{
  "id": 205895,
  "title": "Snapmix aug takes 3X-5X more time to train?",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/205895",
  "author_name": "",
  "post_date": "2020-12-22T11:39:18.076893400Z",
  "votes": 2,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi All,</p>\n<p>I have implemented Snapmix augmentation from the repo here - <a href=\"https://github.com/Shaoli-Huang/SnapMix\" target=\"_blank\">https://github.com/Shaoli-Huang/SnapMix</a></p>\n<p>But the training time (using FP16) seems to go up by about 3X-5X (15 mins per epoch to 1 hour per epoch). Is anyone else seeing similar training times when using Snapmix or is it just me. </p>\n<p>Thanks</p>",
  "messages": [
    {
      "id": "1122364",
      "postDate": "12/22/2020 11:39:18",
      "content": "<p>Hi All,</p>\n<p>I have implemented Snapmix augmentation from the repo here - <a href=\"https://github.com/Shaoli-Huang/SnapMix\" target=\"_blank\">https://github.com/Shaoli-Huang/SnapMix</a></p>\n<p>But the training time (using FP16) seems to go up by about 3X-5X (15 mins per epoch to 1 hour per epoch). Is anyone else seeing similar training times when using Snapmix or is it just me. </p>\n<p>Thanks</p>",
      "rawMarkdown": "Hi All,\n\nI have implemented Snapmix augmentation from the repo here - https://github.com/Shaoli-Huang/SnapMix\n\nBut the training time (using FP16) seems to go up by about 3X-5X (15 mins per epoch to 1 hour per epoch). Is anyone else seeing similar training times when using Snapmix or is it just me. \n\nThanks",
      "votes": null
    },
    {
      "id": "1122403",
      "postDate": "12/22/2020 12:11:31",
      "content": "<p>Hi, Here is my training time per epoch:<br>\nstandard resnet50:  3 min 55 second<br>\nstandard resnet50 with SnapMix: 5 min 13 second</p>",
      "rawMarkdown": "Hi, Here is my training time per epoch:\nstandard resnet50:  3 min 55 second\nstandard resnet50 with SnapMix: 5 min 13 second",
      "votes": null
    },
    {
      "id": "1122410",
      "postDate": "12/22/2020 12:25:15",
      "content": "<p>Interesting, less than 2X increase. I guess my pipeline has some inefficient calculations. Thanks <a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> </p>",
      "rawMarkdown": "Interesting, less than 2X increase. I guess my pipeline has some inefficient calculations. Thanks @shaolihuang",
      "votes": null
    },
    {
      "id": "1122415",
      "postDate": "12/22/2020 12:29:37",
      "content": "<p><a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> sorry forgot to ask, whats the image size you are training with for the reported epoch times?</p>",
      "rawMarkdown": "shaolihuang sorry forgot to ask, whats the image size you are training with for the reported epoch times?",
      "votes": null
    },
    {
      "id": "1122457",
      "postDate": "12/22/2020 12:58:33",
      "content": "<p>input size: 448x448    batch_size: 16</p>",
      "rawMarkdown": "input size: 448x448    batch_size: 16",
      "votes": null
    },
    {
      "id": "1122471",
      "postDate": "12/22/2020 13:11:36",
      "content": "<p>Thankyou <a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> </p>",
      "rawMarkdown": "Thankyou @shaolihuang",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1122403,
      "author_name": "shaolihuang",
      "author_url": "",
      "post_date": "12/22/2020 12:11:31",
      "content": "<p>Hi, Here is my training time per epoch:<br>\nstandard resnet50:  3 min 55 second<br>\nstandard resnet50 with SnapMix: 5 min 13 second</p>",
      "votes": null,
      "replies": [
        {
          "id": 1122410,
          "author_name": "pheadrus",
          "author_url": "",
          "post_date": "12/22/2020 12:25:15",
          "content": "<p>Interesting, less than 2X increase. I guess my pipeline has some inefficient calculations. Thanks <a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1122415,
          "author_name": "pheadrus",
          "author_url": "",
          "post_date": "12/22/2020 12:29:37",
          "content": "<p><a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> sorry forgot to ask, whats the image size you are training with for the reported epoch times?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1122457,
          "author_name": "shaolihuang",
          "author_url": "",
          "post_date": "12/22/2020 12:58:33",
          "content": "<p>input size: 448x448    batch_size: 16</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1122471,
          "author_name": "pheadrus",
          "author_url": "",
          "post_date": "12/22/2020 13:11:36",
          "content": "<p>Thankyou <a href=\"https://www.kaggle.com/shaolihuang\" target=\"_blank\">@shaolihuang</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1122364": "Hi All,\n\nI have implemented Snapmix augmentation from the repo here - https://github.com/Shaoli-Huang/SnapMix\n\nBut the training time (using FP16) seems to go up by about 3X-5X (15 mins per epoch to 1 hour per epoch). Is anyone else seeing similar training times when using Snapmix or is it just me. \n\nThanks",
    "1122403": "Hi, Here is my training time per epoch:\nstandard resnet50:  3 min 55 second\nstandard resnet50 with SnapMix: 5 min 13 second",
    "1122410": "Interesting, less than 2X increase. I guess my pipeline has some inefficient calculations. Thanks @shaolihuang",
    "1122415": "shaolihuang sorry forgot to ask, whats the image size you are training with for the reported epoch times?",
    "1122457": "input size: 448x448    batch_size: 16",
    "1122471": "Thankyou @shaolihuang"
  },
  "source": "meta"
}