ç·åã¬ã€ãUltralytics YOLOv5
Ultralytics' ãžãããã YOLOv5ð ããã¥ã¡ã³ããŒã·ã§ã³ïŒYOLOv5é©åœç㪠"You Only Look Once "ã®ç¬¬5匟ã ãªããžã§ã¯ãæ€åº ã¢ãã«ã¯ãé«éãã€é«ç²ŸåºŠã®çµæããªã¢ã«ã¿ã€ã ã§æäŸããããã«èšèšãããŠããã
PyTorch ããã®åŒ·å㪠ãã£ãŒãã©ãŒãã³ã° ãã¬ãŒã ã¯ãŒã¯ã¯ããã®å€çšéæ§ã䜿ãããããé«ãããã©ãŒãã³ã¹ã§çµ¶å€§ãªäººæ°ãéããŠããŸããç§ãã¡ã®ããã¥ã¡ã³ãã¯ãã€ã³ã¹ããŒã«ããã»ã¹ãã¬ã€ãããã¢ãã«ã®ã¢ãŒããã¯ãã£ã®ãã¥ã¢ã³ã¹ã説æããæ§ã ãªäœ¿çšäŸã玹ä»ããäžé£ã®è©³çŽ°ãªãã¥ãŒããªã¢ã«ãæäŸããŸãããããã®ãªãœãŒã¹ã¯ãYOLOv5 ã®å¯èœæ§ãæ倧éã«æŽ»çšããã®ã«åœ¹ç«ã¡ãŸãã ã³ã³ãã¥ãŒã¿ããžã§ã³ ãããžã§ã¯ãå§ãããïŒ
æ¢çŽ¢ãšåŠç¿
ããã§ã¯ãYOLOv5 ã®ããŸããŸãªåŽé¢ãã¬ã€ãããå æ¬çãªãã¥ãŒããªã¢ã«ã®ã³ã³ãã€ã«ã§ãã
- ã«ã¹ã¿ã ããŒã¿ããã¬ãŒãã³ã°ããð RECOMMENDED: ã«ã¹ã¿ã ããŒã¿ã»ããã§YOLOv5 ã¢ãã«ããã¬ãŒãã³ã°ããæ¹æ³ãåŠã¶ã
- æé«ã®ãã¬ãŒãã³ã°çµæãåŸãããã®ãã³ãâïžïŒã¢ãã«ãã¬ãŒãã³ã°ããã»ã¹ãæé©åããããã®å®è·µçãªãã³ããã芧ãã ããã
- ãã«ãGPU ãã¬ãŒãã³ã°ïŒè€æ°ã®GPUã掻çšããŠãã¬ãŒãã³ã°ãè¿ éåããæ¹æ³ãç解ããã
- PyTorch ããð NEW:PyTorch ãããä»ããŠäºåã«ãã¬ãŒãã³ã°ãããã¢ãã«ãããŒãããããšãåŠã¶ã
- TFLite,ONNX,CoreML,TensorRT ExportðïŒã¢ãã«ãæ§ã ãªãã©ãŒãããã«ãšã¯ã¹ããŒãããæ¹æ³ãç解ããŸãã
- ãã¹ãã¿ã€ã ã»ãªãŒã°ã¡ã³ããŒã·ã§ã³ïŒTTAïŒïŒTTAã䜿çšããŠã¢ãã«ã®äºæž¬ç²ŸåºŠãåäžãããæ¹æ³ãæ¢ããŸãã
- ã¢ãã«ã®ã¢ã³ãµã³ãã«ããã©ãŒãã³ã¹ãåäžãããããã«è€æ°ã®ã¢ãã«ãçµã¿åãããæŠç¥ãåŠã³ãŸãã
- ã¢ãã«ã®æåã/ã¹ããŒã¹ãã«ãŒãã³ã°ãšã¹ããŒã¹æ§ã®æŠå¿µãç解ããããå¹ççãªã¢ãã«ãäœæããæ¹æ³ãç解ããŸãã
- ãã€ããŒãã©ã¡ãŒã¿ã®é²åã¢ãã«æ§èœãåäžãããããã®èªååããããã€ããŒãã©ã¡ãŒã¿ã»ãã¥ãŒãã³ã°ã®ããã»ã¹ãã芧ãã ããã
- åçµã¬ã€ã€ãŒã«ãã転移åŠç¿ïŒYOLOv5 ã§ã¬ã€ã€ãŒãåçµããŠäŒéåŠç¿ãå®è£ ããæ¹æ³ãåŠã¶ã
- ã¢ãŒããã¯ãã£ãŒã»ãµããªãŒðYOLOv5 ã¢ãã«ã®æ§é çãªè©³çŽ°ãæãäžããã
- Roboflow ããŒã¿ã»ããã®ããã«ïŒRoboflow ãããŒã¿ã»ããã®ç®¡çãã©ããªã³ã°ãã¢ã¯ãã£ãã©ãŒãã³ã°ã«æŽ»çšããæ¹æ³ãç解ããã
- ClearML ãã®ã³ã°ðClearML ãçµ±åããŠã¢ãã«ãã¬ãŒãã³ã°äžã«å¹ççã«ãã®ã³ã°ããæ¹æ³ãåŠã³ãŸãã
