First of all thanks to our teammates.
\nOur method is makeup by a tabular model and cnn model.
\nIn the tabular model, we are based on this model
\nsecond, we add (mean, std, max, min) from helmet trajectory data(train_player_tracking.csv) to increase LB from 0.650 to 0.684.
\nThird, we add step_rate (0.684->0.693) from the video.
\nAnd then We apply TTA to cnn model to increase LB from 0.667 to 0.671.
\nFinally, we add cnn model's prediction to the tabular model.
\nAfter doing this, we get 0.724 (public score) and 0.728 (private score).
\nOur code.
\nCheers!
First of all thanks to our teammates.
\nOur method is makeup by a tabular model and cnn model.
\nIn the tabular model, we are based on this model
\nsecond, we add (mean, std, max, min) from helmet trajectory data(train_player_tracking.csv) to increase LB from 0.650 to 0.684.
\nThird, we add step_rate (0.684->0.693) from the video.
\nAnd then We apply TTA to cnn model to increase LB from 0.667 to 0.671.
\nFinally, we add cnn model's prediction to the tabular model.
\nAfter doing this, we get 0.724 (public score) and 0.728 (private score).
\nOur code.
\nCheers!
wow, impressive emsembling of both tabuar and deeplearning, I have never seen that before, is that better than using a model that predicts on both the data and the transformed one ?
", "rawMarkdown": "wow, impressive emsembling of both tabuar and deeplearning, I have never seen that before, is that better than using a model that predicts on both the data and the transformed one ?", "votes": null }, { "id": "2165788", "postDate": "03/02/2023 13:09:19", "content": "We have thought this way too, I think it will be better. we did not have enough time to debug our cnn model(a lot of bugs), so we gave up this way.
", "rawMarkdown": "We have thought this way too, I think it will be better. we did not have enough time to debug our cnn model(a lot of bugs), so we gave up this way.", "votes": null } ], "comments": [ { "id": 2165729, "author_name": "janmpia", "author_url": "", "post_date": "03/02/2023 12:17:19", "content": "wow, impressive emsembling of both tabuar and deeplearning, I have never seen that before, is that better than using a model that predicts on both the data and the transformed one ?
", "votes": null, "replies": [ { "id": 2165788, "author_name": "yoyobar", "author_url": "", "post_date": "03/02/2023 13:09:19", "content": "We have thought this way too, I think it will be better. we did not have enough time to debug our cnn model(a lot of bugs), so we gave up this way.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2165708": "First of all thanks to our teammates.\nOur method is makeup by a tabular model and cnn model.\nIn the tabular model, we are based on [this model](https://www.kaggle.com/code/ahmedelfazouan/nfl-player-contact-detection-helmet-track-ftrs)\nsecond, we add (mean, std, max, min) from helmet trajectory data(train_player_tracking.csv) to increase LB from 0.650 to 0.684.\nThird, we add step_rate (0.684->0.693) from the video.\nAnd then We apply TTA to [cnn model](https://www.kaggle.com/code/zzy990106/nfl-2-5d-cnn-baseline-inference) to increase LB from 0.667 to 0.671.\nFinally, we add cnn model's prediction to the tabular model.\nAfter doing this, we get 0.724 (public score) and 0.728 (private score).\nOur [code](https://www.kaggle.com/code/yoyobar/cnn-with-feature/notebook).\nCheers!", "2165729": "wow, impressive emsembling of both tabuar and deeplearning, I have never seen that before, is that better than using a model that predicts on both the data and the transformed one ?", "2165788": "We have thought this way too, I think it will be better. we did not have enough time to debug our cnn model(a lot of bugs), so we gave up this way." }, "source": "meta" }