{
  "id": 464411,
  "title": "Correlation between in_channels and score?",
  "url": "/competitions/blood-vessel-segmentation/discussion/464411",
  "author_name": "",
  "post_date": "2023-12-30T10:35:49.430861700Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I have noticed that making <code>in_channels</code> smaller (and not bigger) results in better score. <br>\nI am talking about U-Net models with the following setup:</p>\n<pre><code>smp.Unet(\n            encoder_name=CFG.backbone, \n            encoder_weights=weight,\n            in_channels=CFG.in_chans,\n            classes=CFG.target_size,\n            activation=,\n        )\n</code></pre>\n<p>Does anyone have an explanation for this? Thanks!</p>",
  "messages": [
    {
      "id": "2579791",
      "postDate": "12/30/2023 10:35:49",
      "content": "<p>I have noticed that making <code>in_channels</code> smaller (and not bigger) results in better score. <br>\nI am talking about U-Net models with the following setup:</p>\n<pre><code>smp.Unet(\n            encoder_name=CFG.backbone, \n            encoder_weights=weight,\n            in_channels=CFG.in_chans,\n            classes=CFG.target_size,\n            activation=,\n        )\n</code></pre>\n<p>Does anyone have an explanation for this? Thanks!</p>",
      "rawMarkdown": "I have noticed that making `in_channels` smaller (and not bigger) results in better score. \nI am talking about U-Net models with the following setup:\n\n```python\nsmp.Unet(\n            encoder_name=CFG.backbone, \n            encoder_weights=weight,\n            in_channels=CFG.in_chans,\n            classes=CFG.target_size,\n            activation=None,\n        )\n```\n\nDoes anyone have an explanation for this? Thanks!",
      "votes": null
    },
    {
      "id": "2579841",
      "postDate": "12/30/2023 11:37:42",
      "content": "<p>it is already mentioned here<br>\n<a href=\"https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\" target=\"_blank\">https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213</a></p>\n<p>but others may have different results</p>",
      "rawMarkdown": "it is already mentioned here\nhttps://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\n\nbut others may have different results",
      "votes": null
    },
    {
      "id": "2579963",
      "postDate": "12/30/2023 13:26:01",
      "content": "<p>Thanks for the link. 👌</p>",
      "rawMarkdown": "Thanks for the link. 👌",
      "votes": null
    },
    {
      "id": "2581258",
      "postDate": "12/31/2023 15:21:25",
      "content": "<p>So far, some of the insights that I got from this <a href=\"https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\" target=\"_blank\">great thread</a>:</p>\n<ul>\n<li>3D models are hard to train since they take more vRAM and longer to converge</li>\n<li>2.5D models can work if they aren't too deep (a few patches? between 3 and 5?)</li>\n<li>2D models seem to work best for now (patch=1) but there isn't any clear explanation for how?</li>\n</ul>",
      "rawMarkdown": "So far, some of the insights that I got from this [great thread](https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213):\n\n* 3D models are hard to train since they take more vRAM and longer to converge\n* 2.5D models can work if they aren't too deep (a few patches? between 3 and 5?)\n* 2D models seem to work best for now (patch=1) but there isn't any clear explanation for how?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2579841,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "12/30/2023 11:37:42",
      "content": "<p>it is already mentioned here<br>\n<a href=\"https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\" target=\"_blank\">https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213</a></p>\n<p>but others may have different results</p>",
      "votes": null,
      "replies": [
        {
          "id": 2579963,
          "author_name": "yassinealouini",
          "author_url": "",
          "post_date": "12/30/2023 13:26:01",
          "content": "<p>Thanks for the link. 👌</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2581258,
      "author_name": "yassinealouini",
      "author_url": "",
      "post_date": "12/31/2023 15:21:25",
      "content": "<p>So far, some of the insights that I got from this <a href=\"https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\" target=\"_blank\">great thread</a>:</p>\n<ul>\n<li>3D models are hard to train since they take more vRAM and longer to converge</li>\n<li>2.5D models can work if they aren't too deep (a few patches? between 3 and 5?)</li>\n<li>2D models seem to work best for now (patch=1) but there isn't any clear explanation for how?</li>\n</ul>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2579791": "I have noticed that making `in_channels` smaller (and not bigger) results in better score. \nI am talking about U-Net models with the following setup:\n\n```python\nsmp.Unet(\n            encoder_name=CFG.backbone, \n            encoder_weights=weight,\n            in_channels=CFG.in_chans,\n            classes=CFG.target_size,\n            activation=None,\n        )\n```\n\nDoes anyone have an explanation for this? Thanks!",
    "2579841": "it is already mentioned here\nhttps://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213\n\nbut others may have different results",
    "2579963": "Thanks for the link. 👌",
    "2581258": "So far, some of the insights that I got from this [great thread](https://www.kaggle.com/competitions/blood-vessel-segmentation/discussion/461213):\n\n* 3D models are hard to train since they take more vRAM and longer to converge\n* 2.5D models can work if they aren't too deep (a few patches? between 3 and 5?)\n* 2D models seem to work best for now (patch=1) but there isn't any clear explanation for how?"
  },
  "source": "meta"
}