{
  "id": 358193,
  "title": "Update on previously presented ideas",
  "url": "/competitions/tabular-playground-series-oct-2022/discussion/358193",
  "author_name": "",
  "post_date": "2022-10-06T23:25:45.755259200Z",
  "votes": 7,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Some days ago I wrote this <a href=\"https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356916\" target=\"_blank\">post</a> about some possible feature about angles which could help model performance. </p>\n<p>I used as a basis the really well written public notebook by <a href=\"https://www.kaggle.com/paddykb\" target=\"_blank\">@paddykb</a> . In his notebook I calculated some of the proposed features and tweaked the code ( I changed the architecture while trying to fix a nan issue that I found due to a mistake during the implementation).</p>\n<p>I added 5 features for each player and removed the speed feature from each player, even removing some features I saw that the notebook ended up in Out of memory so I had to cut part of the training set and used only 8 files out of 10.</p>\n<p>Using these features and less parts of the dataset I found results not too far from the model with 100% training data and no additional features.</p>\n<p>For now I removed in my proof of concept the mirroring of the field since it would require probably changing the angles (I'm not too sure). I hope my notebook along with the original one can help partcipants to tackle this particularly engaging TPS. </p>\n<p>Everyone keep kaggling and enjoy!</p>",
  "messages": [
    {
      "id": "1975639",
      "postDate": "10/06/2022 23:25:45",
      "content": "<p>Some days ago I wrote this <a href=\"https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356916\" target=\"_blank\">post</a> about some possible feature about angles which could help model performance. </p>\n<p>I used as a basis the really well written public notebook by <a href=\"https://www.kaggle.com/paddykb\" target=\"_blank\">@paddykb</a> . In his notebook I calculated some of the proposed features and tweaked the code ( I changed the architecture while trying to fix a nan issue that I found due to a mistake during the implementation).</p>\n<p>I added 5 features for each player and removed the speed feature from each player, even removing some features I saw that the notebook ended up in Out of memory so I had to cut part of the training set and used only 8 files out of 10.</p>\n<p>Using these features and less parts of the dataset I found results not too far from the model with 100% training data and no additional features.</p>\n<p>For now I removed in my proof of concept the mirroring of the field since it would require probably changing the angles (I'm not too sure). I hope my notebook along with the original one can help partcipants to tackle this particularly engaging TPS. </p>\n<p>Everyone keep kaggling and enjoy!</p>",
      "rawMarkdown": "Some days ago I wrote this [post](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356916) about some possible feature about angles which could help model performance. \n\nI used as a basis the really well written public notebook by @paddykb . In his notebook I calculated some of the proposed features and tweaked the code ( I changed the architecture while trying to fix a nan issue that I found due to a mistake during the implementation).\n\nI added 5 features for each player and removed the speed feature from each player, even removing some features I saw that the notebook ended up in Out of memory so I had to cut part of the training set and used only 8 files out of 10.\n\nUsing these features and less parts of the dataset I found results not too far from the model with 100% training data and no additional features.\n\nFor now I removed in my proof of concept the mirroring of the field since it would require probably changing the angles (I'm not too sure). I hope my notebook along with the original one can help partcipants to tackle this particularly engaging TPS. \n\nEveryone keep kaggling and enjoy!",
      "votes": null
    },
    {
      "id": "1976149",
      "postDate": "10/07/2022 07:17:08",
      "content": "<p><a href=\"https://www.kaggle.com/code/pietromaldini1/tps-2022-10-fastai-proof-of-concept-new-featres\" target=\"_blank\">Great work</a> Pietro. </p>\n<p>If you add the code during the batching after the augmentations you will avoid issues with memory, and you don't need to worry about mirroring (as it has already happened) -- though working directly on the matrices is a little unpleasant. (See how horrible the mirroring code looks :))</p>",
      "rawMarkdown": "[Great work](https://www.kaggle.com/code/pietromaldini1/tps-2022-10-fastai-proof-of-concept-new-featres) Pietro. \n\nIf you add the code during the batching after the augmentations you will avoid issues with memory, and you don't need to worry about mirroring (as it has already happened) -- though working directly on the matrices is a little unpleasant. (See how horrible the mirroring code looks :))",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1976149,
      "author_name": "paddykb",
      "author_url": "",
      "post_date": "10/07/2022 07:17:08",
      "content": "<p><a href=\"https://www.kaggle.com/code/pietromaldini1/tps-2022-10-fastai-proof-of-concept-new-featres\" target=\"_blank\">Great work</a> Pietro. </p>\n<p>If you add the code during the batching after the augmentations you will avoid issues with memory, and you don't need to worry about mirroring (as it has already happened) -- though working directly on the matrices is a little unpleasant. (See how horrible the mirroring code looks :))</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1975639": "Some days ago I wrote this [post](https://www.kaggle.com/competitions/tabular-playground-series-oct-2022/discussion/356916) about some possible feature about angles which could help model performance. \n\nI used as a basis the really well written public notebook by @paddykb . In his notebook I calculated some of the proposed features and tweaked the code ( I changed the architecture while trying to fix a nan issue that I found due to a mistake during the implementation).\n\nI added 5 features for each player and removed the speed feature from each player, even removing some features I saw that the notebook ended up in Out of memory so I had to cut part of the training set and used only 8 files out of 10.\n\nUsing these features and less parts of the dataset I found results not too far from the model with 100% training data and no additional features.\n\nFor now I removed in my proof of concept the mirroring of the field since it would require probably changing the angles (I'm not too sure). I hope my notebook along with the original one can help partcipants to tackle this particularly engaging TPS. \n\nEveryone keep kaggling and enjoy!",
    "1976149": "[Great work](https://www.kaggle.com/code/pietromaldini1/tps-2022-10-fastai-proof-of-concept-new-featres) Pietro. \n\nIf you add the code during the batching after the augmentations you will avoid issues with memory, and you don't need to worry about mirroring (as it has already happened) -- though working directly on the matrices is a little unpleasant. (See how horrible the mirroring code looks :))"
  },
  "source": "meta"
}