{
  "id": 456096,
  "title": "Block in progress? ",
  "url": "/competitions/child-mind-institute-detect-sleep-states/discussion/456096",
  "author_name": "Cody_Null",
  "post_date": "2023-11-18T03:18:22.078000",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Since the release of the 2D unet approach that has been a game changer for the L! I have been struggling to improve scores significantly. I would expect to make progress slowly over the last month but I am definitely running low on ideas outside from a last thought I am working on. The bronze zone which is about 5% of the LB is only spread over a few points! Is anyone else experiencing this? I plan to try to also train the 2D unet myself and see if I can make any progress that way. </p>",
  "messages": [
    {
      "id": 2529197,
      "postDate": "2023-11-18T03:18:22.080Z",
      "content": "<p>Since the release of the 2D unet approach that has been a game changer for the L! I have been struggling to improve scores significantly. I would expect to make progress slowly over the last month but I am definitely running low on ideas outside from a last thought I am working on. The bronze zone which is about 5% of the LB is only spread over a few points! Is anyone else experiencing this? I plan to try to also train the 2D unet myself and see if I can make any progress that way. </p>",
      "rawMarkdown": "Since the release of the 2D unet approach that has been a game changer for the L! I have been struggling to improve scores significantly. I would expect to make progress slowly over the last month but I am definitely running low on ideas outside from a last thought I am working on. The bronze zone which is about 5% of the LB is only spread over a few points! Is anyone else experiencing this? I plan to try to also train the 2D unet myself and see if I can make any progress that way. ",
      "votes": 7
    },
    {
      "id": 2529925,
      "postDate": "2023-11-18T16:39:48.140Z",
      "content": "<p>I'm facing the same problem - I'm trying to modify 2d unet approach, but I get nearly the same score (LB .70~.72) each time.</p>\n<p>You can try using Temporal Fusion Transformer for this task. The problem is that thing requires too much memory for long sequences.</p>\n<p>You can find its implemention in <code>pytorch_forecasting</code> package and here: <a href=\"https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq\" target=\"_blank\">https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq</a></p>",
      "rawMarkdown": "I'm facing the same problem - I'm trying to modify 2d unet approach, but I get nearly the same score (LB .70~.72) each time.\n\nYou can try using Temporal Fusion Transformer for this task. The problem is that thing requires too much memory for long sequences.\n\nYou can find its implemention in `pytorch_forecasting` package and here: https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq",
      "votes": 5
    },
    {
      "id": 2532294,
      "postDate": "2023-11-20T23:20:55.480Z",
      "content": "<p>It seems that we cannot get a lot of improvement by just changing some hyperparameters or adding extra layers based on 2D Unet. I think the best architecture for this competition is still Transformer.😭 </p>",
      "rawMarkdown": "It seems that we cannot get a lot of improvement by just changing some hyperparameters or adding extra layers based on 2D Unet. I think the best architecture for this competition is still Transformer.😭 ",
      "votes": 1,
      "replies": [
        {
          "id": 2532316,
          "postDate": "2023-11-21T00:09:52.127Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2529925,
      "author_name": "Araik Tamazian",
      "author_url": "",
      "post_date": "2023-11-18T16:39:48.140000",
      "content": "<p>I'm facing the same problem - I'm trying to modify 2d unet approach, but I get nearly the same score (LB .70~.72) each time.</p>\n<p>You can try using Temporal Fusion Transformer for this task. The problem is that thing requires too much memory for long sequences.</p>\n<p>You can find its implemention in <code>pytorch_forecasting</code> package and here: <a href=\"https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq\" target=\"_blank\">https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq</a></p>",
      "votes": 5,
      "replies": []
    },
    {
      "id": 2532294,
      "author_name": "Zhecheng Li",
      "author_url": "",
      "post_date": "2023-11-20T23:20:55.480000",
      "content": "<p>It seems that we cannot get a lot of improvement by just changing some hyperparameters or adding extra layers based on 2D Unet. I think the best architecture for this competition is still Transformer.😭 </p>",
      "votes": 1,
      "replies": [
        {
          "id": 2532316,
          "author_name": "",
          "author_url": "",
          "post_date": "2023-11-21T00:09:52.127000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2529197": "Since the release of the 2D unet approach that has been a game changer for the L! I have been struggling to improve scores significantly. I would expect to make progress slowly over the last month but I am definitely running low on ideas outside from a last thought I am working on. The bronze zone which is about 5% of the LB is only spread over a few points! Is anyone else experiencing this? I plan to try to also train the 2D unet myself and see if I can make any progress that way. ",
    "2529925": "I'm facing the same problem - I'm trying to modify 2d unet approach, but I get nearly the same score (LB .70~.72) each time.\n\nYou can try using Temporal Fusion Transformer for this task. The problem is that thing requires too much memory for long sequences.\n\nYou can find its implemention in `pytorch_forecasting` package and here: https://github.com/hyeonbeenlee/TimeSeriesSeq2Seq",
    "2532294": "It seems that we cannot get a lot of improvement by just changing some hyperparameters or adding extra layers based on 2D Unet. I think the best architecture for this competition is still Transformer.😭 "
  }
}