- YOLOv5 withNeural Magic Neural Magic ã® Deepsparse ã䜿ã£ãŠãYOLOv5 ã¢ãã«ãåã蟌ã¿ãéååããæ¹æ³ãçºèŠããã
- Comet ãã®ã³ã°ð NEW:Comet ã掻çšããŠã¢ãã«ãã¬ãŒãã³ã°ã®ãã®ã³ã°ãæ¹åããæ¹æ³ãæ¢ããŸãã
察å¿ç°å¢
Ultralytics ãã¯ãããšããå¿ èŠäžå¯æ¬ ãªäŸåé¢ä¿ãããªã€ã³ã¹ããŒã«ããããããŸããŸãªããã«äœ¿ããç°å¢ãæäŸããã CUDAãCUDNNã Pythonãã㊠PyTorchãªã©ãããªã€ã³ã¹ããŒã«ãããŠããŸãã
- ç¡æGPU ããŒã:
- Google ã¯ã©ãŠã GCPã¯ã€ãã¯ã¹ã¿ãŒãã¬ã€ã
- ã¢ããŸã³ AWSã¯ã€ãã¯ã¹ã¿ãŒãã¬ã€ã
- AzureïŒAzureML ã¯ã€ãã¯ã¹ã¿ãŒãã¬ã€ã
- ããã«ãŒ: Dockerã¯ã€ãã¯ã¹ã¿ãŒãã¬ã€ã
ãããžã§ã¯ãç¶æ³
ãã®ãããžã¯ãYOLOv5 GitHub ActionsContinuous Integration (CI) ãã¹ãããã¹ãŠæ£åžžã«ãã¹ããŠããããšã瀺ããŸãããããã®CIãã¹ãã¯ããã¬ãŒãã³ã°ãæ€èšŒãæšè«ããšã¯ã¹ããŒãããã³ãããŒã¯ãšãã£ãããŸããŸãªéèŠãªåŽé¢ã«ããã£ãŠãYOLOv5 ã®æ©èœãšããã©ãŒãã³ã¹ãå³å¯ã«ãã§ãã¯ããŸãããããã®ãã¹ãã¯ãmacOSãWindowsãUbuntuäžã§ã24æéããšããã³æ°ããã³ãããããšã«å®æœãããäžè²«ããä¿¡é Œæ§ã®é«ãåäœãä¿èšŒããŸãã
ã€ãªãããè²¢ç®ãã
YOLOv5 ã§ã®ããªãã®æ ã¯ãå€ç¬ãªãã®ã§ããå¿ èŠã¯ãããŸãããGitHubã®æŽ»æ°ããã³ãã¥ããã£ã«åå ããLinkedInã§ãããã§ãã·ã§ãã«ãšã€ãªãããTwitterã§çµæãå ±æããYouTubeã§æè²ãªãœãŒã¹ãèŠã€ããŸããããTikTokãš BiliBiliã§ç§ãã¡ããã©ããŒããŠãããé åçãªã³ã³ãã³ããã楜ãã¿ãã ããã
è²¢ç®ããããšã«èå³ããããŸããïŒã³ãŒãã®æ¹åããã°å ±åããããã¥ã¡ã³ãã®æŽæ°ãŸã§ããããã圢ã§ã®è²¢ç®ãæè¿ããŸãã詳ããã¯ã³ã³ããªãã¥ãŒãã»ã¬ã€ãã©ã€ã³ãã芧ãã ããã
ç§ãã¡ã¯ãçãããYOLOv5 ãã©ã®ããã«äœ¿ããããã®é©æ°çãªæ¹æ³ã楜ãã¿ã«ããŠããŸããé£ã³èŸŒãã§ãå®éšããŠãããªãã®ã³ã³ãã¥ãŒã¿ããžã§ã³ãããžã§ã¯ãã«é©åœãèµ·ãããŸãããïŒð
ããããã質å
Ultralytics YOLOv5 ã®äž»ãªç¹åŸŽã¯ïŒ
Ultralytics YOLOv5 ã¯ããã®é«éãã€é«ç²ŸåºŠãªç©äœæ€åºèœåã§ç¥ãããŠãããã§æ§ç¯ãããŠããŸãã PyTorchã§æ§ç¯ãããŠãããæ±çšæ§ãé«ããŠãŒã¶ãŒãã¬ã³ããªãŒã§ãããããæ§ã ãªã³ã³ãã¥ãŒã¿ããžã§ã³ãããžã§ã¯ãã«é©ããŠããŸããäž»ãªæ©èœã«ã¯ããªã¢ã«ã¿ã€ã æšè«ãTTAïŒTest-Time AugmentationïŒãã¢ãã«ã¢ã³ãµã³ãã«ïŒModel EnsemblingïŒãªã©ã®è€æ°ã®ãã¬ãŒãã³ã°ããªãã¯ã®ãµããŒããTFLiteãONNX ãCoreML ãTensorRT ãªã©ã®ãšã¯ã¹ããŒããã©ãŒããããšã®äºææ§ãªã©ããããŸããUltralytics YOLOv5 ãã©ã®ããã«ãããžã§ã¯ããåäžããããã«ã€ããŠã¯ãTFLiteãONNX ã CoreML ã TensorRT ãšã¯ã¹ããŒãã¬ã€ããã芧ãã ããã
