãããããåãç»ååå¼·ãã¾ãï¼ãã£ã¦æãä¸ãã¦ç»åèªè by Deep Learning ã¸ã®åææ¦ã¨ãªãã¾ããã
äºäººããããã Text / Image ã ãã§å¦ç¿ãããã¢ãã«ã pre-trained model ã¨ã㦠multi-modal model ãå¦ç¿ã(åå ãã¼ã ã¯ãããªå¤ããªãã§ãã) çµæã¨ãã¦ç·å1 ä½ãç²å¾ã¨ãããã¡ãã£ã¨ç±ãå±éã§ãã*10ã
4æãããã¼ã çµãã§åå ãã¦ãshared task ãç·å1ä½ã ã£ãã®ã§åã³ãããªã
— 俵 (@tawatawara) June 5, 2018
ãã¨ã¯ãããã©ãã«ãããã«ãå®è£
& å¦ç¿ã»æ¨è«ãããã¨ã«æåãããã®ã®ãå®ã¯ãã®æç¹ã®åã¯ããããç»åèªèã®ãã¯ããã¯ãä¾ãã° Data Augmentation ãå
¨ç¶æ´»ç¨ã§ãã¦ããããæ¢ã«ResNet å
¨çæã ã£ãã«ãé¢ããã VGG16 ã使ã£ã¦ããã¾ãããä»åãã¿ã¹ã¯ãã£ãããã£ã¨æ§è½åºããããããªãããªã£ã¦æãã¾ãã
ã¨ããã§ï¼Kaggle ã¯ãã£ã¦ãªãã£ãã®ï¼
ãèªåèªèº«ã® Kaggle acount ã®å±¥æ´ãè¦ã¦ã¿ãã¨å®ã¯å°ãã ããããã¨ãã形跡ãããã¾ããInstacart Market Basket Analysis ã Mercari Price Suggestion Challenge ãªãããããã§ããã§ã sample submission ã«æ¯ã®çãããããªãã®ãåæã«ã¡ããã£ã¨åºããã ãã§ãã©ã¡ããé·ç¶ããã¦ãã¾ããã
ä½ã¨ãªãã§ããããã®æç¹ã®åã«ã¨ã£ã¦ Kaggle ã¯ã¨ã¦ãé ããã®ã ã£ããã ã¨æãã¾ãã å人ã«ãDiscussion ã¨ãåèã«ãªãæ å ±æ²¢å±±ããããèªãã ãã§ãåå¼·ã«ãªãããã£ã¦è¨ããã¦ã¡ãã£ã¨èªãã ãããã¦ã¿ããã§ãããããã¾ã§åå¼·è³æã¨ãã¦èªãã§ããã ãã§"åå "ã¯ãã¦ãªãã£ããã§ããããã³ã³ãããããã¨æãã¤ã¤ããã¾ã Kaggle ã¨ããã³ãã¥ããã£ã®å å´ã«(å®æ ã¨ãã¦ãæ°æã¡ã¨ãã¦ã)å ¥ãã¦ç¡ãã£ããã ã¨æãã¾ããè±èªãªã®ãå°ãæµææãä¸ãã¦ããé¨åãªã®ããã
RecSys Challenge ã PAN Author Profiling Task ã¯å 輩ãå¾è¼©ã¨çµãã§åºã¦ããããã¡ããã¨åãçµãã¦ãããã ã¨æãã¾ããå¼ã£è¶ãéè¦äºæ¸¬ã¯è²ã ã·ã³ãã«ã ããããé·ç¶ãããæããã§ã Kaggle ã¯... ããã¯ã©ããã¨ããã¨ãããã¨ããããã§ãä¸äººã§åæã³ã³ãã«æããåã身ã«çãããã¨æãããã«ãªãã¾ããã
åæã³ã³ãæ¬æ ¼åæ¦ã決æãã¦ããã®è©±
èæ°´ã®é£ãæ·ãï¼ã¾ãã¯è²·ã£ãããªã³ãã¬ãã·ã³
ãåæã³ã³ãã«ãã©ã¤ãã¼ãã§æ¬æ ¼åæ¦ãããã¨æã£ãã®ã¯ãä¸ã§è¨ã£ã¦ããããã«ã³ã³ãã«èªåèªèº«ã®åã§æããããã«ãªãããã¨ããã®ã¨ãä»äºã ãã§ã¯æ±ããã¼ã¿ã®ç¯å²ãçãããã£ã¨åºãç¯å²ã»ç°ãªãæ¥ç¨®ã®ãã¼ã¿ãæ±ãçµé¨ãããªããã°ãªããªãã¨ããæãããã§ããã
ã¨ã¯ãããããããï¼ããããï¼ãã¨è¨ãã ãã§ã¯åè¿°ãã Instacart ã Mercari ã®äºã®è(ä¸ã®èï¼)ã«ãªã£ã¦ãã¾ãå¯è½æ§ãããã¾ããããã§èãã¾ããããç°å¢ãç¨æãã¦ãã¾ã£ã¦ "使ããªãã®ãå¿ä½ãªã"ã¨èªåã«æãããã°ãããã¨ã
PC注æãã
— 俵 (@tawatawara) July 10, 2018
GPU ã注æãããããéãããã¨ã¯ã§ããªã
— 俵 (@tawatawara) July 11, 2018
åè¨(æ¬ä½ + GTX1080ti x 1) ã§40ä¸å¼±*11 ... ã¡ããã©åºãã°ããã®å¤ã®ãã¼ãã¹ãã»ã¼å
¨é¡çªã£è¾¼ãã§ãããå·éã«èããã¨é ã®ããããè¡çºã ã£ãããããã¾ããããã¼ãããã©ã§è²·ã£ã¦èªä½ããæ¹ãå®ãã¨ã¯æãã¾ãããããã§ãã©ãã£ã¦ã³ã³ãã«ä¸ã
åå ã§ããªãããã¯ããç¨åº¦ãéãæã£ãæ¹ãè¯ãã ããã¨ããå¤æã§ãã*12ã
ä¸å¿ã§ããåæã³ã³ãå°ç¨ã®ã¤ããã¯ãªãããããã家ã«ã¾ã¨ããªè¨ç®è³æºãç¡ãã¦å°ããã¨ãå¤ã ãã£ãã®ããã£ã¦è²·ã£ãé¢ãããã¾ã*13ã
ã¾ãå®ã®ã¨ããGPUã¯ã¤ãã§ã§ã家ã«ã¾ã¨ããªè¨ç®è³æºããªãã£ãã®ã§ä½ã¨ããããã£ãã®ãå
— 俵 (@tawatawara) July 11, 2018
ãå¾ããæãã¨ããã®ãã·ã³ã®ãããã§æããã¨ãåºæ¥ãå°é¢¨ã³ã³ãã¯ãã®å¾ç»åã³ã³ãã«åãçµãä¸ã§ã®åå°ã¨ãªãã¾ããããããã¦æçµçã«ãã®ãã·ã³ã§å¦ç¿ãããã¢ãã«ã...
å è¡æè³ã¨ãã¦ã¯ããåäºåã«ååã§ãã¦ãã¨æãã¾ãããã£ã±ããã®ãã·ã³ãè²·ã£ããã¨ã¯ééãã§ã¯ç¡ãã£ããã¨ãããééãã«ãªããªãããã«åèªèº«ãé å¼µã£ãããã§ãã... ä½æ¦ã¯å¤§æåã¨ãããã¨ã§ãããð
é¾å¤ã§ã¤ã±ã¤ã±ðãã³ãã³ð : SIGNATE å°é¢¨ã³ã³ã
Public 5th â Private æä¸ä½
ããGPU ãè²·ã£ããããã©ã¤ãã¼ãã§ç»åã³ã³ãããããï½ï¼ãã¨æåã«åãçµãã ã®ããã¡ãã§ããçµæã¯ã覧ã®éãð
ãã©ã¤ãã¼ãã§æåã«åºãç´ç²ãªç»åã³ã³ãããã®å°é¢¨ã³ã³ãã ã£ãã®ã¯ãææ¨åé¡ãé¤ãã°ãç»åã³ã³ãåå¿è ã®åã«ã¨ã£ã¦æ¬å½ã«å¹¸éã ã£ãã¨æãã¾ããç»åãµã¤ãºã 64 x 64 ã ã£ããã¨ããè²ããªè©¦è¡ãçãã¹ãã³ã§åºæ¥ãã®ãããã§ãããã¾ã pre-trained model ãå°ã使ãã«ãããé¡ã»ãµã¤ãºæã§ãã£ããã¨ãããèªåã§ä¸ãã ç»å CNN ãå®è£ ããçµé¨ãç©ãããã¨ã§ã*14ããã®æã« ResNet ç³»ã®è«æãããã¤ãèªã¿ããã§ãæ§é ãã©ããªã£ã¦ãããã¨ããå¦ç¿çã¨ãå¦ç¿ iteration æ°ã®è¦æ¨¡ã¨ã scheduling ãã©ããªããã調ã¹ã¾ãã£ãã®ã¯å°å³ã«ã§ããå¾ã ã«æ´»ãã¦ããã¨æãã¾ããåã SGD with NesterovAG æã«å ¥ä¿¡ããã®ããã®ã¨ãã§ã*15ã
ãã®ã³ã³ãã§CNNãå®è£ ããä¸ã§æãåèã«ããã®ã¯ã¡ããªãµã¼ããã£ã¬ã³ã¸ã®åªåè ã®ã¹ã©ã¤ããªã®ã§ããã
Master ã«ææ ¼ããå ±åã Twitter ã«æ稿ããã¨ãã«ãç¥ãã³ã¡ã³ããããã£ã¦ããã£ã¡ãææ
¨æ·±ããã®ãããã¾ããã
ãããã¨ããããã¾ãï¼æãã°ãã¸ã§ã³ããã®æ¼«ç»(ã¡ããªãµã¼ã)ãèªã㧠WideResNet ãå°é¢¨ã³ã³ãç¨ã«ä½ã£ã¦ããä¸å¹´åã§ãã... ã¨ã¦ãæãããã§ãã
— 俵 (@tawatawara) March 18, 2020
ããã°ä¸ã«ç»å ´ããããã«ããã®ã³ã³ã㯠Data Augmentation ã®å¤§åããåã«æãã¦ãããã³ã³ãã§ãããã¾ã*16ãRandomErasing ã¯ãã®ã¨ãããåã«ã¨ã£ã¦ã®å¼·åãªç¸æ£ã§ãããç»åèªèã®ã³ã³ãã§çµ¶å¯¾ã«ä½¿ç¨ãã Augmentation ã¨ãªãã¾ããã
ãã¾ããããä¸ã¤å¾ããã®ãããã¾ãããã®ã³ã³ãã«ã¯å°æ¥ Kaggler*17ã§ããã«ã¬ã¼ã¡ãããããªã©ããã®å¾ãå¥ã®ã³ã³ãã ã£ãã Kaggle é¢é£ã¤ãã³ãçã§ãä¼ãããæ¹ã ãåå ãã¦ãããã¨ã§ãããã®æç¹ã§ã«ã¬ã¼ã¡ãããã㯠Gold ãæã£ã¦ããã®ã§åãããã£ã¨å ãè¡ã£ã¦ãã¾ããããé ã®ä¸ã§ã¯ç¡æèã«ã©ã¤ãã«ãªã¹ãçãªãã¤ã«å ¥ã£ã¦ããã¨æãã¾ãç¬
ãã®ãããããå¾ã ã«ã§ãã Kaggle é¢ä¿ã§ã®äº¤æµãããããã«ãªããã³ãã¥ããã£ã¸ã®åå ã¿ãããªã®ãæèããããã«ãªã£ããããªæ°ããã¾ãã
