{
  "id": 156282,
  "title": "beyond augmentation:   Supervised Contrastive Learning",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/156282",
  "author_name": "",
  "post_date": "2020-06-05T10:02:08.552230600Z",
  "votes": 17,
  "comment_count": 7,
  "views": 0,
  "content": "<p>check this out:</p>\n\n<p>Supervised Contrastive Learning\n<a href=\"https://github.com/HobbitLong/SupContrast\">https://github.com/HobbitLong/SupContrast</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fa77d945fec2c83854442f22caa66365b%2F94707565_1144034099275180_4874217556547731456_o.png?generation=1591351308213633&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": "874802",
      "postDate": "06/05/2020 10:02:08",
      "content": "<p>check this out:</p>\n\n<p>Supervised Contrastive Learning\n<a href=\"https://github.com/HobbitLong/SupContrast\">https://github.com/HobbitLong/SupContrast</a></p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fa77d945fec2c83854442f22caa66365b%2F94707565_1144034099275180_4874217556547731456_o.png?generation=1591351308213633&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "check this out:\n\n Supervised Contrastive Learning\nhttps://github.com/HobbitLong/SupContrast\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fa77d945fec2c83854442f22caa66365b%2F94707565_1144034099275180_4874217556547731456_o.png?generation=1591351308213633&amp;alt=media)",
      "votes": null
    },
    {
      "id": "874815",
      "postDate": "06/05/2020 10:13:14",
      "content": "<p>Wow <a href=\"/hengck23\">@hengck23</a> this is a latest advance in research from just a few weeks ago!</p>",
      "rawMarkdown": "Wow @hengck23 this is a latest advance in research from just a few weeks ago!",
      "votes": null
    },
    {
      "id": "875326",
      "postDate": "06/05/2020 16:52:22",
      "content": "<p>Have you already tried that?</p>",
      "rawMarkdown": "Have you already tried that?",
      "votes": null
    },
    {
      "id": "875393",
      "postDate": "06/05/2020 18:37:56",
      "content": "<p>this ain't for mortals. :)</p>\n\n<p>they used 8192 batch size and contrastive steps are anyway slower :) unless you work for google or nvidia you will likely need to wait to try it out. </p>\n\n<p>I am looking forward to some budget options :)</p>",
      "rawMarkdown": "this ain't for mortals. :)\n\nthey used 8192 batch size and contrastive steps are anyway slower :) unless you work for google or nvidia you will likely need to wait to try it out. \n\nI am looking forward to some budget options :)",
      "votes": null
    },
    {
      "id": "875397",
      "postDate": "06/05/2020 18:42:49",
      "content": "<p>I guess you could simulate big batch size by using Momentum Contrast (see <a href=\"https://arxiv.org/abs/1911.05722\">https://arxiv.org/abs/1911.05722</a> for details), but this is unlikely to make the whole thing more approachable :)</p>",
      "rawMarkdown": "I guess you could simulate big batch size by using Momentum Contrast (see https://arxiv.org/abs/1911.05722 for details), but this is unlikely to make the whole thing more approachable :)",
      "votes": null
    },
    {
      "id": "875409",
      "postDate": "06/05/2020 18:49:33",
      "content": "<p>well i tried this one actually :)  at least it works with smaller batches </p>\n\n<p>it took 36h on 1080Ti with 10% labelled imgs from CIFAR10 (input 32x32 ), hilarious  :D</p>",
      "rawMarkdown": "well i tried this one actually :)  at least it works with smaller batches \n\nit took 36h on 1080Ti with 10% labelled imgs from CIFAR10 (input 32x32 ), hilarious  :D",
      "votes": null
    },
    {
      "id": "875431",
      "postDate": "06/05/2020 19:17:17",
      "content": "<p>he wants us to love him and give up any resistance :) I nearly do :)</p>",
      "rawMarkdown": "he wants us to love him and give up any resistance :) I nearly do :)",
      "votes": null
    },
    {
      "id": "2434930",
      "postDate": "09/12/2023 15:56:28",
      "content": "<p>Thank you for this content. Do you know if there exists a variant for tabular data ?</p>",
      "rawMarkdown": "Thank you for this content. Do you know if there exists a variant for tabular data ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2434930,
      "author_name": "theudbald",
      "author_url": "",
      "post_date": "09/12/2023 15:56:28",
      "content": "<p>Thank you for this content. Do you know if there exists a variant for tabular data ?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 874815,
      "author_name": "sirishks",
      "author_url": "",
      "post_date": "06/05/2020 10:13:14",
      "content": "<p>Wow <a href=\"/hengck23\">@hengck23</a> this is a latest advance in research from just a few weeks ago!</p>",
      "votes": null,
      "replies": [
        {
          "id": 875431,
          "author_name": "romanweilguny",
          "author_url": "",
          "post_date": "06/05/2020 19:17:17",
          "content": "<p>he wants us to love him and give up any resistance :) I nearly do :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 875326,
      "author_name": "ddanevskyi",
      "author_url": "",
      "post_date": "06/05/2020 16:52:22",
      "content": "<p>Have you already tried that?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 875393,
      "author_name": "valanm",
      "author_url": "",
      "post_date": "06/05/2020 18:37:56",
      "content": "<p>this ain't for mortals. :)</p>\n\n<p>they used 8192 batch size and contrastive steps are anyway slower :) unless you work for google or nvidia you will likely need to wait to try it out. </p>\n\n<p>I am looking forward to some budget options :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 875397,
          "author_name": "ddanevskyi",
          "author_url": "",
          "post_date": "06/05/2020 18:42:49",
          "content": "<p>I guess you could simulate big batch size by using Momentum Contrast (see <a href=\"https://arxiv.org/abs/1911.05722\">https://arxiv.org/abs/1911.05722</a> for details), but this is unlikely to make the whole thing more approachable :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 875409,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "06/05/2020 18:49:33",
          "content": "<p>well i tried this one actually :)  at least it works with smaller batches </p>\n\n<p>it took 36h on 1080Ti with 10% labelled imgs from CIFAR10 (input 32x32 ), hilarious  :D</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "874802": "check this out:\n\n Supervised Contrastive Learning\nhttps://github.com/HobbitLong/SupContrast\n\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2Fa77d945fec2c83854442f22caa66365b%2F94707565_1144034099275180_4874217556547731456_o.png?generation=1591351308213633&amp;alt=media)",
    "874815": "Wow @hengck23 this is a latest advance in research from just a few weeks ago!",
    "875326": "Have you already tried that?",
    "875393": "this ain't for mortals. :)\n\nthey used 8192 batch size and contrastive steps are anyway slower :) unless you work for google or nvidia you will likely need to wait to try it out. \n\nI am looking forward to some budget options :)",
    "875397": "I guess you could simulate big batch size by using Momentum Contrast (see https://arxiv.org/abs/1911.05722 for details), but this is unlikely to make the whole thing more approachable :)",
    "875409": "well i tried this one actually :)  at least it works with smaller batches \n\nit took 36h on 1080Ti with 10% labelled imgs from CIFAR10 (input 32x32 ), hilarious  :D",
    "875431": "he wants us to love him and give up any resistance :) I nearly do :)",
    "2434930": "Thank you for this content. Do you know if there exists a variant for tabular data ?"
  },
  "source": "meta"
}