ãã¥ãŒã©ã«ãããã¯ãŒã¯ãã¬ããŒãžãçæããã®ã¯ãªãã§ããïŒããšãã°ããã¹ãŠã®çµæã®å¹³åãã粟床ãéåžžã«äœãã®ã§ããïŒã ãã§ãã¯ãéå§ããå Žæ
ãããã¯ãŒã¯ã¯ãããã€ãã®çç±ã§ãã¬ãŒãã³ã°ãããŠããªãå ŽåããããŸãã å€ãã®ãããã°ã»ãã·ã§ã³ã®çµæãç§ã¯ãã°ãã°åããã§ãã¯ãè¡ãããšã«æ°ä»ããŸããã ããã§ã¯ãååã®æé«ã®ã¢ã€ãã¢ãšãšãã«ãèªåã®çµéšã䟿å©ãªãªã¹ãã«ãŸãšããŸããã ãã®ãªã¹ããã圹ã«ç«ãŠã°å¹žãã§ãã
å 容
0.ãã®ããã¥ã¢ã«ã®äœ¿çšæ¹æ³ã¯ïŒ
I.ããŒã¿ã»ããã®åé¡
IIã ããŒã¿ã®æ£èŠå/æ¡åŒµã®åé¡
IIIã å®è£ ã®åé¡
IVã åŠç¿ã®åé¡
0.ãã®ããã¥ã¢ã«ã®äœ¿çšæ¹æ³ã¯ïŒ
å€ãã®ããšãããŸããããªãããšããããŸãã ããããããã€ãã®åé¡ã¯ä»ã®åé¡ãããäžè¬çã§ãã ç§ã¯éåžžãç·æ¥ããããšããŠãã®å°ããªãªã¹ãããå§ããŸãã
- ãã®ããŒã¿ã¿ã€ãã«å¯ŸããŠæ£ç¢ºã«æ©èœããåçŽãªã¢ãã«ããå§ããŸãïŒäŸïŒç»åã®VGGïŒã å¯èœã§ããã°ãæšæºã®æ倱é¢æ°ã䜿çšããŸãã
- ããŒã¿ã®æ£èŠåãå¢åŒ·ãªã©ããã¹ãŠã®é¯èŠãç¡å¹ã«ããŸãã
- ã¢ãã«ã埮調æŽããå Žåã¯ãååŠçãå確èªããŠãå ã®ã¢ãã«ã®ãã¬ãŒãã³ã°ã«äžèŽãããŸãã
- å ¥åãæ£ããããšã確èªããŠãã ããã
- æ¬åœã«å°ããªããŒã¿ã»ããïŒ2ã20ãµã³ãã«ïŒããå§ããŸãã 次ã«ããããå±éããæ°ããããŒã¿ãåŸã ã«è¿œå ããŸãã
- çç¥ããããã¹ãŠã®ãã©ã°ã¡ã³ããåŸã ã«è¿œå ãå§ããŸãïŒå¢åŒ·/æ£ååãã«ã¹ã¿ã æ倱é¢æ°ãããè€éãªã¢ãã«ãè©ŠããŠãã ããã
ä»ã®ãã¹ãŠã倱æããå Žåããã®é·ããªã¹ãã®èªã¿åããéå§ããåé ç®ã確èªããŸãã
I.ããŒã¿ã»ããã®åé¡

åºå žïŒ http : //dilbert.com/strip/2014-05-07
1.å ¥åã確èªããŸã
å ¥åãé©åã§ããããšã確èªããŸãã ããšãã°ãç»åã®é«ããšå¹ ãç¹°ãè¿ãçµã¿åãããŠ1ã€ã®æã«ããŸããã æã ãééã£ãŠãã¹ãŠã®ãŒãããã¥ãŒã©ã«ãããã¯ãŒã¯ã«äžããŸããã ãŸãã¯ãåãããããäœåºŠã䜿çšããŸããã ãããã£ãŠãå ¥åãšèšç»åºåã®ããã€ãã®ããããå°å·/確èªããŸãããã¹ãŠãæ£åžžã§ããããšã確èªããŠãã ããã
2.ã©ã³ãã ãªå ¥åå€ãè©Šã
