{
  "id": 270382,
  "title": "Transformer ",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/270382",
  "author_name": "",
  "post_date": "2021-09-04T23:42:27.021368400Z",
  "votes": 1,
  "comment_count": 11,
  "views": 0,
  "content": "<p>Hello, have anyone trying to use transformer to solve this challenge? </p>",
  "messages": [
    {
      "id": "1503066",
      "postDate": "09/04/2021 23:42:27",
      "content": "<p>Hello, have anyone trying to use transformer to solve this challenge? </p>",
      "rawMarkdown": "Hello, have anyone trying to use transformer to solve this challenge?",
      "votes": null
    },
    {
      "id": "1504036",
      "postDate": "09/06/2021 02:59:36",
      "content": "<p>I did consider using a transformer for this competition however i believe it is not suitable for this task mainly due to the main requirement as given in the paper 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'<br>\nwhich is you need to train transformers on a very large dataset. You can use pretrained models for this problem however with scarce training samples i dont think transfer learning will be successfull.</p>\n<p>Secondly results for transformers when compared with Resnets is not a very big improvement and i believe this competition is more about the data and not the model.</p>",
      "rawMarkdown": "I did consider using a transformer for this competition however i believe it is not suitable for this task mainly due to the main requirement as given in the paper 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'\nwhich is you need to train transformers on a very large dataset. You can use pretrained models for this problem however with scarce training samples i dont think transfer learning will be successfull.\n\nSecondly results for transformers when compared with Resnets is not a very big improvement and i believe this competition is more about the data and not the model.",
      "votes": null
    },
    {
      "id": "1504086",
      "postDate": "09/06/2021 04:33:33",
      "content": "<p>I also tried a transformer 3D model, but conventional CNN-3D models work well.</p>",
      "rawMarkdown": "I also tried a transformer 3D model, but conventional CNN-3D models work well.",
      "votes": null
    },
    {
      "id": "1504201",
      "postDate": "09/06/2021 07:03:12",
      "content": "<p>I did make a model with the Perceiver Transformer, the issue is they need lots of memory and like <a href=\"https://www.kaggle.com/aryamansharma47\" target=\"_blank\">@aryamansharma47</a> wrote also lots of data.<br>\nI used 3 Steps of convolution + pooling to reduce the memory of the transformer.</p>",
      "rawMarkdown": "I did make a model with the Perceiver Transformer, the issue is they need lots of memory and like @aryamansharma47 wrote also lots of data.\nI used 3 Steps of convolution + pooling to reduce the memory of the transformer.",
      "votes": null
    },
    {
      "id": "1506066",
      "postDate": "09/07/2021 19:37:10",
      "content": "<p>Can you share any resource about 3D transformer? </p>",
      "rawMarkdown": "Can you share any resource about 3D transformer?",
      "votes": null
    },
    {
      "id": "1506069",
      "postDate": "09/07/2021 19:38:24",
      "content": "<p>I find some 2D transformer like Vit but i didn't find any 3D. Do you know anything? </p>",
      "rawMarkdown": "I find some 2D transformer like Vit but i didn't find any 3D. Do you know anything?",
      "votes": null
    },
    {
      "id": "1506089",
      "postDate": "09/07/2021 20:29:55",
      "content": "<p>You have to write it yourself. It is not that hard, just project 3D patches into a vector instead of 2D Patches into a Vector.</p>",
      "rawMarkdown": "You have to write it yourself. It is not that hard, just project 3D patches into a vector instead of 2D Patches into a Vector.",
      "votes": null
    },
    {
      "id": "1506479",
      "postDate": "09/08/2021 09:13:26",
      "content": "<p>Did you see any notebook or code about this?</p>",
      "rawMarkdown": "Did you see any notebook or code about this?",
      "votes": null
    },
    {
      "id": "1506481",
      "postDate": "09/08/2021 09:15:00",
      "content": "<p>Do you know any transformer that trained for medical images? </p>",
      "rawMarkdown": "Do you know any transformer that trained for medical images?",
      "votes": null
    },
    {
      "id": "1507256",
      "postDate": "09/09/2021 03:05:55",
      "content": "<p><a href=\"https://arxiv.org/pdf/2108.09038.pdf\" target=\"_blank\">https://arxiv.org/pdf/2108.09038.pdf</a></p>",
      "rawMarkdown": "https://arxiv.org/pdf/2108.09038.pdf",
      "votes": null
    },
    {
      "id": "1507486",
      "postDate": "09/09/2021 08:40:48",
      "content": "<p>From this article we can understand it's not good for use Transformer instead of CNN. isn't it?</p>",
      "rawMarkdown": "From this article we can understand it's not good for use Transformer instead of CNN. isn't it?",
      "votes": null
    },
    {
      "id": "1508202",
      "postDate": "09/10/2021 02:15:12",
      "content": "<p>For this competition i believe that may be the case</p>",
      "rawMarkdown": "For this competition i believe that may be the case",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1504036,
      "author_name": "aryamansharma47",
      "author_url": "",
      "post_date": "09/06/2021 02:59:36",
