{
  "id": 181149,
  "title": "Saving Melsectograms to speed up training ? ",
  "url": "/competitions/birdsong-recognition/discussion/181149",
  "author_name": "",
  "post_date": "2020-09-07T19:12:11.455582900Z",
  "votes": 2,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Any ideas on how we can save Melspectograms after augmentations to speed up training process ? because in my case training is taking 1 hour per epoch !</p>",
  "messages": [
    {
      "id": "1002043",
      "postDate": "09/07/2020 19:12:11",
      "content": "<p>Any ideas on how we can save Melspectograms after augmentations to speed up training process ? because in my case training is taking 1 hour per epoch !</p>",
      "rawMarkdown": "Any ideas on how we can save Melspectograms after augmentations to speed up training process ? because in my case training is taking 1 hour per epoch !",
      "votes": null
    },
    {
      "id": "1002048",
      "postDate": "09/07/2020 19:16:44",
      "content": "<p>First of all, epoch shouldn't take that much even if you calculate melspoctorgrams with cpu. And second, calculate try to calculate mels on gpu, that's a not heavy operation.</p>",
      "rawMarkdown": "First of all, epoch shouldn't take that much even if you calculate melspoctorgrams with cpu. And second, calculate try to calculate mels on gpu, that's a not heavy operation.",
      "votes": null
    },
    {
      "id": "1002053",
      "postDate": "09/07/2020 19:20:06",
      "content": "<p>Using heavy augmentations, training one epoch can take that much ! ( ~ 1 hour ) ! Maybe I m doing something wrong. But after adding augmentations training took forever .. </p>\n<p>And I am using kaggle's GPU</p>",
      "rawMarkdown": "Using heavy augmentations, training one epoch can take that much ! ( ~ 1 hour ) ! Maybe I m doing something wrong. But after adding augmentations training took forever .. \n\nAnd I am using kaggle's GPU",
      "votes": null
    },
    {
      "id": "1002063",
      "postDate": "09/07/2020 19:31:23",
      "content": "<p>Surely it also depends on your model, image-size, batch-size, evaluation technique etc.</p>\n<p>I need 8 minutes for one epoch (Kaggle GPU) training on EfficientNet-B1 or ResNest-50 using augmentations aswell e.g. RandomErasing </p>\n<p>Do you train on the full audio-file or do you crop 5 random seconds out of it?</p>",
      "rawMarkdown": "Surely it also depends on your model, image-size, batch-size, evaluation technique etc.\n\nI need 8 minutes for one epoch (Kaggle GPU) training on EfficientNet-B1 or ResNest-50 using augmentations aswell e.g. RandomErasing \n\nDo you train on the full audio-file or do you crop 5 random seconds out of it?",
      "votes": null
    },
    {
      "id": "1002082",
      "postDate": "09/07/2020 19:56:07",
      "content": "<p>Do you save the augmented data? Either the data takes a lot of space or the augmentations will repeat every epoch.</p>",
      "rawMarkdown": "Do you save the augmented data? Either the data takes a lot of space or the augmentations will repeat every epoch.",
      "votes": null
    },
    {
      "id": "1002086",
      "postDate": "09/07/2020 19:59:48",
      "content": "<p><a href=\"https://www.kaggle.com/aliabdin1\" target=\"_blank\">@aliabdin1</a> My settings are the same as public notebooks. <br>\nAnd I crop 5 random seconds for training.</p>",
      "rawMarkdown": "aliabdin1 My settings are the same as public notebooks. \nAnd I crop 5 random seconds for training.",
      "votes": null
    },
    {
      "id": "1002088",
      "postDate": "09/07/2020 20:00:41",
      "content": "<p>I am not saving the augmented data but I am asking to see if someone is having my same issue. </p>",
      "rawMarkdown": "I am not saving the augmented data but I am asking to see if someone is having my same issue.",
      "votes": null
    },
    {
      "id": "1002100",
      "postDate": "09/07/2020 20:16:18",
      "content": "<p>Sorry for asking but you also crop 5 seconds for validation?</p>\n<p>What model do you use? Batch-size? Do you use one threshold for validation or do you only use <code>val_loss</code> as metric?</p>",