å®éã®ããŒã¿ã®ä»£ããã«ä¹±æ°ãæž¡ããŠãåããšã©ãŒãæ®ããã©ããã確èªããŠãã ããã ãããããªããããã¯ããæç¹ã§ããªãã®ãããã¯ãŒã¯ãããŒã¿ããŽãã«å€ãããšãã確ããªå åã§ãã ã¬ã€ã€ãŒããšã®ãããã°ïŒæäœããšã®æäœïŒãè©Šè¡ããé害ãçºçããå Žæã確èªããŸãã
3.ããŒã¿ããŒããŒã確èªãã
ãã¹ãŠãããŒã¿ã®é åºã©ããã§ããå¯èœæ§ãããããã¥ãŒã©ã«ãããã¯ãŒã¯ã®å ¥åããŒã¿ãéä¿¡ããã³ãŒãã«ãšã©ãŒããããŸãã æäœãéå§ããåã«ãæåã®ã¬ã€ã€ãŒã®å ¥åããŒã¿ãå°å·ããŠç¢ºèªããŸãã
4.å ¥åãåºåã«æ¥ç¶ãããŠããããšã確èªããŸãã
è€æ°ã®å ¥åãµã³ãã«ãæ£ããã©ãã«ã§ã©ãã«ä»ããããŠããããšã確èªããŸãã ãŸããå ¥åãµã³ãã«ã®äº€æãåºåã©ãã«ã«åœ±é¿ããããšã確èªããŠãã ããã
5.å ¥åãšåºåã®é¢ä¿ã¯ã©ã³ãã ãããŸããïŒ
ãã¶ããå ¥ãå£ãšåºå£ã®éã®é¢ä¿ã®éã©ã³ãã ãªéšåã¯ãã©ã³ãã ãªéšåã«æ¯ã¹ãŠå°ããããŸãïŒèª°ããååŒæã®åŒçšã§ãããšèšããããããŸããïŒã ã€ãŸããå ¥åãšåºåãååã«æ¥ç¶ãããŠããŸããã ã©ã³ãã æ§ã®å°ºåºŠã¯ããŒã¿ã®ã¿ã€ãã«äŸåãããããæ®éçãªæ¹æ³ã¯ãããŸããã
6.ããŒã¿ã»ããã®ãã€ãºãå€ãããŸããïŒ
ãã€ãŠããµã€ãããé£ã¹ç©ã®ç»åã®ã»ãããåãåºãããšãã«ç§ã«èµ·ãããŸããã ããã«ã¯éåžžã«å€ãã®æªãç¹ãããããããã¯ãŒã¯ãèšç·Žã§ããŸããã§ããã äžé£ã®ãµã³ãã«å ¥åå€ãæåã§ç¢ºèªãããã¹ãŠã®ã©ãã«ãé©åã«é 眮ãããŠããããšã確èªããŸãã
ãã®äœæ¥ã¯ãMNISTã«åºã¥ããŠ50ïŒ ãè¶ ããæå·ã¿ã°ã®ç²ŸåºŠã50ïŒ ãè¶ ããããšã瀺ããŠãããããåå¥ã®è°è«ã«å€ããŸãã
7.ããŒã¿ã»ãããã·ã£ããã«ãã
ããŒã¿ãæ··åãããŠããããç¹å®ã®é åºïŒã¿ã°ã§ãœãŒãïŒã§é 眮ãããŠããªãå ŽåãåŠç¿ã«æªåœ±é¿ãäžããå¯èœæ§ããããŸãã ããŒã¿ã»ãããã·ã£ããã«ããŸããå ¥åãšã©ãã«ã®äž¡æ¹ãå¿ ãæ··ããŠãã ããã
8.ã¯ã©ã¹ã®äžåè¡¡ãæžãã
ã¯ã©ã¹Bã®ç»åããšã«ãããŒã¿ã»ããã«ã¯ã©ã¹Aã®ç»åã1000åããã®ã§ããããïŒ æ¬¡ã«ãæ倱é¢æ°ã®ãã©ã³ã¹ãåããã äžåè¡¡ã解æ¶ããããã«ä»ã®ã¢ãããŒããè©Šãå¿ èŠããããŸãã
9.ãã¬ãŒãã³ã°ã«ååãªãµã³ãã«ã¯ãããŸããïŒ
ãããã¯ãŒã¯ããŒããããã¬ãŒãã³ã°ããïŒã€ãŸããæ§æããªãïŒå Žåã倧éã®ããŒã¿ãå¿ èŠã«ãªãå ŽåããããŸãã ããšãã°ãç»åãåé¡ããã«ã¯ ãåã¯ã©ã¹ããšã«1000æ以äžã®ç»åãå¿ èŠã§ãã
10.åäžã®ã©ãã«ãæã€ãããããªãããšã確èªããŸã
