{
  "id": 239313,
  "title": "BirdClef EDA and fastai toy experimenter",
  "url": "/competitions/birdclef-2021/discussion/239313",
  "author_name": "",
  "post_date": "2021-05-15T17:44:38.062384800Z",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Thought I would share my starter notebook containing some <a href=\"https://www.kaggle.com/dryanfurman/birdclef-eda-and-fastai#2.-Audio-EDA\" target=\"_blank\">metadata/audio EDA, as well as my fastai experimental modeling step</a>. </p>\n<p>You will notice the .flac files I use are significantly smaller than the comp .ogg data (~5 GB for all audio). This is because they are pre-processed: Pink noise added, silence removed, cropped to seven seconds, and saved as iunt8. I found <a href=\"https://github.com/fastaudio/fastaudio\" target=\"_blank\">fastaudio</a> transforms to be super helpful here.</p>\n<p>My next step is to deal with the geospatial and temporal representation issue. Has anybody else figured out a good way to get a representative geospatial and temporal training set?  </p>\n<p>Thanks for any and all help,</p>\n<p>Daniel</p>",
  "messages": [
    {
      "id": "1309149",
      "postDate": "05/15/2021 17:44:38",
      "content": "<p>Thought I would share my starter notebook containing some <a href=\"https://www.kaggle.com/dryanfurman/birdclef-eda-and-fastai#2.-Audio-EDA\" target=\"_blank\">metadata/audio EDA, as well as my fastai experimental modeling step</a>. </p>\n<p>You will notice the .flac files I use are significantly smaller than the comp .ogg data (~5 GB for all audio). This is because they are pre-processed: Pink noise added, silence removed, cropped to seven seconds, and saved as iunt8. I found <a href=\"https://github.com/fastaudio/fastaudio\" target=\"_blank\">fastaudio</a> transforms to be super helpful here.</p>\n<p>My next step is to deal with the geospatial and temporal representation issue. Has anybody else figured out a good way to get a representative geospatial and temporal training set?  </p>\n<p>Thanks for any and all help,</p>\n<p>Daniel</p>",
      "rawMarkdown": "Thought I would share my starter notebook containing some [metadata/audio EDA, as well as my fastai experimental modeling step](https://www.kaggle.com/dryanfurman/birdclef-eda-and-fastai#2.-Audio-EDA). \n\nYou will notice the .flac files I use are significantly smaller than the comp .ogg data (~5 GB for all audio). This is because they are pre-processed: Pink noise added, silence removed, cropped to seven seconds, and saved as iunt8. I found [fastaudio](https://github.com/fastaudio/fastaudio) transforms to be super helpful here.\n\nMy next step is to deal with the geospatial and temporal representation issue. Has anybody else figured out a good way to get a representative geospatial and temporal training set?  \n\nThanks for any and all help,\n\nDaniel",
      "votes": null
    },
    {
      "id": "1320395",
      "postDate": "05/24/2021 04:11:34",
      "content": "<p>Hi, thank you for sharing your approach! :)</p>\n<p>Could you also share the script to pre-process the training data to .flac files?</p>",
      "rawMarkdown": "Hi, thank you for sharing your approach! :)\n\nCould you also share the script to pre-process the training data to .flac files?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1320395,
      "author_name": "nguyncaoduy",
      "author_url": "",
      "post_date": "05/24/2021 04:11:34",
      "content": "<p>Hi, thank you for sharing your approach! :)</p>\n<p>Could you also share the script to pre-process the training data to .flac files?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1309149": "Thought I would share my starter notebook containing some [metadata/audio EDA, as well as my fastai experimental modeling step](https://www.kaggle.com/dryanfurman/birdclef-eda-and-fastai#2.-Audio-EDA). \n\nYou will notice the .flac files I use are significantly smaller than the comp .ogg data (~5 GB for all audio). This is because they are pre-processed: Pink noise added, silence removed, cropped to seven seconds, and saved as iunt8. I found [fastaudio](https://github.com/fastaudio/fastaudio) transforms to be super helpful here.\n\nMy next step is to deal with the geospatial and temporal representation issue. Has anybody else figured out a good way to get a representative geospatial and temporal training set?  \n\nThanks for any and all help,\n\nDaniel",
    "1320395": "Hi, thank you for sharing your approach! :)\n\nCould you also share the script to pre-process the training data to .flac files?"
  },
  "source": "meta"
}