Compile-Time Shape Checking

tangent.check_shapes catches the class of bug that finite differences only reveal at run time - rank, broadcast, and matmul-dimension mistakes - before you run anything, and points at the offending line.

import numpy as np
import tangent

def layer(x, w, b):
    return np.tanh(x @ w) + b     # b has the wrong shape

tangent.check_shapes(layer, np.zeros((8, 4)), np.zeros((4, 16)), np.zeros((8,)))
# tangent.errors.ShapeError: Incompatible shapes for add: cannot broadcast
#   (8, 16) and (8,)
#     at layer.py:5: return np.tanh(x @ w) + b

How it works

The function runs on abstract arrays - ShapedArray objects that carry a shape and dtype but no data. NumPy's dispatch protocols (__array_ufunc__/__array_function__) route real operations through the checker, which applies NumPy's actual broadcasting and linear-algebra rules. Because there is no data, it is cheap even for huge shapes, and the error names the source line via the traceback.

No false positives

An operation whose shape rule is not modeled yields an unknown shape (dimensions become None, which is compatible with anything) rather than a guessed one. The checker only ever reports a mismatch it is certain about, so a clean check_shapes never masks a real bug and a modeled op never invents one.

Inputs

Pass real arrays (only their .shape/.dtype are read), tangent.ShapedArray instances, or scalars:

from tangent import ShapedArray
tangent.check_shapes(layer, ShapedArray((8, 4)), ShapedArray((4, 16)), ShapedArray((16,)))

Scope

Covers the forward (primal) function: elementwise ops and broadcasting, matmul/@/dot, reshape, transpose/.T, reductions with axis, concatenate/stack/where, indexing, len(), and iteration. Reductions and ufuncs cover most NumPy code; unmodeled functions degrade to unknown shapes.