ããã¯ããœãŒããããããŒã¿ã»ããã§çºçããŸãïŒã€ãŸããæåã®1äžåã®ãµã³ãã«ã«ã¯åãã¯ã©ã¹ãå«ãŸããŸãïŒã ããŒã¿ã»ãããã·ã£ããã«ããããšã§ç°¡åã«ä¿®æ£ã§ããŸãã
11.ããããµã€ãºãçž®å°ãã
ãã®äœæ¥ã¯ã倧ãããããããã¯ã¢ãã«ã®äžè¬åèœåãäœäžãããå¯èœæ§ãããããšã瀺ããŠããŸãã
è¿œå 1.æšæºããŒã¿ã»ããïŒmnistãcifar10ãªã©ïŒã䜿çšããŸã
hengcherkengã«æè¬ããŸã ïŒ
æ°ãããããã¯ãŒã¯ã¢ãŒããã¯ãã£ããã¹ããããšãããŸãã¯æ°ããã³ãŒããèšè¿°ãããšãã¯ãæåã«æšæºããŒã¿ã»ããã§ã¯ãªãæšæºããŒã¿ã»ããã䜿çšããŸãã ãããã«ã¯ãã§ã«å€ãã®çµæããããã解決å¯èœãã§ããããšãä¿èšŒãããŠããããã§ãã ã©ãã«ã®ãã€ãºããã¬ãŒãã³ã°/ãã¹ãã®ååžã®éããããŒã¿ã»ããã®è€éãã®åé¡ãªã©ã¯ãããŸããã
IIã ããŒã¿ã®æ£èŠå/æ¡åŒµã®åé¡

12.çç¶ã調æŽãã
å ¥åããŒãå¹³åãšåäœåæ£ã«èª¿æŽããŸãããïŒ
13.ããŒã¿å¢åŒ·ã¯åŒ·ãããŸããïŒ
å¢åŒ·ã«ã¯æ£ååå¹æããããŸãã 匷ãããå Žåãããã¯ä»ã®åœ¢åŒã®æ£ååïŒL2æ£ååãããããã¢ãŠããªã©ïŒãšãšãã«ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®æè²äžè¶³ã«ã€ãªããå¯èœæ§ããããŸãã
14.äºååŠç¿æžã¿ã¢ãã«ã®ååŠçã確èªãã
æºåæžã¿ã®ã¢ãã«ã䜿çšããå Žåã¯ããã¬ãŒãã³ã°ããŠããã¢ãã«ãšåãæ£èŠåãšååŠçã䜿çšããŠãã ããã ããšãã°ããã¯ã»ã«ã¯[0ã1]ã[-1ã1]ããŸãã¯[0ã255]ã®ç¯å²ã«ãããŸããïŒ
15.ãã¬ãŒãã³ã°/æ€èšŒ/ãã¹ããããã®ååŠçãæ€èšŒãã
CS231nã¯å žåçãªãã©ãããæããŠããŸãïŒ
ã...ååŠçã®çµ±èšïŒããŒã¿ã®å¹³åãªã©ïŒã¯ããã¬ãŒãã³ã°çšã®ããŒã¿ã§èšç®ããæ€èšŒ/ãã¹ãããŒã¿ã«é©çšããå¿ èŠããããŸãã äŸãã°ãããŒã¿ã»ããå šäœã®å¹³åãèšç®ããŠåç»åãããããæžç®ãããã¬ãŒãã³ã°/æ€èšŒ/ãã¹ãã®ããã«ããŒã¿ããã©ã°ã¡ã³ãã«åå²ããã®ã¯ééãã§ãã
ãŸããåãµã³ãã«ãšãããã®ç°ãªãååŠçã確èªããŸãã
IIIã å®è£ ã®åé¡

ãœãŒã¹ïŒ https : //xkcd.com/1838/
16.åé¡ã®ããåçŽãªããŒãžã§ã³ã解決ããŠã¿ãŠãã ãã
ããã¯ãåé¡ã®å Žæãç¹å®ããã®ã«åœ¹ç«ã¡ãŸãã ããšãã°ãã¿ãŒã²ããåºåããªããžã§ã¯ãã¯ã©ã¹ãšåº§æšã®å Žåãäºæž¬ããªããžã§ã¯ãã¯ã©ã¹ã®ã¿ã«å¶éããŠã¿ãŠãã ããã
