æ©æ¢°åŠç¿ã®ãã¹ããã©ã¯ãã£ã¹ãšã¢ãã«ãã¬ãŒãã³ã°ã®ãã³ã
ã¯ããã«
ã³ã³ãã¥ãŒã¿ã»ããžã§ã³ã®ãããžã§ã¯ãã§æãéèŠãªã¹ãããã®1ã€ã¯ãã¢ãã«ã®ãã¬ãŒãã³ã°ã§ãããã®ã¹ãããã«å°éããåã«ãç®æšãå®çŸ©ããããŒã¿ãåéãã泚éãä»ããå¿ èŠããããŸããããŒã¿ã®ååŠçãè¡ããããŒã¿ãã¯ãªãŒã³ã§äžè²«æ§ãããããšã確èªããããã¢ãã«ã®ãã¬ãŒãã³ã°ã«ç§»ããŸãã
èŠããã ïŒ ã¢ãã«ãã¬ãŒãã³ã°ã®ãã³ãïœå€§èŠæš¡ããŒã¿ã»ããã®æ±ãæ¹ïœããããµã€ãºãGPU å©çšçããã㊠ããã¯ã¹ãã»ãã¬ã·ãžã§ã³
ã§ã¯ãã¢ãã«ã®ãã¬ãŒãã³ã°ãšã¯äœã§ããããïŒã¢ãã«ã®ãã¬ãŒãã³ã°ãšã¯ãèŠèŠçãªãã¿ãŒã³ãèªèããããŒã¿ã«åºã¥ããŠäºæž¬ãè¡ãããã«ã¢ãã«ãæããããã»ã¹ã§ããããã¯ãã¢ããªã±ãŒã·ã§ã³ã®ããã©ãŒãã³ã¹ãšç²ŸåºŠã«çŽæ¥åœ±é¿ããŸãããã®ã¬ã€ãã§ã¯ãã³ã³ãã¥ãŒã¿ããžã§ã³ã¢ãã«ãå¹æçã«ãã¬ãŒãã³ã°ããããã®ãã¹ããã©ã¯ãã£ã¹ãæé©åãã¯ããã¯ããã©ãã«ã·ã¥ãŒãã£ã³ã°ã®ãã³ãã«ã€ããŠèª¬æããŸãã
æ©æ¢°åŠç¿ã¢ãã«ã®ãã¬ãŒãã³ã°æ¹æ³
ã³ã³ãã¥ãŒã¿ã»ããžã§ã³ã®ã¢ãã«ã¯ã誀差ãæå°ã«ãªãããã«å éšãã©ã¡ãŒã¿ã調æŽããããšã§åŠç¿ããããæåã«ãã¢ãã«ã«ã¯ã©ãã«ä»ãããã倧éã®ç»åã»ãããäžããããããããã®ç»åã«äœãåã£ãŠãããäºæž¬ããäºæž¬å€ãå®éã®ã©ãã«ãå 容ãšæ¯èŒããŠèª€å·®ãèšç®ããããã®èª€å·®ã¯ãã¢ãã«ã®äºæž¬ãçã®å€ããã©ãã ããããŠãããã瀺ãã
åŠç¿äžãã¢ãã«ã¯ç¹°ãè¿ãäºæž¬ãè¡ãã誀差ãèšç®ããããã¯ãããã²ãŒã·ã§ã³ãšåŒã°ããããã»ã¹ãéããŠãã©ã¡ãŒã¿ãæŽæ°ããããã®éçšã§ãã¢ãã«ã¯èª€å·®ãæžããããã«å éšãã©ã¡ãŒã¿ (weights and biases) ã調æŽããããã®ãµã€ã¯ã«ãäœåºŠãç¹°ãè¿ãããšã§ãã¢ãã«ã¯åŸã ã«ç²ŸåºŠãé«ããŠããããã®ãã¡ã«ã圢ç¶ãè²ããã¯ã¹ãã£ãªã©ã®è€éãªãã¿ãŒã³ãèªèã§ããããã«ãªãã
ãã®åŠç¿ããã»ã¹ã«ãããã³ã³ãã¥ãŒã¿ããžã§ã³ã¢ãã«ã¯ãç©äœæ€åºãã€ã³ã¹ã¿ã³ã¹åå²ãç»ååé¡ãªã©ãããŸããŸãªã¿ã¹ã¯ãå®è¡ã§ããããã«ãªããæçµçãªç®æšã¯ãå®äžçã®ã¢ããªã±ãŒã·ã§ã³ã§èŠèŠããŒã¿ãæ£ç¢ºã«ç解ã§ããããã«ãæ°ããæªèŠã®ç»åã«åŠç¿ãæ±åã§ããã¢ãã«ãäœæããããšã§ããã
ã¢ãã«ããã¬ãŒãã³ã°ãããšãã«èå°è£ã§äœãèµ·ãã£ãŠããããããã£ããšããã§ãã¢ãã«ããã¬ãŒãã³ã°ãããšãã«èæ ®ãã¹ãç¹ãèŠãŠã¿ããã
倧èŠæš¡ããŒã¿ã»ããã§ã®ãã¬ãŒãã³ã°
