cellmap_flow.norm.safe_expression

Evaluate the small array expressions behind LambdaNormalizer and LambdaPostprocessor.

These expressions arrive inside neuroglancer layer URLs and dashboard requests, and inference servers listen on every interface, so evaluating them with a bare eval() let anyone who could reach a server run code on it. An expression is now parsed once and every node is checked against a whitelist before anything runs:

the array x, numbers, and numeric dtype names as strings (x.astype("uint8"));
arithmetic, comparison and boolean operators, and a if cond else b;
indexing and slicing (x[..., 0]);
np.<name> for the numpy functions, dtypes and constants listed below;
a few array methods and attributes (x.astype, x.clip, x.shape, …);
abs.

Anything else (other names, other attributes, dunders, lambdas, comprehensions, **kwargs) raises ValueError. So does what would let a short expression allocate without bound: exponents must be literal numbers of modest size (no x ** 10**10), a list or tuple cannot be repeated (no [x] * 10**5), keyword arguments are only those in KEYWORDS (no shape=), and the *_like functions take no positional shape either.

That bounds what a whitelisted call can be asked for, not what arithmetic can produce: broadcasting arrays given new axes (x[..., None] * x[None]) still multiplies a chunk’s size, so the whitelist is not a memory limit.

Attributes

`NUMPY_FUNCTIONS`
`NUMPY_DTYPES`
`NUMPY_CONSTANTS`
`ARRAY_METHODS`
`ARRAY_ATTRIBUTES`
`DTYPE_STRINGS`
`KEYWORDS`
`MAX_POSITIONAL`
`MAX_LENGTH`
`MAX_EXPONENT`

Functions

compile_expression(→ Callable[[numpy.ndarray], ...)

Check expression against the whitelist and return a function of x.

Module Contents

cellmap_flow.norm.safe_expression.NUMPY_FUNCTIONS

cellmap_flow.norm.safe_expression.NUMPY_DTYPES

cellmap_flow.norm.safe_expression.NUMPY_CONSTANTS

cellmap_flow.norm.safe_expression.ARRAY_METHODS

cellmap_flow.norm.safe_expression.ARRAY_ATTRIBUTES

cellmap_flow.norm.safe_expression.DTYPE_STRINGS

cellmap_flow.norm.safe_expression.KEYWORDS

cellmap_flow.norm.safe_expression.MAX_POSITIONAL

cellmap_flow.norm.safe_expression.MAX_LENGTH = 500

cellmap_flow.norm.safe_expression.MAX_EXPONENT = 64

cellmap_flow.norm.safe_expression.compile_expression(expression: str) → Callable[[numpy.ndarray], numpy.ndarray]

Check expression against the whitelist and return a function of x.

Raises ValueError, naming the offending construct, if the expression is not allowed.