TensorFlow Meetup (21/03)

Last night we attended, and really enjoyed, the first TensorFlow (TF) London meetup hosted at Twitter. TensorFlow is Google’s 2nd generation open source machine learning framework. To find out about TensorFlow checkout the website or read the whitepaper.

2016-03-21 19.48.40

First up Rebecca Murphy, a Data Scientist at Ocado Technology, gave an overview of TensorFlow. Rebecca showed how easy it is to get going with TF both for installation (just use pip, duh!) and modelling in Python. There are also some really great models online to get started with. Check out Rebecca’s slides here.

This was followed by a fireside chat with Rebecca, Daniel Slater, from Bank of America Merrill Lynch, and Peter Morgan, from the Data Science Partnership. Each spoke about their experiences using TF (even if only in development at the moment), which in general have been very good. The speakers were also asked about what in particular they liked about TF. They answered three things:

  • The fact that it was engineered by Google and is used by them – means it’s probably pretty good.
  • Easy to get going with: documentation is good; tutorials are on point.
  • Great community around the project

Unfortunately the speakers were not able to compare TF with other deep learning packages like Theano (also in Python) or Torch (in Lua), or give much insight to how it performs in production. However, it is known that Google use TF in production for a number of products – so presumably it’s pretty reliable and performant. However, it was noted that performance has actually been an issue because up until very recently (last week?) Google had not released the distributed version — allowing training of models across clusters of nodes rather than only on a single machine.

Next, Daniel Slater gave a great talk on Deep-Q learning with TensorFlow and PyGame. Daniel spoke about replicating Deepmind’s earlier work (2013) using deep reinforcement learning for playing Atari games, specifically he’s been looking at Pong. It’s a really fun project, and Daniel also gave a brief but good introduction to deep reinforcement learning. Using TF running on his standard laptop (using a GPU), Daniel (and few others?) were able to build an AI using deep reinforcement learning that is pretty good at Pong. Really fun stuff. For more information, see Daniels’ blog. If you’re curious, the Python code is here. It’s less than 200 lines.

Also here are Daniel’s slides:

Things we learnt about TF:

  • Super easy to get going. Of course you still need to understand deep learning.
  • Only Python and C++ clients currently supported; more languages to come.
  • Python client currently does not support user-defined (custom) operators.
  • Distributed version of TF is out.
  • TF has great community support – Google engineers prioritising external questions/ issues.
  • Interesting thought: combine TF and Spark. Perhaps for data wrangling/ preprocessing?

Thanks to the organisers and Twitter for hosting.

Comments are closed.