{
  "id": 459938,
  "title": "How to successfully do model ensemble/tta?",
  "url": "/competitions/UBC-OCEAN/discussion/459938",
  "author_name": "阳光开朗大男孩",
  "post_date": "2023-12-07T05:28:00.611000",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I'm curious if anyone has tried model ensemble/tta when using patch/tile-based method? According to my experiments, the inference time for just a single efficientnet_b0 model based on patch has been as high as 7-8 hours (with parallel processing). My question is, is model ensemble/tta feasible in this competition, or should we trust the single model?</p>",
  "messages": [
    {
      "id": 2553299,
      "postDate": "2023-12-08T06:10:25.313Z",
      "content": "<p>Most time consuming tasks are reading images and creating patches which takes approximately 8 hours. You have 4 hours left for inference so you can ensemble plenty of models within that time.</p>",
      "rawMarkdown": "Most time consuming tasks are reading images and creating patches which takes approximately 8 hours. You have 4 hours left for inference so you can ensemble plenty of models within that time.",
      "votes": 1
    },
    {
      "id": 2553277,
      "postDate": "2023-12-08T05:51:00.010Z",
      "content": "<p>Yes, I have tried ensembling and tta using tiles (patches), takes around 7 hours. </p>",
      "rawMarkdown": "Yes, I have tried ensembling and tta using tiles (patches), takes around 7 hours. ",
      "votes": 1
    },
    {
      "id": 2551989,
      "postDate": "2023-12-07T05:28:00.610Z",
      "content": "<p>I'm curious if anyone has tried model ensemble/tta when using patch/tile-based method? According to my experiments, the inference time for just a single efficientnet_b0 model based on patch has been as high as 7-8 hours (with parallel processing). My question is, is model ensemble/tta feasible in this competition, or should we trust the single model?</p>",
      "rawMarkdown": "I'm curious if anyone has tried model ensemble/tta when using patch/tile-based method? According to my experiments, the inference time for just a single efficientnet_b0 model based on patch has been as high as 7-8 hours (with parallel processing). My question is, is model ensemble/tta feasible in this competition, or should we trust the single model?",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 2553299,
      "author_name": "Gunes Evitan",
      "author_url": "",
      "post_date": "2023-12-08T06:10:25.313000",
      "content": "<p>Most time consuming tasks are reading images and creating patches which takes approximately 8 hours. You have 4 hours left for inference so you can ensemble plenty of models within that time.</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 2553277,
      "author_name": "samu2505",
      "author_url": "",
      "post_date": "2023-12-08T05:51:00.010000",
      "content": "<p>Yes, I have tried ensembling and tta using tiles (patches), takes around 7 hours. </p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2553299": "Most time consuming tasks are reading images and creating patches which takes approximately 8 hours. You have 4 hours left for inference so you can ensemble plenty of models within that time.",
    "2553277": "Yes, I have tried ensembling and tta using tiles (patches), takes around 7 hours. ",
    "2551989": "I'm curious if anyone has tried model ensemble/tta when using patch/tile-based method? According to my experiments, the inference time for just a single efficientnet_b0 model based on patch has been as high as 7-8 hours (with parallel processing). My question is, is model ensemble/tta feasible in this competition, or should we trust the single model?"
  }
}