TF-Coder: Program Synthesis for Tensor Manipulations
summary
TF‑Coder turns an input/output example plus optional text into a TensorFlow expression by enumerative search that uses weighted search and operation filtering and ML-guided weight adjustment. The system synthesizes solutions over a 500-op API, solves 63 of 70 tasks from real benchmarks, averages 17 seconds per solved task, and eliminates 98.6% of argument combinations via filters. The tool is public and replaces multi-line handwritten TensorFlow routines with one-line broadcasted expressions.
AI-written summary · about
chapters (12)
- 0:00Coding in TensorFlow can be tricky!
- 3:06TF-Coder Overview
- 4:28Bottom-Up Enumerative Search
- 6:55Weighted Search: Key Takeaways
- 9:382 Operation Filtering: Key Takeaways
- 11:343 Guiding Search with ML: Key Takeaways
- 12:06Experiments: Comparison to Prior Work
- 12:54Experiments: Ablations
- 13:14Experiments: Comparison to Human Performance
- 14:08Public Tool
- 14:40Practical Effectiveness
- 16:20Try TF-Coder yourself!
related talks
12mCoding Agents for Compiler Construction: Beyond the...
01h 09mCodex presentation & demo
34mIt's Alive! Machine Learning Writes Your Code! 🧠
41m"Do What I Mean": The History of AI and Program Synthesis
11mGen AI-Assisted Code Authoring
36mIt's Alive! Machine Learning Writes Your Code! 🧠🔮
35mIt's Alive! Machine Learning Writes Your Code!
21mTowards Neural Synthesis for SMT-Assisted Proof-Oriented Programming