• no external data • data augmentation ◦ pink noise, gaussian noise, gaussian snr, gain (volume adjustment) • model ◦ sedでバックボーンをdensenet121 ◦ attentionのclampをtanhに • training ◦ mixup, specAugmetation ◦ trainining on 30 second clip ◦ bce loss & secondary label • threshold ◦ framewise_output, clipwise_outputと もに0.3 ◦ framewise, clipwiseともに現れた鳥を 予測に ◦ specAugした10TTA • CV vs LB ◦ local CVできなかったのでLBを参考に した • ensemble ◦ LB scoreを参考に13modelの4vote https://www.kaggle.com/c/birdsong-recognition/discussion/183208