{
  "id": 313472,
  "title": "why effnetv2 not beating effnetv1?",
  "url": "/competitions/happy-whale-and-dolphin/discussion/313472",
  "author_name": "",
  "post_date": "2022-03-17T09:25:01.495354700Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I  forked public notebook of effnetv2.</p>\n<p>compared with effnetv1, both big models, almost same pipeline, </p>\n<p>without ensemble, V2 LB  is around 0.03 behind of V1.</p>\n<p>any explanation?</p>",
  "messages": [
    {
      "id": "1725640",
      "postDate": "03/17/2022 09:25:01",
      "content": "<p>I  forked public notebook of effnetv2.</p>\n<p>compared with effnetv1, both big models, almost same pipeline, </p>\n<p>without ensemble, V2 LB  is around 0.03 behind of V1.</p>\n<p>any explanation?</p>",
      "rawMarkdown": "I  forked public notebook of effnetv2.\n\ncompared with effnetv1, both big models, almost same pipeline, \n\nwithout ensemble, V2 LB  is around 0.03 behind of V1.\n\n\nany explanation?",
      "votes": null
    },
    {
      "id": "1725651",
      "postDate": "03/17/2022 09:38:10",
      "content": "<p>effnetv2 beats effnetv1 in my case. Probably it is not a great deal to compare two different models with the same hyperparameters, because both of them have their unique optimal ones.</p>",
      "rawMarkdown": "effnetv2 beats effnetv1 in my case. Probably it is not a great deal to compare two different models with the same hyperparameters, because both of them have their unique optimal ones.",
      "votes": null
    },
    {
      "id": "1725668",
      "postDate": "03/17/2022 09:56:53",
      "content": "<p>thanks for your reply.  </p>\n<p>In other competition, v2 beats v1 easily.</p>\n<p>I will try tune parameters.</p>",
      "rawMarkdown": "thanks for your reply.  \n\nIn other competition, v2 beats v1 easily.\n\nI will try tune parameters.",
      "votes": null
    },
    {
      "id": "1727903",
      "postDate": "03/18/2022 12:35:03",
      "content": "<p>effnet v1 and 2 are very similar wrt hyperparameters. I found no improvement re-optimizing the usual suspects for v2. I also get slightly better results from effnetv2-l than from v1-7b using the same parameters.</p>",
      "rawMarkdown": "effnet v1 and 2 are very similar wrt hyperparameters. I found no improvement re-optimizing the usual suspects for v2. I also get slightly better results from effnetv2-l than from v1-7b using the same parameters.",
      "votes": null
    },
    {
      "id": "1736212",
      "postDate": "03/27/2022 05:28:22",
      "content": "<p>I got 0.729 from V2-XL, up from the 0.699 that the person before me got with a V1-7. (I copied their notebook, changed the model and some lines of code and ran)</p>",
      "rawMarkdown": "I got 0.729 from V2-XL, up from the 0.699 that the person before me got with a V1-7. (I copied their notebook, changed the model and some lines of code and ran)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1725651,
      "author_name": "achilles38",
      "author_url": "",
      "post_date": "03/17/2022 09:38:10",
      "content": "<p>effnetv2 beats effnetv1 in my case. Probably it is not a great deal to compare two different models with the same hyperparameters, because both of them have their unique optimal ones.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1725668,
          "author_name": "dragonzhang",
          "author_url": "",
          "post_date": "03/17/2022 09:56:53",
          "content": "<p>thanks for your reply.  </p>\n<p>In other competition, v2 beats v1 easily.</p>\n<p>I will try tune parameters.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1727903,
          "author_name": "greendolphin",
          "author_url": "",
          "post_date": "03/18/2022 12:35:03",
          "content": "<p>effnet v1 and 2 are very similar wrt hyperparameters. I found no improvement re-optimizing the usual suspects for v2. I also get slightly better results from effnetv2-l than from v1-7b using the same parameters.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1736212,
      "author_name": "dentistdad",
      "author_url": "",
      "post_date": "03/27/2022 05:28:22",
      "content": "<p>I got 0.729 from V2-XL, up from the 0.699 that the person before me got with a V1-7. (I copied their notebook, changed the model and some lines of code and ran)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1725640": "I  forked public notebook of effnetv2.\n\ncompared with effnetv1, both big models, almost same pipeline, \n\nwithout ensemble, V2 LB  is around 0.03 behind of V1.\n\n\nany explanation?",
    "1725651": "effnetv2 beats effnetv1 in my case. Probably it is not a great deal to compare two different models with the same hyperparameters, because both of them have their unique optimal ones.",
    "1725668": "thanks for your reply.  \n\nIn other competition, v2 beats v1 easily.\n\nI will try tune parameters.",
    "1727903": "effnet v1 and 2 are very similar wrt hyperparameters. I found no improvement re-optimizing the usual suspects for v2. I also get slightly better results from effnetv2-l than from v1-7b using the same parameters.",
    "1736212": "I got 0.729 from V2-XL, up from the 0.699 that the person before me got with a V1-7. (I copied their notebook, changed the model and some lines of code and ran)"
  },
  "source": "meta"
}