17.ã確çã§ãæ£ããæ倱é¢æ°ãæ¢ããŸã
åã³ãç¡æ¯ã®CS231nããïŒ æ£èŠåããã«å°ããªãã©ã¡ãŒã¿ãŒã§åæåããŸãã ããšãã°ã10åã®ã¯ã©ã¹ãããå Žåãã確çãã¯ãæ£ããã¯ã©ã¹ãã±ãŒã¹ã®10ïŒ ã§æ±ºå®ãããããšãæå³ããSoftmaxæ倱é¢æ°ã¯æ£ããã¯ã©ã¹ã®ç¢ºçã®é察æ°ã§ãããã€ãŸãã
ãã®åŸãæ£åå匷床ãäžããŠã¿ãŠãã ãããããã«ãããæ倱é¢æ°ãå¢ããã¯ãã§ãã
18.æ倱é¢æ°ã確èªããŸã
ç¬èªã«å®è£ ããå Žåã¯ããã°ã確èªããåäœãã¹ããè¿œå ããŠãã ããã ãããã«äžæ£ç¢ºãªæ倱é¢æ°ããããã¯ãŒã¯ããã©ãŒãã³ã¹ããããã«æãªãããšããããããŸããã
19.æ倱é¢æ°ã®å ¥åã確èªããŸã
ãã¬ãŒã ã¯ãŒã¯ããæ倱é¢æ°ã䜿çšããå Žåãå¿ èŠãªãã®ãå¿ ãæž¡ãããã«ããŠãã ããã ããšãã°ãPyTorchã§ã¯ãNLLLossãšCrossEntropyLossãæ··åããŸããããã¯ã1ã€ç®ã¯softmaxå ¥åãå¿ èŠãšãã2ã€ç®ã¯ããã§ã¯ãªãããã§ãã
20.æžéæ©èœã調æŽãã
æ倱é¢æ°ãè€æ°ã®é¢æ°ã§æ§æãããŠããå Žåã¯ãçžäºã®é¢ä¿ã確èªããŠãã ããã ããã«ã¯ãããŸããŸãªæ¯çã§ã®ãã¹ããå¿ èŠã«ãªãå ŽåããããŸãã
21.ä»ã®ã€ã³ãžã±ãŒã¿ãŒã远跡ãã
æ倱é¢æ°ã¯ããã¥ãŒã©ã«ãããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã®æé©ãªäºæž¬å åã§ã¯ãªãå ŽåããããŸãã å¯èœã§ããã°ã粟床ãªã©ã®ä»ã®ã€ã³ãžã±ãŒã¿ã䜿çšããŸãã
22.åã«ã¹ã¿ã ã¬ã€ã€ãŒã確èªãã
ãããã¯ãŒã¯å±€ãåå¥ã«å®è£ ããŸãããïŒ æåŸ ã©ããã«æ©èœããŠããããšãå確èªããŠãã ããã
23.æ¬ èœããŠãããã¬ã€ã€ãŒããŸãã¯å€æ°ããªãã確èªãã
èŠãŠãã ããããã¶ãããªãã¯ããã€ãã®ã¬ã€ã€ãŒ/å€æ°ã®åŸé æŽæ°ã誀ã£ãŠãªãã«ããŸããã
24.ãããã¯ãŒã¯ãµã€ãºãå¢ãã
ããããããããã¯ãŒã¯ã®è¡šçŸåã§ã¯ç®çé¢æ°ãåžåããã«ã¯äžååã§ãã å®å šã«æ¥ç¶ãããã¬ã€ã€ãŒã«ã¬ã€ã€ãŒãŸãã¯é衚瀺ã®ãŠããããè¿œå ããŠã¿ãŠãã ããã
25.é ãã枬å®ãšã©ãŒãæ¢ã
å ¥åã次ã®ããã«èŠããå Žå ã誀ã£ã枬å®ã«é¢é£ãããšã©ãŒãç°¡åã«ã¹ãããã§ããŸãã ç°åžžãªæ°å€ã䜿çšããŠå ¥åããŒã¿ã枬å®ãïŒããšãã°ã枬å®ããšã«ç°ãªãçŽ æ°ïŒããããããããã¯ãŒã¯å šäœã«ã©ã®ããã«åºãããã確èªããŸãã
