Private LB: 0.825 (13th) 2D UNet (Efficientnet-b3) (CV: 0.763) 2D UNet (CV: 0.786) CenterNet (CV: 0.788) 1D UNet (CV: 0.801) 1D LSTM with Wavelet transform (CV: 0.786) 2D UNet (EfficientNetV2-S) (CV: 0.780) 1D UNet (CV: 0.778) 1D UNet + WaveNet (CV: 0.774) Duplicate flag feature Training Techniques Post Processing Details anglez enmo hour(sin,cos) duplicate flag diff lead Pipeline Overview Tuned duration and downsample rate by each models score_th=0.001 distance = 70 apply following post processing (CV:0.821/public LB: 0.781) 1. 12step unit based pp 2. tolerance based pp 3. remove wakeup event at the beginning of each series 4. remove non pair high peak event 5. score decay at the ending of each series Post processing Find peak Detect duplicate (artificial) wave by 15 min interval and create flag feature weather step is duplicate wave or not. CV and LB were improved by about +0.005~+0.01. 1D UNet +WaveNet (CV: 0.765) tolerance based pp (cv+0.005) pp to bring predicted events in tolerance 12~36 within tolerance 12. Place the score-decayed prediction, -23 and +23 step away from the high peak prediction(score > 0.2). CV:0.812 Team: ricchan Dataset ensemble & low pass filter Multi task learning model sleep state (asleep/awake) onset &wakeup event classification by BCEWithLogitsLoss sleep state → binary label event → gaussian label regression by L1Loss onset → 1 wakeup → -1 other → 0 sleep state diff step 0 1 duplicate wave ◾ other effective techniques warmup, large negative sampling step high peak pred(score > 0.2) -23 23 ×1/50 ×1/50 pred event score step 12 step unit based pp (cv+0.003) if the predicted step is a multiple of 12, the step is shifted by -1 or +1. original pred shifted pred -1 1