倧èŠæš¡ãªããŒã¿ã»ããã䜿ã£ãŠã¢ãã«ãåŠç¿ããå Žåãããã€ãã®ç°ãªãåŽé¢ã«ã€ããŠèããå¿ èŠããããäŸãã°ããããã»ãµã€ãºã®èª¿æŽãGPU ã®å©çšçã®ã³ã³ãããŒã«ããã«ãã¹ã±ãŒã«åŠç¿ã®éžæãªã©ã§ãããããããã®ãªãã·ã§ã³ã«ã€ããŠè©³ãã説æãããã
ããããµã€ãºãšGPU å©çšç
倧èŠæš¡ãªããŒã¿ã»ããã§ã¢ãã«ããã¬ãŒãã³ã°ããå ŽåãGPU ãå¹ççã«å©çšããããšãéµãšãªãããããã»ãµã€ãºã¯éèŠãªèŠçŽ ã§ãããããã¯ãæ©æ¢°åŠç¿ã¢ãã«ã1åã®ãã¬ãŒãã³ã°å埩ã§åŠçããããŒã¿ãµã³ãã«ã®æ°ã§ãã GPU ã§ãµããŒããããŠããæ倧ãããã»ãµã€ãºã䜿çšããã°ããã®æ©èœããã«ã«æŽ»çšããã¢ãã«åŠç¿ã«ãããæéãççž®ããããšãã§ããŸãããã ããGPU ã®ã¡ã¢ãªäžè¶³ã¯é¿ããããã¡ã¢ãªã»ãšã©ãŒãçºçããå Žåã¯ãã¢ãã«ãã¹ã ãŒãºã«ãã¬ãŒãã³ã°ã§ãããŸã§ããããã»ãµã€ãºã段éçã«å°ããããŠãã ããã
YOLO11 ã«é¢ããŠã¯ã次ã®ããã«èšå®ã§ããã batch_size
ãã©ã¡ãŒã¿ã ãã¬ãŒãã³ã°æ§æ ãGPU ã®å®¹éã«åãããããŸã batch=-1
ãèªåçã«æ±ºå®ããŸãã ããããµã€ãº ããããµã€ãºã埮調æŽããããšã§ã ã®ãªãœãŒã¹ãæ倧éã«æŽ»çšãããã¬ãŒãã³ã°ããã»ã¹å
šäœãæ¹åããããšãã§ããŸããããããµã€ãºã埮調æŽããããšã§ãGPU ã®ãªãœãŒã¹ãæ倧éã«æŽ»çšãããã¬ãŒãã³ã°ããã»ã¹å
šäœãæ¹åããããšãã§ããŸãã
ãµãã»ããã»ãã¬ãŒãã³ã°
ãµãã»ãããã¬ãŒãã³ã°ã¯ããã倧ããªããŒã¿ã»ãããè¡šãããå°ããªããŒã¿ã»ããã§ã¢ãã«ããã¬ãŒãã³ã°ããè³¢ãæŠç¥ã§ããç¹ã«åæã®ã¢ãã«éçºãšãã¹ãã«ãããŠãæéãšãªãœãŒã¹ãç¯çŽããããšãã§ããŸããæéããªãå ŽåããŸãã¯æ§ã ãªã¢ãã«æ§æãè©ŠããŠããå Žåããµãã»ãããã¬ãŒãã³ã°ã¯è¯ãéžæè¢ã§ãã
YOLO11 ããµãã»ããã»ãã¬ãŒãã³ã°ãç°¡åã«å®è£
ã§ããã fraction
ãã©ã¡ãŒã¿ã§æå®ããŸãããã®ãã©ã¡ãŒã¿ã§ãåŠç¿ã«äœ¿çšããããŒã¿ã»ããã®å²åãæå®ã§ããŸããäŸãã° fraction=0.1
ã¯ãããŒã¿ã®10%ã§ã¢ãã«ããã¬ãŒãã³ã°ããŸããå®å
šãªããŒã¿ã»ããã䜿ã£ãŠã¢ãã«ããã¬ãŒãã³ã°ããåã«ããã®ãã¯ããã¯ã䜿ã£ãŠã¢ãã«ãçŽ æ©ãå埩ãããã¥ãŒãã³ã°ããããšãã§ããŸãããµãã»ãããã¬ãŒãã³ã°ã¯ãè¿
éãªé²æãšæœåšçãªåé¡ã®æ©æçºèŠã«åœ¹ç«ã¡ãŸãã
ãã«ãã¹ã±ãŒã«ãã¬ãŒãã³ã°
ãã«ãã¹ã±ãŒã«ãã¬ãŒãã³ã°ã¯ãæ§ã ãªãµã€ãºã®ç»åã§ãã¬ãŒãã³ã°ããããšã§ãã¢ãã«ã®æ±åèœåãåäžããããã¯ããã¯ã§ããã¢ãã«ã¯ç°ãªãã¹ã±ãŒã«ãè·é¢ã®ç©äœãæ€åºããããšãåŠç¿ããããããã¹ãã«ãªããŸãã
