[Proposal] KAGGLE COMPETITION (G2Net Detecting Continuous Gravitational Waves)

GPU Support Proposals
#1

Title: KAGGLE COMPETITION (G2Net Detecting Continuous Gravitational Waves)

Author: lkl, lkl#0997

Date posted: 2022/10/27

Summary

  • In the Kaggle competition, I want to train huge model using GPU to get a high score.
  • I want to publish one or more models so that AIN DAO can be advertisedment on Kaggle.

Background

Larger models can be trained with larger memory than when training at the 16GB limit, such as the V100 or P100 borrowed from existing colabs, which is expected to lead to higher scores.

Kaggle competition link: G2Net Detecting Continuous Gravitational Waves | Kaggle

Scope of Work

Achieving a score of 0.660 or higher, exceeding the top 5% grade of 0.655. (2022-10-27, 15/308)
Publish one or more trained models, stating that resources were supported by AIN DAO.

Timeline

1W (Nov 3th, 2022): Train model
2W (Nov 8th, 2022): Publish model and inerence code on Kaggle

Specification

Required Hardware Specifications

  • A100@40GB * 4
  • 200GB system memory

Operation:

  • Train 10+ models.
  • Publish best model.

Targets

Achieving a score of 0.660 or higher.

Participants

Kaggle individual participants interested in AI

Voting

1st voting: 2022-10-27 02:20 ~ 2022-10-29 03:00 on AIN DAO (discord)
2nd voting: on Snapshot when 3 more people participated in the voting.

1 Like
#2

Sorry for the late update. This proposal was accepted. Please share your experience when it’s completed!

  1. Discord
  2. Snapshot
#3

model published here: G2NET large kernel inference | Kaggle

The model trained using AIN DAO resources was ranked #7 on the public leaderboard at the time of publication.

Thanks to the support of the resource, I was able to conduct several experiments, and during experiments, I found the performance of the model could be improved by using a large kernel conv stem.

Thanks again for lending equipment for this proposal.

1 Like