éããåå ð å ¨ç¶ããããð : Kaggle Tokyo Meetup #5
ãå½æã¯ã¡ãã«ãæã£ã¦ããªãã£ãã®ã§ããªãåçãå³ããã£ãã®ã§ãããTwitter å»äººãªããã㧠Threecourse ããã® Tweet ã«ãã¡æ©ãæ°ä»ããã¨ãåºæ¥ã15/70 ã®çãéã«å ¥ããã¨ã«æåããã åå ã§ããã®ã¯ããã£ããã®ã®ãå½æã®åã«ã¨ã£ã¦ã¯çºè¡¨ã²ã¨ã¤ã²ã¨ã¤ã®ã¬ãã«ãé«ããã¦*18ç¥èãå¸åãããã¨å¿ æ»ã§ããã
ãã®ä¸ã§ãç¹ã«ç»åé¢é£ã§æã¡å¸°ãããã®ã¯äºã¤ããã¾ããä¸ã¤ã¯ Cosine Annealing ã§ããPFDet ãã¼ã ã使ã£ã¦ããã®ãç¥ãããã£ãã次ã®ã³ã³ãã§ä½¿ããï¼ãã¨ææ°è¾¼ãã§å¸°ã£ã¦ããå®è£ ãã¾ãããããã¤ã RandomErasing åæ§ã«ç»åã³ã³ãã«ããã¦é ¼ããç¸æ£ã¨ãªã£ã¦ãã¦ãç»åã³ã³ãã§ä½¿ããªãã£ããã¨ã¯ããã¾ãã*19ã
ããä¸ã¤ã¯ãã¡ãã£ã¨ã¡ã¿ã§ãããç»åã¢ãã«ã® backbone ã®æ§é ã¯å¾ã«å¼ããªããã¨ããç¥è¦ã§ããåãä½åº¦ãèªãã§ããç³ã¿è¾¼ã¿ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã®ç 究ååã§è§¦ãããã¦ãã bn - conv - relu ã®é çªã®éãè¦ãã°ãããããã«ããã£ã¡ãç´°ããä¿®æ£ã§æ§è½ãä¸ããå ´åããããã¨ãç¥ããã¦ãã¾ãããããã¸ãæ°ã«ãªã£ã¦ãã¦ãåºæ¬çã«ãªãã§ãå¼·ããå»å¦ç³» Kaggler ã§ãã osciiart ãã*20 ã«æãåã£ã¦èãã¦ã¿ã*21ã¨ãããåçã¯ããããã¸ãã¯ãã¾ãæ°ã«ããããªãæ¹ãè¯ããã¨ãããã¨ã§ããã
ãã¡ããçç±ããã£ã¦å¤ãããªã話ã¯å¥*22ãªã®ã§ãããä½ãããããããã¾ã¾é çªãå ¥ãæ¿ããããã¦æ§è½ãä¸ãããã¨ããè¡çºã¯ä¸æ¯ã ã¨ãããã¨ã§ããããã¯æ§é 以å¤ã«ãè¨ãã話ã ã¨æã£ã¦ãã¦ãä¾ãã°ç»åã®æ§è³ªãä¸åèæ ®ããã« Augmentation ã試ãã¾ãããããªè¡çº*23ã¯ãæ§è½ãåä¸ããå¯è½æ§ã¯ã¼ãã§ãªãã§ããããªãã®æéã¨å´åãè¦ãã¾ãã
ããã®ä¼è©±ããã£ããã¨ãªã£ã¦ãèªåã®ä¸ã§ããç¨åº¦ã®ã«ã¼ã«ãæã£ã¦ã·ã³ãã«ã«ã¢ãã«ãä½ãã¨ããææ³ãåºã¾ã£ã¦è¡ãã¾ããå®è£ ã®æ¹ã§ããç»åèªèã¢ãã«ã¯ backbone â global pooling â classifier head ã¨ãã¦å®è£ ããbackbone ã¯ç¹å¥ãªäºæ (e.g. ãã£ãã«æ°ã pretrained ã¨éã)ãç¡ãéãå¼ããªããã¨ããæ¹éãåºã¾ãã¾ãããã¢ãã«ã®ç´°ããã¨ãããå¼ãããããã¼ã¿ã¨ã¿ã¹ã¯ããã£ããç解ããããã«åãããå¦ç¿ãå¦çãè¡ã£ãæ¹ãè¯ã*24ã¨ããã®ã¯å ¨ã¦ã®ã³ã³ãã«ããã¦è¨ãããã¨ã ã¨æãã¾ãã
ãã¾ãããã®ä¼ã¯å°é¢¨ã³ã³ã被ç½è ã®ä¼ãéå¬ãããããu++ããã® LT(kaggler-ja driven learning) ãè´ã㦠kaggler-ja slack ã«åå ãã¦è³ªåããããã«ãªããªã©ãã³ãã¥ããã£ã¸ã®åå ãã¾ãä¸ã¤é²ãã ä¼ã§ãã£ãããã«æãã¾ã*25ã
åãã¦ã®ãã«ãã©ãã«åé¡ : Human Proteins Atras Image Classification
Public 57th â Private 21stð¥
ããã§ãä½ãæã£ãã社é·ã¯ãè¨ç»ä¸ã®è¾²è¬éçºããã¸ã§ã¯ãã«ä½¿ããããªãã¿ã ã Human Protein Atlas Image Classification ãããã ã¨kaggle ã¸ã®æå ¥ã決æ(â»ããèããã¨ãããã¸ãã...ãã£ã¦æ£ç´æãã¾ãã)ãååè ãåã£ã¦ AgroDesign ãçã¾ãã¾ããã
kaggle ã§åãã¦ã¡ãã«(silver) ãåã£ããã®ã®ããè©«ã³ããããå¾ãªã件 - 俵è¨
AWS 使ãæ¾é¡ã¨ããéåã«ã¯åã¦ãªãã£ãã...*26
ãã¾ãããå 輩ã®èªãã«ãããå 輩ã®ç¥ãåãã®ãè¾²è¬ãã³ãã£ã¼ ã¢ã°ããã¶ã¤ã³ã»ã¹ã¿ã¸ãªã®ç¤¾é· & å¥ã®ãã³ãã£ã¼ã®ã¨ã³ã¸ãã¢ã®ãäºäºº(sheldon ãã, agrwhy ãã) ã¨çµãã§ç»åã³ã³ãã«åºãã¨ããé¢ç½ãçµé¨ã§ããã
ããã§ã¯ä¸åè¡¡ãã¼ã¿ã§ã® multi-label classification ãæ±ããå¾ã åãæ¶ãåããã¨ã«ãªã iMet Collection 2019 ã§å½¹ã«ç«ã¤ç¥è¦ãè²ã å¾ããã¾ãããããªããªã®ãã¡ã¤ã³ç¥èææè (社é·)ã¨ç»åãããªããªãã£ã¦ããã³ãã£ã¼ç¤¾å¡ã®ãäºäººã¨ä¸ç·ã«åå ã§ããã®ãè¯ãã£ãã§ãã
æçµã¢ãã«ã¯åãä½ã£ãã¢ãã«ã«ãªãã¾ãããã社é·ãããããããçè²ã¯æ £ä¾ã§æ±ºã¾ã£ã¦ãã¦ãèªç¶ã«åå¨ããRGBã¨ã¯å ¨ãéããã§ãããã¨ãããã®ã¯ã©ã¹ã¯äººéã§ãèå¥ãå°é£ãªãã§ãããï½ãã¿ãããªéè¦ç¥è¦ãå¾ããããããã£ã¹ã«ãã·ã§ã³ã®èª¿æ»ã»ã«ã¼ãã«ã®è¿½è©¦ã»å¤é¨ãã¼ã¿ã®æºåãåæ ããããªã©ããã¼ã ã¨ãã¦ã¯çµæ§ãã¾ãããã¦ããããããªããã¨æãã¾ãã
ããã¦ããã®ã³ã³ãã§åãã¦å㯠Cosine Annealing & Snapshot Ensemble ã使ç¨ãã¦ãã¾ããã¾ã Stacking ãã©ããã£ã¦ããã®ãããããã£ã¦ããªãã£ãã®ããã試ãã¦ã¿ãã¨ããã®ãèæ¯ã§ããçµæã¨ã㦠Snapshot Ensemble 㨠K-Fold Averaging ãå¤§å¹ ã« Score ã伸ã°ãã¨å ±ã« shakeup ã«è²¢ç®ãããã®ã¨ãã®æåä½é¨ããã¯ãå¾ã ã«æ´»ãããã¨ã«ãªãã¾ãã
ããã Snapshot Ensemble ã¯å¦ç¿ãé·ãåãå¿ è¦ããã£ã¦ããããªãã«æéã³ã¹ãããããã¾ããã¡ãã£ã¨åã«ã«ã¬ã¼ã¡ãããã㨠kaggler-ja slack ã§è°è«ã«ãªã£ãã®ã§ããããã¾ãæ®éã®ãã¼ã¿ãªã K-Fold Averaging ããæç¹ã§å¤æ§æ§ãåºã¦ãã並åã§å¦ç¿åããããè¯ãããã ãªã¼ãã¨ããã®ããã®ã¨ãã®åã®è¦è§£ã§ãããç¹æ®ãªã±ã¼ã¹ãä¾ãã° GM takuoko ãã ã RSNA ã§ãã£ã¦ãããããªããã¼ã¿ãå¤ãã㦠K-Fold Split ããããã éãã¢ãã«ãè¤æ°å¦ç¿ãããæ¹ãè¯ãã¨ããã±ã¼ã¹ãªãã1 cycle ã§ã追å ããã¨æ±åæ§è½ãåä¸ãããããªæ°ããã¾ãã
çæåæ¦ã§ç¡äºæ»äº¡ð: é»ç· & Pet
ãã¿ã³ãã¯è³ªã³ã³ãã®ãã¨ãå¥ã®ã³ã³ããããã¨æãã¤ã¤ãããåãçµããæ°ä»ãã°ãã¨ã¡ãã£ã¨ããæéããªããã¨ããã®ã2åãããã¾ãã*27ãæ¯åº¦æ¯åº¦ãshort-term challenge ã¯ããäºåº¦ã¨ãããªãï¼ãã£ã¦å®£è¨ãããã§ããã©ã å¦ç¿ããªããªï½ãã³ãð*28
short-term challenge ã®å¯ä¸ã®å©ç¹ã¯ãprivate LB å ¬éã®ãç¥ãã«åå ã§ãããã¨...ã§ã¯ãªãã¦ããã¼ã¿ãã¡ããã¨è¦ãä¸ã§ submission ã¾ã§ãã©ãçããã°ããã®ã³ã³ãã§ç¹ã«éè¦ã ã£ãç¥è¦ãçæéã§æã«å ¥ãã¨ãããã¨ã§ããåã«ã¨ã£ã¦ã¯ PetFinder ãããã§ããã
VSB Power Line Fault Detection (1 Day)
Public 1323rd â Private 1267th
- 1D-Conv ãå®è£ ãã¦ã¿ããã£ãã®ã§ãã£ã¦ã¿ããã®ã®å ¨ç¶ Score ãåºãæ»äº¡
- ããã¢ã¯ã«ã©ãã«ãã¿...