äŸãã°ãYOLO11 ããã¬ãŒãã³ã°ããå Žåã以äžã®ããã«èšå®ããããšã§ããã«ãã¹ã±ãŒã«ã»ãã¬ãŒãã³ã°ãæå¹ã«ããããšãã§ããã scale
ãã©ã¡ãŒã¿ã䜿çšããŸãããã®ãã©ã¡ãŒã¿ã¯ããã¬ãŒãã³ã°ç»åã®ãµã€ãºãæå®ãããä¿æ°ã§èª¿æŽããç°ãªãè·é¢ã®ãªããžã§ã¯ããã·ãã¥ã¬ãŒãããŸããäŸãã° scale=0.5
ã¯ç»åãµã€ãºãååã«çž®å°ããã scale=2.0
ã¯2åã«ãªããŸãããã®ãã©ã¡ãŒã¿ãèšå®ããããšã§ãã¢ãã«ã¯ããŸããŸãªç»åã¹ã±ãŒã«ãçµéšããããŸããŸãªãªããžã§ã¯ãã®ãµã€ãºãã·ããªãªã«ããã£ãŠæ€åºèœåãåäžãããããšãã§ããŸãã
ãã£ãã·ã³ã°
ãã£ãã·ã³ã°ã¯ãæ©æ¢°åŠç¿ã¢ãã«ã®åŠç¿å¹çãåäžãããããã®éèŠãªãã¯ããã¯ã§ãããååŠçãããç»åãã¡ã¢ãªã«ä¿åããããšã§ãGPU ããã£ã¹ã¯ããããŒã¿ãããŒãããããŸã§ã®åŸ ã¡æéãççž®ããããã¢ãã«ã¯ããã£ã¹ã¯I/Oæäœã«ããé 延ãªãã«ç¶ç¶çã«ããŒã¿ãåãåãããšãã§ããã
YOLO11 ã cache
ãã©ã¡ãŒã¿ãå¿
èŠã ïŒ
cache=True
:ããŒã¿ã»ããã®ç»åãRAMã«ä¿åããæéã®ã¢ã¯ã»ã¹é床ãæäŸãããããã®ä»£åãšããŠã¡ã¢ãªäœ¿çšéãå¢å ãããcache='disk'
:RAMããé ãããæ¯åæ°ããããŒã¿ãèªã¿èŸŒãããéããcache=False
:ãã£ãã·ã¥ãç¡å¹ã«ããå®å šã«ãã£ã¹ã¯I/Oã«äŸåããŸãã
ããã¯ã¹ç²Ÿå¯ãã¬ãŒãã³ã°
æ··å粟床ãã¬ãŒãã³ã°ã§ã¯ã16 ãããïŒFP16ïŒãš 32 ãããïŒFP32ïŒã®æµ®åå°æ°ç¹åã䜿çšããŸããFP16ãšFP32ã®äž¡æ¹ã®é·æã¯ãFP16ã䜿çšããŠèšç®ãé«éåããFP32ã䜿çšããŠå¿ èŠãªéšåã®ç²ŸåºŠãç¶æããããšã§æŽ»çšãããŸãããã¥ãŒã©ã«ã»ãããã¯ãŒã¯ã®ã»ãšãã©ã®æŒç®ã¯ FP16 ã§è¡ãããèšç®ã®é«éåãšã¡ã¢ãªäœ¿çšéã®äœæžãå®çŸããŠããŸããããããã¢ãã«ã®éã¿ã®ãã¹ã¿ãŒã³ããŒã¯FP32ã§ä¿æãããéã¿ã®æŽæ°ã¹ãããäžã®ç²ŸåºŠãä¿èšŒããŸããåãããŒããŠã§ã¢å¶çŽã®äžã§ããã倧ããªã¢ãã«ããã倧ããªããããµã€ãºãæ±ãããšãã§ããŸãã
æ··å粟床åŠç¿ãå®è£ ããã«ã¯ãåŠç¿ã¹ã¯ãªãããä¿®æ£ããããŒããŠã§ã¢ïŒGPUãªã©ïŒãããããµããŒãããŠããããšã確èªããå¿ èŠããããäŸãã° Tensorflowãªã©ãå€ãã®ææ°ã®ãã£ãŒãã©ãŒãã³ã°ã»ãã¬ãŒã ã¯ãŒã¯ã¯ãæ··å粟床ããã«ãã€ã³ã§ãµããŒãããŠããŸãã
æ··å粟床ãã¬ãŒãã³ã°ã¯ãYOLO11 ã䜿ãã°ç°¡åã§ãã amp
ãã©ã°ãèšå®ããŸããèšå® amp=True
