sound to noise ratio up to 0.5 ◦ background noise ▪ 鳴き声が含まれていない背景音をミックス ◦ modified mixup ▪ 通常のmixupと変えてラベルをそのままの強さで ◦ random clopではなくout-of-foldで正解予測確率が高い部分をクリップ • model ◦ 4 model ensemble ◦ resnext, resnest and external data • 後処理 ◦ しきい値0.5 ◦ 前後の窓のscoreも集約して確率の上位を答えとする ▪ site 1,2は上位3つ ▪ site3は音の長さによって https://www.kaggle.com/c/birdsong-recognition/discussion/183199
extraction ◦ パラメータを変えたlogmelの3チャンネル ◦ secondary labels • model ◦ efficientnetB3~B5 ◦ multi-sample dropout • training ◦ data aug ▪ gain, background noise, low frequency cutoff ◦ mixup • 4 model ensemble https://www.kaggle.com/c/birdsong-recognition/discussion/183339
(all from external data thread) ◦ pink and brown noise ◦ pitch shift ◦ low pass filtering ◦ spec augments (time and frequency masking) • energy base crop & label smoothing (secondary label) • longer input (10~30sec) https://www.kaggle.com/c/birdsong-recognition/discussion/183300