{
  "id": 238466,
  "title": "Is pretrained weight useless?",
  "url": "/competitions/seti-breakthrough-listen/discussion/238466",
  "author_name": "Last Scene",
  "post_date": "2021-05-12T08:42:52.287000",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Since this is 6-channel images. And since none of them in RGB way.<br>\nSo I wonder is pretrained weight(usually trained in ImageNet in RGB way) useless(say if we input image arrays with 6 channels).</p>",
  "messages": [
    {
      "id": 1303800,
      "postDate": "2021-05-12T08:47:03.367Z",
      "content": "<p>Probably useless, yes. There's decent evidence that transfer learning to completely different fields is significantly overhyped: <a href=\"https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf\" target=\"_blank\">https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf</a></p>\n<p>The paper above also shows that networks designed for ImageNet with many many classes are probably not the best designs for different types of problems.</p>",
      "rawMarkdown": "Probably useless, yes. There's decent evidence that transfer learning to completely different fields is significantly overhyped: https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf\n\nThe paper above also shows that networks designed for ImageNet with many many classes are probably not the best designs for different types of problems.\n\n",
      "votes": 3,
      "replies": [
        {
          "id": 1303804,
          "postDate": "2021-05-12T08:52:08.847Z",
          "content": "<p>Oh, thanks~ and now I think that(load model without pretrained weight) is settled down.</p>",
          "rawMarkdown": "Oh, thanks~ and now I think that(load model without pretrained weight) is settled down."
        },
        {
          "id": 1304633,
          "postDate": "2021-05-12T18:44:55.977Z",
          "content": "<p>I also had a similar experience</p>",
          "rawMarkdown": "I also had a similar experience"
        }
      ]
    },
    {
      "id": 1303796,
      "postDate": "2021-05-12T08:42:52.287Z",
      "content": "<p>Since this is 6-channel images. And since none of them in RGB way.<br>\nSo I wonder is pretrained weight(usually trained in ImageNet in RGB way) useless(say if we input image arrays with 6 channels).</p>",
      "rawMarkdown": "Since this is 6-channel images. And since none of them in RGB way.\nSo I wonder is pretrained weight(usually trained in ImageNet in RGB way) useless(say if we input image arrays with 6 channels).",
      "votes": 1
    },
    {
      "id": 1304213,
      "postDate": "2021-05-12T13:38:37.743Z",
      "rawMarkdown": "",
      "votes": 3,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1303800,
      "author_name": "James Howard",
      "author_url": "",
      "post_date": "2021-05-12T08:47:03.367000",
      "content": "<p>Probably useless, yes. There's decent evidence that transfer learning to completely different fields is significantly overhyped: <a href=\"https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf\" target=\"_blank\">https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf</a></p>\n<p>The paper above also shows that networks designed for ImageNet with many many classes are probably not the best designs for different types of problems.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 1303804,
          "author_name": "Last Scene",
          "author_url": "",
          "post_date": "2021-05-12T08:52:08.847000",
          "content": "<p>Oh, thanks~ and now I think that(load model without pretrained weight) is settled down.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1304633,
          "author_name": "Baran Hashemi",
          "author_url": "",
          "post_date": "2021-05-12T18:44:55.977000",
          "content": "<p>I also had a similar experience</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1304213,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-05-12T13:38:37.743000",
      "content": "",
      "votes": 3,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1303800": "Probably useless, yes. There's decent evidence that transfer learning to completely different fields is significantly overhyped: https://papers.nips.cc/paper/2019/file/eb1e78328c46506b46a4ac4a1e378b91-Paper.pdf\n\nThe paper above also shows that networks designed for ImageNet with many many classes are probably not the best designs for different types of problems.\n\n",
    "1303796": "Since this is 6-channel images. And since none of them in RGB way.\nSo I wonder is pretrained weight(usually trained in ImageNet in RGB way) useless(say if we input image arrays with 6 channels).",
    "1304213": ""
  }
}