{
  "id": 237067,
  "title": "PyTorch code for Unsupervised Contrastive Learning of Sound Event Representations",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/237067",
  "author_name": "Eduardo Fonseca",
  "post_date": "2021-05-07T00:31:16.629000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Dear Kaggle people, </p>\n<p>We've released PyTorch code for our ICASSP paper \"Unsupervised Contrastive Learning of Sound Event Representations\". Also the slide deck, poster and a blog post! All can be accessed from <a href=\"https://github.com/edufonseca/uclser20\" target=\"_blank\">https://github.com/edufonseca/uclser20</a></p>\n<p>We propose to learn sound event representations using the proxy task of contrasting differently augmented views of sound events, inspired by the SimCLR framework. The views are computed by sampling time-frequency patches within every clip, mixing them with unrelated backgrounds, and other data augmentations.</p>\n<p>Our results suggest that unsupervised contrastive pre-training can mitigate the impact of data scarcity and increase robustness against noisy labels.</p>\n<p>This work is a collaboration between the <a href=\"https://www.upf.edu/web/mtg/\" target=\"_blank\">Music Technology Group</a> and the <a href=\"https://www.insight-centre.org/\" target=\"_blank\">Insight SFI Centre for Data Analytics</a>.</p>\n<blockquote>\n  <p>Eduardo Fonseca, Diego Ortego, Kevin McGuinness, Noel E. O'Connor, Xavier Serra, <a href=\"https://arxiv.org/pdf/2011.07616.pdf\" target=\"_blank\"><strong>Unsupervised Contrastive Learning of Sound Event Representations.</strong></a>, ICASSP 2021.</p>\n</blockquote>\n<p>Hope this can be useful!</p>\n<p>Eduardo on behalf of the authors</p>",
  "messages": [
    {
      "id": 1296078,
      "postDate": "2021-05-07T00:31:16.630Z",
      "content": "<p>Dear Kaggle people, </p>\n<p>We've released PyTorch code for our ICASSP paper \"Unsupervised Contrastive Learning of Sound Event Representations\". Also the slide deck, poster and a blog post! All can be accessed from <a href=\"https://github.com/edufonseca/uclser20\" target=\"_blank\">https://github.com/edufonseca/uclser20</a></p>\n<p>We propose to learn sound event representations using the proxy task of contrasting differently augmented views of sound events, inspired by the SimCLR framework. The views are computed by sampling time-frequency patches within every clip, mixing them with unrelated backgrounds, and other data augmentations.</p>\n<p>Our results suggest that unsupervised contrastive pre-training can mitigate the impact of data scarcity and increase robustness against noisy labels.</p>\n<p>This work is a collaboration between the <a href=\"https://www.upf.edu/web/mtg/\" target=\"_blank\">Music Technology Group</a> and the <a href=\"https://www.insight-centre.org/\" target=\"_blank\">Insight SFI Centre for Data Analytics</a>.</p>\n<blockquote>\n  <p>Eduardo Fonseca, Diego Ortego, Kevin McGuinness, Noel E. O'Connor, Xavier Serra, <a href=\"https://arxiv.org/pdf/2011.07616.pdf\" target=\"_blank\"><strong>Unsupervised Contrastive Learning of Sound Event Representations.</strong></a>, ICASSP 2021.</p>\n</blockquote>\n<p>Hope this can be useful!</p>\n<p>Eduardo on behalf of the authors</p>",
      "rawMarkdown": "Dear Kaggle people, \n\nWe've released PyTorch code for our ICASSP paper \"Unsupervised Contrastive Learning of Sound Event Representations\". Also the slide deck, poster and a blog post! All can be accessed from https://github.com/edufonseca/uclser20\n\nWe propose to learn sound event representations using the proxy task of contrasting differently augmented views of sound events, inspired by the SimCLR framework. The views are computed by sampling time-frequency patches within every clip, mixing them with unrelated backgrounds, and other data augmentations.\n\nOur results suggest that unsupervised contrastive pre-training can mitigate the impact of data scarcity and increase robustness against noisy labels.\n\nThis work is a collaboration between the <a href=\"https://www.upf.edu/web/mtg/\" target=\"_blank\">Music Technology Group</a> and the <a href=\"https://www.insight-centre.org/\" target=\"_blank\">Insight SFI Centre for Data Analytics</a>.\n\n> Eduardo Fonseca, Diego Ortego, Kevin McGuinness, Noel E. O'Connor, Xavier Serra, <a href=\"https://arxiv.org/pdf/2011.07616.pdf\" target=\"_blank\">**Unsupervised Contrastive Learning of Sound Event Representations.**</a>, ICASSP 2021.\n\nHope this can be useful!\n\nEduardo on behalf of the authors\n\n",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1296078": "Dear Kaggle people, \n\nWe've released PyTorch code for our ICASSP paper \"Unsupervised Contrastive Learning of Sound Event Representations\". Also the slide deck, poster and a blog post! All can be accessed from https://github.com/edufonseca/uclser20\n\nWe propose to learn sound event representations using the proxy task of contrasting differently augmented views of sound events, inspired by the SimCLR framework. The views are computed by sampling time-frequency patches within every clip, mixing them with unrelated backgrounds, and other data augmentations.\n\nOur results suggest that unsupervised contrastive pre-training can mitigate the impact of data scarcity and increase robustness against noisy labels.\n\nThis work is a collaboration between the <a href=\"https://www.upf.edu/web/mtg/\" target=\"_blank\">Music Technology Group</a> and the <a href=\"https://www.insight-centre.org/\" target=\"_blank\">Insight SFI Centre for Data Analytics</a>.\n\n> Eduardo Fonseca, Diego Ortego, Kevin McGuinness, Noel E. O'Connor, Xavier Serra, <a href=\"https://arxiv.org/pdf/2011.07616.pdf\" target=\"_blank\">**Unsupervised Contrastive Learning of Sound Event Representations.**</a>, ICASSP 2021.\n\nHope this can be useful!\n\nEduardo on behalf of the authors\n\n"
  }
}