ã¯ãèªåæ··å粟床ïŒAMPïŒãã¬ãŒãã³ã°ãå¯èœã«ããŸããæ··å粟床ãã¬ãŒãã³ã°ã¯ãã¢ãã«ãã¬ãŒãã³ã°ããã»ã¹ãæé©åããã·ã³ãã«ã§å¹æçãªæ¹æ³ã§ãã
äºåã«èšç·ŽããããŠã§ã€ã
äºååŠç¿æžã¿ã®éã¿ã䜿çšããããšã¯ãã¢ãã«ã®åŠç¿ããã»ã¹ãã¹ããŒãã¢ããããè³¢ãæ¹æ³ã§ããäºåèšç·Žãããéã¿ã¯ããã§ã«å€§èŠæš¡ãªããŒã¿ã»ããã§èšç·Žãããã¢ãã«ããåŸããããã®ã§ãããã¢ãã«ã«å æãäžããŸãã転移åŠç¿ã¯ãäºåã«èšç·Žãããã¢ãã«ãæ°ããé¢é£ã¿ã¹ã¯ã«é©å¿ãããŸããäºåèšç·Žãããã¢ãã«ã埮調æŽããã«ã¯ããããã®éã¿ããéå§ããç¹å®ã®ããŒã¿ã»ããã§èšç·Žãç¶ããŸãããã®åŠç¿æ¹æ³ã§ã¯ãã¢ãã«ãåºæ¬çãªç¹åŸŽããã£ãããšç解ããç¶æ ã§åŠç¿ãéå§ãããããåŠç¿æéãççž®ãããããã©ãŒãã³ã¹ãåäžããããšããããããŸãã
ã«ã€ã㊠pretrained
ãã©ã¡ãŒã¿ã䜿çšãããšãYOLO11 ã§ç§»ç±åŠç¿ãç°¡åã«è¡ãããšãã§ããŸããèšå® pretrained=True
ã¯ããã©ã«ãã®äºååŠç¿æžã¿éã¿ã䜿çšããŸãããã«ã¹ã¿ã äºååŠç¿æžã¿ã¢ãã«ãžã®ãã¹ãæå®ããããšãã§ããŸããäºååŠç¿æžã¿ã®éã¿ãšè»¢ç§»åŠç¿ã䜿çšããããšã§ãã¢ãã«ã®èœåãå¹æçã«é«ããåŠç¿ã³ã¹ããåæžããããšãã§ããŸãã
倧èŠæš¡ããŒã¿ã»ãããæ±ãéã«èæ ®ãã¹ããã®ä»ã®ãã¯ããã¯
倧èŠæš¡ãªããŒã¿ã»ãããæ±ãéã«èæ ®ãã¹ããã¯ããã¯ã¯ä»ã«ãããã€ãããïŒ
- åŠç¿çã¹ã±ãžã¥ãŒã©ãŒ:åŠç¿çã¹ã±ãžã¥ãŒã©ãå®è£
ããããšã§ãåŠç¿äžã«åŠç¿çãåçã«èª¿æŽããããšãã§ãããåŠç¿çãããŸã調æŽããããšã§ãã¢ãã«ã®ãªãŒããŒã·ã¥ãŒããé²ããå®å®æ§ãåäžãããããšãã§ãããYOLO11 ãåŠç¿ãããšã
lrf
ãã©ã¡ãŒã¿ã¯ãæçµçãªåŠç¿ã¬ãŒããåæã¬ãŒãã®äœåã®äžãã«èšå®ããããšã§ãåŠç¿ã¬ãŒãã®ã¹ã±ãžã¥ãŒãªã³ã°ã管çããã®ã«åœ¹ç«ã€ã - åæ£ãã¬ãŒãã³ã°ïŒå€§èŠæš¡ãªããŒã¿ã»ãããæ±ãå Žåãåæ£ãã¬ãŒãã³ã°ã¯å€§ããªå€åãããããããã¬ãŒãã³ã°ã®ã¯ãŒã¯ããŒããè€æ°ã®GPUããã·ã³ã«åæ£ããããšã§ããã¬ãŒãã³ã°æéãççž®ã§ããŸãã
ãã¬ãŒãã³ã°ãããšããã¯ã®æ°
ã¢ãã«ããã¬ãŒãã³ã°ããéããšããã¯ãšã¯ãã¬ãŒãã³ã°ããŒã¿ã»ããå šäœã1åå®å šã«ééããããšãæãããšããã¯ã®éãã¢ãã«ã¯ãã¬ãŒãã³ã°ã»ããã®åäŸã1åãã€åŠçããåŠç¿ã¢ã«ãŽãªãºã ã«åºã¥ããŠãã©ã¡ãŒã¿ãæŽæ°ããŸããã¢ãã«ãæéããããŠåŠç¿ãããã©ã¡ãŒã¿ãæ¹è¯ããããã«ã¯ãéåžžãè€æ°ã®ãšããã¯ãå¿ èŠã§ããã
