![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3381868%2F6d49493565d812fb14f36fded4e4e005%2Fnfl_summary.png?generation=1677776470936082&alt=media\")
Thanks to the organizers and the kaggle team for organizing the contest. EDA(match watching) was a lot of fun. Thanks to all participants for their hard work. I'll be reading and learning from your solutions!
\nAlso, thanks to the team, I could do best until the finish. Thanks @yokuyama @shimishige !
\nDuring the first half of the competition, each team member tried to create models in their own way (3D segmentaion, CenterNet , etc.), but unfortunately, the scores did not increase at all (LB score < 0.7). With 3 weeks remaining, the policy was changed to proceed on the basis of public notebooks.
\nThis is a 2-stage model of Deep Learning (2.5D CNN, Transformer) and GBDT. Each is based on two public notebooks. Thanks @zzy990106 (2.5DCNN) , @columbia2131 (GBDT ).
\nDeep is poor at the level of worrying about the correctness of the CV calculation, but it seems to have been sufficient as a feature to GBDT. deep has CV calculation with dist<2 only.\n
2.5D CNN
\nWe created a 1-class output model that predicts player contact and G in the same class, and a 2-class output model that predicts them separately. Two models were created for Endzone and Sideline, respectively, for a total of 4 models.
Transformer + LSTM
\n↓Adding Deep model predictions (especially CNN) improves the score
\n| \n | CV | \n
|---|---|
| Table only | \n0.7030 | \n
| + 2.5D CNN 1class | \n0.7540 | \n
| + 2.5D CNN 2class | \n0.7681 | \n
| + Transformer+LSTM | \n0.7683 | \n
Not work
\nThanks to the organizers and the kaggle team for organizing the contest. EDA(match watching) was a lot of fun. Thanks to all participants for their hard work. I'll be reading and learning from your solutions!
\nAlso, thanks to the team, I could do best until the finish. Thanks @yokuyama @shimishige !
\nDuring the first half of the competition, each team member tried to create models in their own way (3D segmentaion, CenterNet , etc.), but unfortunately, the scores did not increase at all (LB score < 0.7). With 3 weeks remaining, the policy was changed to proceed on the basis of public notebooks.
\nThis is a 2-stage model of Deep Learning (2.5D CNN, Transformer) and GBDT. Each is based on two public notebooks. Thanks @zzy990106 (2.5DCNN) , @columbia2131 (GBDT ).
\nDeep is poor at the level of worrying about the correctness of the CV calculation, but it seems to have been sufficient as a feature to GBDT. deep has CV calculation with dist<2 only.\n
2.5D CNN
\nWe created a 1-class output model that predicts player contact and G in the same class, and a 2-class output model that predicts them separately. Two models were created for Endzone and Sideline, respectively, for a total of 4 models.
Transformer + LSTM
\n↓Adding Deep model predictions (especially CNN) improves the score
\n| \n | CV | \n
|---|---|
| Table only | \n0.7030 | \n
| + 2.5D CNN 1class | \n0.7540 | \n
| + 2.5D CNN 2class | \n0.7681 | \n
| + Transformer+LSTM | \n0.7683 | \n
Not work
\nthanks for your sharing， may I ask you how you do the shift in the second stage?
\no Tracking data : diff, shift, product\no Deep model prob : shift, cummax, cumsum\nThank you for your comment.
\nThe implementation is as follows
gb = test.groupby([, , ])\n test[] = gb[col_name].shift(-)\n test[] = gb[col_name].shift()\nthanks for your sharing， may I ask you how you do the shift in the second stage?
\no Tracking data : diff, shift, product\no Deep model prob : shift, cummax, cumsum\nThank you for your comment.
\nThe implementation is as follows
gb = test.groupby([, , ])\n test[] = gb[col_name].shift(-)\n test[] = gb[col_name].shift()\n