Tensor flow intro and summit info feb 2017
1 like152 views
TensorFlow is a tool for building computational graphs that can be distributed across CPUs and GPUs. It allows building deep learning models at a low level but also provides higher level APIs like Keras. The key concepts in TensorFlow include the computational graph, operations, sessions, and TensorBoard for visualizing models. TensorFlow 1.0 is stable for building and training models, and upcoming features like XLA will optimize models for faster execution. Higher level APIs and tools like Keras and Estimators make TensorFlow easier to use.
Tensor flow intro and summit info feb 2017
- 1. TensorFlow for Developers By Sam Witteveen
- 2. What is TensorFlow? • A tool for building computational graphs • It allows for the graph to be distributed over many many CPUs/GPUs • Up till now TensorFlow has been a low level Library • TensorFlow has fast become the standard library for Deep Learning
- 3. Pure TF vs Higher Levels • You will always have the most amount of control with Pure TensorFlow • It will usually be more work than Keras/TFLearn etc • It allows you to understand more of what is actually going on in the network • TensorBoard is such a big advantage
- 4. The Big Concepts of TensorFlow • The Graph • Operations • Sessions • TensorBoard
- 5. The Graph • Everything must be built on the graph before you execute • We can do this in a variety of languages • We the graph is run it will be in C++ • With new XLA TF will edit/create an optimized graph
- 6. Operations • Operations are performed on a graph • Standard math operations through to common DL formulas and tools • Allow a high level of granularity in your model
- 7. Sessions • Sessions execute the graph • Nothing is run till you init and run a session
- 8. TensorBoard • Gives us a visual representation of our model • Gives us stats about our training variables like loss accuracy etc • Going to have debugging soon.
- 9. Conclusion • Jump in. Its not as hard as people think and its going to get even easier. • Use TensorBoard to analyze/debug study your models • 1.0.0 is stable and there is no reason to delay
- 10. Summit 2017
- 11. TensorFlow 1.0.0
- 12. XLA • Accelerated Linear Algebra • Domain-specific compiler for linear algebra • ... that optimizes TensorFlow computations • JIT ('Just-in-Time') compilation • AOT ('Ahead-of-Time') compilation
- 13. JIT COMPILATION • Combine Ops in tree into single loops / kernels • eg: on one MNIST forward/backward batch :
- 14. AOT COMPILATION • tfcompile is a standalone tool ... • .. that compiles TF graphs into executable code • ... which can be used as library functions • ... which does not depend on TensorFlow • Target use-case : Mobile devices
- 15. High Level API • Layers (1.1), Estimators(1.2), Canned Estimators • Keras • TFSlim/TFLearn (looks like SKL)
- 16. tf.keras • New TF only version of Keras • Will still have Theano versions and new CNTK Keras • Layers/Models/ Mix & Match • tf.keras.applications.InceptionV3(weights=‘imagine’, include_top=False,pool=‘avg) • Cloud ML support for distributed training and serving • tf.contrib.keras coming in 1.1(mid march) then tf.keras in 1.2
- 17. The End • Sam • email: s@maclea.ai • twitter: sam_witteveen • https://github.com/samwit/TensorFlowTalks • Martin • martin.andrews@redcatlabs.com • http://mdda.net • github: mdda