{
  "id": 162857,
  "title": "Pre-trained CNN V/S Randomly initialize CNN models?",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/162857",
  "author_name": "",
  "post_date": "2020-06-30T08:48:10.055096600Z",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Should i use pre-trained CNN models like VGG, resnet etc. first or i should go for basic CNN model with random initialization?</p>",
  "messages": [
    {
      "id": "907918",
      "postDate": "06/30/2020 08:48:10",
      "content": "<p>Should i use pre-trained CNN models like VGG, resnet etc. first or i should go for basic CNN model with random initialization?</p>",
      "rawMarkdown": "Should i use pre-trained CNN models like VGG, resnet etc. first or i should go for basic CNN model with random initialization?",
      "votes": null
    },
    {
      "id": "908088",
      "postDate": "06/30/2020 11:17:40",
      "content": "<p>Use pre-trained models for better score.</p>",
      "rawMarkdown": "Use pre-trained models for better score.",
      "votes": null
    },
    {
      "id": "909420",
      "postDate": "06/30/2020 15:21:01",
      "content": "<p>The pre-trained ones should deliver a better performance. Just an example for comparison for the ISIC 2017 challenge dataset can be found in this paper(table 4):\n<a href=\"https://www.sciencedirect.com/science/article/abs/pii/S0895611118306050\">https://www.sciencedirect.com/science/article/abs/pii/S0895611118306050</a></p>",
      "rawMarkdown": "The pre-trained ones should deliver a better performance. Just an example for comparison for the ISIC 2017 challenge dataset can be found in this paper(table 4):\nhttps://www.sciencedirect.com/science/article/abs/pii/S0895611118306050",
      "votes": null
    },
    {
      "id": "909709",
      "postDate": "06/30/2020 19:04:33",
      "content": "<p>Hey <a href=\"/anuragkushwah\">@anuragkushwah</a> , if training data is less or processing power is limited then you can go with transfer learning for better results. For an intro to transfer learning have look at this <a href=\"https://www.kaggle.com/manzoormahmood/transfer-learning-using-resnet50-and-pytorch\">notebook</a>.</p>",
      "rawMarkdown": "Hey @anuragkushwah , if training data is less or processing power is limited then you can go with transfer learning for better results. For an intro to transfer learning have look at this [notebook](https://www.kaggle.com/manzoormahmood/transfer-learning-using-resnet50-and-pytorch).",
      "votes": null
    },
    {
      "id": "909889",
      "postDate": "06/30/2020 22:38:16",
      "content": "<p>I have tried and experimented already - starting with a pretrained models gives far better results unless you find some extra data.</p>",
      "rawMarkdown": "I have tried and experimented already - starting with a pretrained models gives far better results unless you find some extra data.",
      "votes": null
    },
    {
      "id": "910159",
      "postDate": "07/01/2020 02:54:42",
      "content": "<p>Okk Thanks, I will definitely check it!</p>",
      "rawMarkdown": "Okk Thanks, I will definitely check it!",
      "votes": null
    },
    {
      "id": "910161",
      "postDate": "07/01/2020 02:56:55",
      "content": "<p>Thanks man !</p>",
      "rawMarkdown": "Thanks man !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 908088,
      "author_name": "ishalgarg",
      "author_url": "",
      "post_date": "06/30/2020 11:17:40",
      "content": "<p>Use pre-trained models for better score.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 909420,
      "author_name": "ipateam",
      "author_url": "",
      "post_date": "06/30/2020 15:21:01",
      "content": "<p>The pre-trained ones should deliver a better performance. Just an example for comparison for the ISIC 2017 challenge dataset can be found in this paper(table 4):\n<a href=\"https://www.sciencedirect.com/science/article/abs/pii/S0895611118306050\">https://www.sciencedirect.com/science/article/abs/pii/S0895611118306050</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 910161,
          "author_name": "anuragkushwah",
          "author_url": "",
          "post_date": "07/01/2020 02:56:55",
          "content": "<p>Thanks man !</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 909709,
      "author_name": "manzoormahmood",
      "author_url": "",
      "post_date": "06/30/2020 19:04:33",
      "content": "<p>Hey <a href=\"/anuragkushwah\">@anuragkushwah</a> , if training data is less or processing power is limited then you can go with transfer learning for better results. For an intro to transfer learning have look at this <a href=\"https://www.kaggle.com/manzoormahmood/transfer-learning-using-resnet50-and-pytorch\">notebook</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 910159,
          "author_name": "anuragkushwah",
          "author_url": "",
          "post_date": "07/01/2020 02:54:42",
          "content": "<p>Okk Thanks, I will definitely check it!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 909889,
      "author_name": "aroraaman",
      "author_url": "",
      "post_date": "06/30/2020 22:38:16",
      "content": "<p>I have tried and experimented already - starting with a pretrained models gives far better results unless you find some extra data.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "907918": "Should i use pre-trained CNN models like VGG, resnet etc. first or i should go for basic CNN model with random initialization?",
    "908088": "Use pre-trained models for better score.",
    "909420": "The pre-trained ones should deliver a better performance. Just an example for comparison for the ISIC 2017 challenge dataset can be found in this paper(table 4):\nhttps://www.sciencedirect.com/science/article/abs/pii/S0895611118306050",
    "909709": "Hey @anuragkushwah , if training data is less or processing power is limited then you can go with transfer learning for better results. For an intro to transfer learning have look at this [notebook](https://www.kaggle.com/manzoormahmood/transfer-learning-using-resnet50-and-pytorch).",
    "909889": "I have tried and experimented already - starting with a pretrained models gives far better results unless you find some extra data.",
    "910159": "Okk Thanks, I will definitely check it!",
    "910161": "Thanks man !"
  },
  "source": "meta"
}