PetFinder.my Adoption Prediction (1 Day)
Public ???th â Private 678th
- ??? ãªã®ã¯ã2nd stage å¶ã ã£ãããã«å ã®é ä½ãããããªããªã£ããã(è¦ãã¦ããªã)ã
- Group K-fold ã®éè¦æ§ãç解ãããã¨ãåºæ¥ã¾ããããã ãããªã« Score ã¯åºãªãã£ã(å®)
- NN ãæ¬çªã®å¦ç¿ã«ä½¿ããã¨ããã«ãã¢ã¼ãã«æ å ±ã§äºæ¸¬ãè¡ããã¯ããã¯*29ãããã®ãç¥ããã¨ãåºæ¥ã¾ãã
ã½ãã§ã®ç»åã³ã³ãåã³ðï¼iMetCollection2019 - FGVC6
Public 18th â Private Lost
ãè²ããªæå³ã§åã«ã¨ã£ã¦ã®ä¸ã¤ã®è»¢æ©ã¨ãªã£ãã³ã³ãã§ããæãé·æé solo ã§åãçµãã ã³ã³ãã§ããã人çåã® EDA ã«ã¼ãã«ã Starter ã«ã¼ãã«ãä¸ããããDiscussion ã«ãåå ãããã¨ããKaggle ãã£ã¦ãããï¼ðªãã¨æ°åãå ¥ãã¾ãã£ã¦ããã³ã³ãã§ããã
ããã§ãã£ããã·ã¥ pic.twitter.com/mncUVlG7jX
— 俵 (@tawatawara) June 6, 2019
åºæ¥ãéãã®ãã¨ããã¦ä¸ä½ã¾ã§æ¥ãã¦æºè¶³ï¼ã®çã§ããã...
ðºãå¤åããã³ãããOOM ãããã
— 俵 (@tawatawara) June 9, 2019
ã¤ãããã®ããããã©è²ã èªå·±è²¬ä»»ãªã®ã¨ç®¡çä¸è¶³ãªé¨åãå¤ãã®ã§ã次ã«æ´»ãããããªããªããã¼ãã¡ã¢ãªã®åããããããªã...
ç»åã¯ä¸æ¦ãã°ãããããªãã ããã¨æã£ã¦ããã©ãæããããã¾ããããã¨ã«ãªãããã
æ£ç´ä»ã§ãæãåºãã¨ã¤ããã§ã*30ãæ¬å½ã«ãã®ã³ã³ãã¯å
¨åã§é å¼µã£ãã ãã«ãèªåã®ãã¹ã¨ã¯ããã£ã¦ã¯ãã¦ã 2nd stage ã§è½ã¡ã¦ãã¾ã£ãã®ãæ²ããã£ãð
ããã ããã£ã¦è¯ãã£ããã¨ã¯ solution ãæ®ãããã¨ã§ãããããããªãã¨ã2nd stage ã§è½ã¡ãããã§ã·ã¹ãã ä¸ã¯åå ãã¦ããªãã£ããã¨ã«ãªã£ã¦ãããã¨ã«ãã¨ã¦ãã¨ã¦ãèããããªãã£ããã§ãã2ã¶æ以ä¸é å¼µã£ã証ãæ®ããã...ã
2 ã主張ããããã ãã«ä¸çªæ°åãå ¥ããé¨å(å³)ã§ãããç´ããã ããã
— 俵 (@tawatawara) June 12, 2019
solution ã«ãæ¸ãã¦ã¾ããå®è£ ãç°¡ç¥åããããã§ä¸å®å ¨ãªé¨åããããæ¹è¯ãã¦ã¾ã次ã®ã³ã³ãã§ä½¿ã£ã¦ã¿ããã¨æã£ã¦ã¾ãã pic.twitter.com/5nuLNkSJrk
ããã«ãã£ã¦è²ããªäººã« solution ãè¦ã¦ãããã¾ããããé ä½ã¯æ®ããªãã£ããã®ã®éå¬ä¸ã«ãã£ã¨é å¼µã£ã¦ãããããã§ãYKãããphalanx ãããtakuoko ãããAppian ãããowruby ãããRK ãããCamaroãã ... ãªã©ãªã©ã®ãæ¥æ¬ã®éã
ããç»åã¤ãã¤ãå¢*31ã¨åããã¼ã¸ã§ç«¶ãåãã¾ãããsolution ã®åé ã«ãæ¸ãã¦ããããã«ã¨ã¦ãç´ æ´ãããçµé¨ã§ä½ãã楽ããã£ãã§ããX5ï¼ãªã¼ã¯ï¼ç¥ããªãåã§ãã....
ãã㦠solution ããã£ããæ¸ãã¦èªåã®è¨¼ãæ®ããããããã次ã®ã³ã³ãã«ç¹ããã®ã§ããã
ç¥ããå±ãð : Instant Gratification (1 week)
Public 66thâ Private 12thð¥
ã壺ã·ã§ãã¯ããç«ã¡ç´ãã¦ããªãä¸ã lost place solution ã¨ãkerneler ã¡ãã£ã¨ã ããã£ã¦ã¿ããããª...ãã¨ããåã® Tweet ãè¦ã gege ãã ãèªã£ã¦ããããã¨ã§ JTC Kagglers season2 ãèªçãã¾ãã
é»éé士ããã¼ãã£ã«å ãã£ã¦ä¸ããã¾ããï¼kerneleråãåãã¹ãï¼åã³JTCéãç«ã¡ä¸ããï¼ï¼ pic.twitter.com/G6XFLs4c5K
— gege (@gegegegegege) June 13, 2019
ããããããªãã°æ¦äºã
— Miyatti (@y4tk38) June 13, 2019
(æªã 0sub) https://t.co/ErdojNroeS
season1ã®ä»²éãæµã¨ãã¦ç¾ãã¦ãã¾ãã¾ããð https://t.co/xu7N0T0i6q
— gege (@gegegegegege) June 13, 2019
ééãèãæãããã¡ã³ãã¼ãå¢ããã
ð
— Miyatti (@y4tk38) June 13, 2019
æ¨æ¥ã®ä»²éã¯ä»æ¥ã®æµã¨æããããã£ã±ã仲é pic.twitter.com/DPZHdO8XZi
— gege (@gegegegegege) June 13, 2019
ããã« tamref ãã 㨠Doraconda ãã ãå ãããJTC Kagglers season 2 ã¯å®å ¨ä½ã¨ãªã£ã¦åãå§ãã¾ããã
1 week ã ãè¡ããã¨ãã¾ã§è¡ãã¾ãããã¨ãã軽ãããªã ã£ã...çãªã®ã§ãããç»åãããã£ã¦ç¡ãã£ãåãä¸ã é²æãç£ããªãä¸æ¹ã§çãããé²æãåºãã¾ããã®ã§ãã¡ããã¡ãç¦ãã¾ãã
åã¾ã ä½ããã¦ãªãã®ã«ã¤ãã¤ããã¼ã ã¡ã¤ãã®åã§ããã¾ã§æ¥ã¾ããããã®ã¾ã¾ã 㨠contribution: 0 ã«ãªã£ã¦ãã¾ã...https://t.co/RD23vMCTuo
— 俵 (@tawatawara) June 17, 2019
ã¾ãããã®ã³ã³ã㯠Miyatti ãããç§å ´*32主åè£ã«ãªããå¦ãã®ç¬æ¸éã§ããã
kaggle expertç§å ´ä¸»ã®æ±äººã誰ãJob Descriptionãæ¸ãã¦æ¬²ããã
— Miyatti (@y4tk38) June 18, 2019
(kernelerã³ã³ããã®ã¾ã¾éã¡ãã«ã ã¨ããã¨é4é 1ã§ããç·ããã®ã§ã¯ï¼)
Miyatti ãããç§å ´ä¸»åè£ã«ãªããå¦ãã¯æã ã®åãã«ããã£ã¦ãã
— 俵 (@tawatawara) June 18, 2019
ä¸ã
é²æãåºããªãã£ãã®ã¯è¾ãã£ãã§ãã1é±é楽ããã¦ãiMet2019 ã®è¾ããå°ãã¥ã¤åããã ããã«æãã¾ã(ãã¼ã ã¡ã¤ãã®çããæ¬å½ã«ãããã¨ããããã¾ããð )ãæçµçã«åãã©ãã«ãããã«ãã¾ã¨ããªãµããä½ããã¨ãåºæ¥ã¦ Public Score ãæ´æ°ããsubmission ã®ä¸ã¤ã«é¸ãã§ããããã¨ã«ãªãã¾ãã(ããä¸ã¤ã¯Local CVãæ大ã®ãµã)ã
ãã¦ããã¨ã¯ç¥ãã ãã§ããã...
