kaggle-aptos2019-blindness-detection

11th place solution for APTOS 2019 Blindness Detection on Kaggle (https://www.kaggle.com/c/aptos2019-blindness-detection).

Solution

Preprocessing

I used only Ben's crop (scale radius).

Augmentation

Output image size is 256x256.

train_transform = transforms.Compose([
    transforms.Resize((288, 288)),
    transforms.RandomAffine(
        degrees=(-180, 180),
        scale=(0.8889, 1.0),
        shear=(-36, 36)),
    transforms.CenterCrop(256),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.RandomVerticalFlip(p=0.5),
    transforms.ColorJitter(contrast=(0.9, 1.1)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

1st-level models (run on local)

• Models: SE-ResNeXt50_32x4d, SE-ResNeXt101_32x4d, SENet154
• Loss: MSE
• Optimizer: SGD (momentum=0.9)
• LR scheduler: CosineAnnealingLR (lr=1e-3 -> 1e-5)
• 30 epochs
• Dataset: 2019 train dataset (5-folds cv) + 2015 dataset (like https://www.kaggle.com/c/aptos2019-blindness-detection/discussion/97860#581042)

2nd-level models (run on kernel)

• Models: SE-ResNeXt50_32x4d, SE-ResNeXt101_32x4d (1st-level models' weights)
• Loss: MSE
• Optimizer: RAdam
• LR scheduler: CosineAnnealingLR (lr=1e-3 -> 1e-5)
• 10 epochs
• Dataset: 2019 train dataset (5-folds cv) + 2019 test dataset (public + private, divided into 5 and used different data each fold. )
• Pseudo labels: weighted average of 1st-level models

Ensemble

Finally, averaged 2nd-level models' predictions.

• PublicLB: 0.826
• PrivateLB: 0.930

Train 1st-level models

To train 1st-level models, run:

python train.py --arch se_resnext50_32x4d
python train.py --arch se_resnext101_32x4d --batch_size 24
python train.py --arch senet154 --batch_size 16

Train 2nd-level models and ensemble

https://www.kaggle.com/uiiurz1/aptos-2019-14th-place-solution