ãããã質åã¯ãã¢ãã«ãèšç·Žãããšããã¯æ°ãã©ã®ããã«æ±ºå®ããããšããããšã§ããè¯ãã¹ã¿ãŒããã€ã³ãã¯300ãšããã¯ã§ããã¢ãã«ãæ©æã«ãªãŒããŒãã£ããããå Žåã¯ããšããã¯æ°ãæžããããšãã§ããŸãã300ãšããã¯ãè¶ ããŠããªãŒããŒãã£ãããçºçããªãå Žåã¯ã600ã1200ããŸãã¯ãã以äžã®ãšããã¯ãŸã§åŠç¿ã延é·ããããšãã§ããŸãã
ããããçæ³çãªãšããã¯æ°ã¯ãããŒã¿ã»ããã®ãµã€ãºããããžã§ã¯ãã®ç®æšã«ãã£ãŠç°ãªããŸãããã倧ããªããŒã¿ã»ããã§ã¯ãã¢ãã«ãå¹æçã«åŠç¿ãããããã«ãããå€ãã®ãšããã¯ãå¿
èŠã«ãªããããããŸããããããå°ããªããŒã¿ã»ããã§ã¯ããªãŒããŒãã£ããã£ã³ã°ãé¿ããããã«ãããå°ãªããšããã¯ãå¿
èŠã«ãªããããããŸãããYOLO11 ã«é¢ããŠã¯ epochs
ãã©ã¡ãŒã¿ã䜿çšããŸãã
ã¢ãŒãªãŒã¹ãããã³ã°
æ©æåæ¢ã¯ãã¢ãã«ã®ãã¬ãŒãã³ã°ãæé©åããããã®è²Žéãªãã¯ããã¯ã§ããæ€èšŒã®ããã©ãŒãã³ã¹ãç£èŠããããšã§ãã¢ãã«ã®æ¹åãæ¢ãŸã£ãããã¬ãŒãã³ã°ãåæ¢ããããšãã§ããŸããèšç®ãªãœãŒã¹ãç¯çŽãããªãŒããŒãã£ããã£ã³ã°ãé²ãããšãã§ããŸãã
ãã®ããã»ã¹ã§ã¯ããã¬ãŒãã³ã°ãåæ¢ããåã«ãæ€èšŒã¡ããªã¯ã¹ã®æ¹åãäœãšããã¯åŸ ã€ãã決å®ããå¿èãã©ã¡ãŒã¿ãèšå®ããããã®ãšããã¯æ°å ã«ã¢ãã«ã®ããã©ãŒãã³ã¹ãåäžããªãå ŽåãæéãšãªãœãŒã¹ã®æµªè²»ãé¿ããããã«ãã¬ãŒãã³ã°ãåæ¢ãããŸãã
YOLO11 ã®å Žåããã¬ãŒãã³ã°èšå®ã§patienceãã©ã¡ãŒã¿ãèšå®ããããšã§ãæ©æåæ¢ãæå¹ã«ããããšãã§ããŸããäŸãã° patience=5
ã€ãŸãã5ãšããã¯é£ç¶ã§æ€èšŒã¡ããªã¯ã¹ã®æ¹åãèŠãããªãå Žåããã¬ãŒãã³ã°ã¯åæ¢ããããã®æ¹æ³ã䜿çšããããšã§ãåŠç¿ããã»ã¹ãå¹ççã«ç¶æããéå°ãªèšç®ãè¡ãããšãªãæé©ãªããã©ãŒãã³ã¹ãéæããããšãã§ããŸãã
ã¯ã©ãŠããšããŒã«ã«ãã¬ãŒãã³ã°ã®éžæ
ã¢ãã«ã®ãã¬ãŒãã³ã°ã«ã¯ãã¯ã©ãŠããã¬ãŒãã³ã°ãšããŒã«ã«ãã¬ãŒãã³ã°ã®2ã€ã®ãªãã·ã§ã³ãããã
ã¯ã©ãŠããã¬ãŒãã³ã°ã¯ã¹ã±ãŒã©ããªãã£ãšåŒ·åãªããŒããŠã§ã¢ãæäŸãã倧èŠæš¡ãªããŒã¿ã»ãããè€éãªã¢ãã«ãæ±ãã®ã«çæ³çã§ããGoogle CloudãAWSãAzureã®ãããªãã©ãããã©ãŒã ã¯ãé«æ§èœGPUãTPUãžã®ãªã³ããã³ãã¢ã¯ã»ã¹ãæäŸãããã¬ãŒãã³ã°æéãççž®ãããã倧ããªã¢ãã«ã®å®éšãå¯èœã«ãããããããã¯ã©ãŠãã»ãã¬ãŒãã³ã°ã¯ãç¹ã«é·æéã®å Žåãã³ã¹ããé«ããªãå¯èœæ§ããããããŒã¿è»¢éã¯ã³ã¹ããšã¬ã€ãã³ã·ãŒãå¢å ãããå¯èœæ§ãããã