shakeup ã®ãç¥ããããããã® discussion ãåºæ¥ã¦ããåããç¥ããã¨ããhttps://t.co/9dPZmTThmn
— 俵 (@tawatawara) June 20, 2019
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããããããããï¼ï¼ï¼ï¼ï¼ï¼ï¼ï¼ï¼ pic.twitter.com/xYn0wYe7J9
— 俵 (@tawatawara) June 21, 2019
ããããç¹æï¼ãç¥ããã®æ以ã ã£ãããã¾ãã
JTCã¯ä»åãç´ æ´ããããã¼ã ã¯ã¼ã¯ã§ããï¼tamrefãããåºæ¬ã®ãã¸ãã¯ãä½ãï¼miyattiãããã¬ãã¬ãé²æãåºãï¼dracondaããã¯æã®å±ããªãã¨ãããããããæ¤è¨¼ãã¦ããã¦ï¼tawaraããã¯privateéã®submitãä½ãï¼ç§ã¯final submitãé¸æãã¾ããï¼
— gege (@gegegegegege) June 21, 2019
çãããããéãæ¹éã§ãµããä½ã£ãä¸ã§ Public 㨠Local CV ã®é«ããã®ã 1ã¤ã¥ã¤é¸ã³ãçæ¹ãéåã«å
¥ãã¾ãããå®å
¨ã«ãã¼ã ã§ã®åå©ã ã£ãã¨æãã¾ãð
ãå ã¿ã«ãã®å¾è²ã
ãã¿ãã¿ãããã㧠solution ãæ¸ããªãã£ãã®ã§ããããã£ãããªã®ã§ããã§è»½ãç´¹ä»ãã¾ããåã®ä½ã£ããµãã¯è¤æ°ç¨®ã® n_component
ã使ã£ã¦ GMM ã fitting ãã¦ããã¤ãã® GMM features ãä½ããããããç¹å¾´éã«ç¨ãã Logistic Regression ã§äºæ¸¬ãè¡ãã¨ãããã®ã§ããã
gege ããã«ãã£ã¦ n_component
ã 3 ã£ã½ãã¨ãããã¨ã¯æ¢ã«å¤æãã¦ããã®ã§ããã ããããã® example ã®ã©ãã«ãã©ã³ãã ã«ããªããããã¦ããã®ã§ãfeature ã«ããã¨ãã¬è¦ç´ ãå
¥ãããããæãã«ãªããããªï½ã¨ãã£ã¦ã¿ãã®ããã¾ãè¡ã£ãã¨ããã®ãäºã®é¡æ«ã§ããçµæçã«ã¯ãã®ããæ¹ã®ãµãã®ã»ã¨ãã©ãéåã«å
¥ã£ã¦ãã®ã§ãä½ãè¯ãã£ãã£ã½ã(ãã¾ãæ·±ã説æã§ããªã)ã
ãã¾ããã®ã³ã³ãã¯ãæ¥æ¬ã代表ãã GM ã®ä¸äººã§ãã Jack ãã ã 5æç®ã® solo gold*33 ãç²å¾ã㦠Grand Master ã«ä¸ããããã³ã³ãã§ãããã¾ãããåãåé¡ã®æ¬è³ªãç解ãåãã¦ããªãã£ãã®ã«å¯¾ããæ¬è³ªãã¡ããã¨ç解ããä¸ã§ããã¡ããã¡ãã·ã³ãã«ãªææ³ã§åã¤ã¨ãã姿å¢ã¯æ¬å½ã«æ§ãããã®ã§ãsmly ããåæ§ã«ããªãå½±é¿ãåãã¦ãã kaggler ã§ãã
ãããªããã§ãããªãã®ãªã®ãªã§ããã Kaggle Tokyo Meetup #6 ã®ç¢ºå®åå 権ãå¾ãã®ã§ããã
é å¼µã£ã¦LTã«ææ¦ : Kaggle Tokyo Meetup #6
ãæ»ãè¾¼ã¿ã§åå 確å®ãã2åç®ã® Kaggle Tokyo Meetup ã§ãã詳細ã«ã¤ãã¦ã¯ããã°è¨äºã«ã¾ã¨ãã¦ãã®ã¨ãçºè¡¨è ã®æ¹ã ãã¹ã©ã¤ããå ¬éãã¦ä¸ãã£ã¦ããã®ã§ãã¡ããã確èªãã ããããã®åå¹´åã«åå ãã#5 ã«æ¯ã¹ãã¨ã(ç¹ã«NNç³»ã¯)大åå 容ãç解ã§ãããããªã£ãã¨æãã¾ãããããã¾ãæé·ããã¨ãããã¨ã«ãããã
å ã¿ã«ãã®ã¤ãã³ãã§ã¯è²ã åºãã£ãè¨èããã£ã¦ã
phalanx ãããå¦ç¿åãã¦ãéæã ã¨ãè¨ã£ã¦ Twitter ãããªããã¨ãè«æãèªã¿ã¾ãããã
ãã£...è³ãçã... èã«éãã¾ã... ãphalanx ããã®ç»åã³ã³ãã«å¯¾ããã¢ããã¼ãã®ä»æ¹ã¨ã¹ãã¤ãã¯ãè¦ç¿ãããé¨åã沢山ãã*34ã¨ã¦ãåèã«ãªãçºè¡¨ã§ããã
smly ãã ãã¢ã³ãµã³ãã«ããªãã¨åã¦ãªãã¨ããã®ã¯çãã
å¼·ããããããã®æ¬å½ã«æ§ãã¾ãããð
Jack ããã¨ãããæ¥æ¬ã®ä¸ä½ kaggler ã¯ã·ã³ãã«ãã¤ã¨ã¬ã¬ã³ããªææ³ã好ã¿ã¾ããè¦ç¿ããããªãã¨å¸¸ã
æã£ã¦ããç¹ã§ãã
ãåã®LT ã¯ãNeural Network ç´ äººãªãã ãã©ä½ã¨ããæ©å«åãããããã ã¨ãããå½æã®åã®æ©ã¿ããã®ã¾ãã¾ç¾ãã¦ããå 容ã§ãããèªå·±é¡ç¤ºæ¬²ãæããããªãã¦èªåèªã (+ å座)ãé·ãããããã«æãã®ã§ããã®ç¹ã¯ãThreecourse ãããæ¬å½ã«ãã¿ã¾ããã§ããð
å 容ã¨ãã¦ã¯å½æã®åãããªãé å¼µã£ã¦ä½ã£ããã®ã§ãä»è¦è¿ãã¦ãå²ã¨ãããããã ã£ããªãã£ã¦ãªãã¾ãã ããã«æ¸ãã¦ãã Batch Size ã決ãã â LR-RangeTest â Cosine (Annealing) Scheduling ã¯åãç»åã³ã³ãã«åãçµãä¸ã§ã®åºæ¬çãªæµãã§ããä»ãå¤ãã£ã¦ãã¾ãããæ¬å½ã«æé©ãªå¤ãã¨ããã¨ããã§ã¯ãªãããããã¾ããããå¦ç¿çããã£ã¨æ±ºãã¦ãã¾ããã®ã¯ã¨ã£ã¦ã楽ã§ã*35ã
ãååã®#5 ã§ã¯å ã ç¥ãåãã ã£ã人 + α ã¨äº¤æµã§ãããããã§ããããLT ãããã®ã¨ lost place solution ãå ¬éãããããã§ãå¤ãã®äººã«ããããã®è¬ã®éæ³é£æ¸ãã¦ã人ã§ãããï¼ãã¨è¨ã£ã¦é ãã¦ãä¼è©±ã®ãã£ããã«ãªã£ãããã¦ãã¾ããããã¨ãé ä½ã失ã£ãã¨ãã¦ããèªåã® solution ãã¡ããã¨å ¬éãã¦æ¬å½ã«è¯ãã£ãã¨æãã¾ããã³ãã¥ããã£ã¸ã®è²¢ç® + èªåã®åå¨ã示ãã¨ããæå³ã§ã¨ã¦ãéè¦ã§ããã
ã㦠Kaggle Tokyo Meetup çµãã£ããã次ã®ã³ã³ãã¸...ã¨ãªãã¨æãããå°ãä¼æ¢ãã¦ãã¾ãã¾ãã
ã¡ãã£ã¨ä¸ä¼ã¿ : 4ãæããã
ãçç±ã¨ãã¦ã¯å¶ç¶ã«ãéã¡ãã«ãåãã¦ãã¾ã£ã¦ã¡ãã£ã¨ã»ã£ã¨ãã¦ãã¾ã£ãã¨ããå½æä»äºã¨äººçã«ã¤ãã¦æ©ã¿åºãã¦ãã¾ã£ã¦ã¨ãã¾ãè²ã ã§ããã¿ã³ãã¯è³ªã®ã¨ãã®ãã¼ã 㧠Recursion ã«åºããã¨ãããéæ»ã§ãªã¼ã¯ãçºçãã¦åºé¼»ãæ«ããã¦ãã¾ãããã...ãã£ã¦ãªã£ãããã¨ã«ãããã°ããã³ã³ãã« submission ãã¦ãã¾ããã§ããã
ããããã®éä½ããã¦ããªãã£ããã¨ããã¨ãããã§ããªãã¦ãä½ããããè«æèªãã ãå®è£ ãããããããã¯ä»ã® Kaggler ã¨äº¤æµããããã¦ã¾ãããã¡ãã£ã¨ãã®æéããåã®ãã®ãå«ãã¦æ¯ãè¿ã£ã¦ã¿ã¾ãã
ãä¾ãã°ãå㯠iMet2019 ã«åå ãã¦ãæéãããã¯ãã®å¾ããããã¤ãã® Kaggle ããããä¼ ã«åå ãã¦ãã¾ããã
