{
  "id": 71393,
  "title": "Question about splitting channels",
  "url": "/competitions/airbus-ship-detection/discussion/71393",
  "author_name": "",
  "post_date": "2018-11-13T10:28:27.018980Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<p>When I looked at previous similar challenges, I see a pattern of splitting each channel into individual classification problem/model. Is it because it’s easier for the model to just focus on a single task? Or is there other better reasons?</p>\n\n<p>Thanks!</p>",
  "messages": [
    {
      "id": "420239",
      "postDate": "11/13/2018 10:28:27",
      "content": "<p>When I looked at previous similar challenges, I see a pattern of splitting each channel into individual classification problem/model. Is it because it’s easier for the model to just focus on a single task? Or is there other better reasons?</p>\n\n<p>Thanks!</p>",
      "rawMarkdown": "When I looked at previous similar challenges, I see a pattern of splitting each channel into individual classification problem/model. Is it because it’s easier for the model to just focus on a single task? Or is there other better reasons?\n\nThanks!",
      "votes": null
    },
    {
      "id": "420254",
      "postDate": "11/13/2018 11:07:41",
      "content": "<p>Are you asking about having one net for classifying ship/no-ship and another one for segmentation? Or having for example an U-Net with multiple output channels and threat each channel individually?</p>",
      "rawMarkdown": "Are you asking about having one net for classifying ship/no-ship and another one for segmentation? Or having for example an U-Net with multiple output channels and threat each channel individually?",
      "votes": null
    },
    {
      "id": "420307",
      "postDate": "11/13/2018 12:44:03",
      "content": "<p>Hey there,\nthanks for replying!</p>\n\n<p>Yes I'm talking about \"having for example an U-Net with multiple output channels and threat each channel individually\"</p>",
      "rawMarkdown": "Hey there,\nthanks for replying!\n\nYes I'm talking about \"having for example an U-Net with multiple output channels and threat each channel individually\"",
      "votes": null
    },
    {
      "id": "420327",
      "postDate": "11/13/2018 13:07:30",
      "content": "<p>Individual prediction is useful when we must decide thresholds. We can adjust them one by one.<br>\nIn this challenge we treat only single class (ship or not ship), so we don't need to split output channels. </p>",
      "rawMarkdown": "Individual prediction is useful when we must decide thresholds. We can adjust them one by one.<br>\nIn this challenge we treat only single class (ship or not ship), so we don't need to split output channels.",
      "votes": null
    },
    {
      "id": "420337",
      "postDate": "11/13/2018 13:25:21",
      "content": "<p>It really depends.. For example if you have multiple classes makes sense to treat all of them together using softmax. </p>\n\n<p>However if you want to predict different things, it makes sense to separate the channels and treat them separately.  Example: One channel predicts the binary mask while other predicts energy transform for watershed like this paper <a href=\"https://arxiv.org/abs/1611.08303\">https://arxiv.org/abs/1611.08303</a></p>",
      "rawMarkdown": "It really depends.. For example if you have multiple classes makes sense to treat all of them together using softmax. \n\nHowever if you want to predict different things, it makes sense to separate the channels and treat them separately.  Example: One channel predicts the binary mask while other predicts energy transform for watershed like this paper https://arxiv.org/abs/1611.08303",
      "votes": null
    },
    {
      "id": "420349",
      "postDate": "11/13/2018 13:36:55",
      "content": "<p>I see where you are coming from. The challenges that I saw was segmenting the nucleus, and there are overlapping problems. Therefore, by predicting the borders, this info could be used for watershed at the post processing stage. Thanks!</p>",
      "rawMarkdown": "I see where you are coming from. The challenges that I saw was segmenting the nucleus, and there are overlapping problems. Therefore, by predicting the borders, this info could be used for watershed at the post processing stage. Thanks!",
      "votes": null
    },
    {
      "id": "420386",
      "postDate": "11/13/2018 14:44:40",
      "content": "<p>Yeah, DSB18 was like this. There are two really nice solutions that I'd recommend: </p>\n\n<p><a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741</a></p>\n\n<p><a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/57709\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/57709</a></p>",
