Contains materials for my talk "You don't know TensorFlow". A recording of the talk can be found here.

In this talk, I present some under-appreciated and under-used features of TensorFlow. I focus on two things: (1) handling variable-length sequences in TensorFlow and (2) XLA.

⚠️Disclaimer 1 ⚠️: Please note that it's my observation that I have found the usage of these two things to be poorer than expected. Someone else may have other findings, and I don't have any immediate conflicts with those.

⚠️Disclaimer 2 ⚠️: The title of this talk is meant to be humorous and is NOT so to hurt anyone’s sentiments.

• notebooks/variable_length_sequences.ipynb: Shows how to use padded_batch() to handle variable-length sequences.
• notebooks/xla_surgery.ipynb: Shows some explorative interactions with tf.function with jit_compile enabled.

Additionally, for XLA, I use this Colab Notebook to demonstrate how XLA can speed up text generation in TensorFlow by 100x.

Link

In the interest of time, I may not be able to even scratch the surface of XLA. I suggest going through the following resources in case XLA in TensorFlow interests you:

• Official documentation
• Faster Text Generation with TensorFlow and XLA
• XLA Integration for TensorFlow Models

But I wanted to cover:

• Exporting TensorFlow SavedModels for deployment (example)
  • How should you export your SavedModels for deployment?
  • Model surgery
  • TF Serving
• Ragged tensors
• Taking advantage of TensorRT to speedup inference (example)