{
  "id": 240135,
  "title": "How to speed up augmentation on Waveform?",
  "url": "/competitions/birdclef-2021/discussion/240135",
  "author_name": "",
  "post_date": "2021-05-18T16:18:52.605891300Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Im sorry if this has been brought up previously, but i cant find it. I am trying to add GaussianNoise, Gain onto the waveform as part of my augmentation pipeline to hopefully improve generalisation. </p>\n<p>However, the use of augmentation and then converting to mel-spectrogram increases my training time significantly. I looked up into some of the top solution codes and most of their augment pipeline is roughly what i am doing here. One hour per epoch is too scary for a colab user like myself.</p>\n<p>I thought of pre-computing the augmented waveforms prior to training, but i wasn't sure how to ensure that the validation set is not augmented since i'm running on KFold as the very same data that is augmented for training would be used as validation on other folds.</p>",
  "messages": [
    {
      "id": "1313620",
      "postDate": "05/18/2021 16:18:52",
      "content": "<p>Im sorry if this has been brought up previously, but i cant find it. I am trying to add GaussianNoise, Gain onto the waveform as part of my augmentation pipeline to hopefully improve generalisation. </p>\n<p>However, the use of augmentation and then converting to mel-spectrogram increases my training time significantly. I looked up into some of the top solution codes and most of their augment pipeline is roughly what i am doing here. One hour per epoch is too scary for a colab user like myself.</p>\n<p>I thought of pre-computing the augmented waveforms prior to training, but i wasn't sure how to ensure that the validation set is not augmented since i'm running on KFold as the very same data that is augmented for training would be used as validation on other folds.</p>",
      "rawMarkdown": "Im sorry if this has been brought up previously, but i cant find it. I am trying to add GaussianNoise, Gain onto the waveform as part of my augmentation pipeline to hopefully improve generalisation. \n\nHowever, the use of augmentation and then converting to mel-spectrogram increases my training time significantly. I looked up into some of the top solution codes and most of their augment pipeline is roughly what i am doing here. One hour per epoch is too scary for a colab user like myself.\n\nI thought of pre-computing the augmented waveforms prior to training, but i wasn't sure how to ensure that the validation set is not augmented since i'm running on KFold as the very same data that is augmented for training would be used as validation on other folds.",
      "votes": null
    },
    {
      "id": "1315181",
      "postDate": "05/19/2021 16:07:26",
      "content": "<p>augmentation on waveform, then use torch-librosa to do STFT and LogMel on GPU.</p>",
      "rawMarkdown": "augmentation on waveform, then use torch-librosa to do STFT and LogMel on GPU.",
      "votes": null
    },
    {
      "id": "1315237",
      "postDate": "05/19/2021 16:45:08",
      "content": "<p>Thank you! I will try it out</p>",
      "rawMarkdown": "Thank you! I will try it out",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1315181,
      "author_name": "superchenhao",
      "author_url": "",
      "post_date": "05/19/2021 16:07:26",
      "content": "<p>augmentation on waveform, then use torch-librosa to do STFT and LogMel on GPU.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1315237,
          "author_name": "angqx95",
          "author_url": "",
          "post_date": "05/19/2021 16:45:08",
          "content": "<p>Thank you! I will try it out</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1313620": "Im sorry if this has been brought up previously, but i cant find it. I am trying to add GaussianNoise, Gain onto the waveform as part of my augmentation pipeline to hopefully improve generalisation. \n\nHowever, the use of augmentation and then converting to mel-spectrogram increases my training time significantly. I looked up into some of the top solution codes and most of their augment pipeline is roughly what i am doing here. One hour per epoch is too scary for a colab user like myself.\n\nI thought of pre-computing the augmented waveforms prior to training, but i wasn't sure how to ensure that the validation set is not augmented since i'm running on KFold as the very same data that is augmented for training would be used as validation on other folds.",
    "1315181": "augmentation on waveform, then use torch-librosa to do STFT and LogMel on GPU.",
    "1315237": "Thank you! I will try it out"
  },
  "source": "meta"
}