SizeCheck: Runtime Shape Validation for Size-Annotated Julia Code

This library provides sizecheck, a macro that automatically adds runtime shape checking to Julia functions based on size-annotated variable names.

When writing Julia code, it's common to use naming conventions that indicate tensor shapes, as in this Medium post. For example, if a tensor weights has shape N × K, you might name the variable weights_NK. This macro adds validation checks that tensors match their annotated shapes at runtime.

SizeCheck.@sizecheck — Macro

Automatically adds runtime shape checking to Julia functions based on size-annotated variable names. Variables with underscores followed by dimension letters (e.g., x_NK) are validated to ensure consistent shapes.

Dimension annotations can contain:

Variable dimensions (uppercase letters): N, K, M - stored in variables of the same name
Constant dimensions (single digits): 3, 4, 2 - checked for exact size

The macro automatically adds shape validation for:

Function arguments with underscores in their names
Variable assignments to names containing underscores, including destructuring assignments

The dimensions are scoped to the function they are defined in. For example, if you define a function foo with a parameter x_NK, the dimension N is only valid within the scope of foo. If you define another function bar with a parameter y_NL, this dimension N can differ from the one in foo, but it is only valid within the scope of bar.

Examples:

@sizecheck function matrix_multiply(a_NK, b_KM)
    result_NM = a_NK * b_KM
    return result_NM, N, K, M  # Dimension variables accessible
end

# This works fine
a_NK = randn(3, 4)  # N=3, K=4
b_KM = randn(4, 5)  # K=4, M=5
result = matrix_multiply(a_NK, b_KM)  # size: (3, 5)

# This raises an error
a_NK = randn(3, 4)
b_KM = randn(5, 6)  # Wrong! K dimensions don't match
result = matrix_multiply(a_NK, b_KM)  # Error!

@sizecheck function with_constants(data_N3, weights_3K)
    # data_N3: first dim variable N, second dim exactly 3
    # weights_3K: first dim exactly 3, second dim variable K
    result_NK = data_N3 * weights_3K
    return result_NK, N, K
end

source