{
  "id": 301766,
  "title": "16G GPU RAM & 9 hours inference Resource Allocation Competition",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/301766",
  "author_name": "",
  "post_date": "2022-01-19T08:38:47.579729Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I tried a large YOLO net (i.e. yolov5l) and it takes about close 9 hours for inference; and my each input test image cannot large than 3500 (because of 9 hours limit) cannot large than 5500 (because RAM limit).</p>\n<p>In this case, I can't increase resolution or using wbf with another model.<br>\nSo, the Key is compare with the result, and make a choice with <strong>\"one single best model\"</strong> and <strong>\"multi-tiny models\"</strong> which means we need to allocate 9 hours. Reduce the inference time will be a useful topic.<br>\nI guess Yolov5s could be the best solution.</p>\n<p>How about your single model inference time? Can you comment it here?</p>",
  "messages": [
    {
      "id": "1656293",
      "postDate": "01/19/2022 08:38:47",
      "content": "<p>I tried a large YOLO net (i.e. yolov5l) and it takes about close 9 hours for inference; and my each input test image cannot large than 3500 (because of 9 hours limit) cannot large than 5500 (because RAM limit).</p>\n<p>In this case, I can't increase resolution or using wbf with another model.<br>\nSo, the Key is compare with the result, and make a choice with <strong>\"one single best model\"</strong> and <strong>\"multi-tiny models\"</strong> which means we need to allocate 9 hours. Reduce the inference time will be a useful topic.<br>\nI guess Yolov5s could be the best solution.</p>\n<p>How about your single model inference time? Can you comment it here?</p>",
      "rawMarkdown": "I tried a large YOLO net (i.e. yolov5l) and it takes about close 9 hours for inference; and my each input test image cannot large than 3500 (because of 9 hours limit) cannot large than 5500 (because RAM limit).\n\nIn this case, I can't increase resolution or using wbf with another model.\nSo, the Key is compare with the result, and make a choice with **\"one single best model\"** and **\"multi-tiny models\"** which means we need to allocate 9 hours. Reduce the inference time will be a useful topic.\nI guess Yolov5s could be the best solution.\n\nHow about your single model inference time? Can you comment it here?",
      "votes": null
    },
    {
      "id": "1656556",
      "postDate": "01/19/2022 12:40:58",
      "content": "<p>According to public LB (<strong>TOP10</strong>) notebook submission time is from 1h to almost 9h … Conclusion -&gt; do not need 9h to jump in TOP10 zone -&gt; … here you can paste your conclusion about inference time (I do not ask about dataset, training way etc):</p>\n<ul>\n<li>TTA/WBF (yes/no?), if yes - TTA (how strong augumentation), WBF (how many models)</li>\n<li>Yolov5/YoloX/YoloR?</li>\n<li>model size - nano/S/M … L … X …..?</li>\n<li>resolution size - 640, 1280, 3600 ………. 90000</li>\n</ul>",
      "rawMarkdown": "According to public LB (**TOP10**) notebook submission time is from 1h to almost 9h ... Conclusion -> do not need 9h to jump in TOP10 zone -> ... here you can paste your conclusion about inference time (I do not ask about dataset, training way etc):\n- TTA/WBF (yes/no?), if yes - TTA (how strong augumentation), WBF (how many models)\n- Yolov5/YoloX/YoloR?\n- model size - nano/S/M ... L ... X .....?\n- resolution size - 640, 1280, 3600 .......... 90000",
      "votes": null
    },
    {
      "id": "1656571",
      "postDate": "01/19/2022 12:57:53",
      "content": "<p>I actually so far, I never succeed in WBF.😂</p>",
      "rawMarkdown": "I actually so far, I never succeed in WBF.😂",
      "votes": null
    },
    {
      "id": "1656939",
      "postDate": "01/19/2022 19:31:26",
      "content": "<p>90000 Lol 🤣</p>",
      "rawMarkdown": "90000 Lol 🤣",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1656556,
      "author_name": "remekkinas",
      "author_url": "",
      "post_date": "01/19/2022 12:40:58",
      "content": "<p>According to public LB (<strong>TOP10</strong>) notebook submission time is from 1h to almost 9h … Conclusion -&gt; do not need 9h to jump in TOP10 zone -&gt; … here you can paste your conclusion about inference time (I do not ask about dataset, training way etc):</p>\n<ul>\n<li>TTA/WBF (yes/no?), if yes - TTA (how strong augumentation), WBF (how many models)</li>\n<li>Yolov5/YoloX/YoloR?</li>\n<li>model size - nano/S/M … L … X …..?</li>\n<li>resolution size - 640, 1280, 3600 ………. 90000</li>\n</ul>",
      "votes": null,
      "replies": [
        {
          "id": 1656571,
          "author_name": "xiaojiu1414",
          "author_url": "",
          "post_date": "01/19/2022 12:57:53",
          "content": "<p>I actually so far, I never succeed in WBF.😂</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1656939,
          "author_name": "shashwatraman",
          "author_url": "",
          "post_date": "01/19/2022 19:31:26",
          "content": "<p>90000 Lol 🤣</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1656293": "I tried a large YOLO net (i.e. yolov5l) and it takes about close 9 hours for inference; and my each input test image cannot large than 3500 (because of 9 hours limit) cannot large than 5500 (because RAM limit).\n\nIn this case, I can't increase resolution or using wbf with another model.\nSo, the Key is compare with the result, and make a choice with **\"one single best model\"** and **\"multi-tiny models\"** which means we need to allocate 9 hours. Reduce the inference time will be a useful topic.\nI guess Yolov5s could be the best solution.\n\nHow about your single model inference time? Can you comment it here?",
    "1656556": "According to public LB (**TOP10**) notebook submission time is from 1h to almost 9h ... Conclusion -> do not need 9h to jump in TOP10 zone -> ... here you can paste your conclusion about inference time (I do not ask about dataset, training way etc):\n- TTA/WBF (yes/no?), if yes - TTA (how strong augumentation), WBF (how many models)\n- Yolov5/YoloX/YoloR?\n- model size - nano/S/M ... L ... X .....?\n- resolution size - 640, 1280, 3600 .......... 90000",
    "1656571": "I actually so far, I never succeed in WBF.😂",
    "1656939": "90000 Lol 🤣"
  },
  "source": "meta"
}