Summary

From all the examples that we have covered in this chapter, we learned how to leverage a distributed training strategy with the TPU and GPU through AI Platform, which runs on TensorFlow Enterprise 2.2 distributions. AI Platform is a service that wraps around TPU or GPU accelerator hardware and manages the configuration and setup for your training job.

Currently, in Google AI Platform, the data ingestion pipeline relies on TFRecordDataset to stream training data in batches into the model training workflow. We also learned how to leverage a prebuilt model downloaded from TensorFlow Hub through the use of the TFHUB_CACHE_DIR environment variable. This is also the means to import your own saved model from an offline estate into Google AI Platform. Overall, in this platform, we used a TensorFlow Enterprise 2.2 distribution to achieve scalable data streaming and distributed training on Google Cloud's TPU or GPU and serialized all the model checkpoints and assets back to...