The Tensors Must Flow
summary
Guildsman ports Google TensorFlow to Clojure via the TensorFlow C API, providing a tiny JNI-backed Java layer and an idiomatic plan-as-data approach that keeps node definitions as maps and separates planning and building. Gradients live in C++ and some Clojure fallbacks; C++ gradients sometimes crash the JVM. A simple optimization learns x=2 in four steps, MNIST demos reach ~83% training accuracy (test dipped near 68% initially), and a custom web-based summary app shows live scalars and histograms for iterative development.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:48Machine Learning
- 3:29Tensorflow
- 5:15Gils Mood
- 6:13Tensorflow Architecture
- 10:39Hello World Example
- 11:25Planning
- 15:21Learning
- 20:21Python and C
- 23:04Graph
- 24:24Tensor Board
- 26:11Closure
- 28:27Live Demo
- 32:40Minh Web App
- 37:18CN NS
related talks
25mMore Clojure Less Complication
01h 00mTensorFlow training on the JVM
42mClojure is Not Afraid of the GPU
43mTensorflow on open source GPUs
26mThe Feynman Machine: AI architecture in Clojure/Neanderthal
34mWorking with Tensorflow from the JVM #bbuzz
26mMachine Learning with TensorFlow
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond