{
  "id": 115782,
  "title": "What is your loss function?",
  "url": "/competitions/understanding_cloud_organization/discussion/115782",
  "author_name": "Bibek",
  "post_date": "2019-11-05T07:25:01.426000",
  "votes": 2,
  "comment_count": 16,
  "views": 0,
  "content": "<p>Mine is: BCEDICELoss. What's yours?</p>",
  "messages": [
    {
      "id": 666898,
      "postDate": "2019-11-06T15:50:29.357Z",
      "content": "<p>currently only BCE. others failed so far.</p>",
      "rawMarkdown": "currently only BCE. others failed so far.",
      "votes": 3
    },
    {
      "id": 665620,
      "postDate": "2019-11-05T07:25:01.427Z",
      "content": "<p>Mine is: BCEDICELoss. What's yours?</p>",
      "rawMarkdown": "Mine is: BCEDICELoss. What's yours?",
      "votes": 2
    },
    {
      "id": 665762,
      "postDate": "2019-11-05T11:21:49.107Z",
      "content": "<p>i tried BCEDiceLoss, BCELoss and DIceLoss .I think they are almost the same</p>",
      "rawMarkdown": "i tried BCEDiceLoss, BCELoss and DIceLoss .I think they are almost the same",
      "votes": 1
    },
    {
      "id": 665624,
      "postDate": "2019-11-05T07:31:29.610Z",
      "content": "<p>any reason for choosing BCEDICELoss? the dataset is not highly imbalanced</p>",
      "rawMarkdown": "any reason for choosing BCEDICELoss? the dataset is not highly imbalanced",
      "votes": 1,
      "replies": [
        {
          "id": 665630,
          "postDate": "2019-11-05T07:39:32.750Z",
          "content": "<p>high imbalanced? sure about that?</p>",
          "rawMarkdown": "high imbalanced? sure about that?",
          "votes": -1
        },
        {
          "id": 665635,
          "postDate": "2019-11-05T07:46:14.113Z",
          "content": "<p>I got better result using your BCEsoftdice ,the one you used in steel comp ..</p>",
          "rawMarkdown": "I got better result using your BCEsoftdice ,the one you used in steel comp ..",
          "votes": 1
        },
        {
          "id": 665785,
          "postDate": "2019-11-05T11:54:10.017Z",
          "content": "<p>yes the dataset is not highly imbalanced in this competition mate,please check the eda kernel of my team mate deis : <a href=\"https://www.kaggle.com/aleksandradeis/understanding-clouds-eda\">https://www.kaggle.com/aleksandradeis/understanding-clouds-eda</a>\nif i understand correctly bcesoftdice/bcedice usually used when the dataset is highly imbalanced</p>",
          "rawMarkdown": "yes the dataset is not highly imbalanced in this competition mate,please check the eda kernel of my team mate deis : https://www.kaggle.com/aleksandradeis/understanding-clouds-eda\nif i understand correctly bcesoftdice/bcedice usually used when the dataset is highly imbalanced",
          "votes": 3
        },
        {
          "id": 665828,
          "postDate": "2019-11-05T12:52:59.797Z",
          "content": "<p><a href=\"/mobassir\">@mobassir</a> did you get better results using a loss function other than BCEDice?</p>",
          "rawMarkdown": "@mobassir did you get better results using a loss function other than BCEDice?",
          "votes": 1
        },
        {
          "id": 665830,
          "postDate": "2019-11-05T12:55:16.073Z",
          "content": "<p>haven't tried much mate,busy working for other competitions,but i think jaccard will work well</p>",
          "rawMarkdown": "haven't tried much mate,busy working for other competitions,but i think jaccard will work well",
          "votes": 1
        },
        {
          "id": 665835,
          "postDate": "2019-11-05T13:02:23.077Z",
          "content": "<p>Well, this opens up a lot of possible approaches to experiment with.</p>",
          "rawMarkdown": "Well, this opens up a lot of possible approaches to experiment with.",
          "votes": 1
        },
        {
          "id": 665844,
          "postDate": "2019-11-05T13:08:37.843Z",
          "content": "<p><a href=\"/axel81\">@axel81</a> \nrecommended link for you : <a href=\"https://lars76.github.io/neural-networks/object-detection/losses-for-segmentation/\">https://lars76.github.io/neural-networks/object-detection/losses-for-segmentation/</a></p>",
