Andrea Suckro, Alexander Höreth, \today
Concepts and Applications of Neural Networks, University of Osnabrück
High level abstractions for efficient low level computations.
- Process:
- Transcribe code into a computation graph
- Optimize the graph
- Outsource computations as a whole to (preferably) the GPU
- Different levels of abstractions:
- Keras is very abstract, better for prototyping
- Theano/Tensorflow is rather low level, more general purpose
- Tensorflow: Provides very expressive visualization tools
Checkout our CANN repository which offers you a stepping stone to actually apply the presented frameworks. Specifically the /playground folder gives you the opportunity to combine intuition earned from playground.tensorflow.org with implementation in different frameworks. Play with this before going all in with one of the frameworks for your next project!
github.com/AndreaSuckro/CANN
{height=30%}