VinBigData Chest X-ray Abnormalities Detection

top 6% solution

CV strategy

image classification

scheduler: cosine annealing loss: arcface loss

• EfficientNet

• NFNet

transformer

sentence-transformer

model 1: sentence-transformers/paraphrase-xlm-r-multilingual-v1

model 2 : sentence-transformers/bert-base-nli-cls-token

indonesian

model 3 : cahya/distilbert-base-indonesian

ensemble

CV best : model 1 + model 3 LB best : model 1 + model 2 + model 3

try

image classification: Vit, resnext

transformer: RoBert

etc: oof stacking, meta model

Most of the top solutions' key points were meta models, but I was sorry to didn't spend much time on meta models