{ "id": 391627, "title": "35th Best CV 750 Public LB 740 Single Tabular model result", "url": "/competitions/nfl-player-contact-detection/discussion/391627", "author_name": "", "post_date": "2023-03-02T03:11:53.744995900Z", "votes": 21, "comment_count": 2, "views": 0, "content": "

First of all, need to thanks our teammates.

\n

@zyhchasel and I working on tabular model training.
\nTotally we rent several 43GB RAM + 3090 GPU serve.

\n

Start from: https://www.kaggle.com/code/ahmedelfazouan/nfl-player-contact-detection-helmet-track-ftrs

\n

We find out that these following features are worked, and helped a lot:
\n \"Check 2 Players are in Same Team\" from 650 LB to 665 LB

\n

\"Check Players A is face to B, and encoding position\" from 665-670 LB

\n

\"Add Cluster number 1 and debug original code \" from 670-685 LB

\n

\"Add Cluster std feature \" from 685-700 LB

\n

\"Add tracking Rolling windows + Shifting statics feature by step\" from 700-717 LB

\n

\"Add helm rolling windows + shifting statics feature by frame\" from 717-735 LB

\n

“Add Count feature by distance, ranking feature…. etc” from 735-740LB

\n

We also try 2 Hidden layer NN, just 718 CV. TabNet just 708 cv, 1D ResNet 717 cv,
\nCatBoost have a good performance with 741 cv, but have high correlation with XGB model.

", "messages": [ { "id": "2165165", "postDate": "03/02/2023 03:11:53", "content": "

First of all, need to thanks our teammates.

\n

@zyhchasel and I working on tabular model training.
\nTotally we rent several 43GB RAM + 3090 GPU serve.

\n

Start from: https://www.kaggle.com/code/ahmedelfazouan/nfl-player-contact-detection-helmet-track-ftrs

\n

We find out that these following features are worked, and helped a lot:
\n \"Check 2 Players are in Same Team\" from 650 LB to 665 LB

\n

\"Check Players A is face to B, and encoding position\" from 665-670 LB

\n

\"Add Cluster number 1 and debug original code \" from 670-685 LB

\n

\"Add Cluster std feature \" from 685-700 LB

\n

\"Add tracking Rolling windows + Shifting statics feature by step\" from 700-717 LB

\n

\"Add helm rolling windows + shifting statics feature by frame\" from 717-735 LB

\n

“Add Count feature by distance, ranking feature…. etc” from 735-740LB

\n

We also try 2 Hidden layer NN, just 718 CV. TabNet just 708 cv, 1D ResNet 717 cv,
\nCatBoost have a good performance with 741 cv, but have high correlation with XGB model.

", "rawMarkdown": "First of all, need to thanks our teammates.\n\n@zyhchasel and I working on tabular model training.\nTotally we rent several 43GB RAM + 3090 GPU serve.\n\nStart from: https://www.kaggle.com/code/ahmedelfazouan/nfl-player-contact-detection-helmet-track-ftrs\n\nWe find out that these following features are worked, and helped a lot:\n \"Check 2 Players are in Same Team\" from 650 LB to 665 LB\n\n \"Check Players A is face to B, and encoding position\" from 665-670 LB\n\n \"Add Cluster number 1 and debug original code \" from 670-685 LB\n\n \"Add Cluster std feature \" from 685-700 LB\n\n \"Add tracking Rolling windows + Shifting statics feature by step\" from 700-717 LB\n\n \"Add helm rolling windows + shifting statics feature by frame\" from 717-735 LB\n\n “Add Count feature by distance, ranking feature.... etc” from 735-740LB\n\nWe also try 2 Hidden layer NN, just 718 CV. TabNet just 708 cv, 1D ResNet 717 cv, \nCatBoost have a good performance with 741 cv, but have high correlation with XGB model.", "votes": null }, { "id": "2165311", "postDate": "03/02/2023 05:29:10", "content": "

If you add a cnn model, you perhaps able to get the gold medal.
\nIt is because after I added the cnn model (0.671lb), my LB increased from 0.693 to 0.724.
\nI didn't even retrain the cnn model.

", "rawMarkdown": "If you add a cnn model, you perhaps able to get the gold medal.\nIt is because after I added the cnn model (0.671lb), my LB increased from 0.693 to 0.724.\nI didn't even retrain the cnn model.", "votes": null }, { "id": "2166717", "postDate": "03/03/2023 01:05:14", "content": "

Congrats, Dewei!

", "rawMarkdown": "Congrats, Dewei!", "votes": null } ], "comments": [ { "id": 2165311, "author_name": "yoyobar", "author_url": "", "post_date": "03/02/2023 05:29:10", "content": "

If you add a cnn model, you perhaps able to get the gold medal.
\nIt is because after I added the cnn model (0.671lb), my LB increased from 0.693 to 0.724.
\nI didn't even retrain the cnn model.

", "votes": null, "replies": [] }, { "id": 2166717, "author_name": "calvchen", "author_url": "", "post_date": "03/03/2023 01:05:14", "content": "

Congrats, Dewei!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2165165": "First of all, need to thanks our teammates.\n\n@zyhchasel and I working on tabular model training.\nTotally we rent several 43GB RAM + 3090 GPU serve.\n\nStart from: https://www.kaggle.com/code/ahmedelfazouan/nfl-player-contact-detection-helmet-track-ftrs\n\nWe find out that these following features are worked, and helped a lot:\n \"Check 2 Players are in Same Team\" from 650 LB to 665 LB\n\n \"Check Players A is face to B, and encoding position\" from 665-670 LB\n\n \"Add Cluster number 1 and debug original code \" from 670-685 LB\n\n \"Add Cluster std feature \" from 685-700 LB\n\n \"Add tracking Rolling windows + Shifting statics feature by step\" from 700-717 LB\n\n \"Add helm rolling windows + shifting statics feature by frame\" from 717-735 LB\n\n “Add Count feature by distance, ranking feature.... etc” from 735-740LB\n\nWe also try 2 Hidden layer NN, just 718 CV. TabNet just 708 cv, 1D ResNet 717 cv, \nCatBoost have a good performance with 741 cv, but have high correlation with XGB model.", "2165311": "If you add a cnn model, you perhaps able to get the gold medal.\nIt is because after I added the cnn model (0.671lb), my LB increased from 0.693 to 0.724.\nI didn't even retrain the cnn model.", "2166717": "Congrats, Dewei!" }, "source": "meta" }