Kaggle G2Net Gravitational Wave Detection : 2nd place solution

Solution writeup: https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275341

Instructions

1. Download data

You have to download the competition dataset from competition website,
and place the files in input/ directory.

┣ input/
┃   ┣ training_labels.csv
┃   ┣ sample_submission.csv
┃   ┣ train/
┃   ┣ test/
┃
┣ configs.py
┣ ...

(Optional:) Add your hardware configurations

# configs.py
HW_CFG = {
    'RTX3090': (16, 128, 1, 24), # CPU count, RAM amount(GB), GPU count, GPU RAM(GB)
    'A100': (9, 60, 1, 40), 
    'Your config', (128, 512, 8, 40) # add your hardware config!
}

2. Setup python environment

conda

conda env create -n kumaconda -f=environment.yaml
conda activate kumaconda

docker

WIP

3. Prepare data

Two new files – input/train.csv and input/test/.csv will be created.

python prep_data.py

(Optional:) Prepare waveform cache

Optionally you can speed up training by making waveform cache.
This is not recommend if your machine has RAM size smaller than 32GB.
input/train_cache.pickle and input/test_cache.pickle will be created.

python prep_data.py --cache

Then, add cache path to Baseline class in configs.py.

# configs.py
class Baseline:
    name = 'baseline'
    seed = 2021
    train_path = INPUT_DIR/'train.csv'
    test_path = INPUT_DIR/'test.csv'
    train_cache = INPUT_DIR/'train_cache.pickle' # here
    test_cache = INPUT_DIR/'test_cache.pickle' # here
    cv = 5

4. Train nueral network

Each experiment class has a name (e.g. name for Nspec16 is nspec_16).
Outputs of an experiment are

  • outoffolds.npy : (train size, 1) np.float32
  • predictions.npy : (cv fold, test size, 1) np.float32
  • {name}_{timestamp}.log : training log
  • foldx.pt : pytorch checkpoint

All outputs will be created in results/{name}/.

python train.py --config {experiment class}
# [Options]
# --progress_bar    : Everyone loves progress bar
# --inference       : Run inference only
# --tta             : Run test time augmentations (FlipWave)
# --limit_fold x    : Train a single fold x. You must run inference again by yourself.

5. Train neural network again (pseudo-label)

For experiments with name starting with Pseudo, you must use train_pseudo.py.
Outputs and options are the same as train.py.
Make sure the dependent experiment (see the table below) was successfully run.

python train_pseudo.py --config {experiment class}

Experiments

# Experiment Dependency Frontend Backend Input size CV Public LB Private LB
1 Pseudo06 Nspec12 CWT efficientnet-b2 256 x 512 0.8779 0.8797 0.8782
2 Pseodo07 Nspec16 CWT efficientnet-b2 128 x 1024 0.87841 0.8801 0.8787
3 Pseudo12 Nspec12arch0 CWT densenet201 256 x 512 0.87762 0.8796 0.8782
4 Pseudo13 MultiInstance04 CWT xcit-tiny-p16 384 x 768 0.87794 0.8800 0.8782
5 Pseudo14 Nspec16arch17 CWT efficientnet-b7 128 x 1024 0.87957 0.8811 0.8800
6 Pseudo18 Nspec21 CWT efficientnet-b4 256 x 1024 0.87942 0.8812 0.8797
7 Pseudo10 Nspec16spec13 CWT efficientnet-b2 128 x 1024 0.87875 0.8802 0.8789
8 Pseudo15 Nspec22aug1 WaveNet efficientnet-b2 128 x 1024 0.87846 0.8809 0.8794
9 Pseudo16 Nspec22arch2 WaveNet efficientnet-b6 128 x 1024 0.87982 0.8823 0.8807
10 Pseudo19 Nspec22arch6 WaveNet densenet201 128 x 1024 0.87831 0.8818 0.8804
11 Pseudo17 Nspec23arch3 CNN efficientnet-b6 128 x 1024 0.87982 0.8823 0.8808
12 Pseudo21 Nspec22arch7 WaveNet effnetv2-m 128 x 1024 0.87861 0.8831 0.8815
13 Pseudo22 Nspec23arch5 CNN effnetv2-m 128 x 1024 0.87847 0.8817 0.8799
14 Pseudo23 Nspec22arch12 WaveNet effnetv2-l 128 x 1024 0.87901 0.8829 0.8811
15 Pseudo24 Nspec30arch2 WaveNet efficientnet-b6 128 x 1024 0.8797 0.8817 0.8805
16 Pseudo25 Nspec25arch1 WaveNet efficientnet-b3 256 x 1024 0.87948 0.8820 0.8803
17 Pseudo26 Nspec22arch10 WaveNet resnet200d 128 x 1024 0.87791 0.881 0.8797
18 PseudoSeq04 Seq03aug3 ResNet1d-18 – 0.87663 0.8804 0.8785
19 PseudoSeq07 Seq12arch4 WaveNet – 0.87698 0.8796 0.8784
20 PseudoSeq03 Seq09 DenseNet1d-121 – 0.86826 0.8723 0.8703

GitHub

https://github.com/analokmaus/kaggle-g2net-public