ããããä¼ã®å ´ã§ã¾ããä¼ãããã«ã¬ã¼ã¡ããããã¨ä½ãããè°è«ããããå®ã¯åãçµã¿å§ããææçã«ã¯çµæ§åæã¿ãããªæãã®ãã¼ã Shirogane(Kurogane) ã®å¾¡ä¸æ¹(ã¢ã©ã¤ãããuratatsu ãããð¸ãã)ãhakubishin ãã ã¨åãã¦é¡ãåããããã¨ã沢山㮠Kaggler*36 ã¨äº¤æµããæ©ä¼ã«ãªãã¾ããããããã£ãã¨ãã«ããæ®æ®µããã°ã¨ãã§ä½ãããçºä¿¡ãã¦ããããã®ãããªãå½¹ã«ç«ã£ã¦ããããã«æãã¾ãã
ãè«æã«é¢ãã¦ã¯ãã³ã³ãã®ããã«ç»åç³»ã®è«æãèªãã§ã¯å¤é¨ã®è«æç´¹ä» LT ä¼ã«åå ããã®ãä½åããã£ã¦ã¾ãããiMet2019 ã®åºç¤ãããããåã® TL ä¸ã§ããè¦ãããããã«ãªã£ãããã§ã³ãã³ãã¢ã¤ã¹ã好ããããªå¥³ã®å*37ã®ç®ã被ã£ãå»çç»åã¤ãã¤ãã¨ã³ã¸ãã¢ã§ãã ãµããããããã*38ã主å¬ããLTä¼ã§ããå½¼ã¯å»çç»åç³»ã³ã³ãã§ã¡ãã«ãçæéã«ç«ã¦ç¶ãã«ã¨ã£ã¦ Master ã¨ãªããå®å ¨ã«å ãè¶ããã¦ãã¾ãã¾ããã
åå ããææ³ãæ¸ããã³ã«æ¯åè¨ã£ã¦ããã§ããã©ããã® LT ä¼ã¯ åã¿ãããªã³ã³ãã«åå ãã人ãå»çé¢ä¿è ãçè«ã¤ãã¤ãå¢ãªã©ã®ãã©ã¨ãã£ã«å¯ãã åå è ãæ°è»½ã«*39è«æç´¹ä»ããããè«æãèªãã¢ããã«ããã«ã¯ã¨ã¦ãè¯ãä¼ã§ãããå ã¿ã«ããã«è²¼ã£ã¦ããã®ã¯ Kaggle Master(4人) ã®ä¾µç¥ãåããä¼ã§ãç¬
ãä¸ã«æãããããªã¤ãã³ãã«ã¯åºåããã ãã©ãä½ããªãã¨ãã¯æ¬å½ã«åæ¥å¯ã¦ãã ãã¿ãããªãããªãã¤ããç¶æ ã«é¥ã£ã¦ããææãããã¾ããããã®ç¶æ³ãæç ´ãããã®ããã㯠... SPA ã§ãã
å¹³æ¥å¿ããã¦ãåæ¥ã«ãªã£ãã kaggle ãããã ...ãã£ã¦æã£ã¦ãã®ã«ãåæã«ãªã£ããããªãããã«ã¤ãªã¼ãkaggle é²ããªãããªã¼ãã¨ãæããªãããã©ãã©ãã¦ä¸æ¥ãçµãã£ã¦ãã¾ãæ¹ãã©ãããªã SPA ã§ãã©ãã©ãã¾ãããï¼ kaggle ããããã©å æ°ããªãã¨ãã«ä¸æ¦æ°æã¡ãå®å ¨ãªã»ãããã¦ãã¡ãã£ã¨ãããã¼ãã¨ããæ°åã«ãªããç¹ã§ SPA ã¯è¯ãã®ã§ã¯ãªããã¨èãã¦ãã¾ãã
SPAã§Kaggleããããã«åã£ãæ½è¨ãç¬æã¨åè¦ã§ç´¹ä»ãã - 俵è¨
ãã®è¨èã¯èªåèªèº«ã«åãããã®ã§ãå®é家ã§åæ¥å¯ãããã«ãªãããã¯å¤åºã®æ©ä¼ãå¢ãã¦ãå¾ã ã«ã§ããããæ¹åã«åãã£ã¦ããããã«æãã¾ããã¾ã SPA ã§ãªãã¦ãè¯ããã§ããã©ãªã«ãããéãç°å¢(ã§ããã°ãªã©ãã¯ã¹ã§ãã)ã«èº«ãç½®ãã¦ãªã»ããããã®ã¯è¯ããã¨ã§ãããã
å ã¿ã«ãã®å¾ã温æ³å¥½ããªã®ã¯å¤ãããã§ãå¾ã«å宿ãããããã¾ããã
ããã®ä»ãã«ã¬ã¼ã¡ãããããå§ãã¨ãããããã¼ã©ã®çãã (wakame ãããtakapy ãããsinchir0 ãã)ãéå¬ããåæ LT ä¼ã«åå ãããããã¾ããã
ãã®ä¼ã®ãæ°è»½ãªã¢ã¦ããããã®å ´ãä½ããã¨ãã主æ¨ã¯ã¨ã¦ãè¯ãã¦ãããããä¼ãå¢ããã¨ãããªã¨æãã¾ãããåèªèº«ãåºããã¢ã¦ããããã®ãããã§èªç¥ãã¦ããããããä¼è©±ãè°è«ã®ãã£ããã«ãªãã±ã¼ã¹ã¯æ²¢å±±ããã¾ããããã¾ãèªåã®æèãæ´çããããã§ã大äºã§ãã ããããã°ãä»å¾ #2 ã¯éå¬ãããã®ã§ããããð
復活ã®ãã£ãã : Kaggle Days Tokyo
ãæ©ãã«ç³ãè¾¼ãã ãããã§ç¡äºã«åå ãããã¨ãåºæ¥ã¾ãããæµç³ã«ããã§ã¯è©³ç´°ãçãã®ã§ãã*40ãæ§ã ãªç¥è¦ãä¸æ°ã«å¾ãããã¨ã¦ãåºæ¿çãªäºæ¥éã ã£ãã¨æãã¾ãï¼Kaggle ã¸ã®ã¢ããã¼ã·ã§ã³ãç¡éã«é«ã¾ã£ã¦ããð
äºæ¥ç®ã® ãªã³ãµã¤ãã³ã³ããã£ã·ã§ã³ã§ã¯ gege ãããuratatsu ããã¨ãã¼ã ãçµã¿ãJTC Kagglers season3 ãèªçãã¾ãã
Adoptionããã¾ãã (AdoptionSpeed=1)
— gege (@gegegegegege) December 11, 2019
ã¤ãã«ä¸å¡ã«ãªããæãæ¥ãã https://t.co/iqLQFmqaPP
— ãã ã㤠(@d1348k) December 11, 2019
çµæçã«ã¯ Public 34th â Private 33th ã§ãããåããTable NN*41 æ å½ãã¾ãï¼ãã£ã¦è±ªèªãã¦ãã®ã«å®¶ã®ãã·ã³ãè½ã¡ã¦*42ããã ã® hand-crafted feature 製é æ©ã«ãªã£ã¦ãã件ã¯æ¬å½ã«ãã¿ã¾ããã§ããð ã¾ããåããã£ã¨æ©ãããã¦ãã¾ã meta data ãå¦çã§ãã¦ããã° ... ã¨ããå¾æããã£ããã ã§ãä¸ç·ã«åå ããã®ã¯ã¨ã¦ã楽ããã£ãã§ãããããã¨ããããã¾ããð
ãæ親ä¼ã§ã¯æ¥æ¬ä¸ã® Kagglerããããã¯ã¤ã¿ãªã¢ããé帰ããã¦ãåæ¦ãã Jun Coda ãã ã¿ããã«æµ·å¤ãã帰ã£ã¦æ¥ãæ¥æ¬äºº Kagglerãããããæ¥æ¬ã¾ã§æ¥ã¦ãããæµ·å¤ã® Kaggler ã®çãããªã©ãæ§ã
ãªäººã
ã¨ã®äº¤æµã®æ©ä¼ã§ããã
ãã¦ãããã§å㯠Kaggle Days Tokyo ã§ã®å人çè£ããã·ã§ã³ãæããã«è¡ãã¾ãã
ãã£ããhttps://t.co/zPW1zlRneG pic.twitter.com/5DbDDNMfnl
— 俵 (@tawatawara) December 12, 2019
ç¹ã«æå¾ã«ãµã¤ã³ãããã£ã Maxwell ãã ãåã® iMet2019 ã® lost place solution ã®ãã¨ãè¦ãã¦ããã¦ãã¦ãããããããé å¼µã£ã¦ãã¨è¨ã£ã¦ããããã®ããã¡ããã¡ãå¬ããã£ãã§ã*43ã
ãããªããã§ãã¾ã Kaggle ãã£ã¦ãããðªãã¨ããæ°ã²ã¼ã¸ã MAX ã«ãªãã¾ããã
ãã¿ããå¼ãã¦éã ããð : ASHRAE - Great Energy Predictor III (2 Hours)
Public 1158th â Private 227thð¥
å³å¯ã«è¨ãã¨ã³ã³ãéå¬ç´å¾ã«ãããã¨ãã¦ã¡ãã£ã¨ã ã EDA ã¨ããã¦ããã®ã§ããããªã¼ã¯ãçºè¦ãã¦ããæ°ã失ãæ¾ç½®ãã¦ãã¾ããã
ä»ã ããè¨ãã®ã§ãããç· ãåãã®æã«ãã¾ã絶対 shake ããã ãããé©å½ãªãã¤æãã¦ã¿ãããã¨å ¬éã«ã¼ãã«ã大éã«éã㦠blending(ã¨ãããåç´ãª averaging) ãã¦æåºãã¾ãããæ¬å½ã¯ãã£ã¡ãè½ã¡ããå¾®å¦ãªçµæã«çµãã£ã¦ãã¾ãããã ããï½ãã£ã¦ãªãã®ãæ³å®ãã¦ããã§ãããæã£ã以ä¸ã« Shakeup ãã¦ãã¾ã£ããä½ã¦ãã£ãã...
