Gradual Tensor Shape Checking

Abstract Tensor shape mismatch is a common source of bugs in deep learning programs. We propose new type-based approach to detect tensor mismatches. One the main features our best-effort inference. As inference problem undecidable general, we allow static type/shape be performed only manner. If cannot guarantee absence inconsistencies, dynamic checks are inserted into program. Another feature gradual typing, where users can improve precision by adding appropriate type annotations formalize and prove that it satisfies criteria typing proposed Siek et al. 2015. have implemented prototype checking tool based on evaluated its effectiveness applying some neural network