Index conf sparkai-feb20-n-pentreath

Index conf sparkai-feb20-n-pentreath

• 1.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Spark and AI
 —
 Fred Reiss
 
 Nick Pentreath

• 2.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Agenda The Machine Learning Workflow Overview of Spark Deep Learning Frameworks Discussion of Future Directions
• 3.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Perception The Machine Learning Workflow
• 4.
  DBG / Feb20, 2018 / © 2018 IBM Corporation In reality, workflow spans teams … The Machine Learning Workflow
• 5.
  DBG / Feb20, 2018 / © 2018 IBM Corporation … and tools The Machine Learning Workflow
• 6.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Major Frameworks Deep Learning on Spark • Deeplearning4J • BigDL • Deep Learning Pipelines • TensorFlowOnSpark • Microsoft Machine Learning on Spark (MMLSpark)
• 7.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Deeplearning4J Deep Learning on Spark • Distributed GPU support for all major deep learning architectures • CPU / Distributed CPU / Single GPU options exist • Supports Convolutional Nets, LSTMs / RNNs, Feedforward Nets, Word2Vec • Supported by Skymind • Backed by its own linear algebra library – ND4J • APIs in Scala, Java, Python • Newer Scala API, Keras-like • Keras import / export for Python API • Production serving is through proprietary layer
• 8.
  DBG / Feb20, 2018 / © 2018 IBM Corporation BigDL Deep Learning on Spark • Distributed CPU with Intel MKL • No GPU support • Most DL models – CNN, RNN • Backed by Intel • Natively integrated with Spark • Scala, Python API • Support for Spark ML pipelines • Uses internal Spark components for distributed training
• 9.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Deep Learning Pipelines Deep Learning on Spark • Created by Databricks • Focus on scoring models (TensorFlow / Keras) and basic transfer learning • No support for training the DL model • Focus on image data & use cases • Natively integrated with Spark • Scala, Python API • Support for Spark ML pipelines • Support for scoring models as a SQL UDF
• 10.
  DBG / Feb20, 2018 / © 2018 IBM Corporation TensorFlowOnSpark Deep Learning on Spark • Created by Yahoo • Scale out TF on Spark clusters • Use Spark executors to launch TF processes • Supports distributed training through TF parameter servers • RDMA / Infiniband improvement to TF to speed up distributed training • Good support for TensorBoard • Good integration with Spark • But only Python API • Some support for Spark ML pipelines • Relatively inactive recently
• 11.
  DBG / Feb20, 2018 / © 2018 IBM Corporation MMLSpark Deep Learning on Spark • Created by Microsoft • Supports training using CNTK including distributed • Image, text data • Good integration with Spark • Scala, Python, R API • Support for Spark ML pipelines • Relatively active, seems quite well supported
• 12.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Other Frameworks Deep Learning on Spark • H20 AI / DeepWater • Apache MXNet Spark integration • TensorFrames • CaffeOnSpark • scalable-deep-learning on Github • MLlib – MLPClassifier only • Sparknet (abandoned)
• 13.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Questions Future Directions in Spark + AI • What are the real-world uses of deep learning for Spark users? • Problems being solved? • DL component relative to the rest of the system? • What ETL and pre-processing is being used? • Transfer learning / using pre-trained models? • Full-scale training, distributed training? • What are the integration pain points between Spark and deep learning framework X? • Serialization issues • Integrating disparate frameworks and systems • Model deployment • Deploying pipelines that span pre-processing, ”traditional” ML and DL?
• 14.
  DBG / Feb20, 2018 / © 2018 IBM Corporation Thank you! Nick Pentreath Principal Engineer — [email protected] @MLnick ibm.com
• 15.
  DBG / Feb20, 2018 / © 2018 IBM Corporation