Leveraging Apache Spark for Scalable Data Prep and Inference in Deep Learning

To scale out deep learning training, a popular approach is to use Distributed Deep Learning Frameworks to parallelize processing and computation across multiple GPUs/CPUs. Distributed Deep Learning Frameworks work well when input training data elements are independent, allowing parallel processing to start immediately. However preprocessing and featurization steps, crucial to Deep Learning development, might involve complex business logic with computations across multiple data elements that the standard Distributed Frameworks cannot handle efficiently. These preprocessing and featurization steps are where Spark can shine, especially with the upcoming support in version 3.0 for binary data formats commonly found in Deep Learning applications. The first part of this talk will cover how Pandas UDFs together with Spark’s support for binary data and Tensorflow’s TFRecord formats can be used to efficiently speed up Deep Learning’s preprocessing and featurization steps. For the second part, the focus will be techniques to efficiently perform batch scoring on large data volume with Deep Learning models where real-time scoring methods do not suffice. Upcoming Spark 3.0’s new Pandas UDFs’ features helpful for Deep Learning inference will be covered.

About:
Databricks provides a unified data analytics platform, powered by Apache Spark™, that accelerates innovation by unifying data science, engineering and business.

Connect with us:

[Skorskimus]

Thanks for showing the code on reading binary audio files. That's going to be really handy! This was a great talk and I loved being able to see some of your code.