{
  "id": 587841,
  "title": "The solution ranked 68th",
  "url": "/competitions/waveform-inversion/discussion/587841",
  "author_name": "Feng",
  "post_date": "2025-07-03T02:12:48.544000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Our team’s solution leverages retraining of public models, specifically <strong>ConvNeXt</strong> and <strong>CAFormer</strong>, as detailed below:</p>\n<ul>\n<li><strong>ConvNeXt</strong> (backbone: <code>convnextv2_huge.fcmae_ft_in22k_in1k_512</code>) was trained on 4×5090 GPUs, achieving a validation MAE of <strong>22.01</strong>.</li>\n<li><strong>CAFormer</strong> (backbone: <code>caformer_b36.sail_in22k_ft_in1k_384</code>) was trained on 2×5090 GPUs, achieving a validation MAE of <strong>23.13</strong>.</li>\n<li>We performed model ensembling (average) using the above two models, and used the public CAFormer model for further validation and tuning.</li>\n<li>We also applied median ensembling of the two models, and further combined this result with the public 25.6 solution.</li>\n</ul>",
  "messages": [
    {
      "id": 3239613,
      "postDate": "2025-07-03T02:12:48.543Z",
      "content": "<p>Our team’s solution leverages retraining of public models, specifically <strong>ConvNeXt</strong> and <strong>CAFormer</strong>, as detailed below:</p>\n<ul>\n<li><strong>ConvNeXt</strong> (backbone: <code>convnextv2_huge.fcmae_ft_in22k_in1k_512</code>) was trained on 4×5090 GPUs, achieving a validation MAE of <strong>22.01</strong>.</li>\n<li><strong>CAFormer</strong> (backbone: <code>caformer_b36.sail_in22k_ft_in1k_384</code>) was trained on 2×5090 GPUs, achieving a validation MAE of <strong>23.13</strong>.</li>\n<li>We performed model ensembling (average) using the above two models, and used the public CAFormer model for further validation and tuning.</li>\n<li>We also applied median ensembling of the two models, and further combined this result with the public 25.6 solution.</li>\n</ul>",
      "rawMarkdown": "Our team’s solution leverages retraining of public models, specifically **ConvNeXt** and **CAFormer**, as detailed below:\n\n- **ConvNeXt** (backbone: `convnextv2_huge.fcmae_ft_in22k_in1k_512`) was trained on 4×5090 GPUs, achieving a validation MAE of **22.01**.\n- **CAFormer** (backbone: `caformer_b36.sail_in22k_ft_in1k_384`) was trained on 2×5090 GPUs, achieving a validation MAE of **23.13**.\n- We performed model ensembling (average) using the above two models, and used the public CAFormer model for further validation and tuning.\n- We also applied median ensembling of the two models, and further combined this result with the public 25.6 solution.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3239613": "Our team’s solution leverages retraining of public models, specifically **ConvNeXt** and **CAFormer**, as detailed below:\n\n- **ConvNeXt** (backbone: `convnextv2_huge.fcmae_ft_in22k_in1k_512`) was trained on 4×5090 GPUs, achieving a validation MAE of **22.01**.\n- **CAFormer** (backbone: `caformer_b36.sail_in22k_ft_in1k_384`) was trained on 2×5090 GPUs, achieving a validation MAE of **23.13**.\n- We performed model ensembling (average) using the above two models, and used the public CAFormer model for further validation and tuning.\n- We also applied median ensembling of the two models, and further combined this result with the public 25.6 solution."
  }
}