{
  "id": 409467,
  "title": "Sugestions on loss functions?",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/409467",
  "author_name": "",
  "post_date": "2023-05-11T06:29:15.416588600Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Which loss function combination do you guys find best possible?</p>",
  "messages": [
    {
      "id": "2254652",
      "postDate": "05/11/2023 06:29:15",
      "content": "<p>Which loss function combination do you guys find best possible?</p>",
      "rawMarkdown": "Which loss function combination do you guys find best possible?",
      "votes": null
    },
    {
      "id": "2254931",
      "postDate": "05/11/2023 10:56:15",
      "content": "<p>(BCELoss(y_pred, y_true)+DiceLoss(y_pred, y_true))/2</p>",
      "rawMarkdown": "(BCELoss(y_pred, y_true)+DiceLoss(y_pred, y_true))/2",
      "votes": null
    },
    {
      "id": "2255024",
      "postDate": "05/11/2023 12:56:44",
      "content": "<p>I'm using BCE + Dice. Tried Lovasz, but the performance was about the same.</p>",
      "rawMarkdown": "I'm using BCE + Dice. Tried Lovasz, but the performance was about the same.",
      "votes": null
    },
    {
      "id": "2255481",
      "postDate": "05/11/2023 18:38:35",
      "content": "<p>Same ratio between BEC and Dice? If not, what are their proportion? Do you just use the out-of-the-box Diceloss from segmentation_models_pytorch? </p>",
      "rawMarkdown": "Same ratio between BEC and Dice? If not, what are their proportion? Do you just use the out-of-the-box Diceloss from segmentation_models_pytorch?",
      "votes": null
    },
    {
      "id": "2255571",
      "postDate": "05/11/2023 20:06:15",
      "content": "<ol>\n<li>Same ratio</li>\n<li>Nope, I have my own implementation based on the metric.</li>\n</ol>",
      "rawMarkdown": "1. Same ratio\n2. Nope, I have my own implementation based on the metric.",
      "votes": null
    },
    {
      "id": "2261226",
      "postDate": "05/16/2023 07:30:19",
      "content": "<p>Thank you for reminding to use the competition metric as a loss. … Keep forgetting. btw BinarySoftF1Loss is in pytorch_toolbelt </p>",
      "rawMarkdown": "Thank you for reminding to use the competition metric as a loss. ... Keep forgetting. btw BinarySoftF1Loss is in pytorch_toolbelt",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2254931,
      "author_name": "yoyobar",
      "author_url": "",
      "post_date": "05/11/2023 10:56:15",
      "content": "<p>(BCELoss(y_pred, y_true)+DiceLoss(y_pred, y_true))/2</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2255024,
      "author_name": "igorkrashenyi",
      "author_url": "",
      "post_date": "05/11/2023 12:56:44",
      "content": "<p>I'm using BCE + Dice. Tried Lovasz, but the performance was about the same.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2255481,
          "author_name": "lionfishy",
          "author_url": "",
          "post_date": "05/11/2023 18:38:35",
          "content": "<p>Same ratio between BEC and Dice? If not, what are their proportion? Do you just use the out-of-the-box Diceloss from segmentation_models_pytorch? </p>",
          "votes": null,
          "replies": [
            {
              "id": 2255571,
              "author_name": "igorkrashenyi",
              "author_url": "",
              "post_date": "05/11/2023 20:06:15",
              "content": "<ol>\n<li>Same ratio</li>\n<li>Nope, I have my own implementation based on the metric.</li>\n</ol>",
              "votes": null,
              "replies": [
                {
                  "id": 2261226,
                  "author_name": "dmitrykonovalov",
                  "author_url": "",
                  "post_date": "05/16/2023 07:30:19",
                  "content": "<p>Thank you for reminding to use the competition metric as a loss. … Keep forgetting. btw BinarySoftF1Loss is in pytorch_toolbelt </p>",
                  "votes": null,
                  "replies": []
                }
              ]
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2254652": "Which loss function combination do you guys find best possible?",
    "2254931": "(BCELoss(y_pred, y_true)+DiceLoss(y_pred, y_true))/2",
    "2255024": "I'm using BCE + Dice. Tried Lovasz, but the performance was about the same.",
    "2255481": "Same ratio between BEC and Dice? If not, what are their proportion? Do you just use the out-of-the-box Diceloss from segmentation_models_pytorch?",
    "2255571": "1. Same ratio\n2. Nope, I have my own implementation based on the metric.",
    "2261226": "Thank you for reminding to use the competition metric as a loss. ... Keep forgetting. btw BinarySoftF1Loss is in pytorch_toolbelt"
  },
  "source": "meta"
}