{
  "id": 234595,
  "title": "Lower Effnet Models CV vs LB",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/234595",
  "author_name": "",
  "post_date": "2021-04-25T04:40:23.714752900Z",
  "votes": 1,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Its suprising to find that lower Effnet having CV at par with higher Model but LB is just worse than what higher models get.<br>\nEg. in my case b0  get 93+ CV , LB &lt;91  i just used imagnet stats and basic set ups.</p>",
  "messages": [
    {
      "id": "1283594",
      "postDate": "04/25/2021 04:40:23",
      "content": "<p>Its suprising to find that lower Effnet having CV at par with higher Model but LB is just worse than what higher models get.<br>\nEg. in my case b0  get 93+ CV , LB &lt;91  i just used imagnet stats and basic set ups.</p>",
      "rawMarkdown": "Its suprising to find that lower Effnet having CV at par with higher Model but LB is just worse than what higher models get.\nEg. in my case b0  get 93+ CV , LB <91  i just used imagnet stats and basic set ups.",
      "votes": null
    },
    {
      "id": "1283617",
      "postDate": "04/25/2021 05:15:29",
      "content": "<p>same here. Deeper networks cope better with noisy labels, which is clearly the case for this hack. Someone shared a link to a paper on this subject, if you look for it.</p>",
      "rawMarkdown": "same here. Deeper networks cope better with noisy labels, which is clearly the case for this hack. Someone shared a link to a paper on this subject, if you look for it.",
      "votes": null
    },
    {
      "id": "1283684",
      "postDate": "04/25/2021 07:17:05",
      "content": "<p>agreed. btw, I think lower and higher models being ensembled could be a nice strategy. But everything could do sometimes means nothing is so good.</p>",
      "rawMarkdown": "agreed. btw, I think lower and higher models being ensembled could be a nice strategy. But everything could do sometimes means nothing is so good.",
      "votes": null
    },
    {
      "id": "1283796",
      "postDate": "04/25/2021 09:32:23",
      "content": "<p><a href=\"https://www.kaggle.com/southsakura\" target=\"_blank\">@southsakura</a>   i left message you can read through, i  have worked in past with people belong to you region, it was nice experience.  just think over and let me know</p>",
      "rawMarkdown": "southsakura   i left message you can read through, i  have worked in past with people belong to you region, it was nice experience.  just think over and let me know",
      "votes": null
    },
    {
      "id": "1283811",
      "postDate": "04/25/2021 09:50:50",
      "content": "<p>I'm actually using a B6 model at the moment which does quite a bit better (0.08 LB) than a B0 model. I need to explore more with B2 and B4.</p>\n<p>I think having enough augmentation is important to provide enough variance that B6 doesn't just overfit. However, surprisingly I found cutmix actually harmful in terms of LB.</p>",
      "rawMarkdown": "I'm actually using a B6 model at the moment which does quite a bit better (0.08 LB) than a B0 model. I need to explore more with B2 and B4.\n\nI think having enough augmentation is important to provide enough variance that B6 doesn't just overfit. However, surprisingly I found cutmix actually harmful in terms of LB.",
      "votes": null
    },
    {
      "id": "1284021",
      "postDate": "04/25/2021 13:52:07",
      "content": "<p>funny you should say that. I discarded cutmix a month ago, as it was giving me horrible results. I'm curious if anyone was successful at it</p>",
      "rawMarkdown": "funny you should say that. I discarded cutmix a month ago, as it was giving me horrible results. I'm curious if anyone was successful at it",
      "votes": null
    },
    {
      "id": "1284028",
      "postDate": "04/25/2021 13:58:55",
      "content": "<p>It's weird because it did improve my CV, or at least didn't harm it, but worsened my LB. I suspect it's because the network learns unnatural shapes like straight edges etc. are acceptable predictions along a cut.</p>",
      "rawMarkdown": "It's weird because it did improve my CV, or at least didn't harm it, but worsened my LB. I suspect it's because the network learns unnatural shapes like straight edges etc. are acceptable predictions along a cut.",
      "votes": null
    },
    {
      "id": "1284195",
      "postDate": "04/25/2021 16:55:33",
      "content": "<p>what's cutmix usage?</p>",
      "rawMarkdown": "what's cutmix usage?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1283617,
      "author_name": "andrasferenczi",
      "author_url": "",
      "post_date": "04/25/2021 05:15:29",
      "content": "<p>same here. Deeper networks cope better with noisy labels, which is clearly the case for this hack. Someone shared a link to a paper on this subject, if you look for it.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1283684,
          "author_name": "southsakura",
          "author_url": "",
          "post_date": "04/25/2021 07:17:05",
          "content": "<p>agreed. btw, I think lower and higher models being ensembled could be a nice strategy. But everything could do sometimes means nothing is so good.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1283796,
          "author_name": "jaideepvalani",
          "author_url": "",
          "post_date": "04/25/2021 09:32:23",
          "content": "<p><a href=\"https://www.kaggle.com/southsakura\" target=\"_blank\">@southsakura</a>   i left message you can read through, i  have worked in past with people belong to you region, it was nice experience.  just think over and let me know</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1283811,
      "author_name": "jamesphoward",
      "author_url": "",
      "post_date": "04/25/2021 09:50:50",
      "content": "<p>I'm actually using a B6 model at the moment which does quite a bit better (0.08 LB) than a B0 model. I need to explore more with B2 and B4.</p>\n<p>I think having enough augmentation is important to provide enough variance that B6 doesn't just overfit. However, surprisingly I found cutmix actually harmful in terms of LB.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1284021,
          "author_name": "andrasferenczi",
          "author_url": "",
          "post_date": "04/25/2021 13:52:07",
          "content": "<p>funny you should say that. I discarded cutmix a month ago, as it was giving me horrible results. I'm curious if anyone was successful at it</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1284028,
          "author_name": "jamesphoward",
          "author_url": "",
          "post_date": "04/25/2021 13:58:55",
          "content": "<p>It's weird because it did improve my CV, or at least didn't harm it, but worsened my LB. I suspect it's because the network learns unnatural shapes like straight edges etc. are acceptable predictions along a cut.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1284195,
          "author_name": "southsakura",
          "author_url": "",
          "post_date": "04/25/2021 16:55:33",
          "content": "<p>what's cutmix usage?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1283594": "Its suprising to find that lower Effnet having CV at par with higher Model but LB is just worse than what higher models get.\nEg. in my case b0  get 93+ CV , LB <91  i just used imagnet stats and basic set ups.",
    "1283617": "same here. Deeper networks cope better with noisy labels, which is clearly the case for this hack. Someone shared a link to a paper on this subject, if you look for it.",
    "1283684": "agreed. btw, I think lower and higher models being ensembled could be a nice strategy. But everything could do sometimes means nothing is so good.",
    "1283796": "southsakura   i left message you can read through, i  have worked in past with people belong to you region, it was nice experience.  just think over and let me know",
    "1283811": "I'm actually using a B6 model at the moment which does quite a bit better (0.08 LB) than a B0 model. I need to explore more with B2 and B4.\n\nI think having enough augmentation is important to provide enough variance that B6 doesn't just overfit. However, surprisingly I found cutmix actually harmful in terms of LB.",
    "1284021": "funny you should say that. I discarded cutmix a month ago, as it was giving me horrible results. I'm curious if anyone was successful at it",
    "1284028": "It's weird because it did improve my CV, or at least didn't harm it, but worsened my LB. I suspect it's because the network learns unnatural shapes like straight edges etc. are acceptable predictions along a cut.",
    "1284195": "what's cutmix usage?"
  },
  "source": "meta"
}