{
  "id": 391727,
  "title": "49th silver, 2.5DCNN solution, public0.736, private0.722 (shakedown from 39 to 49) ",
  "url": "/competitions/nfl-player-contact-detection/discussion/391727",
  "author_name": "suguuuuu",
  "post_date": "2023-03-02T13:33:14.430000",
  "votes": 14,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Thanks to NFL and Kaggle for hosting this interesting competition!!!!<br>\nVery enjoyable competition!<br>\nIm happy to get the my first medal!!</p>\n<ul>\n<li>SUMMARY<ul>\n<li>My solution is based on this notebook. Thank you <a href=\"https://www.kaggle.com/MS-05\" target=\"_blank\">@MS-05</a> ざこ .<ul>\n<li>base : <a href=\"https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/\" target=\"_blank\">https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/</a></li></ul></li>\n<li>I changed some points. The main point is changed to the multitask learning. It makes Learning stabilized and performance improved.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F5bb1b107a5215158f389e398f0702775%2Fimg1.png?generation=1677762394664278&amp;alt=media\" alt=\"\"></li></ul></li>\n</ul>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fe5786238b234c3fd8d26b40d95bb9da2%2F2023-03-02%20220849.png?generation=1677762548232928&amp;alt=media\" alt=\"\"></p>\n<ul>\n<li><p>scores transitions</p>\n<ul>\n<li>1: Baseline: CV0.721, Private 0.691 </li>\n<li>2: Insert TSM :CV 0.743, Private 0.709 </li>\n<li>3: Multitask learning :CV 0.755, Private 0.722</li></ul></li>\n<li><p>preprocessing </p>\n<ul>\n<li>filter data with distance &lt; 2</li>\n<li>Split dataset to  \"player &amp; player\" and  \"player &amp; ground\"</li>\n<li>Crop images around helmet with helmet size. (cropsize = helmet width * 5)</li>\n<li>Oversampling of contact timing data.. (x8)</li></ul></li>\n<li><p>CV</p>\n<ul>\n<li>group K fold. </li>\n<li>Key : game-play</li>\n<li>use only 1 fold</li></ul></li>\n<li><p>modeling</p>\n<ul>\n<li><p>Backbone:<br>\n** Efficientnet B0, B1 (from scratchl.)<br>\n** Insert TSM module ( using camaro-san's solution : <a href=\"https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion/360236\" target=\"_blank\">DFL - Bundesliga Data Shootout | Kaggle</a>)</p></li>\n<li><p>Input <br>\n** image : input size : 128x128,    input frames :13 frames ( +-24frame, 4step), 2view(side and end)<br>\n** Meta : 18 data</p></li>\n<li><p>Output: <br>\n** Target frame Classification : Contact or not<br>\n** All frame Classification : 13frame Contact or not</p></li></ul></li>\n<li><p>postprocessing </p>\n<ul>\n<li>Simple 2models ensemble.</li>\n<li>Implementation of using multi frame estimation was not completed in time.</li></ul></li>\n<li><p>other</p>\n<ul>\n<li>Remove ffmpeg the submit notebook (Est time 7.5h=&gt;2.5h)</li>\n<li>Convert images to numpy for faster training.</li></ul></li>\n<li><p>not worked for me</p>\n<ul>\n<li>focal loss, classweight.</li>\n<li>Input heatmaps at the same time<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F4061bd86ae08d8b8e3f77a38f3052e37%2F222.png?generation=1677763154072742&amp;alt=media\" alt=\"\"></li></ul></li>\n</ul>\n<p>Sorry for my poor English!!<br>\nGGl!</p>\n<p>From Team どんど来い超常現象<br>\n何故ベストを尽くさないのか。（もっとがんばります。）</p>",
  "messages": [
    {
      "id": 2165821,
      "postDate": "2023-03-02T13:33:14.430Z",