26.åŸé ãã§ãã¯ã調ã¹ã
èªåã§ã°ã©ããŒã·ã§ã³ãã£ã»ã³ããå®è£ ããå Žåã¯ãã°ã©ããŒã·ã§ã³ãã§ãã¯ã䜿çšããŠãæ£ãããã£ãŒãããã¯ã確èªã§ããŸãã è¿œå æ å ±ïŒ 1ã2ã3 ã
IVã åŠç¿ã®åé¡

ãœãŒã¹ïŒ http : //carlvondrick.com/ihog/
27.æ¬åœã«å°ããªããŒã¿ã»ããã®åé¡ã解決ãã
å°ããªããŒã¿ã»ããã§ãããã¯ãŒã¯ãåãã¬ãŒãã³ã°ãããããæ©èœããããšã確èªããŸã ã ããšãã°ãã»ãã®1ã2ã€ã®äŸã§åœŒå¥³ãèšç·Žãããããã¯ãŒã¯ããªããžã§ã¯ããåºå¥ã§ãããã©ããã確èªããŸãã åã¯ã©ã¹ã®ãµã³ãã«ãã芧ãã ããã
28.倩ã³ãã®åæåã確èªããŸã
äžæãªå Žåã¯ã XavierãŸãã¯Xeåæåã䜿çšããŸãã ããã«ãåæåãè¡ããšããŒã«ã«ã®æå°å€ãäœäžããå¯èœæ§ããããããå¥ã®åæåãè©ŠããŠãã ããã
29.ãã€ããŒãã©ã¡ãŒã¿ãŒãå€æŽãã
ãã¶ããããªãã¯è²§åŒ±ãªãã€ããŒãã©ã¡ãŒã¿ã®ã»ããã䜿çšããŠããŸãã å¯èœã§ããã°ã ã°ãªããæ€çŽ¢ãè©ŠããŠãã ããã
30.æ£ååãæžãã
æ£ååãå€ãããããããããã¯ãŒã¯ã¯ç¹ã«èšç·ŽãããŠããªãå¯èœæ§ããããŸãã ããããã¢ãŠããããããã«ã ãL2æ£èŠåã®éã¿/ãã€ã¢ã¹ãªã©ã®æ£èŠåãæžãããŸããåªããã³ãŒã¹ã ããã°ã©ããŒåãã®å®è·µçãªæ·±å±€åŠç¿ ãã§ã ãžã§ã¬ããŒãã¯ãŒãã¯ãŸãæè²äžè¶³ã解æ¶ããããšãæšå¥šããŸãã ã€ãŸãããœãŒã¹ããŒã¿ã§ãããã¯ãŒã¯ãåãã¬ãŒãã³ã°ããåãã¬ãŒãã³ã°ãšæŠãã ãã§ååã§ãã
31.æéãäžãã
ãããã¯ãŒã¯ãæå³ã®ããäºæž¬ãå§ããåã«ãåŠç¿ããã®ã«ããå€ãã®æéãå¿ èŠãšãããããããŸããã æ倱é¢æ°ãçå®ã«æžå°ããŠããå Žåã¯ãå°ãé·ãåŠç¿ãããŠãã ããã
32.ãã¬ãŒãã³ã°ã¢ãŒããããã¹ãã¢ãŒãã«åãæ¿ãã
äžéšã®ãã¬ãŒã ã¯ãŒã¯ã§ã¯ãããããã«ã ãããããã¢ãŠããããã³ãã®ä»ã®ã¬ã€ã€ãŒã¯ããã¬ãŒãã³ã°ããã³ãã¹ãäžã«ç°ãªãåäœãããŸãã é©åãªã¢ãŒãã«åãæ¿ãããšããããã¯ãŒã¯ãé©åãªäºæž¬ãè¡ãã®ã«åœ¹ç«ã¡ãŸãã
33.ãã¬ãŒãã³ã°ãèŠèŠåãã
- åã¬ã€ã€ãŒã®ã¢ã¯ãã£ããŒã·ã§ã³ããŠã§ã€ããæŽæ°ã远跡ããŸãã ãããã®æ¯çãäžèŽããããšã確èªããŠãã ããã ããšãã°ããã©ã¡ãŒã¿ãŒïŒéã¿ãšãªãã»ããïŒã®æŽæ°éã®æ¯çã¯1e-3ã§ããå¿ èŠããããŸã ã
