# Public 69th / Private 54th solution

Acknowledgements
Thanks to Kaggle and the hosts for holding this exciting competition! I've learned many things in this competition. Also thanks to all participants especially my teammate schulta(
[@schulta](https://www.kaggle.com/schulta)
) and Chizuchizu(
[@chizuchizu](https://www.kaggle.com/chizuchizu)
)
Models
Scores of each model were as follows;
Model
CV
LB
EfficientB7ns
0.87706
0.8798
EfficientB3ns
0.87267
0.8747
1dCNN
0.87281
0.8769
swin transformer
0.86977
0.8723
Detail of each model was as follows;
CQT
EfficientNetB7ns
cpt parameter: {“sr”: 2048, “fmin”: 20, “fmax”: 500, “hop_length”: 8, “bins_per_octave”: 12, “filter_scale”: 0.7}
5fold validation
Image size: 512 x 512
Tripret Attention (
[paper](https://arxiv.org/abs/2010.03045)
)
Augmentation: shift in x-axis direction
Swin Transformer
cqt parameter {“sr”: 2048, “fmin”: 20, “fmax”: 500, “hop_length”: 8, “bins_per_octave”: 12, “filter_scale”: 0.7}
5fold validation
Image size: 384 x 384
Augmentation: shift in x-axis direction
CWT
EfficientNetB3ns
5fold validation
image size : 256 x 256
Tripret Attention
Augmentation: shift in x-axis direction
1dCNN
We used
[Public Kernel](https://www.kaggle.com/scaomath/g2net-1d-cnn-gem-pool-pytorch-train-inference)
as a baseline.Thanks to Shuhao Cao(
[@scaomath](https://www.kaggle.com/scaomath)
)
5fold validation
add one more conv layer
bandpass_params = dict(lf=25, hf=500)
Other things
Worked
SAM optimizer
Normalization (separately for each gravitational wave observatory)
align spectrograms from each gravitational wave interferometers on channel axis
Not Worked
Mix up
VQT
pretrain using SETI data
Idea
matched filtering
2 stage learning
denoising auto-encoder
If you have a question about this solution, feel free to ask!
Thank you for reading!