CIANNA On-The-Fly

A client–server system for asynchronous, batched deep-learning inference on astronomical FITS images, built on the CIANNA framework and IVOA standards.

Overview

CIANNA On-The-Fly (CIANNA_OTF) is a client–server system designed to run asynchronous, batched inference on astronomical FITS images using the CIANNA deep-learning framework.

The system addresses a concrete operational challenge in scientific ML pipelines: deep-learning models used for radio source detection are expensive to load and initialise. Running inference naïvely — one request triggering one model load — leads to poor GPU utilisation, long response times, and fragile workflows.

CIANNA_OTF solves this by introducing an intermediate execution layer that groups compatible requests into batches, loads the model once, and processes multiple jobs together, while exposing a clean, job-oriented API to users.

The architecture is designed for Virtual Observatory interoperability, following the IVOA Universal Worker Service (UWS) pattern for asynchronous execution and returning results in the IVOA VOTable format.

Key Concepts

Job — a single inference request consisting of an XML description (parameters, model choice, region of interest) and a FITS image file. Each job is assigned a unique identifier and progresses through well-defined execution phases.

Batch — a group of jobs sharing the same model identifier and quantization mode, created automatically by the server according to configurable rules (minimum size, maximum size, waiting time).

Features

  • Asynchronous job submission and polling following IVOA UWS semantics
  • Automatic request batching for efficient GPU utilisation
  • Results returned as IVOA VOTable files
  • Server-side provenance tracking for scientific traceability
  • Three client interfaces: CLI (headless), TTY dashboard (interactive), GUI (PyQt)
  • Conda-based environment, compatible with CPU and GPU servers

Technical Details

  • Language: Python / Shell
  • License: Apache 2.0
  • Status: Active development — functional job lifecycle, stable API
  • IVOA compliance: UWS (partial), VOTable (full), PROV (initial)
  • Dependency: CIANNA framework

Context

Developed at Observatoire de Paris — LUX Team in the framework of the SKA (Square Kilometre Array) project. The system is designed to serve inference from the YOLO-CIANNA source detection pipeline on radio astronomy data, where processing volumes make manual, sequential inference impractical.

  • Cornu, D., Salomé, P., Semelin, B., …, Sainton, G., et al. (2024). YOLO-CIANNA: Galaxy detection with deep learning in radio data. Astronomy & Astrophysics, 690, A211. https://doi.org/10.1051/0004-6361/202449548

References