{
  "id": 516970,
  "title": "what would be the problem?",
  "url": "/competitions/leap-atmospheric-physics-ai-climsim/discussion/516970",
  "author_name": "Simon Beck",
  "post_date": "2024-07-04T10:24:00.957000",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4921379%2Faf9bcecd4bb2a22ee599877eb8573682%2Fasfasf.PNG?generation=1720088605638519&amp;alt=media\"></p>\n<p>somehow train loss and valid scores are not getting better as the epoch goes up…</p>",
  "messages": [
    {
      "id": 2904704,
      "postDate": "2024-07-04T14:27:31.363Z",
      "content": "<p>It's hard to be 100% sure without the source code, but it seems that the signal inside your model vanishes really fast this is due to bugs/issues inside your model.</p>\n<p>Here are some possible causes:</p>\n<ul>\n<li>you are normalizing/regularizing way too much</li>\n<li>somewhere inside your model there is a layer/operation that's not working as intended (e.g. incorrect reshaping or transposing)</li>\n<li>If you are using pytorch, consider to test your architecture with tensorflow, maybe you have an issue in the model training code (e.g. back propagation isn't working due to bugs or non differentiable operations)</li>\n</ul>",
      "rawMarkdown": "It's hard to be 100% sure without the source code, but it seems that the signal inside your model vanishes really fast this is due to bugs/issues inside your model.\n\nHere are some possible causes:\n- you are normalizing/regularizing way too much\n- somewhere inside your model there is a layer/operation that's not working as intended (e.g. incorrect reshaping or transposing)\n- If you are using pytorch, consider to test your architecture with tensorflow, maybe you have an issue in the model training code (e.g. back propagation isn't working due to bugs or non differentiable operations)",
      "votes": 2
    },
    {
      "id": 2904365,
      "postDate": "2024-07-04T10:24:00.957Z",
      "content": "<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4921379%2Faf9bcecd4bb2a22ee599877eb8573682%2Fasfasf.PNG?generation=1720088605638519&amp;alt=media\"></p>\n<p>somehow train loss and valid scores are not getting better as the epoch goes up…</p>",
      "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4921379%2Faf9bcecd4bb2a22ee599877eb8573682%2Fasfasf.PNG?generation=1720088605638519&alt=media)\n\nsomehow train loss and valid scores are not getting better as the epoch goes up...\n",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 2904704,
      "author_name": "Andrey Chubin",
      "author_url": "",
      "post_date": "2024-07-04T14:27:31.363000",
      "content": "<p>It's hard to be 100% sure without the source code, but it seems that the signal inside your model vanishes really fast this is due to bugs/issues inside your model.</p>\n<p>Here are some possible causes:</p>\n<ul>\n<li>you are normalizing/regularizing way too much</li>\n<li>somewhere inside your model there is a layer/operation that's not working as intended (e.g. incorrect reshaping or transposing)</li>\n<li>If you are using pytorch, consider to test your architecture with tensorflow, maybe you have an issue in the model training code (e.g. back propagation isn't working due to bugs or non differentiable operations)</li>\n</ul>",
      "votes": 2,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2904704": "It's hard to be 100% sure without the source code, but it seems that the signal inside your model vanishes really fast this is due to bugs/issues inside your model.\n\nHere are some possible causes:\n- you are normalizing/regularizing way too much\n- somewhere inside your model there is a layer/operation that's not working as intended (e.g. incorrect reshaping or transposing)\n- If you are using pytorch, consider to test your architecture with tensorflow, maybe you have an issue in the model training code (e.g. back propagation isn't working due to bugs or non differentiable operations)",
    "2904365": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4921379%2Faf9bcecd4bb2a22ee599877eb8573682%2Fasfasf.PNG?generation=1720088605638519&alt=media)\n\nsomehow train loss and valid scores are not getting better as the epoch goes up...\n"
  }
}