Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Streamlining AI Prototyping and Deployment with R and MLflow

12 views

Published on

We provide a recap of the MLflow R interface which was announced at Spark+AI Summit Europe and discuss recent developments. The session includes a live demo showcasing the intersection of big data (Spark) and deep learning (via TensorFlow) and how the end-to-end lifecycle from prototyping to deployment can be managed by MLflow.

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Streamlining AI Prototyping and Deployment with R and MLflow

  1. 1. Kevin Kuo @kevinykuo, RStudio Streamlining AI Prototyping and Deployment with R and MLflow #UnifiedAnalytics #SparkAISummit
  2. 2. Daily specials - Quick update on the R ecosystems for AI stuff - Recap of MLflow - Demo - Discussion + Q&A 2#UnifiedAnalytics #SparkAISummit
  3. 3. Sparklyr Update - Arrow integration to massively speed up UDFs - XGBoost - TFRecord read/write - SparkNLP on the way https://spark.rstudio.com/ 3#UnifiedAnalytics #SparkAISummit
  4. 4. TensorFlow Update 4#UnifiedAnalytics #SparkAISummit
  5. 5. TensorFlow Update 5#UnifiedAnalytics #SparkAISummit
  6. 6. TensorFlow Update He looks skeptical, as if you were nothing get it right. It drives me crazy, I can do this reproachful look no longer endure. His name is Olaf. 6#UnifiedAnalytics #SparkAISummit
  7. 7. TensorFlow Update - library(keras) defaults to tf.keras - TensorFlow Probability for probabilistic modeling - Eager execution - Preparing for TF2.0 drop https://tensorflow.rstudio.com/ https://blogs.rstudio.com/tensorflow/ 7#UnifiedAnalytics #SparkAISummit
  8. 8. Quick recap of MLflow Open source platform for - Experiment instrumentation (Tracking) - Reproducible runs (Projects) - Model deployment (Models) 8#UnifiedAnalytics #SparkAISummit
  9. 9. Tracking Keeping track of stuff mlflow_log_param("num_hidden_units", 64) mlflow_log_artifact("training_history.png") mlflow_log_metric("accuracy", metrics$acc) 9#UnifiedAnalytics #SparkAISummit
  10. 10. Projects Packaging up (reproducible) building blocks mlflow_run("data-prep.R") 10#UnifiedAnalytics #SparkAISummit
  11. 11. Models Deployment flavors: keras: version: 2.2.2 data: model.h5 python_function: loader_module: mlflow.keras data: model.h5 env: conda_env.yaml utc_time_created: 19-04-25T01:00:21.21.72 11#UnifiedAnalytics #SparkAISummit mlflow_rfunc_serve( "keras_model", run_uuid = training_run_id )
  12. 12. Demo! 12#UnifiedAnalytics #SparkAISummit
  13. 13. Roadmap How are package dependencies handled for R projects? Conda? Packrat? What if your packages depend on Java/Python libraries? 13#UnifiedAnalytics #SparkAISummit
  14. 14. Quick excursion on dependency management 14#UnifiedAnalytics #SparkAISummit
  15. 15. Renv is in 15#UnifiedAnalytics #SparkAISummit renv::init() renv::restore()
  16. 16. Renv is in 16#UnifiedAnalytics #SparkAISummit Conda support in progress for reticulated packages
  17. 17. What about... What about stuff with Java/rJava dependencies?!?! 17#UnifiedAnalytics #SparkAISummit
  18. 18. Betting on new tech Why MLflow and not something else? 18#UnifiedAnalytics #SparkAISummit
  19. 19. Roadmap Better integration with deployment tech - MLeap (https://github.com/rstudio/mleap) - H2O - TensorFlow Serving - Arbitrary R models (plumber + docker) 19#UnifiedAnalytics #SparkAISummit
  20. 20. Resources - https://mlflow.org/ - https://github.com/mlflow/mlflow - https://tensorflow.rstudio.com/ - https://spark.rstudio.com/ - https://community.rstudio.com/ - Demo Repo: https://github.com/kevinykuo/sais2019-mlflow 20#UnifiedAnalytics #SparkAISummit
  21. 21. DON’T FORGET TO RATE AND REVIEW THE SESSIONS SEARCH SPARK + AI SUMMIT

×