ããŒã«ã«ã»ãã¬ãŒãã³ã°ã¯ããã倧ããªã³ã³ãããŒã«ãšã«ã¹ã¿ãã€ãºãå¯èœã§ãç¹å®ã®ããŒãºã«åãããŠç°å¢ãã«ã¹ã¿ãã€ãºããç¶ç¶çãªã¯ã©ãŠãã³ã¹ããåé¿ããããšãã§ããŸããé·æçãªãããžã§ã¯ãã§ã¯ããçµæžçã§ãããŒã¿ã¯ãªã³ãã¬ãã¹ã«ä¿åããããããããå®å šã§ããããããããŒã«ã«ã®ããŒããŠã§ã¢ã«ã¯ãªãœãŒã¹ã®å¶éããããã¡ã³ããã³ã¹ãå¿ èŠãªå ŽåããããŸãã
ãªããã£ãã€ã¶ãŒã®éžæ
ãªããã£ãã€ã¶ã¯ãã¢ãã«ã®æ§èœã枬å®ããæ倱é¢æ°ãæå°åããããã«ã ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®éã¿ã調æŽããã¢ã«ãŽãªãºã ã§ãããã£ãšç°¡åã«èšãã°ããªããã£ãã€ã¶ã¯èª€å·®ãæžããããã«ãã©ã¡ãŒã¿ã調æŽããããšã§ã ã¢ãã«ã®åŠç¿ãæ¯æŽããŸããé©åãªãªããã£ãã€ã¶ãéžæããããšã¯ãã¢ãã«ã®åŠç¿ã®éããšæ£ç¢ºãã«çŽæ¥åœ±é¿ããŸãã
ãŸããã¢ãã«ã®ããã©ãŒãã³ã¹ãåäžãããããã«ããªããã£ãã€ã¶ã»ãã©ã¡ãŒã¿ã埮調æŽããããšãã§ããŸããåŠç¿çã調æŽããããšã§ããã©ã¡ãŒã¿ãæŽæ°ããéã®ã¹ãããã®å€§ãããèšå®ããŸããå®å®æ§ãä¿ã€ããã«ãæåã¯äžçšåºŠã®åŠç¿çã«èšå®ããé·æçãªåŠç¿ãæ¹åããããã«ãæéã®çµéãšãšãã«åŸã ã«äœäžãããããšãã§ããŸããããã«ãã¢ã¡ã³ã¿ã ãèšå®ããããšã§ãéå»ã®æŽæ°ãçŸåšã®æŽæ°ã«äžãã圱é¿åºŠã決å®ããŸããã¢ã¡ã³ã¿ã ã®äžè¬çãªå€ã¯0.9çšåºŠã§ãããããã¯äžè¬çã«è¯ããã©ã³ã¹ãæäŸããã
äžè¬çãªãªããã£ãã€ã¶ãŒ
ããŸããŸãªãªããã£ãã€ã¶ãŒã«ã¯ãããŸããŸãªé·æãšçæããããŸããäžè¬çãªãªããã£ãã€ã¶ãŒãããã€ãèŠãŠã¿ããã
-
SGDïŒç¢ºççåŸé éäžæ³ïŒïŒ
- ãã©ã¡ãŒã¿ã«å¯Ÿããæ倱é¢æ°ã®åŸé ã䜿çšããŠã¢ãã«ãã©ã¡ãŒã¿ãæŽæ°ããã
- ã·ã³ãã«ã§å¹ççã ããåæã«æéãããããããŒã«ã«ã»ãããã ã«ã¯ãŸã蟌ãå¯èœæ§ãããã
-
ã¢ãã ïŒé©å¿ã¢ãŒã¡ã³ãæšå®ïŒïŒ
- ã¢ã¡ã³ã¿ã ä»ãSGDãšRMSPropã®äž¡æ¹ã®å©ç¹ãå ŒãåããŠããã
- åŸé ã®1次ã¢ãŒã¡ã³ããš2次ã¢ãŒã¡ã³ãã®æšå®å€ã«åºã¥ããŠãåãã©ã¡ãŒã¿ã®åŠç¿çã調æŽããã
- ãã€ãºã®å€ãããŒã¿ãçãªåŸé ã«é©ããŠããã
- å¹ççã§ãäžè¬çã«ãã¥ãŒãã³ã°ãå°ãªããŠæžãã®ã§ãYOLO11 ã«ãå§ãã®ãªããã£ãã€ã¶ã§ãã
-
RMSProp (äºä¹å¹³åå¹³æ¹æ ¹äŒæ)ïŒ
- åŸé ãæè¿ã®åŸé ã®å€§ããã®å®è¡å¹³åã§å²ãããšã«ãã£ãŠãåãã©ã¡ãŒã¿ã®åŠç¿çã調æŽããã
- æ¶å€±åŸé åé¡ã®åŠçã«åœ¹ç«ã¡ããªã«ã¬ã³ãã»ãã¥ãŒã©ã«ã»ãããã¯ãŒã¯ã«æå¹ã§ããã
YOLO11 optimizer
