Cassava-Leaf-Disease-Classification

https://www.kaggle.com/c/cassava-leaf-disease-classification/leaderboard
We got 26th prize.
This is summary and codes.

Summary

1st Stage

・5fold StratifiedKFold
・Using 2020 data
・Cretate predicted train data labels as output file

2nd Stage

・5fold StratifiedKFold
・Using 2020 data
Create new train labels by 1st stage output. Idea is below.
https://www.kaggle.com/c/plant-pathology-2020-fgvc7/discussion/154056

Model

vit_base_patch16_384

・img_size = 384 x 384
・4x TTA

efficientnet_b4_ns

・img_size = 512 x 512
・4x TTA

resnest50d_4s2x40d

・img_size = 512 x 512
・4x TTA

Weighted_Averaging

・Pulic Score 0.9063
・Private Score 0.9010

Some Settings

Augmentaion

def get_train_transforms():
    return Compose([
            RandomResizedCrop(CFG['img_size'], CFG['img_size']),
            Transpose(p=0.5),
            HorizontalFlip(p=0.5),
            VerticalFlip(p=0.5),
            ShiftScaleRotate(p=0.5),
            HueSaturationValue(hue_shift_limit=0.2, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.5),
            RandomBrightnessContrast(brightness_limit=(-0.1,0.1), contrast_limit=(-0.1, 0.1), p=0.5),
            Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),
            CoarseDropout(p=0.5),
            Cutout(p=0.5),
            ToTensorV2(p=1.0),
        ], p=1.)

def get_valid_transforms():
    return Compose([
            CenterCrop(CFG['img_size'], CFG['img_size'], p=1.),
            Resize(CFG['img_size'], CFG['img_size']),
            Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225], max_pixel_value=255.0, p=1.0),
            ToTensorV2(p=1.0),
        ], p=1.)

Loss

TaylorCrossEntropyLoss

LR Scheduler

CosineAnnealingWarmRestarts

Not Worked

・Loss function(BiTemperedLogisticLoss/FocalCosineLoss/CrossEntropyLoss)
・LRscheduler(GradualWarmupScheduler/OneCycleLR/LambdaLR)
・Additional data(2019)