Skip to content
Projects

Deyht/CIANNA has 57 GitHub stars and 4 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 10:16 UTC. Last push 2026-11-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Convolutional Interactive Artificial Neural Networks by/for Astrophysicists

Machine Learning

README.md

Original

cianna_logo

*Logo made by © Sarah E. Anderson*

 

ascl:2501.005

CIANNA - Convolutional Interactive Artificial Neural Networks by/for Astrophysicists

CIANNA is a general-purpose deep learning framework primarily developed and used for astronomical data analysis. Functionalities and optimizations are added based on relevance for astrophysical problem-solving. CIANNA can be used to build and train large neural network models for various tasks and is provided with a high-level Python interface (similar to keras, pytorch, etc.). One of the specificities of CIANNA is its custom implementation of a YOLO-inspired object detector used in the context of galaxy detection in 2D or 3D radio-astronomical data products. The framework is fully GPU-accelerated through low-level CUDA programming.

Development team
David Cornu - creator and lead dev, post-doc researcher, AI Fellow PR[AI]RIE, FR - LUX / Observatoire de Paris, PSL
Gregory Sainton - dev, AI Research engineer, FR - LUX / Observatoire de Paris
Aristide Doussot - dev, HPC Research engineer, FR - LUX / Observatoire de Paris

Preferred contact point: [email protected]

See Copyright © and License terms at the end.

 

Quick access:

 

CIANNA status update (11/2026)

  • The stable branch (now V-1.0.1.4) received a few minor updates and can be installed with the new simplified CMake installer (recommended).
  • The new CIANNA V-1.1.0.4 is available in the experimental branch. Here are a few highlights of the new functionalities:
    • Rework the Python interface. All dimension arguments can now be provided as lists and tuples in addition to the previous int np.array format. All "prev_layer" or "cond_layer" keywords now support both absolute and relative network positions. Moved many default configuration settings from the Python library to the C code.
    • Rework all CIANNA input design. Add a new INPUT layer. Move input size declaration from cnn.init to cnn.input. Support multiple input layers. Create dataset now supports providing a list of NumPy arrays corresponding to the data from each input branch (must match the number of declared input layers in the architecture). Provide support for much more complex architectures.
    • Add support for Adaptive Group Normalization (AdaGN). Added to the existing GN. Takes an extra conditioning layer and declares a hidden dense projection layer to learn gamma and beta. Y = gamma(X, C)*GN(X) + beta(X,C). Gradient propagates (accumulated) to the conditioning layer if it is not an input layer.
    • Rework CIANNA's signature pivot-bias propagation in FLAT style layers to allow more flexible architecture (multi-branch).
    • Improve bin saving format robustness across systems. Also improve partial architecture loading scheme.
    • Rework YOLO-CIANNA assignment process: simplified and removed undesired correlations between refinement steps. Also add a new "regret-based" assignment, as an alternative to best fIoU.
    • Rework merge to allow > 2 previous layers and extend previous layer type compatibility.
    • Rework Conv layer to allow cross-group parameter sharing (similar to a 4D convolution) as opt-in, enable some new architectures like "MLP-MIXER".

CIANNA status update (03/2026)

  • CIANNA now has a dedicated website: cianna.obspm.fr (thanks to the hard work of Aristide Doussot). This website aims to improve the framework's visibility and provide a more flexible interface to host content from the GitHub Wiki page. This includes installation instructions, application examples, tutorials, a dev blog, known publications, contact information, ...
  • A new "experimental" branch has been added. This branch will be used to accumulate new functionalities that have reached a reasonable level of maturity in preparation for the next CIANNA release. This branch can be used for testing and prototyping using new features and reworked interfaces. Highlights of the recently added functionalities (see the experimental patch note for more details):
    • Optimizer choice and a high-level interface to interact with them (SGD, ADAM, RMSprop). Also added decoupled weight decay for all optimizers.
    • New merge layer (Add and concatenate) and reworked gradient flow. Enables the creation of residual blocks and U-Net-style architectures.
    • Weight Exponential Moving Average (WEMA) support.
    • Reworked model saving to keep optimizer and EMA states. Enable clean model training to restart from a saved state.
    • The conv layer is now a grouped convolution. Classical convolution corresponds to a single group. Also provides access to depth-wise convolution.
    • The forward function now returns a NumPy array with the network prediction directly (writing a prediction result file is still possible, but optional).