          "rawMarkdown": "@axel81 \nrecommended link for you : https://lars76.github.io/neural-networks/object-detection/losses-for-segmentation/",
          "votes": 1
        },
        {
          "id": 665914,
          "postDate": "2019-11-05T14:27:31.867Z",
          "content": "<p><a href=\"/mobassir\">@mobassir</a> What a coincidence, I was referring to the same link after seeing this post XD</p>",
          "rawMarkdown": "@mobassir What a coincidence, I was referring to the same link after seeing this post XD",
          "votes": 1
        },
        {
          "id": 665963,
          "postDate": "2019-11-05T15:31:05.743Z",
          "content": "<p>Jaccard did worse for me .. since I had the same thought as you <a href=\"/mobassir\">@mobassir</a> ,I tried . However , since then I have changed few other things . Let me try this out again . Thank you </p>",
          "rawMarkdown": "Jaccard did worse for me .. since I had the same thought as you @mobassir ,I tried . However , since then I have changed few other things . Let me try this out again . Thank you ",
          "votes": 2
        }
      ]
    },
    {
      "id": 666384,
      "postDate": "2019-11-06T03:53:26.917Z",
      "content": "<p>I'm using Dice + BCE for training, but calculate my local using Dice. My local to public is pretty much spot-on. 0.655 local... 0.655 public... 1 fold. </p>",
      "rawMarkdown": "I'm using Dice + BCE for training, but calculate my local using Dice. My local to public is pretty much spot-on. 0.655 local... 0.655 public... 1 fold. ",
      "votes": 2,
      "replies": [
        {
          "id": 666396,
          "postDate": "2019-11-06T04:24:07.493Z",
          "content": "<p>Nice!! Keep digging and you will improve your LB score</p>",
          "rawMarkdown": "Nice!! Keep digging and you will improve your LB score",
          "votes": 1
        },
        {
          "id": 666494,
          "postDate": "2019-11-06T06:23:03.587Z",
          "content": "<p>Definitely. BTW, the backbone is efficientnet-b1.. mainly faster to iterate on ideas as I dig :P</p>",
          "rawMarkdown": "Definitely. BTW, the backbone is efficientnet-b1.. mainly faster to iterate on ideas as I dig :P"
        }
      ]
    },
    {
      "id": 670100,
      "postDate": "2019-11-11T02:25:21.317Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 666898,
      "author_name": "tugstugi",
      "author_url": "",
      "post_date": "2019-11-06T15:50:29.357000",
      "content": "<p>currently only BCE. others failed so far.</p>",
      "votes": 3,
      "replies": []
    },
    {
      "id": 665762,
      "author_name": "llh1818",
      "author_url": "",
      "post_date": "2019-11-05T11:21:49.107000",
      "content": "<p>i tried BCEDiceLoss, BCELoss and DIceLoss .I think they are almost the same</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 665624,
      "author_name": "Mobassir",
      "author_url": "",
      "post_date": "2019-11-05T07:31:29.610000",
      "content": "<p>any reason for choosing BCEDICELoss? the dataset is not highly imbalanced</p>",
      "votes": 1,
      "replies": [
        {
          "id": 665630,
          "author_name": "Bibek",
          "author_url": "",
          "post_date": "2019-11-05T07:39:32.750000",
          "content": "<p>high imbalanced? sure about that?</p>",
          "votes": -1,
          "replies": []
        },
        {
          "id": 665635,
          "author_name": "Nirjhar Roy",
          "author_url": "",
          "post_date": "2019-11-05T07:46:14.113000",
          "content": "<p>I got better result using your BCEsoftdice ,the one you used in steel comp ..</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665785,
          "author_name": "Mobassir",
          "author_url": "",
          "post_date": "2019-11-05T11:54:10.017000",
