{
  "id": 93587,
  "title": "is this approach valid for this competition?",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/93587",
  "author_name": "",
  "post_date": "2019-05-28T12:52:57.613566800Z",
  "votes": 4,
  "comment_count": 4,
  "views": 0,
  "content": "<ol>\n<li><p>use noisy data (because it has lot of data compared to curated) and train with couple of epochs (maybe a lot?) and save weight.  - like using pretrained weight trained with imagenet</p></li>\n<li><p>load the weights that are used to train noisy data</p></li>\n<li><p>use the weights to train curated and predict </p></li>\n</ol>\n\n<p>could this approach work? (I haven't opened the noisy data yet)</p>\n\n<p>how are people using noisy data??</p>",
  "messages": [
    {
      "id": "538346",
      "postDate": "05/28/2019 12:52:57",
      "content": "<ol>\n<li><p>use noisy data (because it has lot of data compared to curated) and train with couple of epochs (maybe a lot?) and save weight.  - like using pretrained weight trained with imagenet</p></li>\n<li><p>load the weights that are used to train noisy data</p></li>\n<li><p>use the weights to train curated and predict </p></li>\n</ol>\n\n<p>could this approach work? (I haven't opened the noisy data yet)</p>\n\n<p>how are people using noisy data??</p>",
      "rawMarkdown": "1. use noisy data (because it has lot of data compared to curated) and train with couple of epochs (maybe a lot?) and save weight.  - like using pretrained weight trained with imagenet\n\n2. load the weights that are used to train noisy data\n\n3. use the weights to train curated and predict \n\ncould this approach work? (I haven't opened the noisy data yet)\n\nhow are people using noisy data??",
      "votes": null
    },
    {
      "id": "538390",
      "postDate": "05/28/2019 14:08:52",
      "content": "<p>There is a topic on that already - <a href=\"https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536\">https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536</a></p>\n\n<p>Pretraining on noisy and fine-tuning on curated didn't make any improvement comparing to curated only for me. The opposite approach did.</p>",
      "rawMarkdown": "There is a topic on that already - https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536\n\nPretraining on noisy and fine-tuning on curated didn't make any improvement comparing to curated only for me. The opposite approach did.",
      "votes": null
    },
    {
      "id": "538480",
      "postDate": "05/28/2019 16:03:11",
      "content": "<p>Pretraining on curated and fine-tuning on noisy  make improvement comparing to curated only ?</p>",
      "rawMarkdown": "Pretraining on curated and fine-tuning on noisy  make improvement comparing to curated only ?",
      "votes": null
    },
    {
      "id": "538497",
      "postDate": "05/28/2019 16:29:29",
      "content": "<p>Training on curated and then on curated plus noisy actually.</p>",
      "rawMarkdown": "Training on curated and then on curated plus noisy actually.",
      "votes": null
    },
    {
      "id": "538533",
      "postDate": "05/28/2019 17:42:12",
      "content": "<p>OK! I understand. thank you.</p>",
      "rawMarkdown": "OK! I understand. thank you.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 538390,
      "author_name": "vzaguskin",
      "author_url": "",
      "post_date": "05/28/2019 14:08:52",
      "content": "<p>There is a topic on that already - <a href=\"https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536\">https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536</a></p>\n\n<p>Pretraining on noisy and fine-tuning on curated didn't make any improvement comparing to curated only for me. The opposite approach did.</p>",
      "votes": null,
      "replies": [
        {
          "id": 538480,
          "author_name": "bossimuimu",
          "author_url": "",
          "post_date": "05/28/2019 16:03:11",
          "content": "<p>Pretraining on curated and fine-tuning on noisy  make improvement comparing to curated only ?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538497,
          "author_name": "vzaguskin",
          "author_url": "",
          "post_date": "05/28/2019 16:29:29",
          "content": "<p>Training on curated and then on curated plus noisy actually.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 538533,
          "author_name": "bossimuimu",
          "author_url": "",
          "post_date": "05/28/2019 17:42:12",
          "content": "<p>OK! I understand. thank you.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "538346": "1. use noisy data (because it has lot of data compared to curated) and train with couple of epochs (maybe a lot?) and save weight.  - like using pretrained weight trained with imagenet\n\n2. load the weights that are used to train noisy data\n\n3. use the weights to train curated and predict \n\ncould this approach work? (I haven't opened the noisy data yet)\n\nhow are people using noisy data??",
    "538390": "There is a topic on that already - https://www.kaggle.com/c/freesound-audio-tagging-2019/discussion/90536\n\nPretraining on noisy and fine-tuning on curated didn't make any improvement comparing to curated only for me. The opposite approach did.",
    "538480": "Pretraining on curated and fine-tuning on noisy  make improvement comparing to curated only ?",
    "538497": "Training on curated and then on curated plus noisy actually.",
    "538533": "OK! I understand. thank you."
  },
  "source": "meta"
}