{
  "id": 571777,
  "title": "Doubt Regarding BirdClef submission",
  "url": "/competitions/birdclef-2025/discussion/571777",
  "author_name": "",
  "post_date": "2025-04-05T15:14:18.752693600Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello everyone, <br>\nI am working on a transformer approach but I am not able to do submission in time can anyone please tell what should I do ? Or it is impossible to run transformer in such type of competitions?</p>\n<p>Please help me to clear it.</p>\n<p>Thank You</p>",
  "messages": [
    {
      "id": "3171330",
      "postDate": "04/05/2025 15:14:18",
      "content": "<p>Hello everyone, <br>\nI am working on a transformer approach but I am not able to do submission in time can anyone please tell what should I do ? Or it is impossible to run transformer in such type of competitions?</p>\n<p>Please help me to clear it.</p>\n<p>Thank You</p>",
      "rawMarkdown": "Hello everyone, \nI am working on a transformer approach but I am not able to do submission in time can anyone please tell what should I do ? Or it is impossible to run transformer in such type of competitions?\n\nPlease help me to clear it.\n\nThank You",
      "votes": null
    },
    {
      "id": "3171332",
      "postDate": "04/05/2025 15:17:21",
      "content": "<p>You can run a quantized version, <br>\nRather than submitting, first try to create dummy or random audios. (Same number of recordings as mentioned in test data.)<br>\nThen try to get predictions and map how much time does it take.</p>\n<p>Another thing you might need to consider is, each 5 sec maps to a vector of size 160000 (32000 x 5)<br>\nWhich is too much.</p>\n<p>Maybe you can try to reduce the Sampling Rate to 16000 and make it half. </p>",
      "rawMarkdown": "You can run a quantized version, \nRather than submitting, first try to create dummy or random audios. (Same number of recordings as mentioned in test data.)\nThen try to get predictions and map how much time does it take.\n\nAnother thing you might need to consider is, each 5 sec maps to a vector of size 160000 (32000 x 5)\nWhich is too much.\n\nMaybe you can try to reduce the Sampling Rate to 16000 and make it half.",
      "votes": null
    },
    {
      "id": "3171338",
      "postDate": "04/05/2025 15:22:21",
      "content": "<p><a href=\"https://www.kaggle.com/salmanahmedtamu\" target=\"_blank\">@salmanahmedtamu</a> Sir, thanks for reply but I tried to generate submission with train_sounscapes that is the simillar data as test_soundscapes that will be generated during submission runtime but I had used 32000 sampling rate to train the transformer approach. and on train_soundscapes it is giving me too much time as we need to so submission on cpu. do you have any suggestions?</p>\n<p>due to this problem I am not able to make submission.</p>",
      "rawMarkdown": "salmanahmedtamu Sir, thanks for reply but I tried to generate submission with train_sounscapes that is the simillar data as test_soundscapes that will be generated during submission runtime but I had used 32000 sampling rate to train the transformer approach. and on train_soundscapes it is giving me too much time as we need to so submission on cpu. do you have any suggestions?\n\ndue to this problem I am not able to make submission.",
      "votes": null
    },
    {
      "id": "3171446",
      "postDate": "04/05/2025 17:28:14",
      "content": "<p>You can either retrain the transformer with 16000 Sampling Rate, or you can convert transformer to an quatinzed version using openvino and inference on that transformer should be much much faster. <br>\n<a href=\"https://huggingface.co/blog/openvino\" target=\"_blank\">https://huggingface.co/blog/openvino</a></p>",
      "rawMarkdown": "You can either retrain the transformer with 16000 Sampling Rate, or you can convert transformer to an quatinzed version using openvino and inference on that transformer should be much much faster. \n[https://huggingface.co/blog/openvino](https://huggingface.co/blog/openvino)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3171332,
      "author_name": "salmanahmedtamu",
      "author_url": "",
      "post_date": "04/05/2025 15:17:21",
      "content": "<p>You can run a quantized version, <br>\nRather than submitting, first try to create dummy or random audios. (Same number of recordings as mentioned in test data.)<br>\nThen try to get predictions and map how much time does it take.</p>\n<p>Another thing you might need to consider is, each 5 sec maps to a vector of size 160000 (32000 x 5)<br>\nWhich is too much.</p>\n<p>Maybe you can try to reduce the Sampling Rate to 16000 and make it half. </p>",
      "votes": null,
      "replies": [
        {
          "id": 3171338,
          "author_name": "mohitgupta12",
          "author_url": "",
          "post_date": "04/05/2025 15:22:21",
          "content": "<p><a href=\"https://www.kaggle.com/salmanahmedtamu\" target=\"_blank\">@salmanahmedtamu</a> Sir, thanks for reply but I tried to generate submission with train_sounscapes that is the simillar data as test_soundscapes that will be generated during submission runtime but I had used 32000 sampling rate to train the transformer approach. and on train_soundscapes it is giving me too much time as we need to so submission on cpu. do you have any suggestions?</p>\n<p>due to this problem I am not able to make submission.</p>",
          "votes": null,
          "replies": [
            {
              "id": 3171446,
              "author_name": "salmanahmedtamu",
              "author_url": "",
              "post_date": "04/05/2025 17:28:14",
              "content": "<p>You can either retrain the transformer with 16000 Sampling Rate, or you can convert transformer to an quatinzed version using openvino and inference on that transformer should be much much faster. <br>\n<a href=\"https://huggingface.co/blog/openvino\" target=\"_blank\">https://huggingface.co/blog/openvino</a></p>",
              "votes": null,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3171330": "Hello everyone, \nI am working on a transformer approach but I am not able to do submission in time can anyone please tell what should I do ? Or it is impossible to run transformer in such type of competitions?\n\nPlease help me to clear it.\n\nThank You",
    "3171332": "You can run a quantized version, \nRather than submitting, first try to create dummy or random audios. (Same number of recordings as mentioned in test data.)\nThen try to get predictions and map how much time does it take.\n\nAnother thing you might need to consider is, each 5 sec maps to a vector of size 160000 (32000 x 5)\nWhich is too much.\n\nMaybe you can try to reduce the Sampling Rate to 16000 and make it half.",
    "3171338": "salmanahmedtamu Sir, thanks for reply but I tried to generate submission with train_sounscapes that is the simillar data as test_soundscapes that will be generated during submission runtime but I had used 32000 sampling rate to train the transformer approach. and on train_soundscapes it is giving me too much time as we need to so submission on cpu. do you have any suggestions?\n\ndue to this problem I am not able to make submission.",
    "3171446": "You can either retrain the transformer with 16000 Sampling Rate, or you can convert transformer to an quatinzed version using openvino and inference on that transformer should be much much faster. \n[https://huggingface.co/blog/openvino](https://huggingface.co/blog/openvino)"
  },
  "source": "meta"
}