boundlab.interp.base

Base ONNX operators shared by every BoundLab abstract domain.

Module Attributes

interpret

The domain-independent base interpreter.

Classes

Add	Addition via Expr.__add__ — exact for every component class.
BroadcastTo	ONNX Expand to a constant shape.
Cast	ONNX Cast; abstract expressions only allow the no-op cast.
DivSimple	Division with a concrete divisor (x * (1/y)) or dividend (reciprocal(y) * x); the fully abstract case belongs to the domain.
Gemm	ONNX Gemm: alpha * op(x) @ op(weight) + beta * bias, lowered to the transpose / matmul / mul / add handlers.
Identity	ONNX Identity — pass through.
MatmulSimple	Matrix product with one concrete operand — exact, a linear map on the abstract one.
MulSimple	Product with at most one abstract operand — exact via einsum.
Neg	Negation — exact.
Reshape	ONNX Reshape with a constant target shape.
ReduceMean	ONNX ReduceMean axis by axis via Expr.mean — exact (linear).
ReduceSum	ONNX ReduceSum axis by axis via Expr.sum — exact (linear).
Squeeze	ONNX Squeeze: drop size-1 axes.
Sub	Subtraction x + (-1) * y — exact; shared symbols cancel.
Transpose	ONNX Transpose; expressions use Expr.transpose, tensors permute.
Unsqueeze	ONNX Unsqueeze: insert size-1 axes.