CV4Code: Sourcecode Understanding via Visual Code Representations

We present CV4Code $$^1$$ , a compact and effective computer vision method for sourcecode understanding. Our leverages the contextual structural information available from code snippet by treating each as two-dimensional image, which naturally encodes context retains underlying through an explicit spatial representation. To codify snippets images, we propose ASCII codepoint-based image representation that facilitates fast generation of images eliminates redundancy in encoding would arise RGB pixel Furthermore, is treated neither lexical analysis (tokenisation) nor syntax tree parsing required, makes proposed agnostic to any particular programming language lightweight application pipeline point view. can even featurise syntactically incorrect not possible methods depend on Abstract Syntax Tree (AST). demonstrate effectiveness learning Convolutional Transformer networks predict functional task, i.e. problem it solves, source directly its representation, using embedding latent space derive similarity score two retrieval setup. Experimental results show our approach achieves state-of-the-art performance comparison other with same task data configurations. For first time benefits understanding form processing task. ( https://github.com/jpmorganchase/cv4code )