{
  "id": 69297,
  "title": "How to handle 4 channel input",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/69297",
  "author_name": "",
  "post_date": "2018-10-22T12:46:49.139115400Z",
  "votes": null,
  "comment_count": 9,
  "views": 0,
  "content": "<p>Hi, guys. How did you handle the 4 channel?</p>\n\n<ul>\n<li>before the pretrained model, bulid a <strong>conv layer that reduce the channel to 3</strong>?</li>\n<li><strong>two branches</strong> ? </li>\n<li>Is that this comptition similar <strong>\"fine-grained image recognition\"</strong>, the BCNN model can help?</li>\n</ul>\n\n<p>If you don't mind, you can share the idea about the <strong>design of model and the Image processing</strong> \n here, thank you.</p>",
  "messages": [
    {
      "id": "408201",
      "postDate": "10/22/2018 12:46:49",
      "content": "<p>Hi, guys. How did you handle the 4 channel?</p>\n\n<ul>\n<li>before the pretrained model, bulid a <strong>conv layer that reduce the channel to 3</strong>?</li>\n<li><strong>two branches</strong> ? </li>\n<li>Is that this comptition similar <strong>\"fine-grained image recognition\"</strong>, the BCNN model can help?</li>\n</ul>\n\n<p>If you don't mind, you can share the idea about the <strong>design of model and the Image processing</strong> \n here, thank you.</p>",
      "rawMarkdown": "Hi, guys. How did you handle the 4 channel?\n\n - before the pretrained model, bulid a **conv layer that reduce the channel to 3**?\n - **two branches** ? \n - Is that this comptition similar **\"fine-grained image recognition\"**, the BCNN model can help?\n\nIf you don't mind, you can share the idea about the **design of model and the Image processing** \n here, thank you.",
      "votes": null
    },
    {
      "id": "408289",
      "postDate": "10/22/2018 16:03:11",
      "content": "<p>I throw away the yellow channel and make a single RGB image with ImageMagick. The model then loads the RGB files.</p>",
      "rawMarkdown": "I throw away the yellow channel and make a single RGB image with ImageMagick. The model then loads the RGB files.",
      "votes": null
    },
    {
      "id": "408481",
      "postDate": "10/23/2018 01:05:05",
      "content": "<p>Thank you. In the data page, the green filter should hence be used to predict the label, and the other filters are used as references. Is that mean, we can just use the green channel image if we use single channel. Or we can throw away one channel of R, B, Y, if we use three channel?</p>",
      "rawMarkdown": "Thank you. In the data page, the green filter should hence be used to predict the label, and the other filters are used as references. Is that mean, we can just use the green channel image if we use single channel. Or we can throw away one channel of R, B, Y, if we use three channel?",
      "votes": null
    },
    {
      "id": "408517",
      "postDate": "10/23/2018 02:59:45",
      "content": "<p>Green is the protein itself. The other colors are other parts of the cell. While they are not required, they can provide useful information. </p>\n\n<p>Not that I know, but I could imagine a pattern in green being the same but having a different meaning depending on where in the cell it is. That location information would come from the other colors.</p>",
      "rawMarkdown": "Green is the protein itself. The other colors are other parts of the cell. While they are not required, they can provide useful information. \n\nNot that I know, but I could imagine a pattern in green being the same but having a different meaning depending on where in the cell it is. That location information would come from the other colors.",
      "votes": null
    },
    {
      "id": "408526",
      "postDate": "10/23/2018 03:17:54",
      "content": "<p>I used two branches (including Y channel). My model now seems to be heavy because of using two backbones. I think it is not a good solution for me. So, I am finding a simple one. </p>",
      "rawMarkdown": "I used two branches (including Y channel). My model now seems to be heavy because of using two backbones. I think it is not a good solution for me. So, I am finding a simple one.",
      "votes": null
    },
    {
      "id": "408529",
      "postDate": "10/23/2018 03:24:47",
      "content": "<blockquote>\n  <p>Green is the protein itself.   </p>\n</blockquote>\n\n<p>You are right.</p>\n\n<blockquote>\n  <p>The other colors are other parts of the cell. While they are not required, they can provide useful information.  </p>\n</blockquote>\n\n<p>I recognize that some patterns appear in other channels. They should be useful.</p>",
      "rawMarkdown": "&gt; Green is the protein itself.   \n\nYou are right.\n&gt; The other colors are other parts of the cell. While they are not required, they can provide useful information.  \n\nI recognize that some patterns appear in other channels. They should be useful.",
      "votes": null
    },
    {
      "id": "408534",
      "postDate": "10/23/2018 03:36:06",
      "content": "<p>very useful information.</p>",
      "rawMarkdown": "very useful information.",
      "votes": null
    },
    {
      "id": "408537",
      "postDate": "10/23/2018 03:37:01",
      "content": "<p>yep, two branches model is too heavy.</p>",
      "rawMarkdown": "yep, two branches model is too heavy.",
      "votes": null
    },
    {
      "id": "408638",
      "postDate": "10/23/2018 07:53:38",
      "content": "<p>Thank you for sharing. Next, I'll do the experiment.</p>",
      "rawMarkdown": "Thank you for sharing. Next, I'll do the experiment.",
      "votes": null
    },
    {
      "id": "408953",
      "postDate": "10/23/2018 16:38:49",
