{
  "id": 506765,
  "title": "Has anyone tried alternatives to spectrograms?",
  "url": "/competitions/birdclef-2024/discussion/506765",
  "author_name": "",
  "post_date": "2024-05-23T06:57:48.217853400Z",
  "votes": 6,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Just saw this one: <a href=\"https://openreview.net/pdf?id=3mlfh6c3dp\" target=\"_blank\">https://openreview.net/pdf?id=3mlfh6c3dp</a> </p>\n<p>Just wondering if anyone in this competition has tried to come up with alternatives to (mel) spectrograms? I would love to see your approaches :) </p>",
  "messages": [
    {
      "id": "2830379",
      "postDate": "05/23/2024 06:57:48",
      "content": "<p>Just saw this one: <a href=\"https://openreview.net/pdf?id=3mlfh6c3dp\" target=\"_blank\">https://openreview.net/pdf?id=3mlfh6c3dp</a> </p>\n<p>Just wondering if anyone in this competition has tried to come up with alternatives to (mel) spectrograms? I would love to see your approaches :) </p>",
      "rawMarkdown": "Just saw this one: https://openreview.net/pdf?id=3mlfh6c3dp \n\nJust wondering if anyone in this competition has tried to come up with alternatives to (mel) spectrograms? I would love to see your approaches :)",
      "votes": null
    },
    {
      "id": "2830390",
      "postDate": "05/23/2024 07:04:38",
      "content": "<p>Seems interesting! Thanks for sharing.</p>",
      "rawMarkdown": "Seems interesting! Thanks for sharing.",
      "votes": null
    },
    {
      "id": "2831213",
      "postDate": "05/23/2024 15:06:42",
      "content": "<p>Besides it takes a lot of time to process a signal on CPU, it generates pretty weird images with Time Frequency Scattering. (That after going through a lot of parameter combinations). Probably, I did not get something. <br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F9000d63b61384f54914f6e9ec7ac3471%2FScreenshot%20from%202024-05-23%2018-04-05.png?generation=1716476669215079&amp;alt=media\"><br>\nMel spectogram for comparison <br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F589556a01987c315c874b9de4896f1b0%2FScreenshot%20from%202024-05-23%2018-05-56.png?generation=1716476780601197&amp;alt=media\"></p>",
      "rawMarkdown": "Besides it takes a lot of time to process a signal on CPU, it generates pretty weird images with Time Frequency Scattering. (That after going through a lot of parameter combinations). Probably, I did not get something. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F9000d63b61384f54914f6e9ec7ac3471%2FScreenshot%20from%202024-05-23%2018-04-05.png?generation=1716476669215079&alt=media)\nMel spectogram for comparison \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F589556a01987c315c874b9de4896f1b0%2FScreenshot%20from%202024-05-23%2018-05-56.png?generation=1716476780601197&alt=media)",
      "votes": null
    },
    {
      "id": "2831226",
      "postDate": "05/23/2024 15:14:59",
      "content": "<p>what package are you using to compute the scattering? I am trying <code>kymatio</code> pytorch and it's definitely slower training so I have my doubts on inference being fast enough.</p>",
      "rawMarkdown": "what package are you using to compute the scattering? I am trying `kymatio` pytorch and it's definitely slower training so I have my doubts on inference being fast enough.",
      "votes": null
    },
    {
      "id": "2831231",
      "postDate": "05/23/2024 15:16:26",
      "content": "<p>I am using the same package </p>",
      "rawMarkdown": "I am using the same package",
      "votes": null
    },
    {
      "id": "2833461",
      "postDate": "05/24/2024 08:02:17",
      "content": "<p>They have some pretty neat examples, like <a href=\"https://www.kymat.io/gallery_1d/classif_keras.html\" target=\"_blank\">this one</a>. Did you use the first order scattering? For me, it does look like a spectrogram…but yeah, inference speed might be an issue.</p>",
      "rawMarkdown": "They have some pretty neat examples, like [this one](https://www.kymat.io/gallery_1d/classif_keras.html). Did you use the first order scattering? For me, it does look like a spectrogram...but yeah, inference speed might be an issue.",
      "votes": null
    },
    {
      "id": "2833505",
      "postDate": "05/24/2024 08:56:51",
      "content": "<p>I tired kymatio time frequency scattering in a past comp, as well as other wavelet transforms, and mel spectrograms worked best for me.</p>\n<p>If we were not compute limited, then maybe these would help in an ensemble.</p>\n<p>Maybe worth spending time on it still, I don't know yet.</p>",
      "rawMarkdown": "I tired kymatio time frequency scattering in a past comp, as well as other wavelet transforms, and mel spectrograms worked best for me.\n\nIf we were not compute limited, then maybe these would help in an ensemble.\n\nMaybe worth spending time on it still, I don't know yet.",
      "votes": null
    },
    {
      "id": "2833521",
      "postDate": "05/24/2024 09:27:49",
      "content": "<p>That's good to know. Maybe this is something for low-data regimes, as it seems to help few-shot learning. When there's enough data, mel specs appear to be the go-to choice.</p>",
