{
  "id": 61966,
  "title": "Some information for our solution",
  "url": "/competitions/freesound-audio-tagging/discussion/61966",
  "author_name": "",
  "post_date": "2018-07-25T16:23:45.998576100Z",
  "votes": 10,
  "comment_count": 4,
  "views": 0,
  "content": "<ul>\n<li><p>We have tested both ImageNet-based pre-trained models, and the models trained from the scratch.</p></li>\n<li><p>Now, using the single NN architecture, we can get about 0.967.</p></li>\n<li><p>Now, we ensemble 3 models to get: 0.973.</p></li>\n<li><p>7 models to get our current score: 0.976.</p></li>\n<li><p>We believe that we can reach about 0.980 without assembling too many models (less than 10 models). But we don't intend to do that, with the goal to maintain the simplicity of our solution.</p></li>\n</ul>",
  "messages": [
    {
      "id": "362077",
      "postDate": "07/25/2018 16:23:46",
      "content": "<ul>\n<li><p>We have tested both ImageNet-based pre-trained models, and the models trained from the scratch.</p></li>\n<li><p>Now, using the single NN architecture, we can get about 0.967.</p></li>\n<li><p>Now, we ensemble 3 models to get: 0.973.</p></li>\n<li><p>7 models to get our current score: 0.976.</p></li>\n<li><p>We believe that we can reach about 0.980 without assembling too many models (less than 10 models). But we don't intend to do that, with the goal to maintain the simplicity of our solution.</p></li>\n</ul>",
      "rawMarkdown": "We have tested both ImageNet-based pre-trained models, and the models trained from the scratch.\n\n - Now, using the single NN architecture, we can get about 0.967.\n\n - Now, we ensemble 3 models to get: 0.973.\n\n - 7 models to get our current score: 0.976.\n\n - We believe that we can reach about 0.980 without assembling too many models (less than 10 models). But we don't intend to do that, with the goal to maintain the simplicity of our solution.",
      "votes": null
    },
    {
      "id": "362667",
      "postDate": "07/26/2018 21:13:04",
      "content": "<p>Thanks for this thread! Awaiting to read more detail about your solution after the end of a competition.</p>",
      "rawMarkdown": "Thanks for this thread! Awaiting to read more detail about your solution after the end of a competition.",
      "votes": null
    },
    {
      "id": "362800",
      "postDate": "07/27/2018 06:23:07",
      "content": "<p>Sure, we will post a short summary and submit a paper to the DCASE workshop.</p>",
      "rawMarkdown": "Sure, we will post a short summary and submit a paper to the DCASE workshop.",
      "votes": null
    },
    {
      "id": "362871",
      "postDate": "07/27/2018 10:05:43",
      "content": "<p>Congratulations for such a wonderful performance! One question. Using the single NN architecture which fetches you 0.967 LB score, are these weights learnt from scratch? Or are these pre-trained weights? The maximum performance I could squeeze out is only ~0.94 on single model, weights trained from scratch and without noisy label correction.</p>",
      "rawMarkdown": "Congratulations for such a wonderful performance! One question. Using the single NN architecture which fetches you 0.967 LB score, are these weights learnt from scratch? Or are these pre-trained weights? The maximum performance I could squeeze out is only ~0.94 on single model, weights trained from scratch and without noisy label correction.",
      "votes": null
    },
    {
      "id": "363296",
      "postDate": "07/28/2018 13:44:51",
      "content": "<p>As aforementioned, both methods are tested. It seems that, sometimes, the model trained from the scratch provides better performance than the pre-trained.</p>",
      "rawMarkdown": "As aforementioned, both methods are tested. It seems that, sometimes, the model trained from the scratch provides better performance than the pre-trained.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 362667,
      "author_name": "opanichev",
      "author_url": "",
      "post_date": "07/26/2018 21:13:04",
      "content": "<p>Thanks for this thread! Awaiting to read more detail about your solution after the end of a competition.</p>",
      "votes": null,
      "replies": [
        {
          "id": 362800,
          "author_name": "kelexu",
          "author_url": "",
          "post_date": "07/27/2018 06:23:07",
          "content": "<p>Sure, we will post a short summary and submit a paper to the DCASE workshop.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 362871,
      "author_name": "gyat2017",
      "author_url": "",
      "post_date": "07/27/2018 10:05:43",
      "content": "<p>Congratulations for such a wonderful performance! One question. Using the single NN architecture which fetches you 0.967 LB score, are these weights learnt from scratch? Or are these pre-trained weights? The maximum performance I could squeeze out is only ~0.94 on single model, weights trained from scratch and without noisy label correction.</p>",
      "votes": null,
      "replies": [
        {
          "id": 363296,
          "author_name": "kelexu",
          "author_url": "",
          "post_date": "07/28/2018 13:44:51",
          "content": "<p>As aforementioned, both methods are tested. It seems that, sometimes, the model trained from the scratch provides better performance than the pre-trained.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "362077": "We have tested both ImageNet-based pre-trained models, and the models trained from the scratch.\n\n - Now, using the single NN architecture, we can get about 0.967.\n\n - Now, we ensemble 3 models to get: 0.973.\n\n - 7 models to get our current score: 0.976.\n\n - We believe that we can reach about 0.980 without assembling too many models (less than 10 models). But we don't intend to do that, with the goal to maintain the simplicity of our solution.",
    "362667": "Thanks for this thread! Awaiting to read more detail about your solution after the end of a competition.",
    "362800": "Sure, we will post a short summary and submit a paper to the DCASE workshop.",
    "362871": "Congratulations for such a wonderful performance! One question. Using the single NN architecture which fetches you 0.967 LB score, are these weights learnt from scratch? Or are these pre-trained weights? The maximum performance I could squeeze out is only ~0.94 on single model, weights trained from scratch and without noisy label correction.",
    "363296": "As aforementioned, both methods are tested. It seems that, sometimes, the model trained from the scratch provides better performance than the pre-trained."
  },
  "source": "meta"
}