          "content": "<p>yes the dataset is not highly imbalanced in this competition mate,please check the eda kernel of my team mate deis : <a href=\"https://www.kaggle.com/aleksandradeis/understanding-clouds-eda\">https://www.kaggle.com/aleksandradeis/understanding-clouds-eda</a>\nif i understand correctly bcesoftdice/bcedice usually used when the dataset is highly imbalanced</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 665828,
          "author_name": "Ram Ramrakhya",
          "author_url": "",
          "post_date": "2019-11-05T12:52:59.797000",
          "content": "<p><a href=\"/mobassir\">@mobassir</a> did you get better results using a loss function other than BCEDice?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665830,
          "author_name": "Mobassir",
          "author_url": "",
          "post_date": "2019-11-05T12:55:16.073000",
          "content": "<p>haven't tried much mate,busy working for other competitions,but i think jaccard will work well</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665835,
          "author_name": "Ram Ramrakhya",
          "author_url": "",
          "post_date": "2019-11-05T13:02:23.077000",
          "content": "<p>Well, this opens up a lot of possible approaches to experiment with.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665844,
          "author_name": "Mobassir",
          "author_url": "",
          "post_date": "2019-11-05T13:08:37.843000",
          "content": "<p><a href=\"/axel81\">@axel81</a> \nrecommended link for you : <a href=\"https://lars76.github.io/neural-networks/object-detection/losses-for-segmentation/\">https://lars76.github.io/neural-networks/object-detection/losses-for-segmentation/</a></p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665914,
          "author_name": "Ram Ramrakhya",
          "author_url": "",
          "post_date": "2019-11-05T14:27:31.867000",
          "content": "<p><a href=\"/mobassir\">@mobassir</a> What a coincidence, I was referring to the same link after seeing this post XD</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 665963,
          "author_name": "Nirjhar Roy",
          "author_url": "",
          "post_date": "2019-11-05T15:31:05.743000",
          "content": "<p>Jaccard did worse for me .. since I had the same thought as you <a href=\"/mobassir\">@mobassir</a> ,I tried . However , since then I have changed few other things . Let me try this out again . Thank you </p>",
          "votes": 2,
          "replies": []
        }
      ]
    },
    {
      "id": 666384,
      "author_name": "Thomas Yokota",
      "author_url": "",
      "post_date": "2019-11-06T03:53:26.917000",
      "content": "<p>I'm using Dice + BCE for training, but calculate my local using Dice. My local to public is pretty much spot-on. 0.655 local... 0.655 public... 1 fold. </p>",
      "votes": 2,
      "replies": [
        {
          "id": 666396,
          "author_name": "Bibek",
          "author_url": "",
          "post_date": "2019-11-06T04:24:07.493000",
          "content": "<p>Nice!! Keep digging and you will improve your LB score</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 666494,
          "author_name": "Thomas Yokota",
          "author_url": "",
          "post_date": "2019-11-06T06:23:03.587000",
          "content": "<p>Definitely. BTW, the backbone is efficientnet-b1.. mainly faster to iterate on ideas as I dig :P</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 670100,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-11-11T02:25:21.317000",
      "content": "",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "666898": "currently only BCE. others failed so far.",
    "665620": "Mine is: BCEDICELoss. What's yours?",
    "665762": "i tried BCEDiceLoss, BCELoss and DIceLoss .I think they are almost the same",
    "665624": "any reason for choosing BCEDICELoss? the dataset is not highly imbalanced",
    "666384": "I'm using Dice + BCE for training, but calculate my local using Dice. My local to public is pretty much spot-on. 0.655 local... 0.655 public... 1 fold. ",
    "670100": ""
  }
}