{
  "id": 95347,
  "title": "Main features of final solution [22 on public]",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/95347",
  "author_name": "Aimoldin Anuar [dsmlkz]",
  "post_date": "2019-06-11T16:01:28.376000",
  "votes": 18,
  "comment_count": 4,
  "views": 0,
  "content": "<p>1.) Using power(0.125) instead of log in Spectrograms;\n2) Selection balanced (or unbalanced) sample of well predicted by current model noisy files (1500, 2500, 5000, 10000). Priority for rarely predicted files on the validation;\n3) Different parameters of the spectrogram extraction (hop, n_mels). We used:\n- rate: 44100, n_mels: {128, 256}, n_fft: 2560, hop_length: {512, 694, 1024}, fmin: 20;\n- dynamic hop preprocessing: hop choose depending on the length of the sound (min_hop:16, max_hop: 1024);</p>\n\n<p>4) Mixup with beta distribution (0.2, 0.2);\n5) Class_weights in BCE calculated by the ratio of the number of classes predicted as top1 on validation;\n6) Train with different seeds and choosing the best model fold by fold ^);\n7) The ensemble of 7 relatively uncorrelated models.</p>",
  "messages": [
    {
      "id": 550425,
      "postDate": "2019-06-11T16:01:28.377Z",
      "content": "<p>1.) Using power(0.125) instead of log in Spectrograms;\n2) Selection balanced (or unbalanced) sample of well predicted by current model noisy files (1500, 2500, 5000, 10000). Priority for rarely predicted files on the validation;\n3) Different parameters of the spectrogram extraction (hop, n_mels). We used:\n- rate: 44100, n_mels: {128, 256}, n_fft: 2560, hop_length: {512, 694, 1024}, fmin: 20;\n- dynamic hop preprocessing: hop choose depending on the length of the sound (min_hop:16, max_hop: 1024);</p>\n\n<p>4) Mixup with beta distribution (0.2, 0.2);\n5) Class_weights in BCE calculated by the ratio of the number of classes predicted as top1 on validation;\n6) Train with different seeds and choosing the best model fold by fold ^);\n7) The ensemble of 7 relatively uncorrelated models.</p>",
      "rawMarkdown": "1.) Using power(0.125) instead of log in Spectrograms;\n2) Selection balanced (or unbalanced) sample of well predicted by current model noisy files (1500, 2500, 5000, 10000). Priority for rarely predicted files on the validation;\n3) Different parameters of the spectrogram extraction (hop, n_mels). We used:\n- rate: 44100, n_mels: {128, 256}, n_fft: 2560, hop_length: {512, 694, 1024}, fmin: 20;\n- dynamic hop preprocessing: hop choose depending on the length of the sound (min\\_hop:16, max\\_hop: 1024);\n\n4) Mixup with beta distribution (0.2, 0.2);\n5) Class_weights in BCE calculated by the ratio of the number of classes predicted as top1 on validation;\n6) Train with different seeds and choosing the best model fold by fold ^);\n7) The ensemble of 7 relatively uncorrelated models.",
      "votes": 18
    },
    {
      "id": 554877,
      "postDate": "2019-06-18T06:07:11.687Z",
      "content": "<p><a href=\"/daisukelab\">@daisukelab</a> all of this hacks help us to build 7 uncorrelated models)</p>\n\n<p>We used power(0.125) almost from the start of the competition and I don’t know how much it improved pipeline in comparison with log.\nBut selected noisy (point 2) and mixup (point 4) boost all of our models.</p>",
      "rawMarkdown": "@daisukelab all of this hacks help us to build 7 uncorrelated models)\n\nWe used power(0.125) almost from the start of the competition and I don’t know how much it improved pipeline in comparison with log.\nBut selected noisy (point 2) and mixup (point 4) boost all of our models.",
      "votes": 3,
      "replies": [
        {
          "id": 555105,
          "postDate": "2019-06-18T12:42:22.693Z",
          "content": "<p>Thanks, sounds good :)</p>",
          "rawMarkdown": "Thanks, sounds good :)",
          "votes": 1
        }
      ]
    },
    {
      "id": 554818,
      "postDate": "2019-06-18T03:43:36.530Z",
      "content": "<p>Thanks for sharing, could I ask which of your 7 points made bigger differences?\nAudio feature tuning looks good, sample selection would also be working good, ...</p>",
      "rawMarkdown": "Thanks for sharing, could I ask which of your 7 points made bigger differences?\nAudio feature tuning looks good, sample selection would also be working good, ...",
      "votes": 1
    },
    {
      "id": 553077,
      "postDate": "2019-06-15T04:58:43.883Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 554877,
      "author_name": "Aimoldin Anuar [dsmlkz]",
      "author_url": "",
      "post_date": "2019-06-18T06:07:11.687000",
      "content": "<p><a href=\"/daisukelab\">@daisukelab</a> all of this hacks help us to build 7 uncorrelated models)</p>\n\n<p>We used power(0.125) almost from the start of the competition and I don’t know how much it improved pipeline in comparison with log.\nBut selected noisy (point 2) and mixup (point 4) boost all of our models.</p>",
      "votes": 3,
      "replies": [
        {
          "id": 555105,
          "author_name": "daisukelab",
          "author_url": "",
          "post_date": "2019-06-18T12:42:22.693000",
          "content": "<p>Thanks, sounds good :)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 554818,
      "author_name": "daisukelab",
      "author_url": "",
      "post_date": "2019-06-18T03:43:36.530000",
      "content": "<p>Thanks for sharing, could I ask which of your 7 points made bigger differences?\nAudio feature tuning looks good, sample selection would also be working good, ...</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 553077,
      "author_name": "",
      "author_url": "",
      "post_date": "2019-06-15T04:58:43.883000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "550425": "1.) Using power(0.125) instead of log in Spectrograms;\n2) Selection balanced (or unbalanced) sample of well predicted by current model noisy files (1500, 2500, 5000, 10000). Priority for rarely predicted files on the validation;\n3) Different parameters of the spectrogram extraction (hop, n_mels). We used:\n- rate: 44100, n_mels: {128, 256}, n_fft: 2560, hop_length: {512, 694, 1024}, fmin: 20;\n- dynamic hop preprocessing: hop choose depending on the length of the sound (min\\_hop:16, max\\_hop: 1024);\n\n4) Mixup with beta distribution (0.2, 0.2);\n5) Class_weights in BCE calculated by the ratio of the number of classes predicted as top1 on validation;\n6) Train with different seeds and choosing the best model fold by fold ^);\n7) The ensemble of 7 relatively uncorrelated models.",
    "554877": "@daisukelab all of this hacks help us to build 7 uncorrelated models)\n\nWe used power(0.125) almost from the start of the competition and I don’t know how much it improved pipeline in comparison with log.\nBut selected noisy (point 2) and mixup (point 4) boost all of our models.",
    "554818": "Thanks for sharing, could I ask which of your 7 points made bigger differences?\nAudio feature tuning looks good, sample selection would also be working good, ...",
    "553077": ""
  }
}