AWS 深局åŠç¿ AMI
深局åŠç¿ã¢ããªã±ãŒã·ã§ã³ããã°ããæ§ç¯ããããã«äºåã«æ§æãããç°å¢
AWS 深局åŠç¿ AMI ã§ã¯ãããŒã¿ãµã€ãšã³ãã£ã¹ããæ©æ¢°åŠç¿ã®å°éå®¶ãç ç©¶è ãã¯ã©ãŠãäžã§ã®æ·±å±€åŠç¿ãããããèŠæš¡ã§å éããããã®ã€ã³ãã©ã¹ãã©ã¯ãã£ãšããŒã«ãæäŸããŠããŸããAmazon EC2 ã€ã³ã¹ã¿ã³ã¹ã«ã¯ Apache MXNetãGluonãTensorFlowãMicrosoft Cognitive ToolkitãCaffeãCaffe2ãTheanoãTorchãPyTorchãChainerãKeras ãšãã£ãäžè¬çãªæ·±å±€åŠç¿ãã¬ãŒã ã¯ãŒã¯ãäºåã€ã³ã¹ããŒã«ãããŠããŸãããã®ãããAmazon EC2 ã€ã³ã¹ã¿ã³ã¹ããã°ããäœæããŠãé«åºŠãªã«ã¹ã¿ã AI ã¢ãã«ã®ãã¬ãŒãã³ã°ãæ°ããã¢ã«ãŽãªãºã ã®å®éšãæ°ããã¹ãã«ãæè¡ã®åŠç¿ãè¡ããŸãã
Amazon EC2 GPU ã€ã³ã¹ã¿ã³ã¹ãå¿ èŠãªå Žåã§ããCPU ã€ã³ã¹ã¿ã³ã¹ãå¿ èŠãªå Žåã§ããæ·±å±€åŠç¿ AMI ã«å¯Ÿããè¿œå æéã¯çºçããŸãããã¢ããªã±ãŒã·ã§ã³ã®ä¿åãå®è¡ã«å¿ èŠãª AWS ãªãœãŒã¹ã«å¯ŸããŠã®ã¿æéãçºçããŸãã
ã¯ã©ãŠãäžã® TensorFlow ãããžã§ã¯ãã® 88% 㯠AWS ã§å®è¡ããŠããŸãã
Nucleus Research ã¯ã深局åŠç¿ã®å°éå®¶ãæ·±å±€åŠç¿ã®ããã«ä»ã®ã¯ã©ãŠããããã€ããŒããã AWS ãéžæãã 5 ã€ã®çç±ã説æããŠããŸãã
AWS 深局åŠç¿ AMI ãšããéžæ
çµéšè±ããªæ©æ¢°åŠç¿ã®å°éå®¶ã§ãã£ãŠããæ©æ¢°åŠç¿ã®äœ¿çšãéå§ããããšã¯æéã®ãããé¢åãªãã®ã§ããAWS ã§ã¯ãéçºè ã®ããŸããŸãªããŒãºããµããŒãããããã« AMI ãçšæããŠããŸãã䜿çšéå§ããã»ã¹ã«ã€ããŠã®è©³çްã¯ãAMI éžæã¬ã€ãããã®ä»ã®æ·±å±€åŠç¿ãªãœãŒã¹ãã芧ãã ããã
Conda AMI
深局åŠç¿ãã¬ãŒã ã¯ãŒã¯ã®äºåã€ã³ã¹ããŒã«ããã pip ããã±ãŒãžãå¥ã®ä»®æ³ç°å¢ã§å¿ èŠãšããéçºè ã¯ãUbuntuãAmazon LinuxãWindows 2016 ã®ããŒãžã§ã³ã§ Conda ããŒã¹ã® AMI ãå©çšã§ããŸãã
Conda AMI ã®å©ç¹ ã®è©³çŽ°ãšæé ã¬ã€ããã芧ã«ãªãã䜿çšãéå§ããŠãã ããã
Base AMI
ãã©ã€ããŒã深局åŠç¿ãšã³ãžã³ãªããžããªãæåããèšå®ããéçºè ããæ·±å±€åŠç¿ãšã³ãžã³ã®ã«ã¹ã¿ã æ§ç¯ãè¡ãéçºè ã¯ãUbuntu ãš Amazon Linux ã®ããŒãžã§ã³ã§ Base AMI ãå©çšã§ããŸãã
Base AMI ã®å©ç¹ ã®è©³çŽ°ãšæé ã¬ã€ããã芧ã«ãªãã䜿çšãéå§ããŠãã ããã
深局åŠç¿ãã¬ãŒã ã¯ãŒã¯ã®ãµããŒã
AWS 深局åŠç¿ AMI ã§ã¯ãã¹ãŠã®äžè¬çãªæ·±å±€åŠç¿ãã¬ãŒã ã¯ãŒã¯ããµããŒãããŠãããããå€§èŠæš¡ãªã¢ãã«ãå®çŸ©ããŠãã¬ãŒãã³ã°ã§ããŸããAmazon Linux ããã³ Ubuntu åãã«æ§ç¯ãããŠãã AMI ã«ã¯ãApache MXNet ãš GluonãTensorFlowãMicrosoft Cognitive ToolkitãCaffeãCaffe2ãTheanoãTorchãPyTorchãChainerãKeras ãäºåã«æ§æãããŠããããããããã®ã©ã®ãã¬ãŒã ã¯ãŒã¯ãå€§èŠæš¡ã«ãã°ãããããã€ããŠå®è¡ã§ããŸãã
