Deep dive in Nx Backends
summary
Shows how Nx represents tensors (binary backend) and the defn expression backend as a shared graph that enables compilation and symbolic transforms. Demonstrates runtime options: EXLA compiles to GPU/CPU with JIT, TorchX (PyTorch) differences like gather semantics, and backend swap via copy. Covers automatic differentiation done symbolically (not finite differences) with examples that yield expected derivatives (one eval returns 42), notes Axon runs unchanged if backend is compatible, and points to backend behavior callbacks to implement.
AI-written summary · about
chapters (6)
- 0:0001:10 - Intro & Table of Contents
- 1:1102:58 - What are the Nx Backends?
- 2:5904:29 - When to Use Different Backends?
- 4:3009:25 - Backends vs Numerical Definitions
- 9:2616:54 - Nx.Defn.Expr Usages
- 16:5522:29 - To +Inf and Beyond
related talks
45mKeynote: Panel(...)
32mTowards new deep learning abstractions on top of exist frameworks
35mTowards Typesafe Deep Learning in Scala
25mApplied Machine Learning in Elixir
42mAnima Anadkumar, Principal Scientist, Amazon Web Services, Endowed Professor, CalTech
28mNetworkX is Fast Now: Graph Analytics Unleashed
25mMore Clojure Less Complication
37mDeep Learning Neural Network Acceleration at the Edge