ä¸å¿ã®æ¹éã¨ãã¦ãleak validation ããã¦ãªããã¤ã ã(確ã9åã10 å)éãã¦ã㦠averaging ããã¾ãããæåéãé試ãã ã£ãããã§ããããã®å¾ã®ãã¨ã示åãã¦ããã®ããç¥ãã¾ããã
念é¡ã® Solo Medal ð : 2019 Data Science Bowl
Public 701th â Private 198thð¥
ãKaggle Days Tokyo ã®ãªã³ãµã¤ãã³ã³ãã§ãã¼ã ã¡ã¤ãã«é ¼ãã£ããã ã£ãåçãããlog data (event data) ã®æ±ããå®è·µã§ãã£ããå¦ã¼ãã¨æã£ã¦åæ¦ãã¾ãããæå¾ã®æå¾ã§ããããããã®ã®ããã£ããåãçµãã ä¸ã§ã®åãã¦ã® Solo Medal *44 ãç²å¾ã¨ããå¬ããçµæã§ããã
ããåãããã»ã©ã¡ã¢ãªã¨ã©ã¼ã«ã¯æ°ãä»ããã¨...ãããæ°ãä»ãã¦ã¯ãããã§ãï¼ãã pandas ã¨ã GroupbyAgg ã§ã®ã¡ã¢ãªæ¶è²»ã®åããææ¡ãåãã¦ããªãã£ãã¨ãããã§ã late sub ããã¦ã¿ãã¨ãå®ã¯æå¾ã® 5 sub ã«ã¯ã¡ãã«åã® sub ãç¡ãã£ããã§ãããã人çã¯ä½ããããããããªããã®ã
çé¢ç®ã«åãçµãã§ãããããã§åå ã§ããåçä¼ããããããªç¥è¦(e.g. lightgbm ã® weight ã¨ãã2nd ãã¼ã ã® meta feature ã®è©³ç´°ã¨ã) ãå¾ããã¦ããã£ãã§ãï¼
å ã¿ã«é åºãã¸ããã¢ãã«ã®è¿½å®é¨ã¯ãã£ã¨åã«çµãã£ã¦ãã®ã§ãããæéãç¡ãã¦ããã°æ¸ãã¦ãªãã®ã§ããã¡ãã£ã¨ãå¾ ã¡ãã ããð
Shakeup åãé ããï¼ : AtmaCup#3
Public 31st â Private 27th
ãã¥ã¼ã¨ã¼ã«ã¼GOTO ããã®ä¸»å¬ãããåã ãããã£ã¨åå ãããã£ã atmaCup ã«åæ¦ãçµæã¨ãã¦ã¯èªåã®æã®é ããå®æãããã¨ã¨ãªãã¾ããããããããç²¾é²ãã¾ãã
ãã®ã¨ããç´åã« DSB2019 ã§è¦äº Solo Gold ãç²å¾ããã pao ãã ã¨ä½ã¨ãªãåããã¼ãã«ã§ä½æ¥ãã¦ãã¦*45ãpao ããã¯ããã§ãè¦äºãª shake control 㧠3ä½ã¨ãªããå㯠27ä½ã¨ããå¾®å¦ãªçµæã§ããã
pao ããã¨ããã°ä»¥ä¸ã® Tweet ãæåã§ãããã
家帰ã£ãã妻ãç¨æãã¦ããã¦ã¾ããð pic.twitter.com/UtRUH3hQDh
— pao (@pppaaaooo) January 23, 2020
ãã®ã¨ãã« shakeup ã¢ãããé ãã... ã¨ããã®ã¯æµç³ã«åè«ã§ãããDSB2019åçä¼ã§å ±æãã¦ããã£ãããã«ã¦ã§ã¢ä»¥éç çãªã¾ã§ã« Shake ã«ã¤ãã¦èãã¦ãã 話ã¯æ¬å½ã«åèã«ãªã£ã¦ãã¦ããvalidation set ãã¡ããã¨ä½ãã®ã¯ãã¡ããã¡ãéè¦ãã¨æ¹ãã¦èªèããã¦ããã£ãç¬éã§ããã
ã¡ãªã¿ã« pao ãã㯠atmaCup#4 ã§ã¯ Public 1st â Private 1st ã§ããã¯ãçè ã®é¢¨æ ¼ãæ¼ãå§ãã¦ãã...
è¦è³ãç¶ãã 4 Daysð / Solo Gold ã¸ã® 4 Daysð : Bengali.AI Handwritten Grapheme Classification
Public 124th â Private 10thð¥
ã人çä½ããããããããªããªã¨å®æãã¾ããããã®ãã¼ãã ããã¡ãã£ã¨è©³ããæç³»åã§ç´¹ä»ãã¾ãã
4 days ã¨ã¯è¨ã£ã¦ãããã®ã®ãèªåã®æå ã®ãã¼ã¯ãã¦ã³*46 ãè¦è¿ãã¦ã¿ãã¨ãæ§æ³ãç«ã¦å§ããã®ã¯ deadline ããä¸ã¶æãããåã® 2020-02-15 ã¨ãªã£ã¦ãã¾ãããã¨ãããåå¼·ã®ããã®ããã°æ¸ãã¦ãã DSB 㮠追å®é¨ãã£ãããã¦ããã¡ã«ã¿ãã¿ãæéãç¡ããªããæ°ã¥ãããä¸é±éåãããã«ãªã£ã¦ã¾ããã*47ã以ä¸ã¯ãã®æã«æ¸ããã¨æãããæç« ã§ãã
æ§é ãæ´çãã. ä½æ ãã£ã¨æ©ãå§ããªãã£ã...ãã®æ¿èªæ¬²æ±ã®å¡ãã...
ãã¦ããã¨ä¸é±éãç¡ãã®ã§ã次ã®ç»åã³ã³ãã«åãã src code æ´ç & Kaggle ã«ä¸ãã¦ã® submission ãåºæ¥ãã°ããããªãããããã®æ°æã¡ã§ãã§ãåãçµãã®ã¯å ¨åã§ãã¨ããæãã§ããããããã src code ã®æ´çã«3æ¥ã»ã©ããã*48ã5æ¥éãåã£ã¦ããç¶æ ã§ã®æåã® submission ã以ä¸ã§ãã
ç¡äºã«åãµãéã£ãããã£ãã4æ¥éé å¼µã£ã¦ãããã pic.twitter.com/Vpi4qcYCb1
— 俵 (@tawatawara) March 12, 2020
solution ã«ãæ¸ãã¦ããããã« graphme ã group ã¨è¦ãªãã multi-label stratified group k-fold 㧠split ã㦠validation ãããvalidation ãããã¤ãã©ã§æå¾ã« å
¨ãã¼ã¿å¦ç¿ + snapshot ensemble ãå½åã®è¨ç»ã§ãããã¨ããããããä¸ã¤ã® graphme ã«ããåå¨ããªããã㪠component ããã£ãããã¦å¦ç¿ããã¾ãè¡ãããè¨ç»ãæ©ã段éã§é æ«ãã¦ãã¾ãã¾ãã
å¾ã«ç¥ãã®ã§ãããããã«é¢ãã¦ã¯ 1st ã® deoxy ãã ã discussion ã«æ¸ãã¦ãã ãã£ã¦ããããæ¹ãã¨ã¦ãè³¢ãã¦ããããæãã¤ãã¦ããã°ãªãã¨ããæãã§ãã
åç´ã« unseen ã validation ã«ããã val loss ãããä¸ãã£ã¡ãã£ã¦ãã¡ãå°ã£ã¦ããã©ãseen validation 㨠unseen validation ã®2種é¡ãä½ãã®ãããªãã»ã©ï½
— 俵 (@tawatawara) March 17, 2020
ãã㦠solution ãã¡ããã¡ãã«ãã³ããï¼https://t.co/8mcalsrVhm
ãã éå¬æéä¸ã¯ãããæãã¤ãããä¸ãããªã validation score ã«ãã£ã¨åè¦å
«è¦ãã¦ã¾ããã
çæ©ã§ããã
— 俵 (@tawatawara) March 13, 2020
I just made my first #kaggle submission to https://t.co/GDKwssBh3A pic.twitter.com/8zbKDh6IEL
æ¬å½ã«ä½ãããããªãã score ãä¸ãããªã
— 俵 (@tawatawara) March 13, 2020
Tweet ãè¦ã¦æ¯ãè¿ãã¨ãããã®ã§ããã段ã
ã¨ããããã®å¢å°ã«å
¥ã£ã¦ããã¾ãã
é«ãé ä½ã¯æã¾ãªãããä½ããä¸ã¤ã§ãç解ãã¦ããçµãããã
— 俵 (@tawatawara) March 14, 2020
ä»åã®ãã£ã¬ã³ã¸ã§ãã³ãã³ã«ããããã¨ã§ã次以éã®ç»åã³ã³ãã®ããã®æºåéåãåºæ¥ã¦ãã(ã¨æããªãã¨ãã£ã¦ãããªã)
— 俵 (@tawatawara) March 14, 2020
ä½ãããå¥ã®è¦³ç¹ã§èªå·±è¯å®ãã¦ç¾å®éé¿ãå§ãã¦ããã
ä»åä¸çªå¾ããããã®ãå®ã¯èªå® ãã·ã³ãè½ã¡ãªããªã£ããã¨ã ã£ãããããç¥ã»é£ç¶ç¨¼å2é±éð
— 俵 (@tawatawara) March 14, 2020
ä½ãããªãã¾ã¾ðºã®éå§è¿ãããæºåã«ãã£ã¡ãæéããã£ãã¨æããããbengali ã¸ã®çæéåå ã§æºåãæ´ã£ãã®ã¯ããªãè¯ãã£ãã¨æã(ãã£ã¡ãç²ãã¦æ¯åããã¡ãã£ããã©)https://t.co/uBvu9KL1AJ
— 俵 (@tawatawara) March 15, 2020
ã§ãããã£ã±ã諦ãããã¾ããã§ãããã ããæå¾ã®è³ãã«åºããã¨ã«ããã¨ããããã§ãã
ã¨ã¯è¨ãã4æ¥éé ãæ©ã¾ããããã ããæå¾ã®æµæã¯ãããã¨ãã
— 俵 (@tawatawara) March 15, 2020
çæ決æ¦ã§ããã£ãã®ãããããªã(äºæ¥è¿·èµ°ãã¦è«¦ãã®å¢å°ã«è³ããã¤ã±ã¯ã½ã®å¦ç¿ãåãã¦ãã)
— 俵 (@tawatawara) March 16, 2020
validation score ãã¡ããã¨åºãã¦ããªãä¸ã§ã®å ¨ãã¼ã¿å¦ç¿ãepoch æ°ã¯ çµäº 3æéåã«çµããããã«ã㦠submission 㨠é¸æãåºæ¥ãããã«ãã¤ã3以ä¸ã® snapshot ãåããããã«(æ¬å½ã¯4è¡ãããã£ã) 35epoch x 3 ã«æ±ºå®ã¨ãããããªãã²ã©ããåã®æªãè³ãã§ããã
Discussion ãè¦ãéã train ã« overfit ãã¦ã public score ã¯ãããªãã«åºããã ã£ãã®ã§ãShake down ãã¦ããããã public ã ãã§ãä¸ããããã¼ã¨ããæ°æã¡ããã£ã¦æå¾ã®å¦ç¿ãå§ãã¾ãã
submit (cycle 1).
ä½ããããããã©ã»ã©ã»ã©ã® score ã«ä¸ãã£ããåºæ¥ãã°ã¡ãã«åã¾ã§è¡ããããåºç¤¾ãããï½
— 俵 (@tawatawara) March 16, 2020
I am a giant nerd, but that's okay. #kaggle - https://t.co/GDKwssBh3A pic.twitter.com/LidZnp8cVG
submit (cycle 2).
ãªãã»ã©ï¼(ä½ãããã£ã¦ãªãé¡)
— 俵 (@tawatawara) March 16, 2020
Doing my part to bring about the singularity. #MachineLearning #kaggle - https://t.co/GDKwssBh3A pic.twitter.com/MGOHzxPC84
submit (cycle 3).
ãã¨ããã¡ãããªãã ããªãã+0.002 ã欲ãã...
