{
  "id": 420079,
  "title": "Increasing image size doesn't work for me on LB",
  "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/420079",
  "author_name": "Tawara",
  "post_date": "2023-06-29T05:46:32.633000",
  "votes": 11,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In several CV tasks, bigger image size often gives us higher performance. I think this competition would be the same and many people got higher score.</p>\n<p>In my case, however, bigger image size got higher score on CV but lower score on LB…  It disappointed me so much 😭</p>\n<p>This is my first semantic segmentation task, maybe I make some mistakes…</p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Image size</th>\n<th>train folder(oof pred)</th>\n<th>validation folder(fold avg)</th>\n<th>LB(fold avg)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>resnet34d</td>\n<td>256</td>\n<td>0.659</td>\n<td>0.642</td>\n<td>0.657</td>\n</tr>\n<tr>\n<td>resnet34d</td>\n<td>512</td>\n<td>0.670</td>\n<td>0.654</td>\n<td>0.642</td>\n</tr>\n<tr>\n<td>resnest26d</td>\n<td>256</td>\n<td>0.660</td>\n<td>0.643</td>\n<td>0.657</td>\n</tr>\n<tr>\n<td>resnest26d</td>\n<td>512</td>\n<td>0.665</td>\n<td>0.648</td>\n<td>0.646</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": 2322155,
      "postDate": "2023-06-29T05:46:32.633Z",
      "content": "<p>In several CV tasks, bigger image size often gives us higher performance. I think this competition would be the same and many people got higher score.</p>\n<p>In my case, however, bigger image size got higher score on CV but lower score on LB…  It disappointed me so much 😭</p>\n<p>This is my first semantic segmentation task, maybe I make some mistakes…</p>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Image size</th>\n<th>train folder(oof pred)</th>\n<th>validation folder(fold avg)</th>\n<th>LB(fold avg)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>resnet34d</td>\n<td>256</td>\n<td>0.659</td>\n<td>0.642</td>\n<td>0.657</td>\n</tr>\n<tr>\n<td>resnet34d</td>\n<td>512</td>\n<td>0.670</td>\n<td>0.654</td>\n<td>0.642</td>\n</tr>\n<tr>\n<td>resnest26d</td>\n<td>256</td>\n<td>0.660</td>\n<td>0.643</td>\n<td>0.657</td>\n</tr>\n<tr>\n<td>resnest26d</td>\n<td>512</td>\n<td>0.665</td>\n<td>0.648</td>\n<td>0.646</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "In several CV tasks, bigger image size often gives us higher performance. I think this competition would be the same and many people got higher score.\n\nIn my case, however, bigger image size got higher score on CV but lower score on LB...  It disappointed me so much 😭\n\nThis is my first semantic segmentation task, maybe I make some mistakes...\n\n| Model | Image size | train folder(oof pred) | validation folder(fold avg) | LB(fold avg) |\n|:------:|:-----------:|:----------------------:|:--------------------------:|:-------------:|\n| resnet34d   | 256 | 0.659 | 0.642  | 0.657 |\n| resnet34d   | 512 | 0.670 | 0.654 | 0.642 |\n| resnest26d | 256 | 0.660 | 0.643 | 0.657 |\n| resnest26d | 512 | 0.665 | 0.648 | 0.646 |\n",
      "votes": 10
    },
    {
      "id": 2322873,
      "postDate": "2023-06-29T14:39:36.993Z",
      "content": "<p><a href=\"https://www.kaggle.com/ttahara\" target=\"_blank\">@ttahara</a> </p>\n<p>For me, 384 works very good compared to 512.<br>\nI used resnest101e</p>\n<p>The best I could get with 512 is 0.658<br>\nThe best I got earlier with 384 is 0.666</p>",
      "rawMarkdown": "@ttahara \n\nFor me, 384 works very good compared to 512.\nI used resnest101e\n\nThe best I could get with 512 is 0.658\nThe best I got earlier with 384 is 0.666",
      "votes": 6,
      "replies": [
        {
          "id": 2322880,
          "postDate": "2023-06-29T14:43:24.497Z",
          "content": "<p>Thanks for your information. I'll try it :)</p>",
          "rawMarkdown": "Thanks for your information. I'll try it :)"
        }
      ]
    },
    {
      "id": 2322988,
      "postDate": "2023-06-29T15:53:17.710Z",