ã¢ãã«ã®ãã¬ãŒãã³ã°ãå é
éçºãã¢ãã«ã®ãã¬ãŒãã³ã°ãä¿é²ããããã«ãAWS Deep Learning AMI ã§ã¯ãäžè¬ç㪠Python ããã±ãŒãžãš Anaconda Platform ã®ã€ã³ã¹ããŒã«ã«å ããŠãäºåã«æ§æããã CUDA ãã©ã€ããŒãcuDNN ãã©ã€ããŒãIntel Math Kernel Library (MKL) ã䜿çšãããææ°ã® NVIDIA GPU ã«ããã¢ã¯ã»ã©ã¬ãŒã·ã§ã³ãå©çšã§ããŸãã
GPU ã€ã³ã¹ã¿ã³ã¹
P3 ã€ã³ã¹ã¿ã³ã¹ã§ã¯ãæ§äžä»£ã® Amazon EC2 GPU ã³ã³ãã¥ãŒãã£ã³ã°ã€ã³ã¹ã¿ã³ã¹ãããæå€§ 14 åé«ãããã©ãŒãã³ã¹ãçºæ®ããŸããNVIDIA Tesla V100 GPU ãæå€§ 8 åæèŒãã P3 ã€ã³ã¹ã¿ã³ã¹ã®æµ®åå°æ°ç¹ããã©ãŒãã³ã¹ã¯ãæ··åç²ŸåºŠã§æå€§ 1 PFLOPSãåç²ŸåºŠã§æå€§ 125 TFLOPSãåç²ŸåºŠã§æå€§ 62 TFLOPS ã§ãã
匷åãªã³ã³ãã¥ãŒãã£ã³ã°
C5 ã€ã³ã¹ã¿ã³ã¹ã¯ 3.0 GHz Intel Xeon ã¹ã±ãŒã©ãã«ããã»ããµãæèŒããŠãããIntel Turbo Boost Technology ã«ããã·ã³ã°ã«ã³ã¢ã§æå€§ 3.5 GHz ãŸã§é«éåã§ããŸããC5 ã€ã³ã¹ã¿ã³ã¹ã¯ vCPU ã«å¯Ÿããã¡ã¢ãªã®å²åãé«ããäŸ¡æ Œæ§èœæ¯ã¯ C4 ã€ã³ã¹ã¿ã³ã¹ã«æ¯ã¹ãŠ 25% åäžããŠãããããèŠæ±ã®å³ããæšè«ã¢ããªã±ãŒã·ã§ã³ã«æé©ã§ãã
Python ããã±ãŒãž
AWS 深局åŠç¿ AMI ã«ã¯ãPython 2.7 ããã³ Python 3.5 ã«ãŒãã«ãæèŒãã Jupyter ããŒãããã¯ãšå ±ã«ãPython åã AWS SDK ãªã©ã®äžè¬ç㪠Python ããã±ãŒãžãã€ã³ã¹ããŒã«ãããŠããŸãã
Anaconda Platform
AWS Deep Learning AMI ã«ã¯ãããã±ãŒãžã®ç®¡çãšãããã€ãç°¡åã«ããããã«ãAnaconda2 ãš Anaconda3 Data Science ãã©ãããã©ãŒã ãã€ã³ã¹ããŒã«ãããŠãããå€§èŠæš¡ããŒã¿åŠçãäºæž¬åæãç§åŠèšç®ã«å¯Ÿå¿ããŠããŸãã
AWS ã§æ·±å±€åŠç¿ã®äœ¿çšãéå§ãã
AWS ã¢ã«ãŠã³ãã«ãµã€ã³ã¢ãããã
AWS 深局åŠç¿ AMI ãå
¥æãã
ãããžã§ã¯ãã«åã£ã AMI ãšã€ã³ã¹ã¿ã³ã¹ã¿ã€ããéžæããŠãã ããã
AWS ã§æ§ç¯ãéå§ãã
Amazon SageMaker ãæ©æ¢°åŠç¿ã§äœ¿çšãã
Amazon SageMaker ã¯ãéçºè
ãããŒã¿ãµã€ãšã³ãã£ã¹ããæ©æ¢°åŠç¿ã¢ãã«ãããããèŠæš¡ã§ãã°ããç°¡åã«æ§ç¯ããã¬ãŒãã³ã°ããããã€ã§ããããã«ãããã«ãããŒãžããµãŒãã¹ã§ããAmazon SageMaker ã䜿çšãããšãéçºè
ã«ããæ©æ¢°åŠç¿ãéåžžæžéããããããªé害ãã¹ãŠãæé€ã§ããŸãã

