Shape Checking for Annotated Tensor Names (github.com)

A new Python library introduces shape checking for annotated tensor names by leveraging Abstract Syntax Tree (AST) transformations to automatically validate tensor shapes at runtime. Designed for popular frameworks like PyTorch, NumPy, and JAX, the library enforces shape consistency based on intuitive variable naming conventions—where underscores separate dimension annotations (e.g., weights_NK represents a tensor with shape N × K). This approach catches mismatches between function arguments and intermediate variables by injecting shape checks directly into the code, providing clear, informative error messages when inconsistencies arise. This development is significant for the AI/ML community as it streamlines debugging for tensor operations, a frequent source of runtime errors in deep learning workflows. By binding tensor shapes to variable names and validating them throughout function scopes, the tool maintains shape correctness without sacrificing flexibility across different backends. Variable dimensions (like N, K, M) are tracked locally within each function, supporting both dynamic and constant size annotations, which helps improve code readability and maintainability. This practical blend of static-style checks and dynamic validation lowers the cognitive load for developers, encouraging best practices in tensor programming and reducing shape-related bugs in complex models.