Mardi 9 mai 2023
- Benchopt: benchmarking optimization algorithms par Jérémy Cohen.
Have you ever had to compare several optimization algorithms for a research paper? This is often a tedious task: languages of implementation may vary (python, Julia, C...), stopping criteria and loss values may not be the same. Moreover, once all this benchmarking work is done, it would be nice to share with others transparently and seamlessly.
Benchopt (website) is a python library that addresses all these issues :
- Each benchmark is a separate repo, that anyone can run locally to benchmark for his own setup.
- Creating a benchmark within benchopt is somewhat easy and avoids bugs in comparisons. Results are highly reproducible.
- Benchopt runs the comparisons, stores the results and publishes them!
I will introduce Benchopt, show how to run an existing benchmark locally or check the results online, and how to build a custom benchmark.
- A Python client for massive data analysis and simulation with VIP par Gaël Vila.
VIP (the Virtual Imaging Platform) is a free, open source platform that provides medical imaging applications as a service. Based on distributed resources, it allows data analysis and simulation workflows to be run with the power of high-throughput computing (HTC).
This talk will show how VIP applications can be launched from any computer using a dedicated Python client.Input data can be managed either :- On VIP servers,- On the same machine as the client,- On a Girder warehouse.
Examples will show how this client can be used to perform transparent research analysis and reproducibility experiments.