SETI-Breakthrough-Listen---E.T.-Signal-Search

https://www.kaggle.com/c/seti-breakthrough-listen/leaderboard
We got 92th prize.
This is summary and codes.

Summary

Data strategy

・5fold StratifiedKFold
・Using new data(3ch)

Model

efficientnet_b0_ns

・img_size = 512 x 512
・optimizer = sam
・epoch = 20
・scheduler = GradualWarmupScheduler + CosineAnnealingLR

efficientnet_b3_ns

・img_size = 512 x 512
・optimizer = AdamP
・epoch = 17
・scheduler = GradualWarmupScheduler + CosineAnnealingLR

eca-nfnetl0

・img_size = 512 x 512
・optimizer = Ranger
・epoch = 15
・scheduler = ReduceLROnPlateau

efficientnet_v2_s

・img_size = 512 x 512
・optimizer = AdamP
・epoch = 15
・scheduler = GradualWarmupScheduler + CosineAnnealingLR

Weight optimazation

I use weight optimazation to maximize cv.
cv = 0.8824
lb = 0.77510

Some Settings

Augmentaion

def get_transforms(*, data):
    
    if data == 'train':
        return A.Compose([
            A.Resize(CFG.size, CFG.size, p=1.0),
            A.HorizontalFlip(p=0.5),
            A.VerticalFlip(p=0.5),
            A.ShiftScaleRotate(rotate_limit=0, p=0.25),
            A.OneOf([
                A.MotionBlur(p=1.0),
                A.GaussianBlur(p=1.0),
                A.GaussNoise(p=1.0),
            ], p=0.2),
            A.OneOf([
                A.OpticalDistortion(distort_limit=1.0, p=1.0),
                A.GridDistortion(num_steps=5, distort_limit=1.0, p=1.0),
                A.ElasticTransform(alpha=3, p=1.0),
            ], p=0.2),
            A.IAASharpen(p=0.25),
            A.Cutout(p=0.3),
            ToTensorV2(),
        ])

    elif data == 'valid':
        return A.Compose([
            A.Resize(CFG.size, CFG.size),
            ToTensorV2(),
        ])

・mixup(alpha=0.5)

Loss

BCEWithlogitsloss(pos_weight=1.5)

Not Worked

・Loss function(FocalCosineLoss/CrossEntropyLoss)
・LRscheduler(OneCycleLR/LambdaLR)
・Additional data(old)
・cutmix