{
  "id": 496851,
  "title": "Why do some models have 3 channels?",
  "url": "/competitions/birdclef-2024/discussion/496851",
  "author_name": "",
  "post_date": "2024-04-22T17:03:46.388948700Z",
  "votes": 1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I'm looking through some models people used, and I keep seeing CNN that take inputs with 3 channels. Where do these 3 channels come from? I thought spectrograms were a 2D image with only 1.</p>",
  "messages": [
    {
      "id": "2768075",
      "postDate": "04/22/2024 17:03:46",
      "content": "<p>I'm looking through some models people used, and I keep seeing CNN that take inputs with 3 channels. Where do these 3 channels come from? I thought spectrograms were a 2D image with only 1.</p>",
      "rawMarkdown": "I'm looking through some models people used, and I keep seeing CNN that take inputs with 3 channels. Where do these 3 channels come from? I thought spectrograms were a 2D image with only 1.",
      "votes": null
    },
    {
      "id": "2768090",
      "postDate": "04/22/2024 17:17:53",
      "content": "<p>Pretrained computer vision models use 3 channels for the usual RGB image channels. You see them in case pretrained models are used.  This <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/494122\" target=\"_blank\">discussion and comments</a> discuss how to best map a batched spectrogram into a batched  3 channel image that can be used by these pretrained models.</p>",
      "rawMarkdown": "Pretrained computer vision models use 3 channels for the usual RGB image channels. You see them in case pretrained models are used.  This [discussion and comments](https://www.kaggle.com/competitions/birdclef-2024/discussion/494122) discuss how to best map a batched spectrogram into a batched  3 channel image that can be used by these pretrained models.",
      "votes": null
    },
    {
      "id": "2768094",
      "postDate": "04/22/2024 17:20:25",
      "content": "<p>Ah that makes sense. But if I wanted to train my own model, it would be ok to use 1 channel?</p>",
      "rawMarkdown": "Ah that makes sense. But if I wanted to train my own model, it would be ok to use 1 channel?",
      "votes": null
    },
    {
      "id": "2768107",
      "postDate": "04/22/2024 17:31:30",
      "content": "<p>You are totally free to do what you find best for you!</p>",
      "rawMarkdown": "You are totally free to do what you find best for you!",
      "votes": null
    },
    {
      "id": "2768414",
      "postDate": "04/22/2024 20:25:04",
      "content": "<p>ok, thanks for the tips</p>",
      "rawMarkdown": "ok, thanks for the tips",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2768090,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "04/22/2024 17:17:53",
      "content": "<p>Pretrained computer vision models use 3 channels for the usual RGB image channels. You see them in case pretrained models are used.  This <a href=\"https://www.kaggle.com/competitions/birdclef-2024/discussion/494122\" target=\"_blank\">discussion and comments</a> discuss how to best map a batched spectrogram into a batched  3 channel image that can be used by these pretrained models.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2768094,
          "author_name": "rfeng12",
          "author_url": "",
          "post_date": "04/22/2024 17:20:25",
          "content": "<p>Ah that makes sense. But if I wanted to train my own model, it would be ok to use 1 channel?</p>",
          "votes": null,
          "replies": [
            {
              "id": 2768107,
              "author_name": "cpmpml",
              "author_url": "",
              "post_date": "04/22/2024 17:31:30",
              "content": "<p>You are totally free to do what you find best for you!</p>",
              "votes": null,
              "replies": [
                {
                  "id": 2768414,
                  "author_name": "rfeng12",
                  "author_url": "",
                  "post_date": "04/22/2024 20:25:04",
                  "content": "<p>ok, thanks for the tips</p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2768075": "I'm looking through some models people used, and I keep seeing CNN that take inputs with 3 channels. Where do these 3 channels come from? I thought spectrograms were a 2D image with only 1.",
    "2768090": "Pretrained computer vision models use 3 channels for the usual RGB image channels. You see them in case pretrained models are used.  This [discussion and comments](https://www.kaggle.com/competitions/birdclef-2024/discussion/494122) discuss how to best map a batched spectrogram into a batched  3 channel image that can be used by these pretrained models.",
    "2768094": "Ah that makes sense. But if I wanted to train my own model, it would be ok to use 1 channel?",
    "2768107": "You are totally free to do what you find best for you!",
    "2768414": "ok, thanks for the tips"
  },
  "source": "meta"
}