ããã«ã¡ã¯ãã»ãããã§ãï¼
Kerasã¯ç°¡åã«å®é¨ãã§ãã¦ä¾¿å©ã§ããï¼ä»åã¯ãKerasã使ã£ã¦æ´»æ§åé¢æ°ãå®éã®ç²¾åº¦ã«ä¸ããå½±é¿ã調ã¹ã¦ã¿ã¾ãã
- 使ç¨ãããã¼ã¿ : MNIST
- 使ç¨ããã©ã¤ãã©ãª : Keras
Kerasã®åããæ¹ã¨ãããããªã人ã¯ãã®è¨äºãåç §ï¼
ææ¨
- training : ãã¼ã¿ãå¦ç¿ãã
- validation : å¦ç¿ãã¼ã¿ã®ä¸é¨ãæãåã£ã¦ãåã¨ããã¯æ¯ã«ãã¹ãããã(éå¦ç¿ãã§ãã¯)
- test : ãã¹ããã¼ã¿ã§ç²¾åº¦ããã§ãã¯(å§ãããåãã¦ãã)
åºæ¬çã«å¦ç¿æã®ç²¾åº¦ã®ä¸ããæ¹ãè¦ããã¨ããæã¯ãtrainingã¨validationãè¦ã(testã¯æå¾ã«è©¦ãç¨ã®ãã®ãªã®ã§ä½¿ããªã)
ä»åã¯ãtraining dataã§ã®æ£è§£ç(accuracy)ã¨validation dataã§ã®æ£è§£çã«ã¤ãã¦åepochãã¨ã«ã©ãå¤ãããè¦ã¦ã¿ãã
使ãæ´»æ§åé¢æ°(activation function)
- Linear
- softplus
- sigmoid
- ReLU (Rectified Linear Unit)
- Leaky ReLU
- PReLU (Parametrized ReLU)
LinearãLeaky ReLUã¾ã§ã®é¢æ°ç³»ã¯ãããæ°å¼ãç¥ããã人ã¯ã°ã°ã£ã¦ã¿ã¦ãã ããããããªã«é£ãããªãã§ãã
PReLUã¨ããã®ã¯ãLeaky ReLUã®è² ã®é¨åã®å¾ããããã¼ã¿ã«ãã£ã¦å¦ç¿ããã¦ããè³¢ãæ¹æ³ãã©ããã£ã¦å¦ç¿ããã¦ãã®ï¼æã«ãªãæ¹ã¯è«æãèªãã§ã¿ã¾ãããâè«æ
æ´»æ§åé¢æ°ã®çµæ
- Linearã¯æ´»æ§åé¢æ°ããããªãç¶æ
ã¤ã¾ãç·å½¢åé¡å¨
- ãã®ã¾ã¾(ç·å½¢åé¡å¨)ã ã¨ã精度ãå ¨ç¶åºã¦ããªã
- éç·å½¢æ§ããã¥ã¼ã©ã«ãããã«éè¦ã¨ãããã¨ãããã
- ReLU系強ããããã¦soft plusããªããªãå¥éãã¦ããã
- 両è ã®é¢æ°ç³»ã¯ä¼¼ã¦ãã
- PReLUããvalidation accuracyã¾ã§è¦ãã¨ãã£ã¨ãåªç§
- æå¤ã«Leaky ReLUãåãçºæ®ãããã¦ããªãï¼
Leaky ReLUã®ãã©ã¡ã¼ã¿ã¨ç²¾åº¦
- trainingãã¼ã¿ã®ç²¾åº¦çã«ã¯ã¯å°ããæ¹ããã
- ããã¯æ£è§£ãã¼ã¿ã«å¯¾ãã¦å½ã¦ã¯ãã¦ããã ã
- å®éã¯éå¦ç¿ãã¦ããå¯è½æ§ãããã