      "content": "<p>I did consider using a transformer for this competition however i believe it is not suitable for this task mainly due to the main requirement as given in the paper 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'<br>\nwhich is you need to train transformers on a very large dataset. You can use pretrained models for this problem however with scarce training samples i dont think transfer learning will be successfull.</p>\n<p>Secondly results for transformers when compared with Resnets is not a very big improvement and i believe this competition is more about the data and not the model.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1506481,
          "author_name": "mohammadhosein1998",
          "author_url": "",
          "post_date": "09/08/2021 09:15:00",
          "content": "<p>Do you know any transformer that trained for medical images? </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1504086,
      "author_name": "hyeonhoonlee",
      "author_url": "",
      "post_date": "09/06/2021 04:33:33",
      "content": "<p>I also tried a transformer 3D model, but conventional CNN-3D models work well.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1506066,
          "author_name": "mohammadhosein1998",
          "author_url": "",
          "post_date": "09/07/2021 19:37:10",
          "content": "<p>Can you share any resource about 3D transformer? </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1504201,
      "author_name": "theudas",
      "author_url": "",
      "post_date": "09/06/2021 07:03:12",
      "content": "<p>I did make a model with the Perceiver Transformer, the issue is they need lots of memory and like <a href=\"https://www.kaggle.com/aryamansharma47\" target=\"_blank\">@aryamansharma47</a> wrote also lots of data.<br>\nI used 3 Steps of convolution + pooling to reduce the memory of the transformer.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1506069,
      "author_name": "mohammadhosein1998",
      "author_url": "",
      "post_date": "09/07/2021 19:38:24",
      "content": "<p>I find some 2D transformer like Vit but i didn't find any 3D. Do you know anything? </p>",
      "votes": null,
      "replies": [
        {
          "id": 1506089,
          "author_name": "skyyy93",
          "author_url": "",
          "post_date": "09/07/2021 20:29:55",
          "content": "<p>You have to write it yourself. It is not that hard, just project 3D patches into a vector instead of 2D Patches into a Vector.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1506479,
          "author_name": "mohammadhosein1998",
          "author_url": "",
          "post_date": "09/08/2021 09:13:26",
          "content": "<p>Did you see any notebook or code about this?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1507256,
      "author_name": "aryamansharma47",
      "author_url": "",
      "post_date": "09/09/2021 03:05:55",
      "content": "<p><a href=\"https://arxiv.org/pdf/2108.09038.pdf\" target=\"_blank\">https://arxiv.org/pdf/2108.09038.pdf</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 1507486,
          "author_name": "mohammadhosein1998",
          "author_url": "",
          "post_date": "09/09/2021 08:40:48",
          "content": "<p>From this article we can understand it's not good for use Transformer instead of CNN. isn't it?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1508202,
          "author_name": "aryamansharma47",
          "author_url": "",
          "post_date": "09/10/2021 02:15:12",
          "content": "<p>For this competition i believe that may be the case</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1503066": "Hello, have anyone trying to use transformer to solve this challenge?",
    "1504036": "I did consider using a transformer for this competition however i believe it is not suitable for this task mainly due to the main requirement as given in the paper 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale'\nwhich is you need to train transformers on a very large dataset. You can use pretrained models for this problem however with scarce training samples i dont think transfer learning will be successfull.\n\nSecondly results for transformers when compared with Resnets is not a very big improvement and i believe this competition is more about the data and not the model.",
    "1504086": "I also tried a transformer 3D model, but conventional CNN-3D models work well.",
    "1504201": "I did make a model with the Perceiver Transformer, the issue is they need lots of memory and like @aryamansharma47 wrote also lots of data.\nI used 3 Steps of convolution + pooling to reduce the memory of the transformer.",
    "1506066": "Can you share any resource about 3D transformer?",
    "1506069": "I find some 2D transformer like Vit but i didn't find any 3D. Do you know anything?",
    "1506089": "You have to write it yourself. It is not that hard, just project 3D patches into a vector instead of 2D Patches into a Vector.",
    "1506479": "Did you see any notebook or code about this?",
    "1506481": "Do you know any transformer that trained for medical images?",
    "1507256": "https://arxiv.org/pdf/2108.09038.pdf",
    "1507486": "From this article we can understand it's not good for use Transformer instead of CNN. isn't it?",
    "1508202": "For this competition i believe that may be the case"
  },
  "source": "meta"
}