      "rawMarkdown": "That's good to know. Maybe this is something for low-data regimes, as it seems to help few-shot learning. When there's enough data, mel specs appear to be the go-to choice.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2830390,
      "author_name": "arindamroy23",
      "author_url": "",
      "post_date": "05/23/2024 07:04:38",
      "content": "<p>Seems interesting! Thanks for sharing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2831213,
      "author_name": "nikitababich",
      "author_url": "",
      "post_date": "05/23/2024 15:06:42",
      "content": "<p>Besides it takes a lot of time to process a signal on CPU, it generates pretty weird images with Time Frequency Scattering. (That after going through a lot of parameter combinations). Probably, I did not get something. <br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F9000d63b61384f54914f6e9ec7ac3471%2FScreenshot%20from%202024-05-23%2018-04-05.png?generation=1716476669215079&amp;alt=media\"><br>\nMel spectogram for comparison <br>\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F589556a01987c315c874b9de4896f1b0%2FScreenshot%20from%202024-05-23%2018-05-56.png?generation=1716476780601197&amp;alt=media\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 2831226,
          "author_name": "willrice",
          "author_url": "",
          "post_date": "05/23/2024 15:14:59",
          "content": "<p>what package are you using to compute the scattering? I am trying <code>kymatio</code> pytorch and it's definitely slower training so I have my doubts on inference being fast enough.</p>",
          "votes": null,
          "replies": [
            {
              "id": 2831231,
              "author_name": "nikitababich",
              "author_url": "",
              "post_date": "05/23/2024 15:16:26",
              "content": "<p>I am using the same package </p>",
              "votes": null,
              "replies": []
            }
          ]
        },
        {
          "id": 2833461,
          "author_name": "stefankahl",
          "author_url": "",
          "post_date": "05/24/2024 08:02:17",
          "content": "<p>They have some pretty neat examples, like <a href=\"https://www.kymat.io/gallery_1d/classif_keras.html\" target=\"_blank\">this one</a>. Did you use the first order scattering? For me, it does look like a spectrogram…but yeah, inference speed might be an issue.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2833505,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "05/24/2024 08:56:51",
      "content": "<p>I tired kymatio time frequency scattering in a past comp, as well as other wavelet transforms, and mel spectrograms worked best for me.</p>\n<p>If we were not compute limited, then maybe these would help in an ensemble.</p>\n<p>Maybe worth spending time on it still, I don't know yet.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2833521,
          "author_name": "stefankahl",
          "author_url": "",
          "post_date": "05/24/2024 09:27:49",
          "content": "<p>That's good to know. Maybe this is something for low-data regimes, as it seems to help few-shot learning. When there's enough data, mel specs appear to be the go-to choice.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2830379": "Just saw this one: https://openreview.net/pdf?id=3mlfh6c3dp \n\nJust wondering if anyone in this competition has tried to come up with alternatives to (mel) spectrograms? I would love to see your approaches :)",
    "2830390": "Seems interesting! Thanks for sharing.",
    "2831213": "Besides it takes a lot of time to process a signal on CPU, it generates pretty weird images with Time Frequency Scattering. (That after going through a lot of parameter combinations). Probably, I did not get something. \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F9000d63b61384f54914f6e9ec7ac3471%2FScreenshot%20from%202024-05-23%2018-04-05.png?generation=1716476669215079&alt=media)\nMel spectogram for comparison \n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6343664%2F589556a01987c315c874b9de4896f1b0%2FScreenshot%20from%202024-05-23%2018-05-56.png?generation=1716476780601197&alt=media)",
    "2831226": "what package are you using to compute the scattering? I am trying `kymatio` pytorch and it's definitely slower training so I have my doubts on inference being fast enough.",
    "2831231": "I am using the same package",
    "2833461": "They have some pretty neat examples, like [this one](https://www.kymat.io/gallery_1d/classif_keras.html). Did you use the first order scattering? For me, it does look like a spectrogram...but yeah, inference speed might be an issue.",
    "2833505": "I tired kymatio time frequency scattering in a past comp, as well as other wavelet transforms, and mel spectrograms worked best for me.\n\nIf we were not compute limited, then maybe these would help in an ensemble.\n\nMaybe worth spending time on it still, I don't know yet.",
    "2833521": "That's good to know. Maybe this is something for low-data regimes, as it seems to help few-shot learning. When there's enough data, mel specs appear to be the go-to choice."
  },
  "source": "meta"
}