{
  "id": 92608,
  "title": "SpecAugment ",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/92608",
  "author_name": "",
  "post_date": "2019-05-18T13:07:15.009963700Z",
  "votes": null,
  "comment_count": 12,
  "views": 0,
  "content": "<p>I am trying to use [SpecAugment ]((<a href=\"https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html\">https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html</a> ) but it doesn't improve my ability to fit better in training set and to score more on LB. Are you trying it? What is your experience with it?</p>",
  "messages": [
    {
      "id": "533102",
      "postDate": "05/18/2019 13:07:15",
      "content": "<p>I am trying to use [SpecAugment ]((<a href=\"https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html\">https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html</a> ) but it doesn't improve my ability to fit better in training set and to score more on LB. Are you trying it? What is your experience with it?</p>",
      "rawMarkdown": "I am trying to use [SpecAugment ]((https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html ) but it doesn't improve my ability to fit better in training set and to score more on LB. Are you trying it? What is your experience with it?",
      "votes": null
    },
    {
      "id": "533232",
      "postDate": "05/18/2019 18:19:21",
      "content": "<p>SpecAugment hasn't helped me at all on this task. As yet.</p>",
      "rawMarkdown": "SpecAugment hasn't helped me at all on this task. As yet.",
      "votes": null
    },
    {
      "id": "534160",
      "postDate": "05/20/2019 19:57:28",
      "content": "<p>great kernel</p>",
      "rawMarkdown": "great kernel",
      "votes": null
    },
    {
      "id": "534527",
      "postDate": "05/21/2019 12:08:45",
      "content": "<p>Same experience on my side</p>",
      "rawMarkdown": "Same experience on my side",
      "votes": null
    },
    {
      "id": "534572",
      "postDate": "05/21/2019 13:30:54",
      "content": "<p>Can confirm, it didn't help me as of now. I think random erasing was used as a data augmentation technique in last year's challenge tho, I would've expected SpecAugment to basically do the same. </p>\n\n<p>This solution mentions using random erasing/cutout for data augmentation, for example.\n<a href=\"https://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914\">https://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914</a></p>\n\n<p>If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?</p>",
      "rawMarkdown": "Can confirm, it didn't help me as of now. I think random erasing was used as a data augmentation technique in last year's challenge tho, I would've expected SpecAugment to basically do the same. \n\nThis solution mentions using random erasing/cutout for data augmentation, for example.\nhttps://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914\n\nIf i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?",
      "votes": null
    },
    {
      "id": "534873",
      "postDate": "05/22/2019 00:50:54",
      "content": "<p>I guess I am the only one where SpecAugment actually helps my score lol...or maybe I should re-test without and I'll end up with a better LB score :)</p>",
      "rawMarkdown": "I guess I am the only one where SpecAugment actually helps my score lol...or maybe I should re-test without and I'll end up with a better LB score :)",
      "votes": null
    },
    {
      "id": "534940",
      "postDate": "05/22/2019 05:07:39",
      "content": "<p>Are you getting a significative improvement?</p>",
      "rawMarkdown": "Are you getting a significative improvement?",
      "votes": null
    },
    {
      "id": "534941",
      "postDate": "05/22/2019 05:08:30",
      "content": "<p>Actually the paper also proposed time warp. I've tried it but it takes forever in the implementation that I've found.</p>",
      "rawMarkdown": "Actually the paper also proposed time warp. I've tried it but it takes forever in the implementation that I've found.",
      "votes": null
    },
    {
      "id": "534987",
      "postDate": "05/22/2019 06:23:40",
      "content": "<p>Actually having thought about it I am Aug-menting the Spec-trogram to some benefit. Just not in the SpecAug way.  Cutouts didn’t help.</p>",
      "rawMarkdown": "Actually having thought about it I am Aug-menting the Spec-trogram to some benefit. Just not in the SpecAug way.  Cutouts didn’t help.",
      "votes": null
    },
    {
      "id": "535011",
      "postDate": "05/22/2019 06:41:24",
      "content": "<blockquote>\n  <p>If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?</p>\n</blockquote>\n\n<p>I tried replacing with median but it didn’t help. Another idea is to try vertically flipping the cutout frequencies.</p>",
      "rawMarkdown": "&gt; If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?\n\nI tried replacing with median but it didn’t help. Another idea is to try vertically flipping the cutout frequencies.",
      "votes": null
    },
    {
      "id": "535321",
      "postDate": "05/22/2019 17:23:19",
      "content": "<p>I got a .015+ LB improvement with the same setup just by adding SpecAugment.</p>",
      "rawMarkdown": "I got a .015+ LB improvement with the same setup just by adding SpecAugment.",
      "votes": null
    },
    {
      "id": "535399",
      "postDate": "05/22/2019 20:34:47",
      "content": "<p>Did you use the implementation to be found among the kernels?</p>",
      "rawMarkdown": "Did you use the implementation to be found among the kernels?",
      "votes": null
    },
    {
      "id": "535414",
      "postDate": "05/22/2019 21:21:00",
      "content": "<p>Yes I'm using the implementation from this kernel <a href=\"https://www.kaggle.com/davids1992/specaugment-quick-implementation\">https://www.kaggle.com/davids1992/specaugment-quick-implementation</a></p>",
      "rawMarkdown": "Yes I'm using the implementation from this kernel https://www.kaggle.com/davids1992/specaugment-quick-implementation",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 533232,