Older status updates have been moved to cianna's dev blog

 

CIANNA application examples

Python scripts and Google-Colab-compatible notebooks are available under the examples directory for most of the following examples.

       Description - Dataset        Visualization Animation or real time
***
Classical computer vision examples
 
***
Image classification
MNIST

Top-1 accuracy ~99.3%
Net. ~LeNet-5
630000 ips @28p*
Open In Colab
mnist_expl
Image classification
Imagenet - 1000 classes

Top-1 acc ~74.7%
Top-5 acc ~91.7%
Net. ~Darknet19
740 ips @448p*
Open In Colab
imagenet_expl imagenet_vid
Object detection
COCO - 1000 classes

mAP@50 ~40.1%
COCO-mAP ~21.9%
Net. ~Darknet19
690 ips @416p*
Open In Colab
coco_expl coco_vid
Real-time on a laptop GPU
***
Astronomical dataset examples
 
***
Source detection
SKA SDC1
2D continuum

560MHz - 1000h
score 479372 pts
Net. 17 conv. layers
500 ips @512p*
Open In Colab
DOI
apparent_flux_distribution sdc1_det_anim
Not real-time here, only animated
Source detection
SKA SDC2
3D HI emission

950-1150MHz - 2000h
score 24664 pts
Net. 23 conv. layers
300 ips @64x64x256*
Open In Colab
DOI
line_flux_distribution sdc2_det_expl
Profile regression
3D Galactic extinction mapping

Net. [C5x5.12-P2-{D3072}x2-D2048-D128]
120000 ips @64p*
DOI
galmap_polar_map_disc
Face-on view of the galactic plane in a 45° "cone" toward the Carina arm (derived from the 3D map)
Per LOS prediction examples
galmap_vid

Integrated extinction skyview
integrated_ext_map
Fake galaxy generation
Based on galaxy zoo 2
Cascaded DDPM

Ensemble of U-Nets
~40M param.
A few ips @192p*
Made with the experimental branch of CIANNA
cascaded_scheme_illust
Cascading pipeline with 3 DDPM models
Generated examples
gen_real_galaxy_comp

*Images (or Inputs) per second (ips) are provided for an RTX 4090 GPU in inference using FP16C_FP32A mixed precision at the specified resolution and with the maximum batch size to saturate performance.

 

Installation

Please take a look at the system requirements and the installation instructions wiki pages.
=> A complete step-by-step installation guide for CIANNA and its dependencies from a fresh Ubuntu 20.04 is available here.

 

How to use

Please read the How to use Wiki page for a minimalistic tour of CIANNA capabilities on a simple example script and dataset.
A full description of all the Python interface functions is available as an API documentation page on the Wiki.
Please also consider consulting the Step-by-step installation guide to verify everything was installed correctly.
Several Python scripts and notebooks are provided as examples for different datasets and applications.

 

Publications

List of known publications that make use of or directly refer to the CIANNA framework.

Preferred citation method

When referring to a specific functionality or application, feel free to cite the relevant publication. In all cases, if your work makes use of any version of CIANNA, please cite the non-version-specific DOI from Zenodo 10.5281/zenodo.12806324.

 

###########################################################################

License

These files are Copyright © 2026-- David Cornu, but released under the Apache2 License.

 

Contributor License Agreement

While you are free to duplicate and modify this repository under the Apache2 License above, by being allowed to submit a contribution to this repository, you agree to the following terms:

  • You grant to the present CIANNA framework (and its Author) your copyright license to reproduce and distribute your contributions and such derivative works.

  • To the fullest extent permitted, you agree not to assert all of your "moral rights" in or relating to your contributions to the benefit of the present CIANNA framework.

  • Your contribution was created in whole or in part by you and you have the right to submit it under the open source license indicated in the LICENSE file; or the contribution is based upon previous work that, to the best of your knowledge, is covered under an appropriate open source license and you have the right to submit that work with modifications.