深層å¦ç¿ãç¨ããç»åèªèåéã§ã¯æ§ã
ãªCNNã®ãããã¯ã¼ã¯æ§é ãææ¡ããã¦ããï¼ImageNetãã¼ã¿ã»ããçãç¨ããäºæ¸¬ç²¾åº¦æ¯è¼ãåºãè¡ããã¦ãã¾ãï¼
ãããã©ã®CNNã¢ãã«ã使ãã¹ããªãã ããâ¦ã¨èãã¦ãã¾ãããï¼æè¿ãããªtweetãç®ã«å
¥ãã¾ããï¼
代表çãªCNNã¢ã¼ããã¯ãã£ã«ã¤ãã¦ããã©ã¡ã¼ã¿æ°ãå®éã®æ¼ç®éã»æéãã¨ãã«ã®ã¼æ¶è²»ãªã©ã精度ã¨ã®é¢ä¿ã¨ã¨ãã«æ¤è¨¼ãç´ç²ã«èªã¿ç©ã¨ãã¦é¢ç½ãã¦ããããèªããããèæè¦çã«NWæ§é ã¨è¨ç®éã®é¢ä¿ãæ´ãã§ããã®ã¯ä»ããµãã¤ãã«è¡ã ã㪠https://t.co/6ZAEkWXKtu
— æ°´ç£å (@sakanazensen) 2017å¹´2æ27æ¥
äºæ¸¬ç²¾åº¦ã«ããæ¯è¼ã¯åºãè¡ããã¦ãã¾ããï¼è¨ç®é度ãã¡ã¢ãªä½¿ç¨éã«ããæ¯è¼ã¯è¦ããã¨ãç¡ãã£ãã®ã§ï¼tweetã§ç´¹ä»ããã¦ããè«æ"An Analysis of Deep Neural Network Models for Practical Applications"ãèªãã§ã¿ã¾ããï¼
æ¯è¼å¯¾è±¡ã®CNNã¢ãã«
ãã®è«æã§ã¯ä»¥ä¸ã®CNNã¢ãã«ãæ¯è¼ãã¦ãã¾ãï¼
- AlexNet
- BN-AlexNet (batch normalized AlexNet)
- BN-NIN (batch normalized Network In Network)
- ENet
- GoogLeNet
- ResNet-18, 34, 50, 101, 152
- VGG-16, 19
- Inception-v3, v4
æ¯è¼æ¹æ³
ãã®è«æã§ã¯ä»¥ä¸ã®æ¯è¼ãè¡ã£ã¦ãã¾ãï¼
- äºæ¸¬ç²¾åº¦ (Top-1 Accuracy, single central-crop)
- å½ä»¤æ°
- äºæ¸¬æé
- æ¶è²»é»å
- ã¡ã¢ãªä½¿ç¨é
äºæ¸¬ç²¾åº¦ã«ããæ¯è¼
ã¾ãã¯è²ã
ãªã¨ããã§ç®ã«ããäºæ¸¬ç²¾åº¦æ¯è¼ãè¦ã¦ã¿ã¾ãï¼
ï¼è«æä¸Figure 1ããå¼ç¨ï¼
縦軸ã¯Top-1 Accuracyãªã®ã§ç¸¦æ£ãé·ãã»ã©è¯ãCNNã¢ãã«ã§ãï¼ResNetãInceptionã¯å¾çºãªã ããã£ã¦äºæ¸¬ç²¾åº¦ãé«ããã¨ãåããã¾ãï¼
äºæ¸¬ç²¾åº¦ã¨äºæ¸¬æéã«ããæ¯è¼
äºæ¸¬ç²¾åº¦ãé«ãã®ã¯ãã¡ããè¯ããã¨ãªã®ã§ããï¼å®å¿ç¨ã§ã¯äºæ¸¬æéãæ°ã«ãªãã±ã¼ã¹ãå¤ã
ããã¾ãï¼
ã³ã³ããã£ã·ã§ã³ã®ããã«äºæ¸¬ã«ãããã§ãæéãããããããªãè¯ãã®ã§ããï¼ãªã¢ã«ã¿ã¤ã æ§ãè¦æ±ãããå¿ç¨å
ã§ã¯ãªãã¹ãé«éã«äºæ¸¬ãè¡ãããæ°æã¡ãããã¾ãï¼
è«æä¸ã§ã¯äºæ¸¬ç²¾åº¦ vs äºæ¸¬æéã«ããæ¯è¼ãè¡ã£ã¦ãã¾ããï¼
ï¼è«æä¸Figure 9ããå¼ç¨ï¼
縦軸ã¯Accuracyï¼æ¨ªè»¸ã¯1ç§ãããã«äºæ¸¬ãããå¹³åç»åæ°ã表ãã¦ãã¾ãï¼å³ä¸ã®å³ä¸ã«åãããã°ï¼ãã®ã¢ãã«ã¯é«éãã¤é«ç²¾åº¦ãªäºæ¸¬ãè¡ãããã¨ãæå³ãã¾ãï¼
ååã®è²ã¯Figure 1ã®ææ³ã¨å¯¾å¿ãã¦ããï¼åã®å¤§ããã¯å½ä»¤æ°ã¨æ¯ä¾ãã¦ãã¾ãï¼
å·¦ã®ã°ã©ãã¯ããããµã¤ãºã1ï¼å³ã®ã°ã©ãã¯ããããµã¤ãºã16ã¨ãã¦ãã¾ãï¼
äºæ¸¬ç²¾åº¦ã¨äºæ¸¬æéã¯ãã¬ã¼ããªãã®é¢ä¿ã«ãããã¨ããããã¾ãï¼å½ããåã¨è¨ãã°å½ããåããããã¾ãããâ¦
ResNetã¨Inceptionã¯VGGã¨åçã®äºæ¸¬æéã«ãé¢ãããé«ç²¾åº¦ãªäºæ¸¬ãè¡ã£ã¦ãããã¨ããï¼VGGãé¸ã¶çç±ã¯ãã¾ãç¡ãã®ããªã¨ããå°è±¡ã§ãï¼
AlexNetããã³NINã®äºæ¸¬æéã¯çãããï¼ç²¾åº¦ãããã¾ã§æ°ã«ããªãã±ã¼ã¹ãªãã°é¸æè¢ã«ä¸ããããã§ãï¼
ãã®ä»
äºæ¸¬ç²¾åº¦ã¨äºæ¸¬æéã«ç¹ã«èå³ããã£ãããï¼ãã®è¨äºã§ã¯ä¸ã®æ¯è¼ãç´¹ä»ãã¾ãããï¼ä»ã«ãæ§ã
ãªæ¯è¼ãè¡ã£ã¦ããã®ã§èå³ã®ããæ¹ã¯è«æãèªãã§ã¿ã¦ãã ããï¼
ï¼è«æã§è§¦ãããã¦ããENetã¯èè
ãã®ææ¡ææ³ã®ããã§ãï¼è«æä¸ã§ãã¡ããã¡ããENetãæ¨ãã¦ããè¨ãåããåºã¦ãã¾ãããâ¦ï¼
åèæç®
- A. Canziani, et al, "An Analysis of Deep Neural Network Models for Practical Applications", 2017. https://arxiv.org/abs/1605.07678