{
  "id": 177817,
  "title": "Cutmix in Tensorflow",
  "url": "/competitions/landmark-recognition-2020/discussion/177817",
  "author_name": "",
  "post_date": "2020-08-27T13:00:29.314658800Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>There are some Cutmix implementations in Tensorflow on Kaggle but most of them were implemented not fully following the original paper (and the original Pytorch code). So I implemented Cutmix in Tensorflow 2.x and my work was recognized by the paper author as a third-party implementation. I think this technique looks very promising for classification tasks. Please visit my github code for details.</p>\n<p>My Tensorflow code:<br>\n<a href=\"https://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix\" target=\"_blank\">https://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix</a></p>\n<p>Original Pytorch code:<br>\n<a href=\"https://github.com/clovaai/CutMix-PyTorch\" target=\"_blank\">https://github.com/clovaai/CutMix-PyTorch</a></p>",
  "messages": [
    {
      "id": "987698",
      "postDate": "08/27/2020 13:00:29",
      "content": "<p>Hi,</p>\n<p>There are some Cutmix implementations in Tensorflow on Kaggle but most of them were implemented not fully following the original paper (and the original Pytorch code). So I implemented Cutmix in Tensorflow 2.x and my work was recognized by the paper author as a third-party implementation. I think this technique looks very promising for classification tasks. Please visit my github code for details.</p>\n<p>My Tensorflow code:<br>\n<a href=\"https://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix\" target=\"_blank\">https://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix</a></p>\n<p>Original Pytorch code:<br>\n<a href=\"https://github.com/clovaai/CutMix-PyTorch\" target=\"_blank\">https://github.com/clovaai/CutMix-PyTorch</a></p>",
      "rawMarkdown": "Hi,\n\nThere are some Cutmix implementations in Tensorflow on Kaggle but most of them were implemented not fully following the original paper (and the original Pytorch code). So I implemented Cutmix in Tensorflow 2.x and my work was recognized by the paper author as a third-party implementation. I think this technique looks very promising for classification tasks. Please visit my github code for details.\n\nMy Tensorflow code:\nhttps://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix\n\nOriginal Pytorch code:\nhttps://github.com/clovaai/CutMix-PyTorch",
      "votes": null
    },
    {
      "id": "987764",
      "postDate": "08/27/2020 13:52:28",
      "content": "<p>Good work.. I would ask why the cutmix is only top left instead of random? </p>",
      "rawMarkdown": "Good work.. I would ask why the cutmix is only top left instead of random?",
      "votes": null
    },
    {
      "id": "988685",
      "postDate": "08/28/2020 07:47:11",
      "content": "<p>That's because I've only showed one batch of cutmix images as an example. The crop location varies by each batch during training. Thanks for asking!</p>",
      "rawMarkdown": "That's because I've only showed one batch of cutmix images as an example. The crop location varies by each batch during training. Thanks for asking!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 987764,
      "author_name": "udaygurugubelli",
      "author_url": "",
      "post_date": "08/27/2020 13:52:28",
      "content": "<p>Good work.. I would ask why the cutmix is only top left instead of random? </p>",
      "votes": null,
      "replies": [
        {
          "id": 988685,
          "author_name": "mark82",
          "author_url": "",
          "post_date": "08/28/2020 07:47:11",
          "content": "<p>That's because I've only showed one batch of cutmix images as an example. The crop location varies by each batch during training. Thanks for asking!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "987698": "Hi,\n\nThere are some Cutmix implementations in Tensorflow on Kaggle but most of them were implemented not fully following the original paper (and the original Pytorch code). So I implemented Cutmix in Tensorflow 2.x and my work was recognized by the paper author as a third-party implementation. I think this technique looks very promising for classification tasks. Please visit my github code for details.\n\nMy Tensorflow code:\nhttps://github.com/jis478/Tensorflow/tree/master/TF2.0/Cutmix\n\nOriginal Pytorch code:\nhttps://github.com/clovaai/CutMix-PyTorch",
    "987764": "Good work.. I would ask why the cutmix is only top left instead of random?",
    "988685": "That's because I've only showed one batch of cutmix images as an example. The crop location varies by each batch during training. Thanks for asking!"
  },
  "source": "meta"
}