semanticlens 0.1.0

A package for mechanistic understanding and validation of large AI model with SemanticLens

An open-source pytorch implementation of SemanticLens.

Installation

pip install semanticlens

Quickstart

import semanticlens as sl

# step 1 - preprocess component-visualizations

act_cv = sl.component_visualization.ActivationComponentVisualizer(
    model,
    dataset,
    layer_names,
)
act_cv.run(batch_size=128, num_workers=16)

# step 2 - compute semantic embeddings

fm = sl.foundation_models.OpenClip("hf-hub:apple/MobileCLIP-S2-OpenCLIP")
lens = sl.Lens(
    dataset=dataset,
    component_visualizer=act_cv,
    foundation_model=fm,
)
lens.compute_semantic_embeddigs(layer_names)

# step 3 - inspect, search, label, evaluate, ...

lens.label(
  ["wall", "small", "sky", "floor", "red",...],
  templates=["a natural image showing {}"],
) # ("layer3", 95): "brown vegetation" ,...

lens.search("watermark") # ("layer4", [63, 21, 362]), ...

lens.evaluate_clarity() # {"layer4" : [[0.64, 0.32, ... ]]}

Project status

Note
The project is currently under active development. 🛠️
Please expect interfaces to change.

The state of development will be updated here.

Contributing

We adhere to the PEP8 standard with a maximum line width of 120 characters.
For linting and style checks, we use ruff, configured to enforce PEP8 compliance along with additional rules.
Our basic tests are implemented using pytest.

License

BSD 3-Clause Clear License

Citation

@article{dreyer2025mechanistic,
  title={Mechanistic understanding and validation of large AI models with SemanticLens},
  author={Dreyer, Maximilian and Berend, Jim and Labarta, Tobias and Vielhaben, Johanna and Wiegand, Thomas and Lapuschkin, Sebastian and Samek, Wojciech},
  journal={arXiv preprint arXiv:2501.05398},
  year={2025}
}