{
  "id": 77368,
  "title": "More than 4 channels?",
  "url": "/competitions/human-protein-atlas-image-classification/discussion/77368",
  "author_name": "Alexey Gavrikov",
  "post_date": "2019-01-11T23:13:20.424000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello everyone and congratulations to the winners!</p>\n\n<p>I am new to deep learning, this is my first competition, so I could not make any meaningful changes in architectures or make complex ensembles (but I learned a lot by reading the discussions, thanks to everyone who shared experimental results and observations). </p>\n\n<p>But I came up with one  simple idea. Since each sample represents the same field and channels are aligned, we can divide the target channel into the others and in some cases get interesting patterns that are not represented in raw data. \n<img src=\"https://i.ibb.co/nMzrQGC/hi1.png\" alt=\"\">\n<img src=\"https://i.ibb.co/12VhxPv/hi4.png\" alt=\"\">\n<img src=\"https://i.ibb.co/TbCsC5d/hi2.png\" alt=\"\">\n<img src=\"https://i.ibb.co/SvgfFvJ/hi3.png\" alt=\"\"></p>\n\n<p>I do not know how meaningful and correct it is in terms of noise or complexity of data. But in my opinion this provides a greater variety of features. <strong>And I'm wondering if anyone did this?</strong></p>\n\n<p>My model was single resnet34 with 6 channel input (additional divided green by blue and green by red)</p>",
  "messages": [
    {
      "id": 454647,
      "postDate": "2019-01-11T23:13:20.423Z",
      "content": "<p>Hello everyone and congratulations to the winners!</p>\n\n<p>I am new to deep learning, this is my first competition, so I could not make any meaningful changes in architectures or make complex ensembles (but I learned a lot by reading the discussions, thanks to everyone who shared experimental results and observations). </p>\n\n<p>But I came up with one  simple idea. Since each sample represents the same field and channels are aligned, we can divide the target channel into the others and in some cases get interesting patterns that are not represented in raw data. \n<img src=\"https://i.ibb.co/nMzrQGC/hi1.png\" alt=\"\">\n<img src=\"https://i.ibb.co/12VhxPv/hi4.png\" alt=\"\">\n<img src=\"https://i.ibb.co/TbCsC5d/hi2.png\" alt=\"\">\n<img src=\"https://i.ibb.co/SvgfFvJ/hi3.png\" alt=\"\"></p>\n\n<p>I do not know how meaningful and correct it is in terms of noise or complexity of data. But in my opinion this provides a greater variety of features. <strong>And I'm wondering if anyone did this?</strong></p>\n\n<p>My model was single resnet34 with 6 channel input (additional divided green by blue and green by red)</p>",
      "rawMarkdown": "Hello everyone and congratulations to the winners!\n\nI am new to deep learning, this is my first competition, so I could not make any meaningful changes in architectures or make complex ensembles (but I learned a lot by reading the discussions, thanks to everyone who shared experimental results and observations). \n\nBut I came up with one  simple idea. Since each sample represents the same field and channels are aligned, we can divide the target channel into the others and in some cases get interesting patterns that are not represented in raw data. \n![][1]\n![][2]\n![][3]\n![][4]\n\nI do not know how meaningful and correct it is in terms of noise or complexity of data. But in my opinion this provides a greater variety of features. **And I'm wondering if anyone did this?**\n\nMy model was single resnet34 with 6 channel input (additional divided green by blue and green by red)\n\n\n  [1]: https://i.ibb.co/nMzrQGC/hi1.png\n  [2]: https://i.ibb.co/12VhxPv/hi4.png\n  [3]: https://i.ibb.co/TbCsC5d/hi2.png\n  [4]: https://i.ibb.co/SvgfFvJ/hi3.png"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "454647": "Hello everyone and congratulations to the winners!\n\nI am new to deep learning, this is my first competition, so I could not make any meaningful changes in architectures or make complex ensembles (but I learned a lot by reading the discussions, thanks to everyone who shared experimental results and observations). \n\nBut I came up with one  simple idea. Since each sample represents the same field and channels are aligned, we can divide the target channel into the others and in some cases get interesting patterns that are not represented in raw data. \n![][1]\n![][2]\n![][3]\n![][4]\n\nI do not know how meaningful and correct it is in terms of noise or complexity of data. But in my opinion this provides a greater variety of features. **And I'm wondering if anyone did this?**\n\nMy model was single resnet34 with 6 channel input (additional divided green by blue and green by red)\n\n\n  [1]: https://i.ibb.co/nMzrQGC/hi1.png\n  [2]: https://i.ibb.co/12VhxPv/hi4.png\n  [3]: https://i.ibb.co/TbCsC5d/hi2.png\n  [4]: https://i.ibb.co/SvgfFvJ/hi3.png"
  }
}