      "rawMarkdown": "Yeah, DSB18 was like this. There are two really nice solutions that I'd recommend: \n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/54741\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/57709",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 420254,
      "author_name": "arc144",
      "author_url": "",
      "post_date": "11/13/2018 11:07:41",
      "content": "<p>Are you asking about having one net for classifying ship/no-ship and another one for segmentation? Or having for example an U-Net with multiple output channels and threat each channel individually?</p>",
      "votes": null,
      "replies": [
        {
          "id": 420307,
          "author_name": "sbongo",
          "author_url": "",
          "post_date": "11/13/2018 12:44:03",
          "content": "<p>Hey there,\nthanks for replying!</p>\n\n<p>Yes I'm talking about \"having for example an U-Net with multiple output channels and threat each channel individually\"</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 420337,
          "author_name": "arc144",
          "author_url": "",
          "post_date": "11/13/2018 13:25:21",
          "content": "<p>It really depends.. For example if you have multiple classes makes sense to treat all of them together using softmax. </p>\n\n<p>However if you want to predict different things, it makes sense to separate the channels and treat them separately.  Example: One channel predicts the binary mask while other predicts energy transform for watershed like this paper <a href=\"https://arxiv.org/abs/1611.08303\">https://arxiv.org/abs/1611.08303</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 420349,
          "author_name": "sbongo",
          "author_url": "",
          "post_date": "11/13/2018 13:36:55",
          "content": "<p>I see where you are coming from. The challenges that I saw was segmenting the nucleus, and there are overlapping problems. Therefore, by predicting the borders, this info could be used for watershed at the post processing stage. Thanks!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 420386,
          "author_name": "arc144",
          "author_url": "",
          "post_date": "11/13/2018 14:44:40",
          "content": "<p>Yeah, DSB18 was like this. There are two really nice solutions that I'd recommend: </p>\n\n<p><a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741</a></p>\n\n<p><a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/57709\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/57709</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 420327,
      "author_name": "toshik",
      "author_url": "",
      "post_date": "11/13/2018 13:07:30",
      "content": "<p>Individual prediction is useful when we must decide thresholds. We can adjust them one by one.<br>\nIn this challenge we treat only single class (ship or not ship), so we don't need to split output channels. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "420239": "When I looked at previous similar challenges, I see a pattern of splitting each channel into individual classification problem/model. Is it because it’s easier for the model to just focus on a single task? Or is there other better reasons?\n\nThanks!",
    "420254": "Are you asking about having one net for classifying ship/no-ship and another one for segmentation? Or having for example an U-Net with multiple output channels and threat each channel individually?",
    "420307": "Hey there,\nthanks for replying!\n\nYes I'm talking about \"having for example an U-Net with multiple output channels and threat each channel individually\"",
    "420327": "Individual prediction is useful when we must decide thresholds. We can adjust them one by one.<br>\nIn this challenge we treat only single class (ship or not ship), so we don't need to split output channels.",
    "420337": "It really depends.. For example if you have multiple classes makes sense to treat all of them together using softmax. \n\nHowever if you want to predict different things, it makes sense to separate the channels and treat them separately.  Example: One channel predicts the binary mask while other predicts energy transform for watershed like this paper https://arxiv.org/abs/1611.08303",
    "420349": "I see where you are coming from. The challenges that I saw was segmenting the nucleus, and there are overlapping problems. Therefore, by predicting the borders, this info could be used for watershed at the post processing stage. Thanks!",
    "420386": "Yeah, DSB18 was like this. There are two really nice solutions that I'd recommend: \n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/54741\n\nhttps://www.kaggle.com/c/data-science-bowl-2018/discussion/57709"
  },
  "source": "meta"
}