Workshop Challenges
Timeline of the challenges:
Phase I starts: April 12, 2023
Phase II starts: June 12, 2023
Phase II ends: July 30, 2023
Final submission of code: Aug 01, 2023
Notification of results: Aug 05, 2023
Submission of report: Aug 09, 2023
FINAL LEADERBOARD (updated 7 June, 2023)
Track I: Budgeted Model Training challenge
The aim of this competition track is to be train a deep learning model for the task of classification with a constraint of 6 GB GPU memory limit. This implies that the maximum GPU memory usage at any point in the entire training process should not exceed 6 GB. Any solution that violates this constraint will not qualify. The training time is limited to 9 GPU hours.
The following machine configuration will be used for training your models/frameworks -
GPU : Quadro RTX 8000
GPU Mem. Limit : 6GB
Total Time Limit : 9 Hours
CPU Cores : 4
RAM : 32GB
Track II: Budgeted Model Inference challenge
This competition requires that the resultant model is light-weight. Solutions will be ranked in terms of accuracy metric and latency.
The following machine configuration will be used for inference -
GPU : Quadro RTX 8000
GPU Mem. Limit : 16GB
CPU Cores : 4
RAM : 32GB
PRIZES:
For both Tracks winning solutions will receive travel grants of up to -
1st Place - 8000 NOK (~750 USD)
2nd Place - 5500 NOK (~520 USD)
The prizes for these challenges are sponsored by Norwegian Artificial Intelligence Research Consortium (NORA).
RULES
Both the competitions will be hosted as Kaggle community competitions and the participants need to adhere to the standard rules of Kaggle. See the challenge pages for more details.
Each challenge will be organized in two phases. Phase I corresponds to the standard classification problem and the Kaggle leaderboard of the respective competitions will be used to rank the solutions. Phase II of the competitions will be using a different evaluation metric that also considers the training time/inference time. This phase will be the final phase and the leaderboard of phase two will be hosted at FINAL LEADERBOARD.
All participants would be required to submit the code for their solution.
The top 10 solutions on the final leaderboard would be required to submit their final code and a report of 2 pages describing their solution.
Workshop publication on the challenges
Based on the results and discussions around the two challenges, we will compile challenge papers that will be either submitted for inclusion in ICCV 2023 proceedings or will be submitted as a technical article to a good journal such as TMLR or IJCV.