      "author_name": "robga",
      "author_url": "",
      "post_date": "05/18/2019 18:19:21",
      "content": "<p>SpecAugment hasn't helped me at all on this task. As yet.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 534160,
      "author_name": "afanasiivanovs",
      "author_url": "",
      "post_date": "05/20/2019 19:57:28",
      "content": "<p>great kernel</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 534527,
      "author_name": "lftuwujie",
      "author_url": "",
      "post_date": "05/21/2019 12:08:45",
      "content": "<p>Same experience on my side</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 534572,
      "author_name": "mfluegge",
      "author_url": "",
      "post_date": "05/21/2019 13:30:54",
      "content": "<p>Can confirm, it didn't help me as of now. I think random erasing was used as a data augmentation technique in last year's challenge tho, I would've expected SpecAugment to basically do the same. </p>\n\n<p>This solution mentions using random erasing/cutout for data augmentation, for example.\n<a href=\"https://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914\">https://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914</a></p>\n\n<p>If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?</p>",
      "votes": null,
      "replies": [
        {
          "id": 534941,
          "author_name": "lucamassaron",
          "author_url": "",
          "post_date": "05/22/2019 05:08:30",
          "content": "<p>Actually the paper also proposed time warp. I've tried it but it takes forever in the implementation that I've found.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 535011,
          "author_name": "robga",
          "author_url": "",
          "post_date": "05/22/2019 06:41:24",
          "content": "<blockquote>\n  <p>If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?</p>\n</blockquote>\n\n<p>I tried replacing with median but it didn’t help. Another idea is to try vertically flipping the cutout frequencies.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 534873,
      "author_name": "jamesrequa",
      "author_url": "",
      "post_date": "05/22/2019 00:50:54",
      "content": "<p>I guess I am the only one where SpecAugment actually helps my score lol...or maybe I should re-test without and I'll end up with a better LB score :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 534940,
          "author_name": "lucamassaron",
          "author_url": "",
          "post_date": "05/22/2019 05:07:39",
          "content": "<p>Are you getting a significative improvement?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 534987,
          "author_name": "robga",
          "author_url": "",
          "post_date": "05/22/2019 06:23:40",
          "content": "<p>Actually having thought about it I am Aug-menting the Spec-trogram to some benefit. Just not in the SpecAug way.  Cutouts didn’t help.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 535321,
          "author_name": "jamesrequa",
          "author_url": "",
          "post_date": "05/22/2019 17:23:19",
          "content": "<p>I got a .015+ LB improvement with the same setup just by adding SpecAugment.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 535399,
          "author_name": "lucamassaron",
          "author_url": "",
          "post_date": "05/22/2019 20:34:47",
          "content": "<p>Did you use the implementation to be found among the kernels?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 535414,
          "author_name": "jamesrequa",
          "author_url": "",
          "post_date": "05/22/2019 21:21:00",
          "content": "<p>Yes I'm using the implementation from this kernel <a href=\"https://www.kaggle.com/davids1992/specaugment-quick-implementation\">https://www.kaggle.com/davids1992/specaugment-quick-implementation</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "533102": "I am trying to use [SpecAugment ]((https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html ) but it doesn't improve my ability to fit better in training set and to score more on LB. Are you trying it? What is your experience with it?",
    "533232": "SpecAugment hasn't helped me at all on this task. As yet.",
    "534160": "great kernel",
    "534527": "Same experience on my side",
    "534572": "Can confirm, it didn't help me as of now. I think random erasing was used as a data augmentation technique in last year's challenge tho, I would've expected SpecAugment to basically do the same. \n\nThis solution mentions using random erasing/cutout for data augmentation, for example.\nhttps://www.kaggle.com/c/freesound-audio-tagging/discussion/64262#latest-523914\n\nIf i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?",
    "534873": "I guess I am the only one where SpecAugment actually helps my score lol...or maybe I should re-test without and I'll end up with a better LB score :)",
    "534940": "Are you getting a significative improvement?",
    "534941": "Actually the paper also proposed time warp. I've tried it but it takes forever in the implementation that I've found.",
    "534987": "Actually having thought about it I am Aug-menting the Spec-trogram to some benefit. Just not in the SpecAug way.  Cutouts didn’t help.",
    "535011": "&gt; If i remember correctly, the SpecAugment paper proposes to simply set the value to 0, maybe we can do better by repeating/interpolating between adjacent bins/timesteps to make the resulting spectrogram more realistic?\n\nI tried replacing with median but it didn’t help. Another idea is to try vertically flipping the cutout frequencies.",
    "535321": "I got a .015+ LB improvement with the same setup just by adding SpecAugment.",
    "535399": "Did you use the implementation to be found among the kernels?",
    "535414": "Yes I'm using the implementation from this kernel https://www.kaggle.com/davids1992/specaugment-quick-implementation"
  },
  "source": "meta"
}