{
  "id": 156038,
  "title": "Melanoma Starter",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/156038",
  "author_name": "",
  "post_date": "2020-06-04T06:55:48.985741100Z",
  "votes": 21,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi everyone!</p>\n\n<p>I would like to share with you my research pipeline for this competition:</p>\n\n<ul>\n<li><a href=\"https://www.kaggle.com/shonenkov/training-cv-melanoma-starter\">[Training CV] Melanoma Starter</a></li>\n<li><a href=\"https://www.kaggle.com/shonenkov/inference-single-model-melanoma-starter\">[Inference Single Model] Melanoma Starter</a></li>\n</ul>\n\n<p>I have got for single model:</p>\n\n<p>0.927 (LB), 0.94976(Val)</p>\n\n<p>Welcome!</p>",
  "messages": [
    {
      "id": "873457",
      "postDate": "06/04/2020 06:55:48",
      "content": "<p>Hi everyone!</p>\n\n<p>I would like to share with you my research pipeline for this competition:</p>\n\n<ul>\n<li><a href=\"https://www.kaggle.com/shonenkov/training-cv-melanoma-starter\">[Training CV] Melanoma Starter</a></li>\n<li><a href=\"https://www.kaggle.com/shonenkov/inference-single-model-melanoma-starter\">[Inference Single Model] Melanoma Starter</a></li>\n</ul>\n\n<p>I have got for single model:</p>\n\n<p>0.927 (LB), 0.94976(Val)</p>\n\n<p>Welcome!</p>",
      "rawMarkdown": "Hi everyone!\n\nI would like to share with you my research pipeline for this competition:\n\n- [[Training CV] Melanoma Starter](https://www.kaggle.com/shonenkov/training-cv-melanoma-starter)\n- [[Inference Single Model] Melanoma Starter](https://www.kaggle.com/shonenkov/inference-single-model-melanoma-starter)\n\n\nI have got for single model:\n\n0.927 (LB), 0.94976(Val)\n\nWelcome!",
      "votes": null
    },
    {
      "id": "873468",
      "postDate": "06/04/2020 07:06:29",
      "content": "<p>Great work. Thanks for sharing</p>",
      "rawMarkdown": "Great work. Thanks for sharing",
      "votes": null
    },
    {
      "id": "873635",
      "postDate": "06/04/2020 10:11:48",
      "content": "<p>hey <a href=\"/shonenkov\">@shonenkov</a> thanks for sharing this clean pipeline.</p>\n\n<p>It seems that you are simply scaling images by dividing by 255, but pretrained efficientnet models expect this : \n<code>normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], \n                                     std=[0.229, 0.224, 0.225])</code> </p>\n\n<p>Since you are training all the weight I guess it's no big deal but I was wondering if this is on purpose? Does it help to this instead of normalizing?</p>",
      "rawMarkdown": "hey @shonenkov thanks for sharing this clean pipeline.\n\nIt seems that you are simply scaling images by dividing by 255, but pretrained efficientnet models expect this : \n`    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], \n                                     std=[0.229, 0.224, 0.225])` \n\nSince you are training all the weight I guess it's no big deal but I was wondering if this is on purpose? Does it help to this instead of normalizing?",
      "votes": null
    },
    {
      "id": "873675",
      "postDate": "06/04/2020 10:57:04",
      "content": "<p><a href=\"/optimo\">@optimo</a> thank you! </p>\n\n<p>I think imagenet normalization is right approach for using pretrained weights. But I like dividing by 255, as I don't notice differences when I train all layers. Main thing to use same transformation in inference stage :)</p>",
      "rawMarkdown": "optimo thank you! \n\nI think imagenet normalization is right approach for using pretrained weights. But I like dividing by 255, as I don't notice differences when I train all layers. Main thing to use same transformation in inference stage :)",
      "votes": null
    },
    {
      "id": "873734",
      "postDate": "06/04/2020 11:44:43",
      "content": "<p>fair enough, thank you!</p>",
      "rawMarkdown": "fair enough, thank you!",
      "votes": null
    },
    {
      "id": "895020",
      "postDate": "06/21/2020 03:32:30",
      "content": "<p>greate work buddy</p>",
      "rawMarkdown": "greate work buddy",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 873468,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "06/04/2020 07:06:29",
      "content": "<p>Great work. Thanks for sharing</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 873635,
      "author_name": "optimo",
      "author_url": "",
      "post_date": "06/04/2020 10:11:48",
      "content": "<p>hey <a href=\"/shonenkov\">@shonenkov</a> thanks for sharing this clean pipeline.</p>\n\n<p>It seems that you are simply scaling images by dividing by 255, but pretrained efficientnet models expect this : \n<code>normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], \n                                     std=[0.229, 0.224, 0.225])</code> </p>\n\n<p>Since you are training all the weight I guess it's no big deal but I was wondering if this is on purpose? Does it help to this instead of normalizing?</p>",
      "votes": null,
      "replies": [
        {
          "id": 873675,
          "author_name": "shonenkov",
          "author_url": "",
          "post_date": "06/04/2020 10:57:04",
          "content": "<p><a href=\"/optimo\">@optimo</a> thank you! </p>\n\n<p>I think imagenet normalization is right approach for using pretrained weights. But I like dividing by 255, as I don't notice differences when I train all layers. Main thing to use same transformation in inference stage :)</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 873734,
          "author_name": "optimo",
          "author_url": "",
          "post_date": "06/04/2020 11:44:43",
          "content": "<p>fair enough, thank you!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 895020,
      "author_name": "muralidhar123",
      "author_url": "",
      "post_date": "06/21/2020 03:32:30",
      "content": "<p>greate work buddy</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "873457": "Hi everyone!\n\nI would like to share with you my research pipeline for this competition:\n\n- [[Training CV] Melanoma Starter](https://www.kaggle.com/shonenkov/training-cv-melanoma-starter)\n- [[Inference Single Model] Melanoma Starter](https://www.kaggle.com/shonenkov/inference-single-model-melanoma-starter)\n\n\nI have got for single model:\n\n0.927 (LB), 0.94976(Val)\n\nWelcome!",
    "873468": "Great work. Thanks for sharing",
    "873635": "hey @shonenkov thanks for sharing this clean pipeline.\n\nIt seems that you are simply scaling images by dividing by 255, but pretrained efficientnet models expect this : \n`    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], \n                                     std=[0.229, 0.224, 0.225])` \n\nSince you are training all the weight I guess it's no big deal but I was wondering if this is on purpose? Does it help to this instead of normalizing?",
    "873675": "optimo thank you! \n\nI think imagenet normalization is right approach for using pretrained weights. But I like dividing by 255, as I don't notice differences when I train all layers. Main thing to use same transformation in inference stage :)",
    "873734": "fair enough, thank you!",
    "895020": "greate work buddy"
  },
  "source": "meta"
}