      "content": "<blockquote>\n  <p>before the pretrained model, bulid a conv layer that reduce the channel to 3?</p>\n</blockquote>\n\n<p>I tested this idea. It gives me same result as two branches with a half of complexities. </p>",
      "rawMarkdown": "&gt; before the pretrained model, bulid a conv layer that reduce the channel to 3?\n\nI tested this idea. It gives me same result as two branches with a half of complexities.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 408289,
      "author_name": "ldm314",
      "author_url": "",
      "post_date": "10/22/2018 16:03:11",
      "content": "<p>I throw away the yellow channel and make a single RGB image with ImageMagick. The model then loads the RGB files.</p>",
      "votes": null,
      "replies": [
        {
          "id": 408481,
          "author_name": "garybios",
          "author_url": "",
          "post_date": "10/23/2018 01:05:05",
          "content": "<p>Thank you. In the data page, the green filter should hence be used to predict the label, and the other filters are used as references. Is that mean, we can just use the green channel image if we use single channel. Or we can throw away one channel of R, B, Y, if we use three channel?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 408517,
          "author_name": "ldm314",
          "author_url": "",
          "post_date": "10/23/2018 02:59:45",
          "content": "<p>Green is the protein itself. The other colors are other parts of the cell. While they are not required, they can provide useful information. </p>\n\n<p>Not that I know, but I could imagine a pattern in green being the same but having a different meaning depending on where in the cell it is. That location information would come from the other colors.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 408529,
          "author_name": "backaggle",
          "author_url": "",
          "post_date": "10/23/2018 03:24:47",
          "content": "<blockquote>\n  <p>Green is the protein itself.   </p>\n</blockquote>\n\n<p>You are right.</p>\n\n<blockquote>\n  <p>The other colors are other parts of the cell. While they are not required, they can provide useful information.  </p>\n</blockquote>\n\n<p>I recognize that some patterns appear in other channels. They should be useful.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 408534,
          "author_name": "garybios",
          "author_url": "",
          "post_date": "10/23/2018 03:36:06",
          "content": "<p>very useful information.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 408526,
      "author_name": "backaggle",
      "author_url": "",
      "post_date": "10/23/2018 03:17:54",
      "content": "<p>I used two branches (including Y channel). My model now seems to be heavy because of using two backbones. I think it is not a good solution for me. So, I am finding a simple one. </p>",
      "votes": null,
      "replies": [
        {
          "id": 408537,
          "author_name": "garybios",
          "author_url": "",
          "post_date": "10/23/2018 03:37:01",
          "content": "<p>yep, two branches model is too heavy.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 408953,
          "author_name": "backaggle",
          "author_url": "",
          "post_date": "10/23/2018 16:38:49",
          "content": "<blockquote>\n  <p>before the pretrained model, bulid a conv layer that reduce the channel to 3?</p>\n</blockquote>\n\n<p>I tested this idea. It gives me same result as two branches with a half of complexities. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 408638,
      "author_name": "zhangeng",
      "author_url": "",
      "post_date": "10/23/2018 07:53:38",
      "content": "<p>Thank you for sharing. Next, I'll do the experiment.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "408201": "Hi, guys. How did you handle the 4 channel?\n\n - before the pretrained model, bulid a **conv layer that reduce the channel to 3**?\n - **two branches** ? \n - Is that this comptition similar **\"fine-grained image recognition\"**, the BCNN model can help?\n\nIf you don't mind, you can share the idea about the **design of model and the Image processing** \n here, thank you.",
    "408289": "I throw away the yellow channel and make a single RGB image with ImageMagick. The model then loads the RGB files.",
    "408481": "Thank you. In the data page, the green filter should hence be used to predict the label, and the other filters are used as references. Is that mean, we can just use the green channel image if we use single channel. Or we can throw away one channel of R, B, Y, if we use three channel?",
    "408517": "Green is the protein itself. The other colors are other parts of the cell. While they are not required, they can provide useful information. \n\nNot that I know, but I could imagine a pattern in green being the same but having a different meaning depending on where in the cell it is. That location information would come from the other colors.",
    "408526": "I used two branches (including Y channel). My model now seems to be heavy because of using two backbones. I think it is not a good solution for me. So, I am finding a simple one.",
    "408529": "&gt; Green is the protein itself.   \n\nYou are right.\n&gt; The other colors are other parts of the cell. While they are not required, they can provide useful information.  \n\nI recognize that some patterns appear in other channels. They should be useful.",
    "408534": "very useful information.",
    "408537": "yep, two branches model is too heavy.",
    "408638": "Thank you for sharing. Next, I'll do the experiment.",
    "408953": "&gt; before the pretrained model, bulid a conv layer that reduce the channel to 3?\n\nI tested this idea. It gives me same result as two branches with a half of complexities."
  },
  "source": "meta"
}