Testing

JLNN uses a comprehensive test suite to ensure the mathematical correctness of Łukasiewicz logic operations, the stability of neural modules, and the reliability of export pipelines.

Running Tests Locally

To run the tests, you need to install the development dependencies. It is recommended to use a virtual environment.

# Install JLNN with test dependencies
pip install -e ".[test]"

# Run the full suite using pytest
python -m pytest tests -vv

Test Suite Structure

The test suite is organized into several modules, mirroring the library structure:

  • Core Logic (tests/core/): Verifies interval arithmetic and the fundamental Łukasiewicz t-norms/t-conorms.

  • Neural Components (tests/nn/): Ensures Flax NNX modules (gates, predicates) behave correctly during forward passes and handle constraints.

  • Export Pipelines (tests/export/): Validates the consistency of models when exported to:

    • StableHLO: Precision-matching between JAX and OpenXLA.

    • ONNX: Structural integrity of the generated computation graphs.

    • PyTorch: Numerical parity after JAX → ONNX → PyTorch conversion.

  • Symbolic Reasoning (tests/symbolic/): Tests the parser and compiler for translating logical strings into neural architectures.

  • Storage & Utilities (tests/storage/, tests/utils/): Verifies checkpointing, metadata extraction, and Xarray integration.

Continuous Integration

We use GitHub Actions to automatically run the entire test suite on every push and pull request to the main branch.

The CI environment mirrors the production setup using: * Ubuntu Latest * Python 3.12 * CPU-based JAX and PyTorch runtimes

Automated Verification

For critical operations like PyTorch conversion, JLNN includes built-in verification utilities:

from jlnn.export.torch_map import verify_pytorch_conversion

# This utility (tested in tests/export/test_torch_map.py)
# automatically compares JAX and PyTorch outputs.
results = verify_pytorch_conversion(jax_model, torch_model, sample_input)
print(f"Max difference: {results['max_diff']}")