knshnb / kaggle-g2net2-2nd-place

2nd place solution for Kaggle "G2Net Detecting Continuous Gravitational Waves"

2nd Place Solution of Kaggle G2Net2 Competition

This is the Preferred Wave's solution for G2Net Detecting Continuous Gravitational Waves.

Dataset

We have a single file input/test_real.csv that lists the test data with real noise, which we detected in a similar way as a public notebook. Please add the competition dataset under input/.

$ ls -F input
sample_submission.csv  test/  test_real.csv  train/  train_labels.csv

Usage

Our solution does not require any training. You can make predictions for the test data by simply running the following one command.

python predict.py --data_name test --config_path config/default.yaml --seed 0 --out_dir result/seed0

It saves the results under result/seed0/. You can use pred.csv as a prediction. You can see the parameters of --topk (100 by default) highest scores for each data in score.csv.

By specifying --data_name train, you can run validation on train data.

python predict.py --data_name test --config_path config/default.yaml --seed 0 --out_dir result/seed0

It took around 20 seconds to predict single data on NVIDIA V100 (=around 3 GPU hours and 2 GPU days for the execution of all train data and test data, respectively).

For higher scores

The prediction by the above command scores around 0.825 in the private leaderboard. Averaging the results of 2 seeds raises the score to around 0.828, which is enough to win 2nd place. You can increase the score to 0.832 by averaging more seeds (~5) and even to 0.836 by ensembling different configurations (config/freq4.yaml and config/freq6.yaml).

Links

• For an overview of our key ideas and detailed explanation, please also refer to 2nd Place Solution: GPU-Accelerated Random Search in Kaggle discussion.