{
  "id": 179290,
  "title": "Is denoise in only test data effective?",
  "url": "/competitions/birdsong-recognition/discussion/179290",
  "author_name": "",
  "post_date": "2020-09-02T03:49:43.299564600Z",
  "votes": 4,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I shared notebook which apply denoise in test time.<br>\n(<a href=\"https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757\" target=\"_blank\">https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757</a>)</p>\n<p>At First, I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)</p>\n<p>Is denoise in only test data effective?<br>\nDenoise processing needs lot of time, so I want to apply to only test, if  test only is effective enough.</p>",
  "messages": [
    {
      "id": "994898",
      "postDate": "09/02/2020 03:49:43",
      "content": "<p>I shared notebook which apply denoise in test time.<br>\n(<a href=\"https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757\" target=\"_blank\">https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757</a>)</p>\n<p>At First, I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)</p>\n<p>Is denoise in only test data effective?<br>\nDenoise processing needs lot of time, so I want to apply to only test, if  test only is effective enough.</p>",
      "rawMarkdown": "I shared notebook which apply denoise in test time.\n(https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757)\n\nAt First, I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)\n\nIs denoise in only test data effective?\nDenoise processing needs lot of time, so I want to apply to only test, if  test only is effective enough.",
      "votes": null
    },
    {
      "id": "995121",
      "postDate": "09/02/2020 07:19:13",
      "content": "<blockquote>\n  <p>I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)</p>\n</blockquote>\n<p>The result may indicate that the average SNR level is lower in test dataset than that in train dataset.</p>",
      "rawMarkdown": "> I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)\n\nThe result may indicate that the average SNR level is lower in test dataset than that in train dataset.",
      "votes": null
    },
    {
      "id": "995155",
      "postDate": "09/02/2020 07:51:43",
      "content": "<p>Sorry I may be misunderstanding something but just a quick thought - 0.544 is the baseline / 100% 'nocall' prediction I think, and can be an indicator of an error or similar in the prediction process - or the model failing to identify any calls?</p>\n<p>Apologies if I've misunderstood…</p>",
      "rawMarkdown": "Sorry I may be misunderstanding something but just a quick thought - 0.544 is the baseline / 100% 'nocall' prediction I think, and can be an indicator of an error or similar in the prediction process - or the model failing to identify any calls?\n\nApologies if I've misunderstood...",
      "votes": null
    },
    {
      "id": "995182",
      "postDate": "09/02/2020 08:06:56",
      "content": "<p>Thank you share your insight.</p>\n<p>I think it's unlikely that my model predict all test data as nocall.<br>\nThe reason is that my model predict other than nocall at sample test data. (Please see <a href=\"https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise/output?scriptVersionId=41807757&amp;select=submission.csv\" target=\"_blank\">my notebook submittion file</a>.)</p>\n<p>But the possibility is not zero, so I try other model.<br>\nThanks.</p>",
      "rawMarkdown": "Thank you share your insight.\n\nI think it's unlikely that my model predict all test data as nocall.\nThe reason is that my model predict other than nocall at sample test data. (Please see [my notebook submittion file](https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise/output?scriptVersionId=41807757&select=submission.csv).)\n\nBut the possibility is not zero, so I try other model.\nThanks.",
      "votes": null
    },
    {
      "id": "995191",
      "postDate": "09/02/2020 08:14:26",
      "content": "<p>Thank you comment.</p>\n<blockquote>\n  <p>The result may indicate that the average SNR level is lower in test dataset than that in train dataset.</p>\n</blockquote>\n<p>I think so too.</p>\n<p>I expect that this is one of the causes of gap of LB and local CV. (Of course there are many other reasons…)</p>\n<p>Now I'll try more high score model.</p>",
      "rawMarkdown": "Thank you comment.\n\n> The result may indicate that the average SNR level is lower in test dataset than that in train dataset.\n\nI think so too.\n\nI expect that this is one of the causes of gap of LB and local CV. (Of course there are many other reasons...)\n\nNow I'll try more high score model.",
      "votes": null
    },
    {
      "id": "995234",
      "postDate": "09/02/2020 09:06:34",
      "content": "<p>Was there any timeout for the submission? Denoising can take much longer time than ordinary inference</p>",
      "rawMarkdown": "Was there any timeout for the submission? Denoising can take much longer time than ordinary inference",