— 俵 (@tawatawara) March 16, 2020
Doing my part to bring about the singularity. #MachineLearning #kaggle - https://t.co/GDKwssBh3A pic.twitter.com/3FE3OTKGZC
ä½ãæã£ããã score ãä¸ãã£ã¦ãã¦ãèãã å¿ãå復ãã¦ãã¾ããããã¡ãã public ãä¸ãã£ãã ãã§ãããå
ã
ã¯å¦ç¿ããã¾ãã§ããªã㦠1700 ä½ã¨ãã«å±
ãããã§ããããèããã¨ãè¯ãã£ãï½ðãã£ã¦æãããããªãã¨(â)ãè¨ããããã«ã¯å¿ã®ä½è£ãæ»ãã¾ããã
4æ¥éé å¼µã£ã(â»éä¸æ¯åããã)ãããã§ãã®ãç¥ãã«åå ã§ããã®ã§å¬ããã
— 俵 (@tawatawara) March 16, 2020
ä¹ ã ã«ç»åèªèã®åºç¤ã®å¾©ç¿ã§ããã®ã楽ããã£ããã©ãæå¾ã®ç¬éãä¸ç·ã«å¾ ã¤ã®ãã³ã³ãã®ééå³ã£ããã
submit (Snapshot Ensemble)
æå¾ã¯ä¼¸ã³ãããç²ãæ§ã§ããï½ pic.twitter.com/fI26J3pTt3
— 俵 (@tawatawara) March 16, 2020
ã¾ããããªãããªã§ã4æ¥éã«ãã¦ã¯é å¼µã£ãæ¹ã ããã¨ãããã¨ã§ããªãæ°æ¥½ãªæ°æã¡ã§ private LB ã®å ¬éãå¾ ã¡ã¾ããããã¨24ä½ä¸ããã°éåã§ããããã° Master ã«...ããã§ã Validation åºæ¥ã¦ãªãããªð ãã£ã¦ããã§ãã¾ãé ã¡ãã«åããã°å¾¡ã®åããªï½ã¨ããæ°æã¡ã ã£ãã®ãè¦ãã¦ãã¾ãã
ããããæéããª
— 俵 (@tawatawara) March 16, 2020
å®ã¯ãã®æç¹ã§åèªèº«ã®çµæã¸ã®èå³ãã»ã¨ãã©å¤±ã£ã¦ããã®ã§ãéåéè¿ã«å±
ãããã¤ãã®æ¥æ¬äººãã¼ã ã®åå¿ãè¦ããã®ã¨ããã shakeup ãã¦éåã«å
¥ãããããç¥ããããã¨ãããã¨ã§ TL ã«å¾
æ©ãã¦ã¾ããããããã...
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
Tawaraããéãã
— mocobtð (@mocobt) March 17, 2020
... ã¯ï¼ ä½ã... ä½ãè¨ã£ã¦...ï¼
ãããã§ããæµç³ã«ããå¾ãªããããã¨æãã¤ã¤ LB ãè¦ã«è¡ãã¨ã
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ãããããããããããã.
ããããããã....?????(è¨èãåºãªã) pic.twitter.com/WzMUYlOitO
— 俵 (@tawatawara) March 17, 2020
æãéãã...æ¯ãä¹±ãã...ãªãã ããã¯..?
— 俵 (@tawatawara) March 17, 2020
æ£ç´è¨³ãåããã¾ããã§ããããã®æã¯å¬ããã¨ãããã·ã§ãã¯ã¨ãããã¨ã«ããå¿ã®ä½è£ããªããªããåããå¤§å¹ ã« shakeup ãã ããã¿ãããã¾ãã ãã¼ã ã®ãã¨ã«æ°ä»ãããç¥ãããã®ãé ãããããã§ãã
ãã¾ã æ¤è¨¼ä¸ã§ãããè¦å ã¨ãã¦ã¯ãã¢ãã«çãªæå³ã§ã¯ sSE-Pooling 㨠Data Augmentation ã®ä»æ¹ãèã ã£ãããã§ãã³ã³ãçãªæå³ã§ã¯æã£ããããå¤ãã®ãã¼ã ã unseen grapheme ãã¬ã³ç¡è¦ãã¦ããã¨ãããã¨ã§ããã以ä¸ã¯ ããã ã¿ããã«ãã解説ã§ãã
æ£ç´ã0.98ãããããä¸ã¯ä½ãã対å¦ãã¦ããã¨æã£ã¦ããããã©ãrecallæé©åãã¦ãªãã©ãããunseenã¬ã³ç¡è¦ã§graphemeåä½ã§åé¡ãã¦ãã人ããããããããã§ãããããã¨ãã¼ã¬ã¼ãã§è½ã¡ãä¸ã«recallã«æé©åãã¦ãªãåé¡ãunseenã«ããã¹ã³ã¢å¸¯ä½ä¸ã«ãã極ãã¦æ·±å»ã«é¡å¨åããã
— ããã ã¿ (@nejumi_dqx) March 20, 2020
ããã¦ããªãKaggleréã¯ãã®åé¡ã«ï¼æä¸ä½ã®ä¸é¨ãé¤ãã¦ï¼é©åã«å¯¾å¦ã§ããªãã£ãã®ãï¼ããããã¯Publicã«ãã¾ãã«å°æ°ã®unseenããå«ã¾ããªãã£ããããKagglerã¯åºæ¬çã«ã¯CV+Publicãå¼·çã«æé©åããã½ã«ãã¼ãªã®ã§unseenãåé¡ãããã£ããpublicã«ããããªãã«å ¥ããªãã¨motivateãããªãã
— ããã ã¿ (@nejumi_dqx) March 20, 2020
å㯠unseen grapheme å°ç¨ã«ä½ãããããã§ã¯ç¡ãã§ããã seen grapheme ã« overfit ãã¦ããããã§ããªãã®ã§çãæ®ã£ãã£ã¦ãã¨ã ã£ãã¿ããã§ããData ã®ãã¼ã¸ã« ãtest set 㯠unseen grapheme å«ãã§ãããã£ã¦æ¸ããã¦ããã§ããã©ããæµç³ã«ããã¾ã§ public 㨠private 㧠unseen ã®æ¯çã«å·®ãããã¨ã¯...ãã¨ããã®ãåå ãã Kaggler ã®æ°æã¡ã ã¨æãã¾ãã
ã¨ããããã§ãé·ãé·ãæ è·¯ã¯ããã«çµãããéã« Kaggle Master ã«ãªã£ãã®ã§ããã
å¥è·¡ã® solo gold ãç²å¾ããéã« Kaggle Master ã«ãªãã¾ããï¼
— 俵 (@tawatawara) 2020å¹´3æ18æ¥
Kaggle ã¨åºä¼ã£ã¦ãã4å¹´åãä¸å¿µçºèµ·ãã¦åæã³ã³ãã«æ¬æ°ã§åãçµã¿å§ãã¦ãã1å¹´åãé·ãæ è·¯ã§ããã
ä»ã¾ã§ãã¼ã ãçµãã ãè°è«ããããã¦ä¸ãã£ãçãããæ¬å½ã«ãããã¨ããããã¾ããï¼ããããããããããé¡ãè´ãã¾ãï¼ pic.twitter.com/IzWin2rioa
ã¾ã¨ã
ãè³æ¥µå½ããåã®ãã¨ãåèªèããã®ã§ãããåã¯åãç©ã¿éãããã®ã§åºæ¥ã¦ãã¾ããæå¾ã® Bengali ã§ç¨ããææ³ã¯ã©ãããããéå»ã®ã³ã³ãã§å¦ãã ãã®ã§ããããããæ´»ãã¦ãã®é ä½ãåãã¾ããããã¡ããä¸ä½é£ã unseen grapheme ã¸ã®å¯¾çãæ ã£ã¦ããªããã°éåã¾ã§ä¸ãããã¨ã¯ç¡ãã£ããã§ããã©ãç´ç²ã«ãã®ã¹ã³ã¢ãåããã¨ãåºæ¥ãã®ã¯åèªèº«ãç©ã¿éãã¦ãããã®ã«ãããã®ã§ãã
ãå ã¿ã«å¯ä¸åãã¦ä½¿ã£ããã®ã¯ sSE Module ã§ãããå®ã¯ç»åè«æ LT ä¼ã§ç¥ãåã£ãã¨ãã GAN ä¸çã®ä½äººã®ããã°*49ã§ç¥ã£ããã®ã§ããèªãã ã¨ãã¯ãã¾ããã®ã¢ãã«ã® pre-trained ããã£ãã使ãããããããã©ãèªå㧠scratch ã§ã®å¦ç¿ã¯ããªãã ãããªãã¨æã£ã¦ã¾ãããç¥èã£ã¦ãã®ã¨ãå½¹ã«ç«ããªãã¨æã£ãã¨ãã¦ãæãã¬ã¨ããã§å½¹ç«ã¤å ´åãå¾ã
ã«ãã¦ãã£ã¦ãä»åãããã ã£ãã¨ããããã§ãã
ãåã ãããã£ã¦ Kaggle ã«ã¢ããã¼ã·ã§ã³é«ãæã¡è¾¼ããã®ã¯ kaggler-ja ã®ã³ãã¥ããã£ã®ãããã§ãããã¾ããæ¬æ ¼åæ¦åã® Instacart ã Mercari ã®è©±ã§ãã£ãããã«ãã³ã³ãã¯å¤ç¬ã«åå ããã®ã¯é常ã«è¾ããã®ããã£ã¦ããã¼ã ãçµã¾ãªãã«ãã¦ãå
·ä½çãªç«¶ãåãç¸æãå±
ããã¨ã¯ã¨ã¦ãéè¦ã§ããåã®å ´åãåææã«å§ãã Kaggler ã®æ¹ã
ãçµæ§ããã£ããã£ããããã§ãäºãã«åç£ç¢ç£¨ãããã¨ãåºæ¥ãã¨æã£ã¦ãã¾ãã
ãããä¸ã¤ã³ãã¥ããã£ã§æ¬å½ã«ãããããã£ãã®ã¯åå¿è
ã®è³ªåã«ãä¸å¯§ã«çãã¦ãã ããä¸ä½è