      "content": "<p>Thanks to NFL and Kaggle for hosting this interesting competition!!!!<br>\nVery enjoyable competition!<br>\nIm happy to get the my first medal!!</p>\n<ul>\n<li>SUMMARY<ul>\n<li>My solution is based on this notebook. Thank you <a href=\"https://www.kaggle.com/MS-05\" target=\"_blank\">@MS-05</a> ざこ .<ul>\n<li>base : <a href=\"https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/\" target=\"_blank\">https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/</a></li></ul></li>\n<li>I changed some points. The main point is changed to the multitask learning. It makes Learning stabilized and performance improved.<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F5bb1b107a5215158f389e398f0702775%2Fimg1.png?generation=1677762394664278&amp;alt=media\" alt=\"\"></li></ul></li>\n</ul>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fe5786238b234c3fd8d26b40d95bb9da2%2F2023-03-02%20220849.png?generation=1677762548232928&amp;alt=media\" alt=\"\"></p>\n<ul>\n<li><p>scores transitions</p>\n<ul>\n<li>1: Baseline: CV0.721, Private 0.691 </li>\n<li>2: Insert TSM :CV 0.743, Private 0.709 </li>\n<li>3: Multitask learning :CV 0.755, Private 0.722</li></ul></li>\n<li><p>preprocessing </p>\n<ul>\n<li>filter data with distance &lt; 2</li>\n<li>Split dataset to  \"player &amp; player\" and  \"player &amp; ground\"</li>\n<li>Crop images around helmet with helmet size. (cropsize = helmet width * 5)</li>\n<li>Oversampling of contact timing data.. (x8)</li></ul></li>\n<li><p>CV</p>\n<ul>\n<li>group K fold. </li>\n<li>Key : game-play</li>\n<li>use only 1 fold</li></ul></li>\n<li><p>modeling</p>\n<ul>\n<li><p>Backbone:<br>\n** Efficientnet B0, B1 (from scratchl.)<br>\n** Insert TSM module ( using camaro-san's solution : <a href=\"https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion/360236\" target=\"_blank\">DFL - Bundesliga Data Shootout | Kaggle</a>)</p></li>\n<li><p>Input <br>\n** image : input size : 128x128,    input frames :13 frames ( +-24frame, 4step), 2view(side and end)<br>\n** Meta : 18 data</p></li>\n<li><p>Output: <br>\n** Target frame Classification : Contact or not<br>\n** All frame Classification : 13frame Contact or not</p></li></ul></li>\n<li><p>postprocessing </p>\n<ul>\n<li>Simple 2models ensemble.</li>\n<li>Implementation of using multi frame estimation was not completed in time.</li></ul></li>\n<li><p>other</p>\n<ul>\n<li>Remove ffmpeg the submit notebook (Est time 7.5h=&gt;2.5h)</li>\n<li>Convert images to numpy for faster training.</li></ul></li>\n<li><p>not worked for me</p>\n<ul>\n<li>focal loss, classweight.</li>\n<li>Input heatmaps at the same time<br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F4061bd86ae08d8b8e3f77a38f3052e37%2F222.png?generation=1677763154072742&amp;alt=media\" alt=\"\"></li></ul></li>\n</ul>\n<p>Sorry for my poor English!!<br>\nGGl!</p>\n<p>From Team どんど来い超常現象<br>\n何故ベストを尽くさないのか。（もっとがんばります。）</p>",
      "rawMarkdown": "Thanks to NFL and Kaggle for hosting this interesting competition!!!!\nVery enjoyable competition!\nIm happy to get the my first medal!!\n\n- SUMMARY\n - My solution is based on this notebook. Thank you @MS-05 ざこ .\n     - base : https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/\n - I changed some points. The main point is changed to the multitask learning. It makes Learning stabilized and performance improved.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F5bb1b107a5215158f389e398f0702775%2Fimg1.png?generation=1677762394664278&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fe5786238b234c3fd8d26b40d95bb9da2%2F2023-03-02%20220849.png?generation=1677762548232928&alt=media)\n\n- scores transitions\n - 1: Baseline: CV0.721, Private 0.691 \n - 2: Insert TSM :CV 0.743, Private 0.709 \n - 3: Multitask learning :CV 0.755, Private 0.722\n\n- preprocessing \n - filter data with distance < 2\n - Split dataset to  \"player & player\" and  \"player & ground\"\n - Crop images around helmet with helmet size. (cropsize = helmet width * 5)\n - Oversampling of contact timing data.. (x8)\n\n- CV\n - group K fold. \n - Key : game-play\n - use only 1 fold\n\n- modeling\n *  Backbone:\n ** Efficientnet B0, B1 (from scratchl.)