GCPNext'19 で発表された、AI/ML関連のまとめ資料 @社内勉強会
Google Cloud Next’19 Summary- AI/ML -Ryosuke MashikoAI System Dept. Machine Learning Engineering TeamDeNA Co,. Ltd
View Slide
■ Next ■ NDA
Google Cloud Next’19■ : San Francisco■ : 4/9 ~ 4/11■ GCP■ : 34000 500
■ !, $✓ *) ✓ $&"#■ Uber/Lyft(■ +%'FordBikeAmazon Go
Cloud Next AI/ML Summary
tl;dr■ AI&1+03 (0 $% .■ Keyword: Digital Transformation( ‘18 Democratizing AI )✓ %'*4)/2: !%"%%etc✓ $# 5-■ Anthos✓ GKE both on GCP, ON-PREM, AWS■ Microsoft,6✓ Managed Active Directory✓ Cloud SQL with Microsoft SQL server
tl;dr■ AI/MLAI ✓ AI Platform✓ AI Hub✓ AutoML○ AutoML Table○ AutoML Video API✓ BigQueryML
: https://www.youtube.com/watch?v=OHIEZ-Scek8
■ BigQuery ML■ AutoML■ AI Platform
BigQuery ML
BigQuery ML■ SQL /■ Update■ BQ
BigQuery ML: ~~CREATE OR REPLACE MODEL `test_dataset.boston_predict_model`OPTIONS(model_type='linear_reg', input_label_cols=['price']) ASSELECT*FROM `test_dataset.boston_train`: http://soymsk.hatenablog.com/entry/2018/09/08/144846
BigQuery ML: ~ ~SELECT*FROMML.TRAINING_INFO(MODEL `test_dataset.boston_predict_model`)
BigQuery ML: ~~SELECT*FROMML.EVALUATE(MODEL `test_dataset.boston_predict_model`, (SELECT*FROM`test_dataset.boston_test`))
BigQuery ML: ~ ~SELECTprice,predicted_priceFROMML.PREDICT(MODEL `soymsk-gcp.test_dataset.boston_predict_model`,(SELECT*FROM`test_dataset.boston_test`))
NEW: https://www.youtube.com/watch?v=OHIEZ-Scek8
BigQueryDWH PreprocessingModel TrainPredictBigQueryDataModel TrainPredictScheduled QueryBQMLScheduled QueryBQML■ BQ
AutoML
AutoML■ 20189Cloud Next FK +$■ -6))+7 S■ ,4+#(#0 >P%"*I5B;✓ '++8GC"+C:✓ %"*GCP0 B;+! &QJ✓ '++■ APIE✓ Vision … D13RA.?2✓ NL … [email protected]✓ Translation … HM
NEW: AutoML Vision Edge■ AutoML #"API$■ AutoML Vision Edge✓ Vision ! ○ Tensorflow Lite format○ Edge TPU compiled TensorFlow Lite format○ TensorFlow■ Edge TPU : https://cloud.google.com/edge-tpu/
NEW: AutoML Video API■ Vision APIVideo API ■ ✓ ✓ ✓ : https://cloud.google.com/vision/automl/docs/?hl=ja
NEW: AutoML Table■ Structured Data( ) API✓ BG/GCS■ AdaNet✓ https://github.com/tensorflow/adanet : https://cloud.google.com/automl-tables/?hl=ja
AutoML TableDEMO
AutoML Table:
AutoML Table: ■ 1000 ■ > 1 hour ■ →
AutoML Table: ■ ✓ IN: BQ/GIS✓ OUT: BQ/GIS■ ✓ API
AI Platform
AI Platform■ AI Platform (Cloud ML Engine $" + α)■ &0-#✓ : Jupyter Notebook ( ≒ SageMaker @ AWS)✓ AI Hub: %,+ !(✓ Kubeflow: *)%,✓ Training/Prediction: %,/'1&/■ CMLE*.
AI Platform: Notebook■ Managed Jupyter (JaaS)■ DL Tensorflow/Pytorch&!✓ "$■ $#&%✓ GPU
AI Platform: AI Hub■ " !$) "■ Google+*#' " (&%: https://aihub.cloud.google.com
AI Platform: Kubeflow (Pipeline)■ *($/#3.%1 OSS✓ TFX (Tensorflow Extended) +-✓ Kubernetes() ✓ JupyterHub, Katib, Pipeline■ Kubeflow Pipeline✓ ML#!,)& ✓ AI Hub2'○ AI Hub "0https://github.com/kubeflow/kubeflow
AI Platform: Training/Prediction■ $)/&+ %*!(✓ $)/&+#'Docker GPU/TPU"#'%* ✓ Multi-GPU✓ TPU
TPU■ GoogleTensor Processing Unit■ GPU ■ https://storage.googleapis.com/nexttpu/index.html■ Update:✓ TPU v3 (2018) GA
()" #TPUMulti GPUAI Platform%DWHBigQueryCloudStorageCloud MachineLearning#& !Kubeflow PipelineEDAModel $" #Single GPU$AI Notebook( 4 0. 1 ) 422 0. ,AI Hub
■ NEW: Recommendation API✓ Personalized Recommendation API○ ■ NEW: Cloud Inference API✓ β✓ ○
■ AI)80%/%AGPRQ✓ @IE$ !■ Tensorflow')+,/7?!✓ TF2.0 Eager Execution ModeKJNFMLD✓ 5<-(1AC4;>7=>(#SH!T■ 6:.(9&2✓ AIGCP*>3,AWS(9&2OB"
: Cloud Dataflow■ Dataflow✓ %*0 Apache Beam !2✓ Python/Java&MapReduce*!1 ,✓ %*4/(#()) !2 ■ Update✓ Preemptible VM(Spot [email protected]) +3.○ 5" -$✓ ' 70%OFF