ãã©ã¡ãŒã¿ã§ã¯ãSGDãAdamãAdamWãNAdamãRAdamãRMSProp ãªã©ã®ããŸããŸãªãªããã£ãã€ã¶ããéžæã§ããŸãã auto
ã¢ãã«æ§æã«åºã¥ãèªåéžæã®ããã
å°å瀟äŒãšã®ã€ãªãã
ã³ã³ãã¥ãŒã¿ã»ããžã§ã³æ奜家ã®ã³ãã¥ããã£ã«åå ããããšã§ãåé¡ã解決ããããéãåŠã¶ããšãã§ããŸããããã§ã¯ãã€ãªãããå©ããåŸãã¢ã€ãã¢ãå ±æããæ¹æ³ãããã€ã玹ä»ããŸãã
å°åè³æº
- GitHub IssuesïŒ YOLO11 GitHub ãªããžããªã«ã¢ã¯ã»ã¹ããIssues ã¿ãã§è³ªåããã°å ±åãæ°æ©èœã®ææ¡ãè¡ã£ãŠãã ãããã³ãã¥ããã£ãšã¡ã³ãããŒã¯ãšãŠã掻çºã§ãæå©ãããæºåãã§ããŠããŸãã
- Ultralytics DiscordãµãŒããŒïŒ Ultralytics Discord ãµãŒããŒã«åå ããŠãä»ã®ãŠãŒã¶ãŒãéçºè ãšãã£ãããããããµããŒããåããããçµéšãå ±æããŸãããã
å ¬åŒææž
- Ultralytics YOLO11 ããã¥ã¡ã³ãæ§ã ãªã³ã³ãã¥ãŒã¿ããžã§ã³ãããžã§ã¯ãã«é¢ãã詳现ãªã¬ã€ãã圹ç«ã€ãã³ãã«ã€ããŠã¯ã YOLO11 ã®å ¬åŒããã¥ã¡ã³ããã芧ãã ããã
ãããã®ãªãœãŒã¹ã䜿çšããããšã§ã課é¡ã解決ããã³ã³ãã¥ãŒã¿ããžã§ã³ã³ãã¥ããã£ã«ãããææ°ã®ãã¬ã³ãããã©ã¯ãã£ã¹ãåžžã«ææ¡ããããšãã§ããŸãã
èŠç¹
ã³ã³ãã¥ãŒã¿ããžã§ã³ã¢ãã«ã®ãã¬ãŒãã³ã°ã«ã¯ãåªãããã©ã¯ãã£ã¹ã«åŸãããšãæŠç¥ãæé©åããããšãåé¡ãçºçãããšãã«è§£æ±ºããããšãå«ãŸããŸããããããµã€ãºã®èª¿æŽãæ··å粟床ãã¬ãŒãã³ã°ãäºåãã¬ãŒãã³ã°æžã¿ã®éã¿ããã®éå§ãªã©ã®ãã¯ããã¯ã¯ãã¢ãã«ãããè¯ãåäœãããããéããã¬ãŒãã³ã°ããããšãã§ããŸãããµãã»ããåŠç¿ãæ©æåæ¢ãªã©ã®ææ³ã¯ãæéãšãªãœãŒã¹ã®ç¯çŽã«åœ¹ç«ã¡ãŸããã³ãã¥ããã£ãšã®ã€ãªãããç¶æããæ°ãããã¬ã³ãã«ã€ããŠããããšã¯ãã¢ãã«ãã¬ãŒãã³ã°ã®ã¹ãã«ãåäžããç¶ããã®ã«åœ¹ç«ã¡ãŸãã
ããããã質å
Ultralytics YOLO ã§å€§èŠæš¡ãªããŒã¿ã»ããããã¬ãŒãã³ã°ããéãGPU ã®å©çšçãåäžãããã«ã¯ïŒ
GPU ã batch_size
ãã©ã¡ãŒã¿ããGPU ããµããŒãããæ倧ãµã€ãºã«èšå®ããŸããããã«ãããGPU ã®æ©èœããã«ã«æŽ»çšãããã¬ãŒãã³ã°æéãççž®ããããšãã§ããŸããã¡ã¢ãªãšã©ãŒãçºçããå Žåã¯ããã¬ãŒãã³ã°ãã¹ã ãŒãºã«å®è¡ããããŸã§ãããããµã€ãºãåŸã
ã«å°ããããŠãã ãããYOLO11 ã®å Žå batch=-1
ããã¬ãŒãã³ã°ã¹ã¯ãªããã«è¿œå ãããšãå¹ççãªåŠçã®ããã®æé©ãªããããµã€ãºãèªåçã«æ±ºå®ãããŸãã詳现ã«ã€ããŠã¯ ãã¬ãŒãã³ã°æ§æ.
