Below you can find a outline of how to reproduce our 1D Model solution for the G2Net-Gravitational-Wave-Detection competition. All shell commands used in each step are run from current directory

1. Download kaggle competition data (assumes the Kaggle API is installed)
   1. cd ../../data/
   2. kaggle competitions download -c g2net-gravitational-wave-detection
   3. unzip -q g2net-gravitational-wave-detection
2. Generate whiten wave from competition data
   1. run notebook 1D_Model/notebooks/generate_whiten_wave.ipynb)

1. To train a single model using a config listed config.py, run python train.py --model_config <config_name>
2. To perform inference on a single model, run python infer.py --model_config <config_name> --gen_oof 1 --gen_test 1

1. very fast prediction: run notebook/stacking.ipynb

   1. uses precomputed predictions
   2. runs in 2 hours (on hardware similar to ours)

2. ordinary prediction: run reproduce_infer_test.sh

   1. uses pretrained models
   2. expect this to run for 6 hours

3. retrain models: run python reproduce_train.sh

   1. trains all models from scratch
   2. If we want to use ensemble, we need to run python reproduce_infer_oof_test.sh. Then we need to change the directory inside notebook/stacking.ipynb correspondingly to point to each oof prediction and submission file.
   3. expect this to run about a week

1 x NVIDIA Tesla A100 or 2 x NVIDIA GeForce RTX 2080Ti.

• notebooks/:
  • stacking.ipynb: for stacking, use hacking-lb method
  • SyntheticSignal.ipynb: for GW simulation
  • get_avg_w0.ipynb: for avg_w0 generation
  • generate_whiten_wave.ipynb: for whiten wave generation: ../data/1D_Model/whiten-train-w0/ and ../data/1D_Model/whiten-test-w0/
  • Richard_Models/: folder contains the original notebook for model generation from Richard
• src/:
  • augmentation.py: augmentation functions
  • config.py: Model configuration
  • dataset.py: dataset preparation
  • infer_helper.py: helper functions for inference
  • loss.py: related loss functions
  • lrfinder.py: learning rate finder class
  • models.py: interface to decide which model to use
  • models_1d.py: 1D model structure
  • models_2d.py: 2D model structure
  • models_3d.py: 3D model structure
  • optim.py: optimizer class
  • train_helper.py: helper functions for training
  • TTA.py: class for test time augmentation
  • util.py: utility functions
• infer.py: inference interface
• train.py: training interface
• reproduce_trian.sh: script file for model reproduction
• reproduce_infer.sh: script file for model prediction (OOF, test) reproduction