TF-Coder: Program Synthesis for Tensor Manipulations

Kensen Shi, David Bieber, Rishabh Singh

Keywords: Program Synthesis, TensorFlow

TL;DR: We present a program synthesizer for tensor manipulation tasks in TensorFlow, using an efficient bottom-up enumerative search.

Abstract: Deep learning frameworks such as TensorFlow and PyTorch come with steep learning curves. We present a tool called TF-Coder for programming by example in TensorFlow. It uses a bottom-up weighted enumerative search with learned models that prioritize relevant operations. TF-Coder solves 63 of 70 real-world tasks within 5 minutes, often achieving superhuman performance -- finding solutions that are simpler than those written by TensorFlow experts, in less time.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:2003.09040/code)