{
  "id": 210550,
  "title": "Does anyone try focal loss？",
  "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/210550",
  "author_name": "",
  "post_date": "2021-01-11T08:46:59.638045700Z",
  "votes": 2,
  "comment_count": 9,
  "views": 0,
  "content": "<p>When I use focal loss，the CV is higher，But the Lb is lower</p>",
  "messages": [
    {
      "id": "1148599",
      "postDate": "01/11/2021 08:46:59",
      "content": "<p>When I use focal loss，the CV is higher，But the Lb is lower</p>",
      "rawMarkdown": "When I use focal loss，the CV is higher，But the Lb is lower",
      "votes": null
    },
    {
      "id": "1148603",
      "postDate": "01/11/2021 08:53:17",
      "content": "<p>trust your local CV</p>",
      "rawMarkdown": "trust your local CV",
      "votes": null
    },
    {
      "id": "1149080",
      "postDate": "01/11/2021 15:17:36",
      "content": "<p>i estimate shakeup is -/+ 0.005.<br>\nAnything within that is with statistical error</p>",
      "rawMarkdown": "i estimate shakeup is -/+ 0.005.\nAnything within that is with statistical error",
      "votes": null
    },
    {
      "id": "1149518",
      "postDate": "01/11/2021 22:48:05",
      "content": "<p>Same results with you, focal loss, better CV but slightly(around 0.005) worse LB. </p>",
      "rawMarkdown": "Same results with you, focal loss, better CV but slightly(around 0.005) worse LB.",
      "votes": null
    },
    {
      "id": "1151842",
      "postDate": "01/13/2021 15:49:48",
      "content": "<p>how can you estimate the shakeup? right now my local cv still keep improving but public lb almost cant move anymore so I think there might be some randomness, distribution shift or ceiling because of labelling noise</p>",
      "rawMarkdown": "how can you estimate the shakeup? right now my local cv still keep improving but public lb almost cant move anymore so I think there might be some randomness, distribution shift or ceiling because of labelling noise",
      "votes": null
    },
    {
      "id": "1152805",
      "postDate": "01/14/2021 13:06:43",
      "content": "<p>but LB is also part of the validation. good if all cv metric and LB increases. But if discrepancy starts to show, you may need to be a little bit careful.</p>\n<p>it seems that the metric will plateau for quite a bit after overfitting occurs, in my observation.</p>",
      "rawMarkdown": "but LB is also part of the validation. good if all cv metric and LB increases. But if discrepancy starts to show, you may need to be a little bit careful.\n\nit seems that the metric will plateau for quite a bit after overfitting occurs, in my observation.",
      "votes": null
    },
    {
      "id": "1154843",
      "postDate": "01/16/2021 01:47:40",
      "content": "<p>In my case, focal loss improves both CV and Public LB by +0.001.</p>",
      "rawMarkdown": "In my case, focal loss improves both CV and Public LB by +0.001.",
      "votes": null
    },
    {
      "id": "1224513",
      "postDate": "03/02/2021 19:33:02",
      "content": "<p>Hi can you share your focal loss class and alpha and gamma value you used? Thanks in advanced. </p>",
      "rawMarkdown": "Hi can you share your focal loss class and alpha and gamma value you used? Thanks in advanced.",
      "votes": null
    },
    {
      "id": "1225997",
      "postDate": "03/04/2021 06:14:18",
      "content": "<p>worse LB when training with focal loss, I was hoping it would make a improvement, even if a little bit</p>",
      "rawMarkdown": "worse LB when training with focal loss, I was hoping it would make a improvement, even if a little bit",
      "votes": null
    },
    {
      "id": "1236352",
      "postDate": "03/13/2021 03:28:21",
      "content": "<p>I tried  alpha=0.25,0.5,0.75 and gamma=2,  no help :(</p>",
      "rawMarkdown": "I tried  alpha=0.25,0.5,0.75 and gamma=2,  no help :(",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1148603,
      "author_name": "moewie94",
      "author_url": "",
      "post_date": "01/11/2021 08:53:17",
      "content": "<p>trust your local CV</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1149080,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "01/11/2021 15:17:36",
      "content": "<p>i estimate shakeup is -/+ 0.005.<br>\nAnything within that is with statistical error</p>",
      "votes": null,
      "replies": [
        {
          "id": 1151842,
          "author_name": "moewie94",
          "author_url": "",
          "post_date": "01/13/2021 15:49:48",
          "content": "<p>how can you estimate the shakeup? right now my local cv still keep improving but public lb almost cant move anymore so I think there might be some randomness, distribution shift or ceiling because of labelling noise</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1152805,
          "author_name": "yl1202",
          "author_url": "",
          "post_date": "01/14/2021 13:06:43",
          "content": "<p>but LB is also part of the validation. good if all cv metric and LB increases. But if discrepancy starts to show, you may need to be a little bit careful.</p>\n<p>it seems that the metric will plateau for quite a bit after overfitting occurs, in my observation.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1149518,
      "author_name": "andy1010",
      "author_url": "",
      "post_date": "01/11/2021 22:48:05",
      "content": "<p>Same results with you, focal loss, better CV but slightly(around 0.005) worse LB. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1154843,
      "author_name": "ttahara",
      "author_url": "",
      "post_date": "01/16/2021 01:47:40",
      "content": "<p>In my case, focal loss improves both CV and Public LB by +0.001.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1224513,
      "author_name": "durbin164",
      "author_url": "",
      "post_date": "03/02/2021 19:33:02",
      "content": "<p>Hi can you share your focal loss class and alpha and gamma value you used? Thanks in advanced. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1236352,
          "author_name": "c82447085",
          "author_url": "",
          "post_date": "03/13/2021 03:28:21",
          "content": "<p>I tried  alpha=0.25,0.5,0.75 and gamma=2,  no help :(</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1225997,
      "author_name": "plugin1689",
      "author_url": "",
      "post_date": "03/04/2021 06:14:18",
      "content": "<p>worse LB when training with focal loss, I was hoping it would make a improvement, even if a little bit</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1148599": "When I use focal loss，the CV is higher，But the Lb is lower",
    "1148603": "trust your local CV",
    "1149080": "i estimate shakeup is -/+ 0.005.\nAnything within that is with statistical error",
    "1149518": "Same results with you, focal loss, better CV but slightly(around 0.005) worse LB.",
    "1151842": "how can you estimate the shakeup? right now my local cv still keep improving but public lb almost cant move anymore so I think there might be some randomness, distribution shift or ceiling because of labelling noise",
    "1152805": "but LB is also part of the validation. good if all cv metric and LB increases. But if discrepancy starts to show, you may need to be a little bit careful.\n\nit seems that the metric will plateau for quite a bit after overfitting occurs, in my observation.",
    "1154843": "In my case, focal loss improves both CV and Public LB by +0.001.",
    "1224513": "Hi can you share your focal loss class and alpha and gamma value you used? Thanks in advanced.",
    "1225997": "worse LB when training with focal loss, I was hoping it would make a improvement, even if a little bit",
    "1236352": "I tried  alpha=0.25,0.5,0.75 and gamma=2,  no help :("
  },
  "source": "meta"
}