èªåã®ããŒã¿ã»ããã§ã«ã¹ã¿ã ã¢ãã«ïŒYOLOv5 ïŒããã¬ãŒãã³ã°ããã«ã¯ïŒ
ããªãã®ããŒã¿ã»ããã§ã«ã¹ã¿ã ã»ã¢ãã«ïŒYOLOv5 ïŒããã¬ãŒãã³ã°ããã«ã¯ãããã€ãã®éèŠãªã¹ããããããããŸããã©ãã«ã§æ³šéãä»ããããŒã¿ã»ãããå¿
èŠãªåœ¢åŒã§æºåããã次ã«ãYOLOv5 ã®åŠç¿ãã©ã¡ãŒã¿ãèšå®ãã ã® train.py
ã¹ã¯ãªããã䜿çšããŸãããã®ããã»ã¹ã®è©³çŽ°ãªãã¥ãŒããªã¢ã«ã«ã€ããŠã¯ãç§ãã¡ã® ã«ã¹ã¿ã ããŒã¿ã»ã¬ã€ã.ç¹å®ã®ãŠãŒã¹ã±ãŒã¹ã«æé©ãªçµæãåŸããããããã¹ãããã»ãã€ã»ã¹ãããã§èª¬æããŸãã
ãªãRCNNã®ãããªä»ã®ç©äœæ€åºã¢ãã«ã§ã¯ãªããUltralytics YOLOv5 ã
Ultralytics YOLOv5 YOLOv5 ã¯ãç»åå šäœãäžåºŠã«åŠçãããããè€æ°åã®ãã¹ãå¿ èŠãª RCNN ã®é åããŒã¹ã®ã¢ãããŒããšæ¯èŒããŠå€§å¹ ã«é«éã§ããããã«ã ã®æ§ã ãªãšã¯ã¹ããŒããã©ãŒããããšã®ã·ãŒã ã¬ã¹ãªçµ±åãšè±å¯ãªããã¥ã¡ã³ãã«ãããåå¿è ã«ããããã§ãã·ã§ãã«ã«ãæé©ãªéžæè¢ãšãªã£ãŠããŸããã¢ãŒããã¯ãã£ã®å©ç¹ã«ã€ããŠã¯ãYOLOv5ã¢ãŒããã¯ãã£ã®æŠèŠãã芧ãã ããã
ãã¬ãŒãã³ã°äžã«YOLOv5 ã¢ãã«ã®ããã©ãŒãã³ã¹ãæé©åããã«ã¯ïŒ
YOLOv5 ã¢ãã«ã®ããã©ãŒãã³ã¹ãæé©åããã«ã¯ãæ§ã ãªãã€ããŒãã©ã¡ãŒã¿ããã¥ãŒãã³ã°ããããŒã¿å¢åŒ·ã転移åŠç¿ã®ãããªãã¯ããã¯ãåãå ¥ããå¿ èŠããããŸããUltralytics ã¯ãã¢ãã«ã®å¹çãæ¹åããããã®ãã€ããŒãã©ã¡ãŒã¿ã®é²åãšãã«ãŒãã³ã°/ã¹ããŒã¹ã«é¢ããå æ¬çãªãªãœãŒã¹ãæäŸããŸãããã¬ãŒãã³ã°äžã«æé©ãªããã©ãŒãã³ã¹ãéæããããã®å®çšçãªæŽå¯ãæäŸããTips for Best Training Resultsã¬ã€ãã§ãå®çšçãªãã³ããèŠã€ããããšãã§ããŸãã
YOLOv5 ãã©ã®ãããªç°å¢ã§ã¢ããªã±ãŒã·ã§ã³ãå®è¡ã§ããŸããïŒ
Ultralytics YOLOv5 ã¯ãGradientãGoogle ColabãKaggleäžã®ç¡æã®GPU ããŒãããã¯ããGoogle CloudãAmazon AWSãAzureãªã©ã®äž»èŠãªã¯ã©ãŠããã©ãããã©ãŒã ãªã©ãããŸããŸãªç°å¢ããµããŒãããŠããŸããã»ããã¢ããã«äŸ¿å©ãªDockerã€ã¡ãŒãžãçšæãããŠããŸãããããã®ç°å¢ã®ã»ããã¢ããã«é¢ãã詳现ãªã¬ã€ãã«ã€ããŠã¯ãåãã©ãããã©ãŒã ã®ã¹ããããã€ã¹ãããã®æé ãå«ãããµããŒããããŠããç°å¢ã®ã»ã¯ã·ã§ã³ãã芧ãã ããã