{
  "id": 593918,
  "title": "Why are the model structures of public notebooks almost the same? ",
  "url": "/competitions/cmi-detect-behavior-with-sensor-data/discussion/593918",
  "author_name": "",
  "post_date": "2025-07-31T10:38:23.291194400Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Is it because they are all based on the same notebook, or is this model a \"fixed template\"?<br>\nMost of them include <code>SEBlock</code>, <code>ResidualBlock</code>, <code>AttentionLayer</code>, and so on.<br>\nWho can tell me why this is😂</p>",
  "messages": [
    {
      "id": "3258856",
      "postDate": "07/31/2025 10:38:23",
      "content": "<p>Is it because they are all based on the same notebook, or is this model a \"fixed template\"?<br>\nMost of them include <code>SEBlock</code>, <code>ResidualBlock</code>, <code>AttentionLayer</code>, and so on.<br>\nWho can tell me why this is😂</p>",
      "rawMarkdown": "Is it because they are all based on the same notebook, or is this model a \"fixed template\"?\nMost of them include `SEBlock`, `ResidualBlock`, `AttentionLayer`, and so on.\nWho can tell me why this is😂",
      "votes": null
    },
    {
      "id": "3258922",
      "postDate": "07/31/2025 13:44:23",
      "content": "<blockquote>\n  <p>Is it because they are all based on the same notebook,</p>\n</blockquote>\n<p>Pretty much this. You can sort by \"Recently Created\" and see the older notebooks. IIRC it was TF arch based on the 3 blocks you mentioned which was top notebook for a while</p>",
      "rawMarkdown": "> Is it because they are all based on the same notebook,\n\nPretty much this. You can sort by \"Recently Created\" and see the older notebooks. IIRC it was TF arch based on the 3 blocks you mentioned which was top notebook for a while",
      "votes": null
    },
    {
      "id": "3261701",
      "postDate": "08/02/2025 03:34:41",
      "content": "<p>I will try different architectures, but the best accuracy comes from the <code>SE-Res-Attn model (SEBlock, ResidualBlock, AttentionLayer)</code>. I also experimented with changing the structure, such as putting the Attention layer first, the ResidualBlock last, or even starting with an RNN, but those did not work. I believe this is the most optimized architecture for this competition. This is my final answer!</p>",
      "rawMarkdown": "I will try different architectures, but the best accuracy comes from the `SE-Res-Attn model (SEBlock, ResidualBlock, AttentionLayer)`. I also experimented with changing the structure, such as putting the Attention layer first, the ResidualBlock last, or even starting with an RNN, but those did not work. I believe this is the most optimized architecture for this competition. This is my final answer!",
      "votes": null
    },
    {
      "id": "3262038",
      "postDate": "08/02/2025 19:05:39",
      "content": "<p>As others have said, probably just because someone got a really good score with it initially and people ran with it and found the same. Plus, probably historical competition have the same thing and it just works? Would be cool if you crushed everyone else with something else though.</p>",
      "rawMarkdown": "As others have said, probably just because someone got a really good score with it initially and people ran with it and found the same. Plus, probably historical competition have the same thing and it just works? Would be cool if you crushed everyone else with something else though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3258922,
      "author_name": "gerwynng",
      "author_url": "",
      "post_date": "07/31/2025 13:44:23",
      "content": "<blockquote>\n  <p>Is it because they are all based on the same notebook,</p>\n</blockquote>\n<p>Pretty much this. You can sort by \"Recently Created\" and see the older notebooks. IIRC it was TF arch based on the 3 blocks you mentioned which was top notebook for a while</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3261701,
      "author_name": "sangrampatil5150",
      "author_url": "",
      "post_date": "08/02/2025 03:34:41",
      "content": "<p>I will try different architectures, but the best accuracy comes from the <code>SE-Res-Attn model (SEBlock, ResidualBlock, AttentionLayer)</code>. I also experimented with changing the structure, such as putting the Attention layer first, the ResidualBlock last, or even starting with an RNN, but those did not work. I believe this is the most optimized architecture for this competition. This is my final answer!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3262038,
      "author_name": "msthil",
      "author_url": "",
      "post_date": "08/02/2025 19:05:39",
      "content": "<p>As others have said, probably just because someone got a really good score with it initially and people ran with it and found the same. Plus, probably historical competition have the same thing and it just works? Would be cool if you crushed everyone else with something else though.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3258856": "Is it because they are all based on the same notebook, or is this model a \"fixed template\"?\nMost of them include `SEBlock`, `ResidualBlock`, `AttentionLayer`, and so on.\nWho can tell me why this is😂",
    "3258922": "> Is it because they are all based on the same notebook,\n\nPretty much this. You can sort by \"Recently Created\" and see the older notebooks. IIRC it was TF arch based on the 3 blocks you mentioned which was top notebook for a while",
    "3261701": "I will try different architectures, but the best accuracy comes from the `SE-Res-Attn model (SEBlock, ResidualBlock, AttentionLayer)`. I also experimented with changing the structure, such as putting the Attention layer first, the ResidualBlock last, or even starting with an RNN, but those did not work. I believe this is the most optimized architecture for this competition. This is my final answer!",
    "3262038": "As others have said, probably just because someone got a really good score with it initially and people ran with it and found the same. Plus, probably historical competition have the same thing and it just works? Would be cool if you crushed everyone else with something else though."
  },
  "source": "meta"
}