{
  "id": 584978,
  "title": "TeamUp - Best CV for dataset wise [ 7 days left ]",
  "url": "/competitions/waveform-inversion/discussion/584978",
  "author_name": "",
  "post_date": "2025-06-17T08:08:29.472995600Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<table>\n<thead>\n<tr>\n<th>Dataset</th>\n<th>CV</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>CurveFault_A</td>\n<td>8.06</td>\n</tr>\n<tr>\n<td>CurveFault_B</td>\n<td>70.68</td>\n</tr>\n<tr>\n<td>CurveVel_A</td>\n<td>15.71</td>\n</tr>\n<tr>\n<td>CurveVel_B</td>\n<td>44.71</td>\n</tr>\n<tr>\n<td>FlatFault_A</td>\n<td>5.77</td>\n</tr>\n<tr>\n<td>FlatFault_B</td>\n<td>34.15</td>\n</tr>\n<tr>\n<td>FlatVel_A</td>\n<td>3.41</td>\n</tr>\n<tr>\n<td>FlatVel_B</td>\n<td>10.94</td>\n</tr>\n<tr>\n<td>Style_A</td>\n<td>31.25 (beat Bartley's cv)</td>\n</tr>\n<tr>\n<td>Style_B</td>\n<td>36.22 (beat Bartley's cv)</td>\n</tr>\n<tr>\n<td><strong>Average</strong></td>\n<td><strong>26.57</strong> at <strong>Epoch 57</strong>  --&gt; not yet submitted</td>\n</tr>\n</tbody>\n</table>\n<blockquote>\n  <p>As 1 weeks left, its good to merge teams if we have better datasets wise cv merge teams!</p>\n</blockquote>",
  "messages": [
    {
      "id": "3226063",
      "postDate": "06/17/2025 08:08:29",
      "content": "<table>\n<thead>\n<tr>\n<th>Dataset</th>\n<th>CV</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>CurveFault_A</td>\n<td>8.06</td>\n</tr>\n<tr>\n<td>CurveFault_B</td>\n<td>70.68</td>\n</tr>\n<tr>\n<td>CurveVel_A</td>\n<td>15.71</td>\n</tr>\n<tr>\n<td>CurveVel_B</td>\n<td>44.71</td>\n</tr>\n<tr>\n<td>FlatFault_A</td>\n<td>5.77</td>\n</tr>\n<tr>\n<td>FlatFault_B</td>\n<td>34.15</td>\n</tr>\n<tr>\n<td>FlatVel_A</td>\n<td>3.41</td>\n</tr>\n<tr>\n<td>FlatVel_B</td>\n<td>10.94</td>\n</tr>\n<tr>\n<td>Style_A</td>\n<td>31.25 (beat Bartley's cv)</td>\n</tr>\n<tr>\n<td>Style_B</td>\n<td>36.22 (beat Bartley's cv)</td>\n</tr>\n<tr>\n<td><strong>Average</strong></td>\n<td><strong>26.57</strong> at <strong>Epoch 57</strong>  --&gt; not yet submitted</td>\n</tr>\n</tbody>\n</table>\n<blockquote>\n  <p>As 1 weeks left, its good to merge teams if we have better datasets wise cv merge teams!</p>\n</blockquote>",
      "rawMarkdown": "| Dataset        | CV   |\n|----------------|--------|\n| CurveFault_A   | 8.06   |\n| CurveFault_B   | 70.68  |\n| CurveVel_A     | 15.71  |\n| CurveVel_B     | 44.71  |\n| FlatFault_A    | 5.77   |\n| FlatFault_B    | 34.15  |\n| FlatVel_A      | 3.41   |\n| FlatVel_B      | 10.94  |\n| Style_A        | 31.25 (beat Bartley's cv) |\n| Style_B        | 36.22 (beat Bartley's cv) |\n| **Average**    | **26.57** at **Epoch 57**  --> not yet submitted |\n\n> As 1 weeks left, its good to merge teams if we have better datasets wise cv merge teams!",
      "votes": null
    },
    {
      "id": "3226118",
      "postDate": "06/17/2025 09:11:17",
      "content": "<p>Which model did you use to achieve this result? Is it CAFormer?</p>",
      "rawMarkdown": "Which model did you use to achieve this result? Is it CAFormer?",
      "votes": null
    },
    {
      "id": "3226191",
      "postDate": "06/17/2025 11:13:55",
      "content": "<p>you can try this experiment below, i not sure if it would work.</p>\n<ol>\n<li>let's assume some family are difficult because </li>\n</ol>\n<ul>\n<li>there are some conflicting features</li>\n<li>conflicting because there is not enough parameters in the network. hence different \"labels\" are forced to be predicted using same features (of different input)</li>\n</ul>\n<ol>\n<li><p>now we have current model m</p></li>\n<li><p>we finetune m to m1 using all train samples but using weighted mae loss. here family-1 has the most weight. we want to model m1 to focus on family-1, sacrify the accuracy of others if necessary</p></li>\n<li><p>repeat for m2,m3 … weighing more only fmaily-2,3 …</p></li>\n</ol>\n<p>if that works, then you need team member that can improve specific family and also predict family label</p>",
