My peer in the Bay Area, Christine Matheney (@Matheney), put together a bunch of information around TensorFlow today, and I thought it was relevant to my audience. I’ve been doing a lot of work with CUDA and DirectX on GPUs lately, so it was right down my alley!
Christine has chatted with a number of people across the team about TensorFlow, Caffe, Deep Learning and GPUs on Azure, and here’s what she was able to gather:
What is it?
TensorFlow is an open source software library for machine learning across a range of tasks. It was originally developed by Google to meet their needs for systems capable of building and training neural networks to detect and decipher patterns and correlations, analogous to the learning and reasoning which humans use. It became open source in Nov. 2015.
Top 5 use cases:
- Voice/Sound Recognition
- Text Based Applications
- Image Recognition
- Time Series
- Video Detection
What are alternatives?
A few I’m familiar with, but there are others I’m sure I missed.
- Caffe (used by UC Berkeley)
- CNTK (created by Microsoft)
What’s the easiest way to get started on Azure?
- Deep Learning toolkit for Data Science VM – this has everything, running on Windows
- This deep learning toolkit provides Windows GPU versions of mxnet, CNTK, TensorFlow, and Keras for use on Azure GPU N-series instances
- Data Science Virtual Machine for Linux (Ubuntu) – runs everything on Ubuntu! (released 4/13)
- CNTK, TensorFlow, MXNet, Caffe, Caffe2, DIGITS, H2O, Keras, Theano, and Torch are built, installed, and configured so they are ready to run immediately. The NVIDIA driver, CUDA, and cuDNN are also included. All frameworks are the GPU versions but work on the CPU as well. Many sample Jupyter notebooks are included.
GPUs (NV/NC machines) are only for HDD VMs in the following regions: US East, US North Central, US South Central, West Europe, Japan East, and Asia Southeast
Want to play with it?
TensorFlow has a playground if you want to know the basics – check this out! Learn about how Neural Networks work for classification and regression.