What it does

TensorSensor wraps your NumPy, PyTorch, JAX, or TensorFlow code and rewrites the exception messages you get when matrix dimensions collide. Instead of matmul: Input operand 1 has a mismatch in its core dimension 0, you get Cause: @ on tensor operand W w/shape (764, 100) and operand X.T w/shape (764, 200). It also draws AST diagrams showing every sub-expression’s shape.

The interesting bit

The author built a custom parser that re-evaluates your assignment statements to capture intermediate shapes, then injects that context back into the error. It’s a small act of interpreter-level revenge against one of deep learning’s most tedious debugging rituals.

Key highlights

• Works across NumPy, PyTorch, JAX, TensorFlow, Keras, and fastai with optional dependency installs ([torch], [tensorflow], [jax], [all])
• Two modes: tsensor.clarify() for augmented exceptions, tsensor.astviz() for full computation-tree diagrams
• Requires actual Graphviz dot binary (not just the Python package) for AST visualization
• Version 1.0 as of December 2021; tested against TensorFlow 2.5.0, PyTorch 1.10.0, JAX 0.2.20

Caveats

• M1 Macs hit illegal instruction errors in many Anaconda-installed tensor libraries as of September 2021 (PyTorch excepted)
• Custom parser only handles assignment statements and expressions; inline method bodies and \ line continuations break it
• Side effects in re-evaluated statements will execute twice; not thread-safe with threading

Verdict

Grab it if you’re teaching deep learning or still squinting at gufunc signature errors. Skip if your debugging is already handled by a typed tensor library or you live entirely in method-chained one-liners.