flamehaven-tensor-canon 0.1.3

Universal tensor validation and drift detection for PyTorch (NumPy optional).

Flamehaven Tensor Canon

"Data is not just numbers; it is a structural covenant."

Flamehaven Tensor Canon is a small library for:

• Structural validation (shapes and optional dtypes)
• Distribution drift detection (MMD-based)

PyTorch is required. NumPy support is optional.

Installation

pip install flamehaven-tensor-canon

Enable NumPy support:

pip install flamehaven-tensor-canon[numpy]

Quick start (validation)

import torch
from tensor_canon import validate

spec = "batch channels 224 224"
data = torch.randn(32, 3, 224, 224)

validate(data, spec, key="input_layer")

Drift detection

import torch
from tensor_canon.validator import engine

torch.manual_seed(42)
golden = torch.randn(100, 5)
engine.check_drift("embeddings", golden, register_golden=True)

current = golden + torch.randn_like(golden) * 0.01
drift = engine.check_drift("embeddings", current)
print("drift:", drift)

Development

pip install -e .[dev,numpy]
ruff check .
pytest -q

Docker

docker build -t tensor-canon .
docker run --rm tensor-canon

Release notes and security

• See CHANGELOG.md for release notes.
• See SECURITY.md for vulnerability reporting.

License

MIT. See LICENSE.