25/xx レシピ_exp05 V1 精度:0.462 ① Pretrain ds2 (only fold0) ② Finetuning ds1 w/ ① (5fold) ③ Pseudo label to ds3 – Ds3 : Pickup tile w/ average confidence >0.6、 bbox conf >0.4 ④ Finetuning ① w/ ③ & ds1(5fold) Sartorius - Cell Instance Segmentation | Kaggle Kaggle Wheatコンペからの学び 〜 物体検出コンペで当たり前に行われている(っぽい)こと - オットセイの経営日誌 (hatenablog.com) V3精度:0.494 ① Pretrain ds2 (only fold0) ② Finetuning ds1 w/ ① (5fold) ③ Pseudo label to ds3、ds2 – Pickup tile w/ average confidence >0.4Ds2: Pickup tile w/ average confidence >0.5, bbox conf > 0.5 ④ Finetuning ① w/ ③ data (only fold0) ⑤ Finetuning ④ w/ ds1 (5fold) V2 精度:0.442 ① Pretrain ds2 (only fold0) ② Finetuning ds1 w/ ① (5fold) ③ Pseudo label to ds3, ds2 – Ds3 : Pickup tile w/ average confidence >0.6、 bbox conf >0.4 – Ds2: Pickup tile w/ average confidence >0.5, bbox conf > 0.5 ④ Finetuning ① w/ ③ & ds1 (5fold) epoch: 84 metrics/mAP_0.5(M): 0.67136 epoch: 79.0 metrics/mAP_0.5(M): 0.6343 epoch: 83 metrics/mAP_0.5(M): 0.60875 epoch: 94.0 metrics/mAP_0.5(M): 0.62476 epoch: 67 metrics/mAP_0.5(M): 0.66475 average: 0.640784 V1 ①終了:model(fold0) ②開始:model(fold0)➡終了:model(fold0~4) ③開始:model(fold0~4)➡ds3疑似 ④開始:model(fold0)➡終了:model(fold0~4) V3 ①終了:model(fold0) ②開始:model(fold0)➡終了:model(fold0~4) ③開始:model(fold0~4) ➡ds3、ds2疑似 ④開始:model(fold0)➡model(fold0) ⑤提出:model(fold0)➡終了: model(fold0~4)