Shopee---Price-Match-Guarantee

We got 829th prize.
This is summary and codes.

Summary

CNN Model

NFNetl0

・img_size = 512 x 512
・activation = silu→mish
・optimizer = ranger
・loss = CurricularFace

NFNetl1

・img_size = 512 x 512
・activation = silu→mish
・optimizer = ranger
・loss = CurricularFace

Efficientnetb5

・img_size = 512 x 512
・activation = relu→mish
・optimizer = ranger
・loss = CurricularFace

Resnest101e

・img_size = 512 x 512
・activation = relu→mish
・optimizer = ranger
・loss = ArcFace
・pooling = GeM

Weighted_Averaging

・Pulic Score 0.9063
・Private Score 0.9010

Augmentaion

def get_train_transforms():
    return albumentations.Compose(
        [   albumentations.Resize(Config.IMG_SIZE,Config.IMG_SIZE,always_apply=True),
            albumentations.HorizontalFlip(p=0.5),
            albumentations.VerticalFlip(p=0.5),
            albumentations.Rotate(limit=120, p=0.8),
            albumentations.RandomBrightness(limit=(0.09, 0.6), p=0.5),
            albumentations.Normalize(mean = Config.MEAN, std = Config.STD),
            ToTensorV2(p=1.0),
        ]
    

LR Scheduler

CosineAnnealingWarmRestarts

NLP Model

Roberta-base

・input = Indo → English
・optimizer = Adam
・loss = ArcFace

paraphrase-xlm-r-multilingual-v1

・optimizer = Adam
・loss = ArcFace

distilbert-base-indonesian

・input = English → Indo
・optimizer = Adam
・loss = ArcFace

Not Worked

・model: Seresnexts、other resnests、other Efficientnets、other NFnets、regnets
・loss: AdaCos、Focalloss
・activation: relu→mish、silu→mish
・optimizer: SGD、adamw
・augmentation: Cutout、RandAugment、Autoaugment
・other: AMP、 emmbedding concat