\n ** Insert TSM module ( using camaro-san's solution : [DFL - Bundesliga Data Shootout | Kaggle](https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion/360236))\n\t\t\t\n - Input \n ** image : input size : 128x128,    input frames :13 frames ( +-24frame, 4step), 2view(side and end)\n ** Meta : 18 data\n - Output: \n ** Target frame Classification : Contact or not\n ** All frame Classification : 13frame Contact or not\n\n- postprocessing \n - Simple 2models ensemble.\n  - Implementation of using multi frame estimation was not completed in time.\n\n- other\n  - Remove ffmpeg the submit notebook (Est time 7.5h=>2.5h)\n  - Convert images to numpy for faster training.\n\n- not worked for me\n  - focal loss, classweight.\n  - Input heatmaps at the same time\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F4061bd86ae08d8b8e3f77a38f3052e37%2F222.png?generation=1677763154072742&alt=media)\n\nSorry for my poor English!!\nGGl!\n\nFrom Team どんど来い超常現象\n何故ベストを尽くさないのか。（もっとがんばります。）",
      "votes": 14
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2165821": "Thanks to NFL and Kaggle for hosting this interesting competition!!!!\nVery enjoyable competition!\nIm happy to get the my first medal!!\n\n- SUMMARY\n - My solution is based on this notebook. Thank you @MS-05 ざこ .\n     - base : https://www.kaggle.com/code/royalacecat/nfl-2-5d-cnn/\n - I changed some points. The main point is changed to the multitask learning. It makes Learning stabilized and performance improved.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F5bb1b107a5215158f389e398f0702775%2Fimg1.png?generation=1677762394664278&alt=media)\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fe5786238b234c3fd8d26b40d95bb9da2%2F2023-03-02%20220849.png?generation=1677762548232928&alt=media)\n\n- scores transitions\n - 1: Baseline: CV0.721, Private 0.691 \n - 2: Insert TSM :CV 0.743, Private 0.709 \n - 3: Multitask learning :CV 0.755, Private 0.722\n\n- preprocessing \n - filter data with distance < 2\n - Split dataset to  \"player & player\" and  \"player & ground\"\n - Crop images around helmet with helmet size. (cropsize = helmet width * 5)\n - Oversampling of contact timing data.. (x8)\n\n- CV\n - group K fold. \n - Key : game-play\n - use only 1 fold\n\n- modeling\n *  Backbone:\n ** Efficientnet B0, B1 (from scratchl.)\n ** Insert TSM module ( using camaro-san's solution : [DFL - Bundesliga Data Shootout | Kaggle](https://www.kaggle.com/competitions/dfl-bundesliga-data-shootout/discussion/360236))\n\t\t\t\n - Input \n ** image : input size : 128x128,    input frames :13 frames ( +-24frame, 4step), 2view(side and end)\n ** Meta : 18 data\n - Output: \n ** Target frame Classification : Contact or not\n ** All frame Classification : 13frame Contact or not\n\n- postprocessing \n - Simple 2models ensemble.\n  - Implementation of using multi frame estimation was not completed in time.\n\n- other\n  - Remove ffmpeg the submit notebook (Est time 7.5h=>2.5h)\n  - Convert images to numpy for faster training.\n\n- not worked for me\n  - focal loss, classweight.\n  - Input heatmaps at the same time\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F4061bd86ae08d8b8e3f77a38f3052e37%2F222.png?generation=1677763154072742&alt=media)\n\nSorry for my poor English!!\nGGl!\n\nFrom Team どんど来い超常現象\n何故ベストを尽くさないのか。（もっとがんばります。）"
  }
}