ããã¯ã¹ãã»ãã¬ã·ãžã§ã³ã»ãã¬ãŒãã³ã°ãšã¯äœã§ããïŒãŸããYOLO11 ã
æ··å粟床ãã¬ãŒãã³ã°ã¯ãèšç®é床ãšç²ŸåºŠã®ãã©ã³ã¹ããšãããã«ã16ãããïŒFP16ïŒãš32ãããïŒFP32ïŒã®äž¡æ¹ã®æµ®åå°æ°ç¹åãå©çšããŸãããã®ã¢ãããŒãã¯ãã¢ãã«ãç ç²ã«ããããšãªãããã¬ãŒãã³ã°ãé«éåããã¡ã¢ãªäœ¿çšéãåæžããŸãã 粟床.YOLO11 ã§æ··å粟床ãã¬ãŒãã³ã°ãæå¹ã«ããã«ã¯ã次ã®ããã«èšå®ããŸãã amp
ãã©ã¡ãŒã¿ã True
ããã¬ãŒãã³ã°èšå®ã«è¿œå ããŸããããã«ãããAutomatic Mixed PrecisionïŒAMPïŒãã¬ãŒãã³ã°ãæå¹ã«ãªããŸãããã®æé©åææ³ã®è©³çŽ°ã«ã€ããŠã¯ ãã¬ãŒãã³ã°æ§æ.
ãã«ãã¹ã±ãŒã«ãã¬ãŒãã³ã°ã¯ãYOLO11 ã¢ãã«ã®æ§èœãã©ã®ããã«åäžãããã®ãïŒ
ãã«ãã¹ã±ãŒã«ãã¬ãŒãã³ã°ã¯ãæ§ã
ãªãµã€ãºã®ç»åã§ãã¬ãŒãã³ã°ãè¡ãããšã§ãã¢ãã«ã®ããã©ãŒãã³ã¹ãåäžãããã¢ãã«ãæ§ã
ãªã¹ã±ãŒã«ãè·é¢ã§ããè¯ãäžè¬åã§ããããã«ããŸããYOLO11 㧠scale
ãã©ã¡ãŒã¿ã䜿çšãããäŸãã° scale=0.5
ã¯ç»åãµã€ãºãååã«çž®å°ããã scale=2.0
åã«ããããã®ãã¯ããã¯ã¯ãããŸããŸãªè·é¢ã®ãªããžã§ã¯ããã·ãã¥ã¬ãŒãããã¢ãã«ãããŸããŸãªã·ããªãªã«å¯ŸããŠããããã¹ãã«ãããèšå®ã詳现ã«ã€ããŠã¯ ãã¬ãŒãã³ã°æ§æ.
YOLO11 ãäºååŠç¿ãããéã¿ã䜿çšããŠãã¬ãŒãã³ã°ãé«éåããã«ã¯ã©ãããã°ããã§ããïŒ
äºåã«ãã¬ãŒãã³ã°ãããéã¿ã䜿çšããããšã§ããã¬ãŒãã³ã°æéã倧å¹
ã«ççž®ããåºæ¬çãªç¹åŸŽããã§ã«ç解ããŠããã¢ãã«ããå§ããããšã§ãã¢ãã«ã®ããã©ãŒãã³ã¹ãåäžãããããšãã§ããŸããYOLO11 ã§ã¯ pretrained
ãã©ã¡ãŒã¿ã True
ãŸãã¯ããã¬ãŒãã³ã°èšå®ã«ã«ã¹ã¿ã äºåãã¬ãŒãã³ã°éã¿ãžã®ãã¹ãæå®ããŸãã転移åŠç¿ãšããŠç¥ããããã®ã¢ãããŒãã§ã¯ã倧èŠæš¡ãªããŒã¿ã»ããããã®ç¥èã掻çšããŠãç¹å®ã®ã¿ã¹ã¯ã«é©å¿ãããŸããäºååŠç¿ãããéã¿ãšãã®å©ç¹ã«ã€ããŠãã£ãšç¥ã ãã.
ã¢ãã«ã®ãã¬ãŒãã³ã°ã«æšå¥šããããšããã¯æ°ãšãYOLO11 ã§ãããèšå®ããæ¹æ³ã¯ïŒ
ãšããã¯æ°ãšã¯ãã¢ãã«åŠç¿äžã«åŠç¿ããŒã¿ã»ãããå®å
šã«ééããåæ°ã®ããšã§ãããå
žåçãªéå§ç¹ã¯300ãšããã¯ã§ããã¢ãã«ãæ©æã«ãªãŒããŒãã£ããããå Žåã¯ããã®æ°ãæžããããšãã§ããŸãããŸãããªãŒããŒãã£ããã£ã³ã°ãèŠãããªãå Žåã¯ã600ã1200ããŸãã¯ãã以äžã®ãšããã¯æ°ãŸã§åŠç¿ã延é·ããããšãã§ããŸãããã®èšå®ã¯YOLO11 㧠epochs
ãã©ã¡ãŒã¿ã䜿çšããŸããçæ³çãªãšããã¯æ°ã®æ±ºå®ã«é¢ãããã®ä»ã®ã¢ããã€ã¹ã«ã€ããŠã¯ã以äžã®ã»ã¯ã·ã§ã³ãåç
§ããŠãã ããã ãšããã¯æ°.