Deep Learning을 위한 AWS 기반 인공 지능(AI) 서비스 (윤석찬)
46 likes11,575 views
The document discusses various tools and technologies related to deep learning, including MXNet, GPU and FPGA computing, and AWS services. It highlights the efficiency of GPU-accelerated computing and provides links to resources for developers, including the AWS Activate package offering credits and online training. Key topics include deep reinforcement learning applications and model training using cloud resources.
Deep Learning을 위한 AWS 기반 인공 지능(AI) 서비스 (윤석찬)
- 1. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
- 12. More compute Accuracy Scale (data size, model size) neural networks other approaches Now © Jeff Dean, Trends and Developments in Deep Learning Research http://www.slideshare.net/AIFrontiers/jeff-dean-trends-and-developments-in-deep-learning-research
- 14. BlindTool by Joseph Paul Cohen on Nexus 4 Mobile Application • https://github.com/dmlc/mxnet.js/ • http://rupeshs.github.io/machineye MXNetJS in Web Browser Web Applications • https://www.youtube.com/watch ?v=UHUC4ueEiwM • https://play.google.com/store/ap ps/details?id=the.blindtool
- 15. Deep Drone: Object Detection and Tracking for Smart Drones on Embedded System TX1 with customized board Drone • https://web.stanford.edu/class/cs231a/prev_projects_20 16/deep-drone-object__2_.pdf Deep RL | Playing Flappy Birds • https://github.com/li-haoran/DRL-FlappyBird • https://github.com/devsisters/DQN-tensorflow Human-Level Control through Deep Reinforcement Learning
- 24. Spot Instances (80% ↓) = $30 per hour
- 25. $aws ec2-run-instances ami-b232d0db --instance-count 20 --instance-type p2.8xlarge --region us-east-1 $aws ec2-stop-instances i-10a64379 i-10a64280 ...
- 26. 1GiB GPU Memory 2 GiB 4 GiB 8 GiB
- 27. NVIDIA Tesla GPU Card P2: GPU-accelerated computing § Enabling a high degree of parallelism – each GPU has thousands of cores § Consistent, well documented set of APIs (CUDA, OpenACC, OpenCL) § Supported by a wide variety of ISVs and open source frameworks Xilinx UltraScale+ FPGA F1: FPGA-accelerated computing § Massively parallel – each FPGA includes millions of parallel system logic cells § Flexible – no fixed instruction set, can implement wide or narrow datapaths § Programmable using available, cloud-based FPGA development tools
- 33. • • • •
- 35. 3
- 38. • • 1 4.75 8.5 12.25 16 1 4.75 8.5 12.25 16 Speedup(x) # GPUs Resnet 152 Inceptin V3 Alexnet Ideal 91% Efficiency 88% Efficiency # GPUs • EC2 16x P2.16xlarge by AWS CloudFormation • Mounted on Amazon EFS
- 39. ../../tools/launch.py -n $DEEPLEARNING_WORKERS_COUNT -H $DEEPLEARNING_WORKERS_PATH python train_mnist.py --gpus $(seq -s , 0 1 $ (($DEEPLEARNING_WORKER_GPU_COUNT - 1))) --network lenet --kv-store dist_sync parameter server network choice update policy
- 47. import mxnet as mx bucket = 'my-model-bucket’ s3_client = boto3.client('s3') f_params = 'resnet-18-0000.params' f_symbol = 'resnet-18-symbol.json' #params f_params_file = tempfile.NamedTemporaryFile() s3_client.download_file(bucket, f_params, f_params_file.name) #symbol f_symbol_file = tempfile.NamedTemporaryFile() s3_client.download_file(bucket, f_symbol, f_symbol_file.name) def lambda_handler(event, context): model = load_model(f_params_file, f_symbol_file) • •
- 48. $ wrk -t40 -c120 -d60s https://48l7awor09.execute-api.us- east-1.amazonaws.com/dev/predict?url=img_url Thread Stats Avg Stdev Max +/- Stdev Latency 1.18s 96.82ms 1.98s 81.23% Req/Sec 2.39 2.87 10.00 87.91% 4520 requests in 1.00m, 3.16MB read Socket errors: connect 0, read 0, write 0, timeout 0 Requests/sec: 75.23 Transfer/sec: 53.85KB
- 50. • •
- 52. 딥러닝 개발자를 위한 AWS 크레딧 제공! http://bit.ly/awskr-feedback AWS Activate 패키지 100달러 무료 크레딧 + 80 달러 Qwiklab Credit 600달러 온라인 강좌 수강권+ 100달러 1개월 비지니스 서포트 등록하시면 패키지를 받으실 수 있는 URL 및 AWS 학습 정보를 이메일로 보내드립니다!
- 53. • • • •