      "rawMarkdown": "Sorry for asking but you also crop 5 seconds for validation?\n\nWhat model do you use? Batch-size? Do you use one threshold for validation or do you only use `val_loss ` as metric?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1002048,
      "author_name": "yaroshevskiy",
      "author_url": "",
      "post_date": "09/07/2020 19:16:44",
      "content": "<p>First of all, epoch shouldn't take that much even if you calculate melspoctorgrams with cpu. And second, calculate try to calculate mels on gpu, that's a not heavy operation.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1002053,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "09/07/2020 19:20:06",
          "content": "<p>Using heavy augmentations, training one epoch can take that much ! ( ~ 1 hour ) ! Maybe I m doing something wrong. But after adding augmentations training took forever .. </p>\n<p>And I am using kaggle's GPU</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1002063,
          "author_name": "aliabdin1",
          "author_url": "",
          "post_date": "09/07/2020 19:31:23",
          "content": "<p>Surely it also depends on your model, image-size, batch-size, evaluation technique etc.</p>\n<p>I need 8 minutes for one epoch (Kaggle GPU) training on EfficientNet-B1 or ResNest-50 using augmentations aswell e.g. RandomErasing </p>\n<p>Do you train on the full audio-file or do you crop 5 random seconds out of it?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1002086,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "09/07/2020 19:59:48",
          "content": "<p><a href=\"https://www.kaggle.com/aliabdin1\" target=\"_blank\">@aliabdin1</a> My settings are the same as public notebooks. <br>\nAnd I crop 5 random seconds for training.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1002100,
          "author_name": "aliabdin1",
          "author_url": "",
          "post_date": "09/07/2020 20:16:18",
          "content": "<p>Sorry for asking but you also crop 5 seconds for validation?</p>\n<p>What model do you use? Batch-size? Do you use one threshold for validation or do you only use <code>val_loss</code> as metric?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1002082,
      "author_name": "squarenabla",
      "author_url": "",
      "post_date": "09/07/2020 19:56:07",
      "content": "<p>Do you save the augmented data? Either the data takes a lot of space or the augmentations will repeat every epoch.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1002088,
          "author_name": "nacirbouazizi",
          "author_url": "",
          "post_date": "09/07/2020 20:00:41",
          "content": "<p>I am not saving the augmented data but I am asking to see if someone is having my same issue. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1002043": "Any ideas on how we can save Melspectograms after augmentations to speed up training process ? because in my case training is taking 1 hour per epoch !",
    "1002048": "First of all, epoch shouldn't take that much even if you calculate melspoctorgrams with cpu. And second, calculate try to calculate mels on gpu, that's a not heavy operation.",
    "1002053": "Using heavy augmentations, training one epoch can take that much ! ( ~ 1 hour ) ! Maybe I m doing something wrong. But after adding augmentations training took forever .. \n\nAnd I am using kaggle's GPU",
    "1002063": "Surely it also depends on your model, image-size, batch-size, evaluation technique etc.\n\nI need 8 minutes for one epoch (Kaggle GPU) training on EfficientNet-B1 or ResNest-50 using augmentations aswell e.g. RandomErasing \n\nDo you train on the full audio-file or do you crop 5 random seconds out of it?",
    "1002082": "Do you save the augmented data? Either the data takes a lot of space or the augmentations will repeat every epoch.",
    "1002086": "aliabdin1 My settings are the same as public notebooks. \nAnd I crop 5 random seconds for training.",
    "1002088": "I am not saving the augmented data but I am asking to see if someone is having my same issue.",
    "1002100": "Sorry for asking but you also crop 5 seconds for validation?\n\nWhat model do you use? Batch-size? Do you use one threshold for validation or do you only use `val_loss ` as metric?"
  },
  "source": "meta"
}