{
  "id": 168649,
  "title": "How many efficientDet epochs are you running?",
  "url": "/competitions/global-wheat-detection/discussion/168649",
  "author_name": "",
  "post_date": "2020-07-21T11:27:20.344890300Z",
  "votes": 2,
  "comment_count": 15,
  "views": 0,
  "content": "<p>I ran 40 times and got lb: 0.7219. (EfficientDet-D5)</p>",
  "messages": [
    {
      "id": "938156",
      "postDate": "07/21/2020 11:27:20",
      "content": "<p>I ran 40 times and got lb: 0.7219. (EfficientDet-D5)</p>",
      "rawMarkdown": "I ran 40 times and got lb: 0.7219. (EfficientDet-D5)",
      "votes": null
    },
    {
      "id": "939304",
      "postDate": "07/22/2020 06:47:06",
      "content": "<p>near 40-50. Effdet-d5 too.\nI think it depends of learning rate and scheduler params</p>",
      "rawMarkdown": "near 40-50. Effdet-d5 too.\nI think it depends of learning rate and scheduler params",
      "votes": null
    },
    {
      "id": "940199",
      "postDate": "07/22/2020 18:21:56",
      "content": "<p>Have you experimented with Effdet-d6？</p>",
      "rawMarkdown": "Have you experimented with Effdet-d6？",
      "votes": null
    },
    {
      "id": "940881",
      "postDate": "07/23/2020 06:02:28",
      "content": "<p>I can't actually train more than 40 epochs with D5-D7.  I'm computing the score during validation in each epoch; this is taking up time.  Got at best ~0.68.</p>",
      "rawMarkdown": "I can't actually train more than 40 epochs with D5-D7.  I'm computing the score during validation in each epoch; this is taking up time.  Got at best ~0.68.",
      "votes": null
    },
    {
      "id": "941419",
      "postDate": "07/23/2020 08:03:14",
      "content": "<p>It takes about 8 hours to run 40 epochs using D5 on a Keggle notebook. \nTherefore, there is not enough time to try d6.</p>",
      "rawMarkdown": "It takes about 8 hours to run 40 epochs using D5 on a Keggle notebook. \nTherefore, there is not enough time to try d6.",
      "votes": null
    },
    {
      "id": "944015",
      "postDate": "07/24/2020 18:33:50",
      "content": "<p>I train 40 epochs on D7, got 0.7209</p>",
      "rawMarkdown": "I train 40 epochs on D7, got 0.7209",
      "votes": null
    },
    {
      "id": "944032",
      "postDate": "07/24/2020 18:57:17",
      "content": "<p>100 epoch for D7,  I got 0.7399 (inference with TTA)\nBest model weigths I got near epoch 66</p>",
      "rawMarkdown": "100 epoch for D7,  I got 0.7399 (inference with TTA)\nBest model weigths I got near epoch 66",
      "votes": null
    },
    {
      "id": "944033",
      "postDate": "07/24/2020 18:58:40",
      "content": "<p>do you use many tweaks?</p>",
      "rawMarkdown": "do you use many tweaks?",
      "votes": null
    },
    {
      "id": "944034",
      "postDate": "07/24/2020 18:58:44",
      "content": "<p>I will suggest try running for 80~90 epoch. </p>",
      "rawMarkdown": "I will suggest try running for 80~90 epoch.",
      "votes": null
    },
    {
      "id": "944035",
      "postDate": "07/24/2020 18:59:18",
      "content": "<p>Nope I used alex kernal. I only change the augmentations</p>",
      "rawMarkdown": "Nope I used alex kernal. I only change the augmentations",
      "votes": null
    },
    {
      "id": "944038",
      "postDate": "07/24/2020 19:01:49",
      "content": "<p>I tried some change and get best loss on 29/40 epochs\nlr=0.00015. Not good enough</p>",
      "rawMarkdown": "I tried some change and get best loss on 29/40 epochs\nlr=0.00015. Not good enough",
      "votes": null
    },
    {
      "id": "944570",
      "postDate": "07/25/2020 07:34:01",
      "content": "<p>How much time does it take for d7 100 epoch?</p>",
      "rawMarkdown": "How much time does it take for d7 100 epoch?",
      "votes": null
    },
    {
      "id": "944628",
      "postDate": "07/25/2020 08:09:28",
      "content": "<p>26min/1epochs *100</p>",
      "rawMarkdown": "26min/1epochs *100",
      "votes": null
    },
    {
      "id": "946154",
      "postDate": "07/26/2020 11:57:08",
      "content": "<p>d4 100 epoch no pl: 0.725</p>",
      "rawMarkdown": "d4 100 epoch no pl: 0.725",
      "votes": null
    },
    {
      "id": "949452",
      "postDate": "07/28/2020 16:36:09",
      "content": "<p><a href=\"/alexwang2019\">@alexwang2019</a>  is 72.5 cv mAP ?</p>",
      "rawMarkdown": "alexwang2019  is 72.5 cv mAP ?",
      "votes": null
    },
    {
      "id": "949534",
      "postDate": "07/28/2020 17:37:09",
      "content": "<p>single model with wbf LB score</p>",
      "rawMarkdown": "single model with wbf LB score",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 939304,
      "author_name": "kupchanski",
      "author_url": "",
      "post_date": "07/22/2020 06:47:06",
      "content": "<p>near 40-50. Effdet-d5 too.\nI think it depends of learning rate and scheduler params</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 940199,