- TensorboardãCrayonãªã©ã®èŠèŠåã©ã€ãã©ãªãæ€èšããŠãã ããã æåŸã®æ段ãšããŠãåçŽã«éã¿/ã·ãã/ã¢ã¯ãã£ããŒã·ã§ã³ã®å€ãå°å·ã§ããŸãã
- å¹³åããŒãããã¯ããã«å€§ãããããã¯ãŒã¯ã®ã¢ã¯ãã£ããŒã·ã§ã³ã«ã¯æ³šæããŠãã ããã Batch NormãŸãã¯ELUãè©ŠããŠãã ããã
- Deeplearning4jã¯ãéã¿ãšã·ããã®ãã¹ãã°ã©ã ã§äœãèŠãã¹ããã瀺ããŸããã
ãéã¿ã«ã€ããŠã¯ããã°ãããããšããããã®ãã¹ãã°ã©ã ã¯ã»ãŒã¬ãŠã¹ïŒæ£èŠïŒååžã«ãªãã¯ãã§ãã ã·ãããã¹ãã°ã©ã ã¯éåžžãŒãããå§ãŸããéåžžã¬ãŠã¹ååžçšåºŠã§çµãããŸãïŒå¯äžã®äŸå€ã¯LSTMã§ãïŒã ãã©ã¹/ãã€ãã¹ã®ç¡é倧ã«éžè±ãããã©ã¡ãŒã¿ãŒã«æ³šæããŠãã ããã 倧ãããªããããã·ããã«æ³šæããŠãã ããã ã¯ã©ã¹ã®ååžãäžåè¡¡ãããå Žåãåé¡ã®åºåå±€ã§ãããçºçããããšããããŸãã
- ã¬ã€ã€ãŒã®æŽæ°ã確èªããŠãã ããããããã«ã¯æ£èŠååžãå¿ èŠã§ãã
34.å¥ã®ãªããã£ãã€ã¶ãŒãè©Šã
ãªããã£ãã€ã¶ãŒã®éžæã¯ãç¹ã«äžé©åãªãã€ããŒãã©ã¡ãŒã¿ãŒãéžæããŠããªãéãããã¥ãŒã©ã«ãããã¯ãŒã¯ã®åŠç¿ã劚ããã¹ãã§ã¯ãããŸããã ããããã¿ã¹ã¯ã«æé©ãªãªããã£ãã€ã¶ãŒã䜿çšãããšãæçæéã§å¯èœãªéãæé«ã®ãã¬ãŒãã³ã°ãåããããšãã§ããŸãã 䜿çšããã¢ã«ãŽãªãºã ã説æããç§åŠèšäºã«ã¯ããªããã£ãã€ã¶ãŒãèšèŒããå¿ èŠããããŸãã ããã§ãªãå Žåã¯ãAdamãŸãã¯ãã¬ãŒã³SGDã䜿çšããããšã奜ã¿ãŸãã
åŸé éäžãªããã£ãã€ã¶ãŒã®è©³çŽ°ã«ã€ããŠã¯ã Sebastian Ruderã®åªããèšäºãã芧ãã ããã
35.åŸé ã®ççº/æ¶å€±
- éåžžã«å€§ããªå€ã¯åŸé ã®ããŒã¹ãã瀺ãå¯èœæ§ããããããã¬ã€ã€ãŒã®æŽæ°ã確èªããŠãã ããã åŸé ã¯ãªããã³ã°ã圹ç«ã€å ŽåããããŸãã
- ã¬ã€ã€ãŒã®ã¢ã¯ãã£ãåã確èªããŸãã Deeplearning4jã¯ãåªããã¢ããã€ã¹ãæäŸããŸãã ãã¢ã¯ãã£ããŒã·ã§ã³ã®æšæºåå·®ã¯ã0.5ã2.0çšåºŠã§ãã ãããè¶ ããéèŠãªã¹ãããã¯ãã¢ã¯ãã£ããŒã·ã§ã³ã®ççºãŸãã¯æ¶å€±ã瀺ãå ŽåããããŸãã
36.ãã¬ãŒãã³ã°ã®ã¹ããŒãã¢ãã/ã¹ããŒããŠã³
åŠç¿é床ãé ããšãã¢ãã«ã®åæãéåžžã«é ããªããŸãã
åŠç¿é床ãéããšãæåã«æ倱é¢æ°ãããã«æžå°ãããã®åŸãé©åãªè§£æ±ºçãèŠã€ããããšãé£ãããªããŸãã
åŠç¿é床ã10åéããããé ããããããŠãå®éšããŠãã ããã
