ADELT: Transpilation Between Deep Learning Frameworks

Linyuan Gong, Jiayi Wang, Alvin Cheung

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Code Transpilation, Deep Learning, Deep Learning Frameworks, PyTorch, Application, Adversarial Training

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: We propose the Adversarial DEep Learning Transpiler (ADELT), a novel approach to source-to-source transpilation between deep learning frameworks. ADELT uniquely decouples code skeleton transpilation and API keyword mapping. For code transpilation, it uses few-shot prompting on large language models, while for API keyword mapping, it employs contextual embeddings from a code-specific BERT. These embeddings are trained in a domain-adversarial setup to generate a keyword translation dictionary. ADELT is trained on an unlabeled web-crawled deep learning corpus, eschewing hand-crafted rules and parallel data. It outperforms state-of-the-art transpilers, improving exact match scores by 15.9 pts and 12.0 pts for PyTorch-Keras and PyTorch-MXNet transpilation pairs respectively. We provide open access to our code, corpus, and evaluation benchmarks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7810