{
  "id": 164766,
  "title": "Any alternative to EfficientNets? ",
  "url": "/competitions/alaska2-image-steganalysis/discussion/164766",
  "author_name": "",
  "post_date": "2020-07-07T13:03:48.324729200Z",
  "votes": 2,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi, I was wondering has anybody managed to get a decent public LB score without using any efficientnet model? I personally tried Densenet variants and SeResnext variants, but surprisingly, even with the exact similar setup,  some of these models never converged at all while the remaining models converged at a very poor public CV score. I was wondering what might be the reason for this. </p>",
  "messages": [
    {
      "id": "918738",
      "postDate": "07/07/2020 13:03:48",
      "content": "<p>Hi, I was wondering has anybody managed to get a decent public LB score without using any efficientnet model? I personally tried Densenet variants and SeResnext variants, but surprisingly, even with the exact similar setup,  some of these models never converged at all while the remaining models converged at a very poor public CV score. I was wondering what might be the reason for this. </p>",
      "rawMarkdown": "Hi, I was wondering has anybody managed to get a decent public LB score without using any efficientnet model? I personally tried Densenet variants and SeResnext variants, but surprisingly, even with the exact similar setup,  some of these models never converged at all while the remaining models converged at a very poor public CV score. I was wondering what might be the reason for this.",
      "votes": null
    },
    {
      "id": "918792",
      "postDate": "07/07/2020 13:52:15",
      "content": "<p>How poor a score seresnext50 gave?</p>",
      "rawMarkdown": "How poor a score seresnext50 gave?",
      "votes": null
    },
    {
      "id": "918807",
      "postDate": "07/07/2020 14:05:16",
      "content": "<p>Tried SeResnext50. After the first epoch, the training AUC improved from 0.56 to 0.65, but after that, the AUC almost stopped improving for the subsequent epochs. I stopped training at 5th epoch when the AUC was around 0.67</p>",
      "rawMarkdown": "Tried SeResnext50. After the first epoch, the training AUC improved from 0.56 to 0.65, but after that, the AUC almost stopped improving for the subsequent epochs. I stopped training at 5th epoch when the AUC was around 0.67",
      "votes": null
    },
    {
      "id": "918853",
      "postDate": "07/07/2020 14:41:09",
      "content": "<p>See <a href=\"https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008\">https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008</a></p>",
      "rawMarkdown": "See https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008",
      "votes": null
    },
    {
      "id": "919138",
      "postDate": "07/07/2020 18:18:57",
      "content": "<p>Thanks for the direction! :D </p>",
      "rawMarkdown": "Thanks for the direction! :D",
      "votes": null
    },
    {
      "id": "919257",
      "postDate": "07/07/2020 19:15:50",
      "content": "<p>My SE-Resnext reaches LB .894 now. And its metric is still improving.</p>",
      "rawMarkdown": "My SE-Resnext reaches LB .894 now. And its metric is still improving.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 918792,
      "author_name": "mightyrains",
      "author_url": "",
      "post_date": "07/07/2020 13:52:15",
      "content": "<p>How poor a score seresnext50 gave?</p>",
      "votes": null,
      "replies": [
        {
          "id": 918807,
          "author_name": "udaykamal",
          "author_url": "",
          "post_date": "07/07/2020 14:05:16",
          "content": "<p>Tried SeResnext50. After the first epoch, the training AUC improved from 0.56 to 0.65, but after that, the AUC almost stopped improving for the subsequent epochs. I stopped training at 5th epoch when the AUC was around 0.67</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 919257,
          "author_name": "cttsai",
          "author_url": "",
          "post_date": "07/07/2020 19:15:50",
          "content": "<p>My SE-Resnext reaches LB .894 now. And its metric is still improving.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 918853,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "07/07/2020 14:41:09",
      "content": "<p>See <a href=\"https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008\">https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 919138,
          "author_name": "udaykamal",
          "author_url": "",
          "post_date": "07/07/2020 18:18:57",
          "content": "<p>Thanks for the direction! :D </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "918738": "Hi, I was wondering has anybody managed to get a decent public LB score without using any efficientnet model? I personally tried Densenet variants and SeResnext variants, but surprisingly, even with the exact similar setup,  some of these models never converged at all while the remaining models converged at a very poor public CV score. I was wondering what might be the reason for this.",
    "918792": "How poor a score seresnext50 gave?",
    "918807": "Tried SeResnext50. After the first epoch, the training AUC improved from 0.56 to 0.65, but after that, the AUC almost stopped improving for the subsequent epochs. I stopped training at 5th epoch when the AUC was around 0.67",
    "918853": "See https://www.kaggle.com/c/alaska2-image-steganalysis/discussion/162177#918008",
    "919138": "Thanks for the direction! :D",
    "919257": "My SE-Resnext reaches LB .894 now. And its metric is still improving."
  },
  "source": "meta"
}