{
  "id": 228905,
  "title": "[NewBaseLine] => ViT for Segmentation",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/228905",
  "author_name": "",
  "post_date": "2021-03-27T05:54:16.029343600Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello,</p>\n<p>I try to do something new using the <strong>Self attention mechanisms</strong>. So i create a new code using <strong>Vision Transformer</strong> as encoder for <strong>UNet</strong>. </p>\n<p><strong><a href=\"https://www.kaggle.com/elcaiseri/hubmap-pytorch-vit-for-segmentation-train\">HuBMAP | Pytorch | ViT for Segmentation [Train]</a></strong></p>\n<p>The training results get improved than the code i forked with light hyperparameters.<br>\n<strong>Take a look and I waiting for your feedbacks</strong></p>\n<p>Thanks</p>",
  "messages": [
    {
      "id": "1253875",
      "postDate": "03/27/2021 05:54:16",
      "content": "<p>Hello,</p>\n<p>I try to do something new using the <strong>Self attention mechanisms</strong>. So i create a new code using <strong>Vision Transformer</strong> as encoder for <strong>UNet</strong>. </p>\n<p><strong><a href=\"https://www.kaggle.com/elcaiseri/hubmap-pytorch-vit-for-segmentation-train\">HuBMAP | Pytorch | ViT for Segmentation [Train]</a></strong></p>\n<p>The training results get improved than the code i forked with light hyperparameters.<br>\n<strong>Take a look and I waiting for your feedbacks</strong></p>\n<p>Thanks</p>",
      "rawMarkdown": "Hello,\n\nI try to do something new using the **Self attention mechanisms**. So i create a new code using **Vision Transformer** as encoder for **UNet**. \n\n**<a href='https://www.kaggle.com/elcaiseri/hubmap-pytorch-vit-for-segmentation-train'>HuBMAP | Pytorch | ViT for Segmentation [Train]</a>**\n\nThe training results get improved than the code i forked with light hyperparameters.\n**Take a look and I waiting for your feedbacks**\n\nThanks",
      "votes": null
    },
    {
      "id": "1254100",
      "postDate": "03/27/2021 10:52:41",
      "content": "<p>It's a good approach. One suggestion would be to reduce the probability of the augmentations; make each group of augmentations like the 'weak' augmentations only apply 50% of the time. So at least half of the time, the image is not distorted. Otherwise your model wouldn't train properly because it never gets to see the true data. Something to keep in mind is that different architectures have certain level of resilience to distortions so long as their training data is good quality.</p>",
      "rawMarkdown": "It's a good approach. One suggestion would be to reduce the probability of the augmentations; make each group of augmentations like the 'weak' augmentations only apply 50% of the time. So at least half of the time, the image is not distorted. Otherwise your model wouldn't train properly because it never gets to see the true data. Something to keep in mind is that different architectures have certain level of resilience to distortions so long as their training data is good quality.",
      "votes": null
    },
    {
      "id": "1254435",
      "postDate": "03/27/2021 17:11:26",
      "content": "<p>Thanks for your recommendations, I would apply it as soon as possible </p>",
      "rawMarkdown": "Thanks for your recommendations, I would apply it as soon as possible",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1254100,
      "author_name": "erikdali",
      "author_url": "",
      "post_date": "03/27/2021 10:52:41",
      "content": "<p>It's a good approach. One suggestion would be to reduce the probability of the augmentations; make each group of augmentations like the 'weak' augmentations only apply 50% of the time. So at least half of the time, the image is not distorted. Otherwise your model wouldn't train properly because it never gets to see the true data. Something to keep in mind is that different architectures have certain level of resilience to distortions so long as their training data is good quality.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1254435,
          "author_name": "elcaiseri",
          "author_url": "",
          "post_date": "03/27/2021 17:11:26",
          "content": "<p>Thanks for your recommendations, I would apply it as soon as possible </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1253875": "Hello,\n\nI try to do something new using the **Self attention mechanisms**. So i create a new code using **Vision Transformer** as encoder for **UNet**. \n\n**<a href='https://www.kaggle.com/elcaiseri/hubmap-pytorch-vit-for-segmentation-train'>HuBMAP | Pytorch | ViT for Segmentation [Train]</a>**\n\nThe training results get improved than the code i forked with light hyperparameters.\n**Take a look and I waiting for your feedbacks**\n\nThanks",
    "1254100": "It's a good approach. One suggestion would be to reduce the probability of the augmentations; make each group of augmentations like the 'weak' augmentations only apply 50% of the time. So at least half of the time, the image is not distorted. Otherwise your model wouldn't train properly because it never gets to see the true data. Something to keep in mind is that different architectures have certain level of resilience to distortions so long as their training data is good quality.",
    "1254435": "Thanks for your recommendations, I would apply it as soon as possible"
  },
  "source": "meta"
}