{
  "id": 417432,
  "title": "37 Rank model notebooks",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/417432",
  "author_name": "",
  "post_date": "2023-06-15T17:10:01.329658900Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<h2>My top 4 single model results</h2>\n<ol>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/best-model\" target=\"_blank\">Private LB 0.62</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/2nd-best-model\" target=\"_blank\">Private LB 0.61- Dynamic Pool</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/attention-model\" target=\"_blank\">Private LB 0.63 - Attention</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/attention-model\" target=\"_blank\">Private LB 0.61 - Large Receptor Fields </a></li>\n</ol>\n<h2>Hyperparameter</h2>\n<ul>\n<li>Image Resolution: 320x320</li>\n<li>Batch Size:  32</li>\n<li>LR: 6e-4</li>\n<li>Epochs: 14</li>\n<li>Scheduler: Cosine Anealing  + Step Lr </li>\n<li>Weight Decay: 4e-6</li>\n<li>Minimum Lr: 1e-7</li>\n<li>Channels used: 28 to 40</li>\n<li>Cutmix</li>\n<li>Validation Fragment: 1</li>\n<li>Tile Size: 320/4</li>\n</ul>\n<h2>Did not work for me</h2>\n<h2>Hyperparameter</h2>\n<ul>\n<li><p>Image Resolution: 224 x 224 , 256x256 , 384,384</p></li>\n<li><p>LR: 1e-4 , 3e-4 , 9e-4</p></li>\n<li><p>Epochs: 7, 10 , 30 </p></li>\n<li><p>Minimum Lr: 1e-6</p></li>\n<li><p>Channels used: any channels below 25 and above 40 decrease the model performance</p></li>\n<li><p>Validation Fragment: 3</p></li>\n</ul>",
  "messages": [
    {
      "id": "2304090",
      "postDate": "06/15/2023 17:10:01",
      "content": "<h2>My top 4 single model results</h2>\n<ol>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/best-model\" target=\"_blank\">Private LB 0.62</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/2nd-best-model\" target=\"_blank\">Private LB 0.61- Dynamic Pool</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/attention-model\" target=\"_blank\">Private LB 0.63 - Attention</a></li>\n<li><a href=\"https://www.kaggle.com/code/arunodhayan/attention-model\" target=\"_blank\">Private LB 0.61 - Large Receptor Fields </a></li>\n</ol>\n<h2>Hyperparameter</h2>\n<ul>\n<li>Image Resolution: 320x320</li>\n<li>Batch Size:  32</li>\n<li>LR: 6e-4</li>\n<li>Epochs: 14</li>\n<li>Scheduler: Cosine Anealing  + Step Lr </li>\n<li>Weight Decay: 4e-6</li>\n<li>Minimum Lr: 1e-7</li>\n<li>Channels used: 28 to 40</li>\n<li>Cutmix</li>\n<li>Validation Fragment: 1</li>\n<li>Tile Size: 320/4</li>\n</ul>\n<h2>Did not work for me</h2>\n<h2>Hyperparameter</h2>\n<ul>\n<li><p>Image Resolution: 224 x 224 , 256x256 , 384,384</p></li>\n<li><p>LR: 1e-4 , 3e-4 , 9e-4</p></li>\n<li><p>Epochs: 7, 10 , 30 </p></li>\n<li><p>Minimum Lr: 1e-6</p></li>\n<li><p>Channels used: any channels below 25 and above 40 decrease the model performance</p></li>\n<li><p>Validation Fragment: 3</p></li>\n</ul>",
      "rawMarkdown": "##My top 4 single model results \n1.  [Private LB 0.62](https://www.kaggle.com/code/arunodhayan/best-model)\n2.  [Private LB 0.61- Dynamic Pool](https://www.kaggle.com/code/arunodhayan/2nd-best-model)\n3. [Private LB 0.63 - Attention](https://www.kaggle.com/code/arunodhayan/attention-model)\n4. [Private LB 0.61 - Large Receptor Fields ](https://www.kaggle.com/code/arunodhayan/attention-model)\n\n##Hyperparameter\n- Image Resolution: 320x320\n- Batch Size:  32\n- LR: 6e-4\n- Epochs: 14\n- Scheduler: Cosine Anealing  + Step Lr \n- Weight Decay: 4e-6\n- Minimum Lr: 1e-7\n- Channels used: 28 to 40\n- Cutmix\n- Validation Fragment: 1\n- Tile Size: 320/4\n\n##Did not work for me \n##Hyperparameter\n- Image Resolution: 224 x 224 , 256x256 , 384,384\n- LR: 1e-4 , 3e-4 , 9e-4\n- Epochs: 7, 10 , 30 \n- Minimum Lr: 1e-6\n- Channels used: any channels below 25 and above 40 decrease the model performance\n\n- Validation Fragment: 3",
      "votes": null
    },
    {
      "id": "2304442",
      "postDate": "06/16/2023 02:13:37",
      "content": "<p>but these link to notebook is private?</p>",
      "rawMarkdown": "but these link to notebook is private?",
      "votes": null
    },
    {
      "id": "2304669",
      "postDate": "06/16/2023 06:43:09",
      "content": "<p>no i made it public</p>",
      "rawMarkdown": "no i made it public",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2304442,
      "author_name": "cutebomb",
      "author_url": "",
      "post_date": "06/16/2023 02:13:37",
      "content": "<p>but these link to notebook is private?</p>",
      "votes": null,
      "replies": [
        {
          "id": 2304669,
          "author_name": "arunodhayan",
          "author_url": "",
          "post_date": "06/16/2023 06:43:09",
          "content": "<p>no i made it public</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2304090": "##My top 4 single model results \n1.  [Private LB 0.62](https://www.kaggle.com/code/arunodhayan/best-model)\n2.  [Private LB 0.61- Dynamic Pool](https://www.kaggle.com/code/arunodhayan/2nd-best-model)\n3. [Private LB 0.63 - Attention](https://www.kaggle.com/code/arunodhayan/attention-model)\n4. [Private LB 0.61 - Large Receptor Fields ](https://www.kaggle.com/code/arunodhayan/attention-model)\n\n##Hyperparameter\n- Image Resolution: 320x320\n- Batch Size:  32\n- LR: 6e-4\n- Epochs: 14\n- Scheduler: Cosine Anealing  + Step Lr \n- Weight Decay: 4e-6\n- Minimum Lr: 1e-7\n- Channels used: 28 to 40\n- Cutmix\n- Validation Fragment: 1\n- Tile Size: 320/4\n\n##Did not work for me \n##Hyperparameter\n- Image Resolution: 224 x 224 , 256x256 , 384,384\n- LR: 1e-4 , 3e-4 , 9e-4\n- Epochs: 7, 10 , 30 \n- Minimum Lr: 1e-6\n- Channels used: any channels below 25 and above 40 decrease the model performance\n\n- Validation Fragment: 3",
    "2304442": "but these link to notebook is private?",
    "2304669": "no i made it public"
  },
  "source": "meta"
}