      "author_name": "irinali",
      "author_url": "",
      "post_date": "07/22/2020 18:21:56",
      "content": "<p>Have you experimented with Effdet-d6？</p>",
      "votes": null,
      "replies": [
        {
          "id": 941419,
          "author_name": "wonjunpark",
          "author_url": "",
          "post_date": "07/23/2020 08:03:14",
          "content": "<p>It takes about 8 hours to run 40 epochs using D5 on a Keggle notebook. \nTherefore, there is not enough time to try d6.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 940881,
      "author_name": "jackchungchiehyu",
      "author_url": "",
      "post_date": "07/23/2020 06:02:28",
      "content": "<p>I can't actually train more than 40 epochs with D5-D7.  I'm computing the score during validation in each epoch; this is taking up time.  Got at best ~0.68.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 944015,
      "author_name": "doanquanvietnamca",
      "author_url": "",
      "post_date": "07/24/2020 18:33:50",
      "content": "<p>I train 40 epochs on D7, got 0.7209</p>",
      "votes": null,
      "replies": [
        {
          "id": 944032,
          "author_name": "ahmmadrashid",
          "author_url": "",
          "post_date": "07/24/2020 18:57:17",
          "content": "<p>100 epoch for D7,  I got 0.7399 (inference with TTA)\nBest model weigths I got near epoch 66</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 944033,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/24/2020 18:58:40",
          "content": "<p>do you use many tweaks?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 944035,
          "author_name": "ahmmadrashid",
          "author_url": "",
          "post_date": "07/24/2020 18:59:18",
          "content": "<p>Nope I used alex kernal. I only change the augmentations</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 944038,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/24/2020 19:01:49",
          "content": "<p>I tried some change and get best loss on 29/40 epochs\nlr=0.00015. Not good enough</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 944570,
          "author_name": "wonjunpark",
          "author_url": "",
          "post_date": "07/25/2020 07:34:01",
          "content": "<p>How much time does it take for d7 100 epoch?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 944628,
          "author_name": "doanquanvietnamca",
          "author_url": "",
          "post_date": "07/25/2020 08:09:28",
          "content": "<p>26min/1epochs *100</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 944034,
      "author_name": "ahmmadrashid",
      "author_url": "",
      "post_date": "07/24/2020 18:58:44",
      "content": "<p>I will suggest try running for 80~90 epoch. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 946154,
      "author_name": "alexwang2019",
      "author_url": "",
      "post_date": "07/26/2020 11:57:08",
      "content": "<p>d4 100 epoch no pl: 0.725</p>",
      "votes": null,
      "replies": [
        {
          "id": 949452,
          "author_name": "jaideepvalani",
          "author_url": "",
          "post_date": "07/28/2020 16:36:09",
          "content": "<p><a href=\"/alexwang2019\">@alexwang2019</a>  is 72.5 cv mAP ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 949534,
          "author_name": "alexwang2019",
          "author_url": "",
          "post_date": "07/28/2020 17:37:09",
          "content": "<p>single model with wbf LB score</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "938156": "I ran 40 times and got lb: 0.7219. (EfficientDet-D5)",
    "939304": "near 40-50. Effdet-d5 too.\nI think it depends of learning rate and scheduler params",
    "940199": "Have you experimented with Effdet-d6？",
    "940881": "I can't actually train more than 40 epochs with D5-D7.  I'm computing the score during validation in each epoch; this is taking up time.  Got at best ~0.68.",
    "941419": "It takes about 8 hours to run 40 epochs using D5 on a Keggle notebook. \nTherefore, there is not enough time to try d6.",
    "944015": "I train 40 epochs on D7, got 0.7209",
    "944032": "100 epoch for D7,  I got 0.7399 (inference with TTA)\nBest model weigths I got near epoch 66",
    "944033": "do you use many tweaks?",
    "944034": "I will suggest try running for 80~90 epoch.",
    "944035": "Nope I used alex kernal. I only change the augmentations",
    "944038": "I tried some change and get best loss on 29/40 epochs\nlr=0.00015. Not good enough",
    "944570": "How much time does it take for d7 100 epoch?",
    "944628": "26min/1epochs *100",
    "946154": "d4 100 epoch no pl: 0.725",
    "949452": "alexwang2019  is 72.5 cv mAP ?",
    "949534": "single model with wbf LB score"
  },
  "source": "meta"
}