      "votes": null
    },
    {
      "id": "995254",
      "postDate": "09/02/2020 09:45:59",
      "content": "<p>If use full data in noise extraction, I got timeout error. but when use only 2 second, I did.</p>",
      "rawMarkdown": "If use full data in noise extraction, I got timeout error. but when use only 2 second, I did.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 995121,
      "author_name": "hidehisaarai1213",
      "author_url": "",
      "post_date": "09/02/2020 07:19:13",
      "content": "<blockquote>\n  <p>I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)</p>\n</blockquote>\n<p>The result may indicate that the average SNR level is lower in test dataset than that in train dataset.</p>",
      "votes": null,
      "replies": [
        {
          "id": 995191,
          "author_name": "takamichitoda",
          "author_url": "",
          "post_date": "09/02/2020 08:14:26",
          "content": "<p>Thank you comment.</p>\n<blockquote>\n  <p>The result may indicate that the average SNR level is lower in test dataset than that in train dataset.</p>\n</blockquote>\n<p>I think so too.</p>\n<p>I expect that this is one of the causes of gap of LB and local CV. (Of course there are many other reasons…)</p>\n<p>Now I'll try more high score model.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 995155,
      "author_name": "davidedwards1",
      "author_url": "",
      "post_date": "09/02/2020 07:51:43",
      "content": "<p>Sorry I may be misunderstanding something but just a quick thought - 0.544 is the baseline / 100% 'nocall' prediction I think, and can be an indicator of an error or similar in the prediction process - or the model failing to identify any calls?</p>\n<p>Apologies if I've misunderstood…</p>",
      "votes": null,
      "replies": [
        {
          "id": 995182,
          "author_name": "takamichitoda",
          "author_url": "",
          "post_date": "09/02/2020 08:06:56",
          "content": "<p>Thank you share your insight.</p>\n<p>I think it's unlikely that my model predict all test data as nocall.<br>\nThe reason is that my model predict other than nocall at sample test data. (Please see <a href=\"https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise/output?scriptVersionId=41807757&amp;select=submission.csv\" target=\"_blank\">my notebook submittion file</a>.)</p>\n<p>But the possibility is not zero, so I try other model.<br>\nThanks.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 995234,
          "author_name": "alanchn31",
          "author_url": "",
          "post_date": "09/02/2020 09:06:34",
          "content": "<p>Was there any timeout for the submission? Denoising can take much longer time than ordinary inference</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 995254,
          "author_name": "takamichitoda",
          "author_url": "",
          "post_date": "09/02/2020 09:45:59",
          "content": "<p>If use full data in noise extraction, I got timeout error. but when use only 2 second, I did.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "994898": "I shared notebook which apply denoise in test time.\n(https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise?scriptVersionId=41807757)\n\nAt First, I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)\n\nIs denoise in only test data effective?\nDenoise processing needs lot of time, so I want to apply to only test, if  test only is effective enough.",
    "995121": "> I had thought that we should apply denoise to training as well, but I got improving score even if I only apply to test.(LB: 0.471→0.544)\n\nThe result may indicate that the average SNR level is lower in test dataset than that in train dataset.",
    "995155": "Sorry I may be misunderstanding something but just a quick thought - 0.544 is the baseline / 100% 'nocall' prediction I think, and can be an indicator of an error or similar in the prediction process - or the model failing to identify any calls?\n\nApologies if I've misunderstood...",
    "995182": "Thank you share your insight.\n\nI think it's unlikely that my model predict all test data as nocall.\nThe reason is that my model predict other than nocall at sample test data. (Please see [my notebook submittion file](https://www.kaggle.com/takamichitoda/birdcall-nocall-prediction-with-denoise/output?scriptVersionId=41807757&select=submission.csv).)\n\nBut the possibility is not zero, so I try other model.\nThanks.",
    "995191": "Thank you comment.\n\n> The result may indicate that the average SNR level is lower in test dataset than that in train dataset.\n\nI think so too.\n\nI expect that this is one of the causes of gap of LB and local CV. (Of course there are many other reasons...)\n\nNow I'll try more high score model.",
    "995234": "Was there any timeout for the submission? Denoising can take much longer time than ordinary inference",
    "995254": "If use full data in noise extraction, I got timeout error. but when use only 2 second, I did."
  },
  "source": "meta"
}