37. NaNæ¡ä»¶ã®é€å»
éæ°å€ïŒNaNïŒç¶æ ã¯ãRNNãåŠç¿ãããšãã«ã¯ããã«äžè¬çã§ãïŒç§ãèããéãïŒã ãããã解決ããããã€ãã®æ¹æ³ïŒ
- ç¹ã«æåã®100åã®ç¹°ãè¿ãã§NaNã衚瀺ãããå Žåã¯ãåŠç¿é床ãäžããŸãã
- ãŒããŸãã¯èªç¶æ°ã®èªç¶å¯Ÿæ°ãåããŒãã«ããé€ç®ã«ãããéæ°å€ãçºçããå¯èœæ§ããããŸãã
- ã©ãã»ã«ã¹ãã¥ã¯ãŒãã¯ãNaNãçºçããå Žåã®å¯ŸåŠæ¹æ³ã«ã€ããŠã®è¯ããã³ããæäŸããŸãã
- ãããã¯ãŒã¯ãã¬ã€ã€ãŒããšã«è©äŸ¡ããNaNã衚瀺ãããå Žæã確èªããŠãã ããã
ãœãŒã¹
cs231n.github.io/neural-networks-3
russellsstewart.com/notes/0.html
stackoverflow.com/questions/41488279/neural-network-always-predicts-the-same-class
deeplearning4j.org/visualization
www.reddit.com/r/MachineLearning/comments/46b8dz/what_does_debugging_a_deep_net_look_like
www.researchgate.net/post/why_the_prediction_or_the_output_of_neural_network_does_not_change_during_the_test_phase
book.caltech.edu/bookforum/showthread.php?t=4113
gab41.lab41.org/some-tips-for-debugging-deep-learning-3f69e56ea134
www.quora.com/How-do-I-debug-an-artificial-neural-network-algorithm
russellsstewart.com/notes/0.html
stackoverflow.com/questions/41488279/neural-network-always-predicts-the-same-class
deeplearning4j.org/visualization
www.reddit.com/r/MachineLearning/comments/46b8dz/what_does_debugging_a_deep_net_look_like
www.researchgate.net/post/why_the_prediction_or_the_output_of_neural_network_does_not_change_during_the_test_phase
book.caltech.edu/bookforum/showthread.php?t=4113
gab41.lab41.org/some-tips-for-debugging-deep-learning-3f69e56ea134
www.quora.com/How-do-I-debug-an-artificial-neural-network-algorithm