# 40th place solution(4 model ensemble)

First of all,I'd like to say thanks to
[@qi0239](https://www.kaggle.com/qi0239)
[@bilogo](https://www.kaggle.com/bilogo)
[@ageage](https://www.kaggle.com/ageage)
[@hsdtlp](https://www.kaggle.com/hsdtlp)
,this is our first match in kaggle,they spend much time finding suitable method and persisted till the end.
score
Public Leaderboard   0.9050
Private Leaderboard  0.9007
training details
data:  2020 data with 5fold
Image_size:  512x512
Models: efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50
Loss function: bi-tempered loss with label smooth
Optimizer: AdamW
Lr scheduler: Cosine Annealing with Warm Restarts
Augmentations: HorizontalFlip, VerticalFlip,ShiftScaleRotate,mixup,snapmix(when we using HueSaturationValue and RandomBrightnessContrast,it improve our cv,but lower lb)
TTA: ShiftScaleRotate,Flip
ensemble
we using 4 models :efficientnet b4,efficientnet b3,repvgg b2g4,se_resnext50.like other teams,single model didn't performance as good as ensemble.At first,we using efficientnet b3 & se_resnext50,by this way,we get higest public lb 0.900.when finding ensemble working,we continue test efficientnet-b4 and add it in our ensemble,in this discussion
[https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/213183](url)
,we notice repvgg model,and we tried testing type A0-B3,find type B2g4 have best performance in cv and public lb,by adding repvgg,it help our public lb +0.003.
methods that didn't work
remove noisy images,we using PHash to find same images and remove similar images,both cv and lb become lower
Pseudo-Labelling
concat avgpool feature & maxpool feature
in this match,I have learnt a lot,people here are enthusiasm,their Reply in the discussion gave me a lot of help  in the game.Best wish!