2. About me
• Google Developer Expert for Machine Learning & Deep
Learning
• Deep Learning - Language & Dialogue
• Multiple Startups- B2B, B2C
• Red Dragon AI
3. • Deep Learning Consulting and Prototyping
• Education & Training
• Products:
• Conversational Computing
• Natural Voice Generation - multiple languages
• Knowledge Base Creation & Reasoning
7. PyTorch Roadmap
• Path to 1.0 and to Production
• Imperative and declarative
• Easy for research, easy for production
• Folding Caffe2 into PyTorch
• ONNX/Caffe for mobile support
8. PyTorch Roadmap
• Fused kernels for declarative graphs
• Tools, pre-trained models & Datasets
• ELF (game playing), Glow (compiler), FAISS etc
• Currently in use in Facebook
• Azure & AWS Support
• Coming in the next few months.
9. ONNX
(Open Neural Network Exchange)
• Converting models to Caffe, CNTK, others
• 0.4 adds RNN support for PyTorch
• TensorFlow support is experimental currently
12. Fast.ai Library
• Wrapper for PyTorch
• Fast.ai - Jeremy Howard & Rachel Thomas
• Adds features beyond PyTorch Vision and Text
• Abstracts a lot of the lower level PyTorch code
• Wraps Pandas, Spacy, TorchVision, TorchText to create
a set of often used data and ML functions
• Focuses on data handling and cutting edge techniques
13. Fast.ai Library
• date handling, text look ups and embeddings, tabular
data, categorical data etc
• Latest things like Stochastic Weight Averaging
• Very opinionated lib, often well explained in the
courses
• Function names are not always the easiest to decipher
• Documentation ….
18. How to Choose?
• Worth learning both
• If your are building for production go with
TensorFlow
• If you are doing research or trying to make
something new work go with PyTorch