      "content": "<p>Based on public lb size I would not trust it.</p>\n<p>I have done some experiments and I got the following:</p>\n<p>Effb7 val folder CV: 0.6682<br>\nEffb7 val folder LB: 0.644</p>\n<p>Effb3 val folder CV: 0.6506<br>\nEffb3 val folder LB: 0.656</p>\n<p>I have more experiments, for example another effb7 that got the following:<br>\nEffb7 val folder CV: 0.6729<br>\nEffb7 val folder LB: 0.662</p>\n<p>In this competition I would trust CV over LB, your CV strategy is nice, get out of folds CV and validation folder CV. In my opinion if both of them increase, you are good. </p>",
      "rawMarkdown": "Based on public lb size I would not trust it.\n\nI have done some experiments and I got the following:\n\nEffb7 val folder CV: 0.6682\nEffb7 val folder LB: 0.644\n\nEffb3 val folder CV: 0.6506\nEffb3 val folder LB: 0.656\n\nI have more experiments, for example another effb7 that got the following:\nEffb7 val folder CV: 0.6729\nEffb7 val folder LB: 0.662\n\nIn this competition I would trust CV over LB, your CV strategy is nice, get out of folds CV and validation folder CV. In my opinion if both of them increase, you are good. \n",
      "votes": 4
    }
  ],
  "comments": [
    {
      "id": 2322873,
      "author_name": "Balaji Selvaraj",
      "author_url": "",
      "post_date": "2023-06-29T14:39:36.993000",
      "content": "<p><a href=\"https://www.kaggle.com/ttahara\" target=\"_blank\">@ttahara</a> </p>\n<p>For me, 384 works very good compared to 512.<br>\nI used resnest101e</p>\n<p>The best I could get with 512 is 0.658<br>\nThe best I got earlier with 384 is 0.666</p>",
      "votes": 6,
      "replies": [
        {
          "id": 2322880,
          "author_name": "Tawara",
          "author_url": "",
          "post_date": "2023-06-29T14:43:24.497000",
          "content": "<p>Thanks for your information. I'll try it :)</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 2322988,
      "author_name": "Martin Kovacevic Buvinic",
      "author_url": "",
      "post_date": "2023-06-29T15:53:17.710000",
      "content": "<p>Based on public lb size I would not trust it.</p>\n<p>I have done some experiments and I got the following:</p>\n<p>Effb7 val folder CV: 0.6682<br>\nEffb7 val folder LB: 0.644</p>\n<p>Effb3 val folder CV: 0.6506<br>\nEffb3 val folder LB: 0.656</p>\n<p>I have more experiments, for example another effb7 that got the following:<br>\nEffb7 val folder CV: 0.6729<br>\nEffb7 val folder LB: 0.662</p>\n<p>In this competition I would trust CV over LB, your CV strategy is nice, get out of folds CV and validation folder CV. In my opinion if both of them increase, you are good. </p>",
      "votes": 4,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2322155": "In several CV tasks, bigger image size often gives us higher performance. I think this competition would be the same and many people got higher score.\n\nIn my case, however, bigger image size got higher score on CV but lower score on LB...  It disappointed me so much 😭\n\nThis is my first semantic segmentation task, maybe I make some mistakes...\n\n| Model | Image size | train folder(oof pred) | validation folder(fold avg) | LB(fold avg) |\n|:------:|:-----------:|:----------------------:|:--------------------------:|:-------------:|\n| resnet34d   | 256 | 0.659 | 0.642  | 0.657 |\n| resnet34d   | 512 | 0.670 | 0.654 | 0.642 |\n| resnest26d | 256 | 0.660 | 0.643 | 0.657 |\n| resnest26d | 512 | 0.665 | 0.648 | 0.646 |\n",
    "2322873": "@ttahara \n\nFor me, 384 works very good compared to 512.\nI used resnest101e\n\nThe best I could get with 512 is 0.658\nThe best I got earlier with 384 is 0.666",
    "2322988": "Based on public lb size I would not trust it.\n\nI have done some experiments and I got the following:\n\nEffb7 val folder CV: 0.6682\nEffb7 val folder LB: 0.644\n\nEffb3 val folder CV: 0.6506\nEffb3 val folder LB: 0.656\n\nI have more experiments, for example another effb7 that got the following:\nEffb7 val folder CV: 0.6729\nEffb7 val folder LB: 0.662\n\nIn this competition I would trust CV over LB, your CV strategy is nice, get out of folds CV and validation folder CV. In my opinion if both of them increase, you are good. \n"
  }
}