# 12th place solution [SiN Nakaism924]

First of all, I would like to thank my teammates(
[@naoism](https://www.kaggle.com/naoism)
,
[@hidehisaarai1213](https://www.kaggle.com/hidehisaarai1213)
,
[@yasufuminakama](https://www.kaggle.com/yasufuminakama)
,
[@sinpcw](https://www.kaggle.com/sinpcw)
) for competing with me. I would also like to thank kaggle and EGO for organizing an interesting competition.
I explain my approach. My teammates will add their own approaches in the comments!
Y.Nakama part (
[https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529297](https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529297)
)
SiNpcw part (
[https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529331](https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529331)
)
Naoism part (
[https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529412](https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529412)
)
Hidehisa part (
[https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529639](https://www.kaggle.com/c/g2net-gravitational-wave-detection/discussion/275356#1529639)
)
hirune part intro
I joined this competition after the SETI competition was over. This task is very similar to the SETI competition task. This task is very similar to the SETI competition task, so I reused most of the SETI code.
preprocessing
Divide the all wave form by 4.6152116213830774e-20 (max(abs) of the entire train and test data).
Use nnAudio to run CQT. Randomly select one of flattop, blackmanharris, or nuttall to the window. This can be used in TTA. Finally, I also added CWT to this.
The spectrograms are combined in the frequency direction and input to the model as a single-channel image
Resize to 384x512 before inputting into the model, and normalize the spectrogram of the entire data set in mean, std
Augmentation
Before converting to a spectrogram, mixup as follows.
x = x1 + x2
y = y1 + y2 -(y1*y2)
Randomly roll shift in time direction
train setup
model: timm tf_efficientnet_b4_ap
optimizer: Adam
Lr: 0.001
scheduler: CosineAnnealingLR
pseudo label
Using pseudo-labels showed some improvement, but not a lot.
stacking
Using other team members models, and finally stacking 137 models with NN or XGB, there was a significant score increase. Other models include 1dcnn, swin transformer and efficientnet b5-8.
Team members will explain these details later.