- 大äºãªã®ã¯validation accuracyã§ãå®éã«ãã®epochã§trainingããªãã£ããã®ã«å¯¾ãã¦ã©ã®ãããã®å¹æãçºæ®ãããã
- validation accuracyãè¦ãã¨ã ããã㧠Leaky ReLUã調ç¯ããã°ããããã
- ã ããå¤§å¹ ãªç²¾åº¦æ¹åã¨ããããã«ã¯ãããªã
- ä»åã¯å±¤ãä¸å±¤ã ããå·®ããã¾ããªãã®ãããæ·±ãããã«ãªã£ã¦ããã¨PReLUãLeaky ReLUãéå¦ç¿ãé²ãæ¹åã«æ´»èºãã¦ããããããã
RReLU(Randomized ReLU)
ããã¡ãã£ã¨ç²¾åº¦åä¸ã«è²¢ç®ã§ããããªã®ããRandomized ReLUã ä½ãéããã¨ããã¨ãLeaky ReLUã®å¾ãã«ã©ã³ãã æ§ãä»ä¸ãããã¨ã§ãå¾ãã«å¹ ãä¸ãã¦ããã
ç»å:è«æ(Empirical Evaluation of Rectified Activations in Convolutional Network)
è«æã®ä¸ã§ã¯ReLUãPReLUããé«ç²¾åº¦ã«ãªãã¨æ¸ããã¦ããããæ®å¿µãªããä»ã®Kerasã§ã¯å®è£ ããã¦ããªãã
ã ã£ããä½ã£ã¦ããï¼ï¼ ã¨æã£ãã®ã§ãããæ å¼±ã®ããæ念*1ããã®ãã¡å®è£ ãããããããªããããããããæã¯ãå½ããåã§ããèªåã§æ¸ãã¦ããæã®æ¹ãæ¡å¼µããããã§ããã
ã¾ã¨ã
ä»æ¥ã®ãã¤ã³ãã¯
ã»éç·å½¢æ§ã¯ç²¾åº¦ã«å¤§ããå¯ä¸ãã¦ãã
ã»ReLUå¼·ãããã®æ´¾çãã©ãã©ãåºã¦ãã¦ãã
ã»RReLUã¯å¼·ãããããå®è£
ã§ãªãããª(ä»åæ¬é¡)
ä»åº¦ã¯å±¤ã®æ·±ãã¨ãã«ã注ç®ãã¦èª¿ã¹ã¦ã¿ããã§ããï¼ã§ã¯ã§ã¯ï¼
ã¼ãããä½ãDeep Learning âPythonã§å¦ã¶ãã£ã¼ãã©ã¼ãã³ã°ã®çè«ã¨å®è£
- ä½è : æè¤åº·æ¯
- åºç社/ã¡ã¼ã«ã¼: ãªã©ã¤ãªã¼ã¸ã£ãã³
- çºå£²æ¥: 2016/09/24
- ã¡ãã£ã¢: åè¡æ¬ï¼ã½ããã«ãã¼ï¼
- ãã®ååãå«ãããã° (14件) ãè¦ã
深層å¦ç¿ Deep Learning (ç£ä¿®:人工ç¥è½å¦ä¼)
- ä½è : 麻çè±æ¨¹,å®ç°å®æ¨¹,åç°æ°ä¸,岡éå大è¼,岡谷貴ä¹,ä¹ ä¿é½å¤ªé,ãã¬ã¬ã©ããã·ã«,人工ç¥è½å¦ä¼,ç¥å¶æå¼
- åºç社/ã¡ã¼ã«ã¼: è¿ä»£ç§å¦ç¤¾
- çºå£²æ¥: 2015/11/05
- ã¡ãã£ã¢: åè¡æ¬
- ãã®ååãå«ãããã° (1件) ãè¦ã
*1:å ¥åãã¼ã¿æ¯ã«ã©ã³ãã ã«ä¿æ°ãã»ããããå¿ è¦ããããå²ã¨æ ¹ã£ãã®é¨åãããããªãã¨ãããªãã®ããªããã¨æã£ãã