{
  "id": 204631,
  "title": "Training ResNet50 with SnapMix (single fold, notta, LB=0.891)",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/204631",
  "author_name": "David",
  "post_date": "2020-12-16T04:37:16.269000",
  "votes": 19,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I simply run the release code of SnapMix (find the link below) and would like to share some results with you. </p>\n<p>Input Image size:  448</p>\n<p>Training :<br>\nlearning rate: use the default value<br>\nBackbone: Resnet-50<br>\nepoch number: 40  (decay the learning rate after 20 epochs)<br>\nCV folds: 5</p>\n<p>SnapMix parameters:<br>\nI used the default values (prob=1,beta=1)</p>\n<p>Testing:<br>\nCenter Crop (no TTA)</p>\n<p>**Results:<br>\nSingle fold:   val: 0.899, LB: 0.891<br>\n5 folds avg:  LB: 0.897<br>\n**</p>\n<p>SnapMix overview</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F299388%2F746cc9901746284761ec3f3f962556af%2Foverview.jpg?generation=1608093300974260&amp;alt=media\" alt=\"\"></p>\n<p>Paper: SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021) <a href=\"url\" target=\"_blank\">https://arxiv.org/abs/2012.04846</a><br>\nCode: <a href=\"https://github.com/Shaoli-Huang/SnapMix\" target=\"_blank\">https://github.com/Shaoli-Huang/SnapMix</a></p>",
  "messages": [
    {
      "id": 1115197,
      "postDate": "2020-12-16T04:37:16.270Z",
      "content": "<p>I simply run the release code of SnapMix (find the link below) and would like to share some results with you. </p>\n<p>Input Image size:  448</p>\n<p>Training :<br>\nlearning rate: use the default value<br>\nBackbone: Resnet-50<br>\nepoch number: 40  (decay the learning rate after 20 epochs)<br>\nCV folds: 5</p>\n<p>SnapMix parameters:<br>\nI used the default values (prob=1,beta=1)</p>\n<p>Testing:<br>\nCenter Crop (no TTA)</p>\n<p>**Results:<br>\nSingle fold:   val: 0.899, LB: 0.891<br>\n5 folds avg:  LB: 0.897<br>\n**</p>\n<p>SnapMix overview</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F299388%2F746cc9901746284761ec3f3f962556af%2Foverview.jpg?generation=1608093300974260&amp;alt=media\" alt=\"\"></p>\n<p>Paper: SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021) <a href=\"url\" target=\"_blank\">https://arxiv.org/abs/2012.04846</a><br>\nCode: <a href=\"https://github.com/Shaoli-Huang/SnapMix\" target=\"_blank\">https://github.com/Shaoli-Huang/SnapMix</a></p>",
      "rawMarkdown": "I simply run the release code of SnapMix (find the link below) and would like to share some results with you. \n\nInput Image size:  448\n\nTraining :\nlearning rate: use the default value\nBackbone: Resnet-50\nepoch number: 40  (decay the learning rate after 20 epochs)\nCV folds: 5\n\nSnapMix parameters:\nI used the default values (prob=1,beta=1)\n\nTesting:\nCenter Crop (no TTA)\n\n**Results:\nSingle fold:   val: 0.899, LB: 0.891\n5 folds avg:  LB: 0.897\n**\n\n\nSnapMix overview\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F299388%2F746cc9901746284761ec3f3f962556af%2Foverview.jpg?generation=1608093300974260&alt=media)\n\n\nPaper: SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021) [https://arxiv.org/abs/2012.04846](url)\nCode: https://github.com/Shaoli-Huang/SnapMix\n\n",
      "votes": 19
    },
    {
      "id": 1115237,
      "postDate": "2020-12-16T05:44:24.113Z",
      "content": "<p>40 epochs how much time did it take per fold ??</p>",
      "rawMarkdown": "40 epochs how much time did it take per fold ??",
      "replies": [
        {
          "id": 1115242,
          "postDate": "2020-12-16T05:55:31.450Z",
          "content": "<p>It took around 4 hour for training one fold.</p>",
          "rawMarkdown": "It took around 4 hour for training one fold."
        }
      ]
    },
    {
      "id": 1115241,
      "postDate": "2020-12-16T05:55:06.067Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1115237,
      "author_name": "Atharva Phatak",
      "author_url": "",
      "post_date": "2020-12-16T05:44:24.113000",
      "content": "<p>40 epochs how much time did it take per fold ??</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1115242,
          "author_name": "David",
          "author_url": "",
          "post_date": "2020-12-16T05:55:31.450000",
          "content": "<p>It took around 4 hour for training one fold.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1115241,
      "author_name": "",
      "author_url": "",
      "post_date": "2020-12-16T05:55:06.067000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1115197": "I simply run the release code of SnapMix (find the link below) and would like to share some results with you. \n\nInput Image size:  448\n\nTraining :\nlearning rate: use the default value\nBackbone: Resnet-50\nepoch number: 40  (decay the learning rate after 20 epochs)\nCV folds: 5\n\nSnapMix parameters:\nI used the default values (prob=1,beta=1)\n\nTesting:\nCenter Crop (no TTA)\n\n**Results:\nSingle fold:   val: 0.899, LB: 0.891\n5 folds avg:  LB: 0.897\n**\n\n\nSnapMix overview\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F299388%2F746cc9901746284761ec3f3f962556af%2Foverview.jpg?generation=1608093300974260&alt=media)\n\n\nPaper: SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data (AAAI 2021) [https://arxiv.org/abs/2012.04846](url)\nCode: https://github.com/Shaoli-Huang/SnapMix\n\n",
    "1115237": "40 epochs how much time did it take per fold ??",
    "1115241": ""
  }
}