ã®åå¨ã§ãNNã®å¦ç¿é¢é£ãªã©é常ã«ãä¸è©±ã«ãªãã¾ãããKaggle Tokyo Meetup #6 ã®æ親ä¼ã§ããèãããããã¨ããã£ã¦ smly ããã¨è©±ããã¨ãã¯ãããä»å smly ããã¨è©±ãã¦ãããªï¼ãã¨çµæ§ä¸æè°ãªæ°æã¡ã«ãªãã¾ãã*50ãKaggle Tokyo Meetup ã¯ãã¬ãã«ãåãã Kaggle ã楽ãããã¨ãã趣æ¨ã§éå¬ããã¦ããã kaggler-ja ãæ¹éã¨ãã¦ã¯åãã ã¨æãã¾ããå£æ ¹ç¡ã交æµã§ããã¨è¨ãã®ã¯æ¬å½ã«ç´ æ´ããããã¨ã§ãã
ãæå¾ã«æ¹ãã¦æããã¨ã¯ Twitter ã§ãããã°ã§ã Kaggle Discussion ã§ãããã®ã§èªåããæ å ±ãçºä¿¡ãããã¨ã®éè¦æ§ã§ããæ å ±ãæ´çãããã¨ããã®ããã¾ãããèªåã®ãã¨ãç¥ã£ã¦ãããããã¨ã§ç«¶ãåãç¸æãå ·ä½çã«è¦ãã¦ããããç¸æãããæç¨ãªæ å ±ãããããããã¦ãé常ã«è¯ã循ç°ãä½ããã¨ãåºæ¥ã¾ãã
以ä¸ããKaggle ã«åºä¼ã£ã¦ãã 5å¹´å¼±ã®äººçã§æã£ããã¨ã§ããç¹ã«ã³ãã¥ããã£ã«æ¬å½ã«å©ããããã¨æã£ã¦ãã¾ãããããã¨ããããã¾ããï¼
ããããã Kaggle ããã®ï¼
ãã¾ãã
ããä»®ã«ä»åã® Bengali 㧠Silver ãåã£ã¦ Master ã«ãªã£ã¦ãã¾ã£ãå ´åã¯ã¡ãã£ã¨ãã¼ã³ãã¦ã³ããããããªãããªã¼ã¨æãã®ã§ãããæ³å®å¤ãªãã¨ã« Solo Gold ãæã«å
¥ãã¦ãã¾ãã¾ããããã㯠Grand Master ã¸ã®éãéãã¦ãã¾ã£ãã¨ãããã¨ã§ããè¦ããã«éããããªãã
ããã«ä»åã®ã³ã³ãã§ãåã¯ãã¾ãããåç㪠shakeup ã¨è¯ãæ績ãåãã¨ããä¸æ¯ç©è³ªãæåãã¦ãã¾ãã¾ãããã¾ã ã¾ã ã³ã³ãããããã¨ã«ãªãã§ãããã
ã¦ããã§ä»å¾ã¯ããã¤å§ã¾ããããããªã iMet2020 ã¸ã®ãªãã³ã¸ã¨ã
M5 ã® Uncertainty ã¸ã®ææ¦ããããã¨æã£ã¦ãã¾ãã
åè ã¯å®å ¨ã«è¶£å³ & å»å¹´ã®ãªãã³ã¸ã§ãå¾è ã¯å®åçãªè¦³ç¹ã§æ¯éåãçµã¿ããã¨æã£ã¦ããå 容ã§ããä¸ç¢ºå®æ§ããã¼ãã¨ãã¦åãçµããã³ã³ããããã®ã¯ãã¡ããã¡ããããããð
ããã¾ã§èªãã§ãã ãã£ãæ¹ã ã(ãã®è¨äºã®)é·ãé·ãæ è·¯ãç²ãæ§ã§ãã & ãããã¨ããããã¾ããã Kaggler ã®çæ§ãããããã対æ¦ãããããé¡ããã¾ãï¼ï¼ï¼
*1:æåæ°çãªæå³ã§ã a long long journey ã«ãªã£ã¦ãã¾ã£ã...
*2:ä¸å¹´ãããåã«ãä¸åº¦æ話ãã¨ã ãæ¸ãã¾ããããããããã¯ãã£ã¨ãã£ã¨æãã話ã§ã
*3:ããã¯é常ã«ã¾ã£ã¨ããªè©±ã§ãç¹ã«ãã¡ã®ç 究室ãå¿ç¨åéã ã£ãã ãã«ããã å¦ç¿ã»äºæ¸¬ããã ãã§ã¯ç 究ã¨ãã¦ãã¡ã ããã¨ãããã¨ã§ããã©ã®ãããªç¥è¦ãå¾ãããã©ããªã³ã³ã»ããã示ããããéè¦ã
*4:https://hkashima.github.io/index.html
*5:ãã®ã¨ãã«ãè¿·æããããå çæ¹ãç 究室ã®çæ§ãæ¬å½ã«ç³ã訳ããã¾ããã大å¤ãä¸è©±ã«ãªãã¾ããã
*6:https://dl.acm.org/doi/10.1145/3298689.3347018 ã® First Authorããã¡ãããã人ã
*7:è»æ¬ã«ããã£ã¡ããä¸è©±ã«ãªãã¾ãã
*8:表ç¾ã¯å¤å°éã£ã¦ãããããããªãã§ãã大æã¯åã£ã¦ããã¨æãã¾ãã
*9:ãã ãå½æéä¸ããæµ·å¤ã«è¡ã£ã¦ãã¾ã£ãã®ã§ã¢ããã¤ã¶çå½¹å²
*10:å ã¿ã«çµæçºè¡¨ãä¸åº¦JSAI2018ã«åå ãã¦ããã¨ãã§ãçºè¡¨ãè´ããªãããã£ã¡ããã¤ãã¤ãã¦ãæªãããã£ããã ã£ãã¨æãã¾ãã
*11:å½æã¡ã¢ãªãé«ãã£ãããããã£ã¦å°ãé«ã
*12:ãã¾ã«ä½¿ãã¿ã¯ã·ã¼ã¿ãããªæã
*13:çµå±åæã³ã³ãå°ç¨ã«ãªã£ã¦ã¾ããã
*14:ChainerCV ã®ã¬ãã¸ããªã¨ããã£ã¡ããä¸è©±ã«ãªã£ã
*15:ã¨ããçµæ§å¹´ä¸ã®ç»åã«å¼·ãå 輩社å¡ã®ãAdam 㯠Loss ãæªãã...ãã¨ãããè¨èãæªã ã«é ããé¢ããªã...
*16:å®è£ ãã¹ã§éä¸ã¾ã§ Data Augmentation ãåãã¦ããªãã£ã
*17:ããããå°±è·ãã
*18:ãã¼ãã«ã³ã³ããç¹ã« Home Credit ããããã£ã
*19:ãã¡ããããã¯å¥½ã¿ã®åé¡ã§ãReduceLRonPlateu ã¨ããæç¨ãã¦ãæ¹ãããããããã£ãããã¾ã
*20:å½å®¶è©¦é¨åæ ¼ããã§ã¨ããããã¾ãï¼
*21:ã³ãã¥éãªã®ã§å²ã¨å¼·å¼ã ã£ãã¨æãã¾ãããã¿ã¾ãã...
*22:ä¾ãã°ç»åãµã¤ãºãå°ãããªããFeature Map ãå°ãããªããããã®ãé²ãããã« ResNet ã® first-conv or ãã®å¾ã® max-pool ãå¤æ´ãããªã©
*23:人å AutoAugment ããªï¼
*24:ãã¡ããã¢ãã«ãå¼ãå ´åãããã¾ãããã¿ã¹ã¯ã«åãããã«ããã¾ãããã¨ãããã¨ã§ãã
*25:å¶ç¶ã«ãããã®ã¨ãåã£ã人ã®ä¸äººã¯å¾ã«ãã¼ã ãçµã tamref ããã ã£ããããã
*26:ããã¦ãã®ãããããããªã³ãã¬ãã·ã³ãã¸ããæ²ããã®ã調åãæªããªãã ãã¾ã
*27:ãã«ããªï¼
*28:ãã®å¾ãæ²ããã« 3 åã»ã©ãã£ã¦ãã¾ããã¨ã«ãªãã¾ã
*29:e.g. å¦ç¿æ¸ã¿ã¢ãã«ã§åæ£è¡¨ç¾ã«è½ã¨ã â SVD ãªã©ã§å§ç¸®
*30:ã¡ãã£ã¨æ¶ãããããããã®ãã¼ããæ¸ãã®ãä¸çªããã©ãã£ãã
*31:ä»ã«ãã¾ã ã¾ã ããã£ããã£ããã§ããã¹ãã¼ã¹ã®é½åä¸å²æã
*32:æè¿ã¡ãã£ã¨ã¥ã¤æ»èªã«ãªã£ã¦æ¥ã¦ã¾ããã
*33:ãããã»ã¼ã»ã¼ã©ãããããç°å¢ã§ç²å¾ãå®å ¨ã«ã¬ã¸ã§ã³ãã
*34:autophalanx ã¯çä¼¼ã§ããæ°ãããªãã§ããã
*35:ãã æè¿ã®è«æèªãã§ãªãã®ã§ããã£ã¨è¯ãããæ¹ãåºã¦ãããããããªãã§ãã
*36:ãããå®éã¯ãã£ã¨æ²¢å±±ã®æ¹ã ã¨äº¤æµãã¦ã¾ããå²æã
*37:ç´èå§å¦¹ã¯è¯ãã
*38:å®ã¯å²ã¨åææã«åæã³ã³ããå§ãã¦ããã¿ããã§ãï¼https://fam-taro.hatenablog.com/entry/2019/11/23/175328
*39:ããããªããåå è ã®ã¢ãããé«ããã質ãé«ã
*40:æ¤ç´¢ããã°æ§ã ãªæ¹ã®åå è¨ãåºã¦ããã®ã§æ¯éã覧ä¸ããï¼
*41:SPA ã§å®è£ ãé²ãã¦ãããã®ã®ä¸ã¤ã
*42:ãã®æç¹ã§ã¯ã¾ã 調åãæªãã£ãã
*43:ãã®3ã«æå¾ãããã«åããã¼ã¸ã«ä¸¦ã¶ãã¨ã«ãªãã¨ã¯æãããã¾ããã§ãããç¬
*44:ãï¼ãASHRAE ã¯ï¼ãã£ã¦ï¼ããã¯å®å ¨ã«ãã¼ã«ã¦ã³ãã§ããå®åãããªãã®ã§ã
*45:ãã¼ã ãçµãã§ããããã§ã¯ç¡ã
*46:ã¢ã¤ãã¢ã¨ããæ¯ãè¿ããããã«ã³ã³ããã¨ã«æ¥ä»ã¨ä½ããã£ããã®æ¥è¨ã¿ãããªã®ãã¤ããããã«ãã¦ãã¾ã
*47:ä¸å¿ Discussion ã¨ãã¯ãã©ãã©è¦ã¦ã
*48:ãã ããã価å¤ã¯ååããã¾ãã
*49:https://blog.shikoan.com/concurrent_spatial_and_channel_squeeze_and_excitation/
*50:æµç³ã«æ°æ¥ããããã®ã§ãããªãããã£ãã㧠Kaggle ãå§ãã¾ãããã¨ã¯è¨ãã¾ããã§ããã