      "rawMarkdown": "you can try this experiment below, i not sure if it would work.\n\n1. let's assume some family are difficult because \n- there are some conflicting features\n- conflicting because there is not enough parameters in the network. hence different \"labels\" are forced to be predicted using same features (of different input)\n\n2. now we have current model m\n\n3. we finetune m to m1 using all train samples but using weighted mae loss. here family-1 has the most weight. we want to model m1 to focus on family-1, sacrify the accuracy of others if necessary\n4. repeat for m2,m3 ... weighing more only fmaily-2,3 ...\n\nif that works, then you need team member that can improve specific family and also predict family label",
      "votes": null
    },
    {
      "id": "3226244",
      "postDate": "06/17/2025 12:16:29",
      "content": "<p>Thanks, i will try <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> </p>",
      "rawMarkdown": "Thanks, i will try @hengck23",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3226118,
      "author_name": "everythingbeok",
      "author_url": "",
      "post_date": "06/17/2025 09:11:17",
      "content": "<p>Which model did you use to achieve this result? Is it CAFormer?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3226191,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "06/17/2025 11:13:55",
      "content": "<p>you can try this experiment below, i not sure if it would work.</p>\n<ol>\n<li>let's assume some family are difficult because </li>\n</ol>\n<ul>\n<li>there are some conflicting features</li>\n<li>conflicting because there is not enough parameters in the network. hence different \"labels\" are forced to be predicted using same features (of different input)</li>\n</ul>\n<ol>\n<li><p>now we have current model m</p></li>\n<li><p>we finetune m to m1 using all train samples but using weighted mae loss. here family-1 has the most weight. we want to model m1 to focus on family-1, sacrify the accuracy of others if necessary</p></li>\n<li><p>repeat for m2,m3 … weighing more only fmaily-2,3 …</p></li>\n</ol>\n<p>if that works, then you need team member that can improve specific family and also predict family label</p>",
      "votes": null,
      "replies": [
        {
          "id": 3226244,
          "author_name": "seshurajup",
          "author_url": "",
          "post_date": "06/17/2025 12:16:29",
          "content": "<p>Thanks, i will try <a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3226063": "| Dataset        | CV   |\n|----------------|--------|\n| CurveFault_A   | 8.06   |\n| CurveFault_B   | 70.68  |\n| CurveVel_A     | 15.71  |\n| CurveVel_B     | 44.71  |\n| FlatFault_A    | 5.77   |\n| FlatFault_B    | 34.15  |\n| FlatVel_A      | 3.41   |\n| FlatVel_B      | 10.94  |\n| Style_A        | 31.25 (beat Bartley's cv) |\n| Style_B        | 36.22 (beat Bartley's cv) |\n| **Average**    | **26.57** at **Epoch 57**  --> not yet submitted |\n\n> As 1 weeks left, its good to merge teams if we have better datasets wise cv merge teams!",
    "3226118": "Which model did you use to achieve this result? Is it CAFormer?",
    "3226191": "you can try this experiment below, i not sure if it would work.\n\n1. let's assume some family are difficult because \n- there are some conflicting features\n- conflicting because there is not enough parameters in the network. hence different \"labels\" are forced to be predicted using same features (of different input)\n\n2. now we have current model m\n\n3. we finetune m to m1 using all train samples but using weighted mae loss. here family-1 has the most weight. we want to model m1 to focus on family-1, sacrify the accuracy of others if necessary\n4. repeat for m2,m3 ... weighing more only fmaily-2,3 ...\n\nif that works, then you need team member that can improve specific family and also predict family label",
    "3226244": "Thanks, i will try @hengck23"
  },
  "source": "meta"
}