{
  "id": 146773,
  "title": "Does color of the sample matter?",
  "url": "/competitions/prostate-cancer-grade-assessment/discussion/146773",
  "author_name": "Kaushal Shah",
  "post_date": "2020-04-28T12:44:37.757000",
  "votes": 0,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I trained a super basic DenseNet121 with colored patches of size 256*256 and got around ~0.47 but as mentioned in the competition description that biopsy samples may have different colors based on the procedure followed, I tried to train the same model with grayscale images and the score went down to 0.00.</p>\n\n<p>So I'm wondering that what might be the reason? Is the color that much important?</p>",
  "messages": [
    {
      "id": 824699,
      "postDate": "2020-04-28T14:41:46.377Z",
      "content": "<p>Image nets still need as input all RGB channels. I would just say that you have a bug in your code.</p>",
      "rawMarkdown": "Image nets still need as input all RGB channels. I would just say that you have a bug in your code.",
      "replies": [
        {
          "id": 824731,
          "postDate": "2020-04-28T15:02:25.320Z",
          "content": "<p><a href=\"/somnambwl\">@somnambwl</a> I'm not using ImageNet weights. I've trained model from scratch</p>",
          "rawMarkdown": "@somnambwl I'm not using ImageNet weights. I've trained model from scratch"
        }
      ]
    },
    {
      "id": 824544,
      "postDate": "2020-04-28T12:44:37.757Z",
      "content": "<p>I trained a super basic DenseNet121 with colored patches of size 256*256 and got around ~0.47 but as mentioned in the competition description that biopsy samples may have different colors based on the procedure followed, I tried to train the same model with grayscale images and the score went down to 0.00.</p>\n\n<p>So I'm wondering that what might be the reason? Is the color that much important?</p>",
      "rawMarkdown": "I trained a super basic DenseNet121 with colored patches of size 256*256 and got around ~0.47 but as mentioned in the competition description that biopsy samples may have different colors based on the procedure followed, I tried to train the same model with grayscale images and the score went down to 0.00.\n\nSo I'm wondering that what might be the reason? Is the color that much important?"
    }
  ],
  "comments": [
    {
      "id": 824699,
      "author_name": "somnambWl",
      "author_url": "",
      "post_date": "2020-04-28T14:41:46.377000",
      "content": "<p>Image nets still need as input all RGB channels. I would just say that you have a bug in your code.</p>",
      "votes": 0,
      "replies": [
        {
          "id": 824731,
          "author_name": "Kaushal Shah",
          "author_url": "",
          "post_date": "2020-04-28T15:02:25.320000",
          "content": "<p><a href=\"/somnambwl\">@somnambwl</a> I'm not using ImageNet weights. I've trained model from scratch</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "824699": "Image nets still need as input all RGB channels. I would just say that you have a bug in your code.",
    "824544": "I trained a super basic DenseNet121 with colored patches of size 256*256 and got around ~0.47 but as mentioned in the competition description that biopsy samples may have different colors based on the procedure followed, I tried to train the same model with grayscale images and the score went down to 0.00.\n\nSo I'm wondering that what might be the reason? Is the color that much important?"
  }
}