TF-Coder: Program Synthesis for Tensor Manipulations

The success and popularity of deep learning is on the rise, partially due to powerful frameworks such as TensorFlow PyTorch, which make it easier develop models. However, these libraries also come with steep curves, since programming in quite different from traditional imperative explicit loops conditionals. In this work, we present a tool called TF-Coder for by example TensorFlow. uses bottom-up weighted enumerative search, value-based pruning equivalent expressions flexible type- filtering ensure that adhere various requirements imposed library. We train models predict operations features input output tensors natural language descriptions tasks prioritize relevant during search. solves 63 70 real-world within 5 minutes, sometimes finding simpler solutions less time compared experienced human programmers.