🚀 Getting Started - 📚 Documentation - 💻 User guide

[!CAUTION] Alpha software: PyC is currently under active development. Public APIs may change and be unstable between releases.

PyC is a library built upon PyTorch and PyTorch Lightning to easily implement interpretable and causally transparent deep learning models. The library provides primitives for annotated tensors, interpretable layers, interventions, interpretable probabilistic graphical models, and APIs for running experiments at scale.

The name of the library stands for both

• PyTorch Concepts: as concepts are essential building blocks for interpretable deep learning.
• $P(y|C)$: as the main purpose of the library is to support sound probabilistic modeling of the conditional distribution of targets $y$ given concepts $C$.

Quick Start

Install PyC from PyPI:

pip install --pre "pytorch-concepts[data]"

Use pip install --pre pytorch-concepts for core-only (no data dependencies), or see full installation options for conda setup.

Design Principles

A few basic elements let you name concepts, compute them, train them and act on them. Here is what each one lets you do:

import torch_concepts as pyc

ann = pyc.Annotations(
    labels=["red", "shape", "size"],
    types=[
        "binary",
        "categorical",
        "continuous",
    ],
    cardinalities=[1, 3, 1],
)

x = torch.randn(8, 5)
c = pyc.AnnotatedTensor(x, ann)

c["red"]     # slice by concept
c.binary()   # slice by type

from torch_concepts.nn import (
    LinearEmbeddingToConcept,
)

layer = LinearEmbeddingToConcept(
    in_embeddings=16,
    out_concepts=ann,
)
emb = torch.randn(8, 16)
out = layer.annotate(layer(emb))

from torch import nn
from torch_concepts.nn import ConceptLoss

loss_fn = ConceptLoss(
    binary=nn.BCEWithLogitsLoss(),
    categorical=nn.CrossEntropyLoss(),
    continuous=nn.MSELoss(),
)
loss = loss_fn(out, target)

from torch_concepts.nn import DoIntervention, UniformPolicy, intervention

with intervention(layer, DoIntervention(1.0), UniformPolicy(), ["red"]):
    out = layer(emb)  # "red" is set to 1.0

The quickstart trains a first model end to end, and the user guide covers the rest of PyC.

PyC Software Stack

The library is organized to be modular and accessible at different levels of abstraction:

• Conceptarium (No-code API): applications and benchmarking. These APIs allow to easily run large-scale experiments by interfacing only with configuration files. Built on top of Hydra and WandB.
• High-level APIs: use out-of-the-box models. These APIs allow to instantiate models with 1 line of code. Models are available both as plain PyTorch modules (implementing forward, allowing custom training loops) and as PyTorch Lightning modules.
• Mid-level APIs: interpretable probabilistic graphical models. These APIs allow to define variables (concepts and embeddings), connect them via conditional distributions parametrized by interpretable layers, and perform probabilistic inference on the resulting graphical model.
• Low-level APIs: interpretable layers. These APIs allow to build architectures from basic interpretable layers in a plain PyTorch-like interface. These APIs also include annotated tensors, interventions, metrics, losses, and datasets.

Contributing

Contributions are welcome! Please check our contributing guidelines to get started.

Thanks to all contributors! 🧡

External Contributors

• Sonia Laguna, ETH Zurich (CH).
• Moritz Vandenhirtz, ETH Zurich (CH).

Cite this Library

If you found this library useful for your research article, blog post, or product, we would be grateful if you would cite it using the following bibtex entry:

@software{pycteam2025concept,
    author = {De Felice, Giovanni and Barbiero, Pietro and De Santis, Francesco and Gabrielli, Edoardo and Casanova Flores, Arianna and Ciravegna, Gabriele and Debot, David and Espinosa Zarlenga, Mateo and Jamnik, Mateja and Marra, Giuseppe},
    license = {Apache 2.0},
    month = {3},
    title = {{PyTorch Concepts}},
    url = {https://github.com/pyc-team/pytorch_concepts},
    year = {2025}
}

Reference authors: Giovanni De Felice, Pietro Barbiero, and Mateo Espinosa Zarlenga.

Funding

This project is supported by the following organizations: