{
  "id": 111423,
  "title": "FPN or Unet: Which one is better?",
  "url": "/competitions/understanding_cloud_organization/discussion/111423",
  "author_name": "",
  "post_date": "2019-10-05T13:50:55.400412600Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I believe most of us are experimenting with Unet models(+different encoders). Did anybody try <code>FPN</code> based models? How does it perform on LB?\nI have been training with <code>FPN</code> model and so far it looks promising on validation...will update the results once tested on LB</p>",
  "messages": [
    {
      "id": "642048",
      "postDate": "10/05/2019 13:50:55",
      "content": "<p>I believe most of us are experimenting with Unet models(+different encoders). Did anybody try <code>FPN</code> based models? How does it perform on LB?\nI have been training with <code>FPN</code> model and so far it looks promising on validation...will update the results once tested on LB</p>",
      "rawMarkdown": "I believe most of us are experimenting with Unet models(+different encoders). Did anybody try `FPN` based models? How does it perform on LB?\nI have been training with `FPN` model and so far it looks promising on validation...will update the results once tested on LB",
      "votes": null
    },
    {
      "id": "642363",
      "postDate": "10/06/2019 00:38:08",
      "content": "<p>I have never experimented FPN, but for better discussion, I leave the comment about the difference between them.</p>\n\n<p>The main difference is that there are multiple prediction layers: one for each upsampling layer. Like the U-Net, the FPN has laterals connection between the bottom-up pyramid (left) and the top-down pyramid (right). But, where U-net only copies the features and append them, FPN apply an a1x1 convolution layer before adding them. This allows the bottom-up pyramid called “backbone” to be pretty much whatever you want.</p>",
      "rawMarkdown": "I have never experimented FPN, but for better discussion, I leave the comment about the difference between them.\n\nThe main difference is that there are multiple prediction layers: one for each upsampling layer. Like the U-Net, the FPN has laterals connection between the bottom-up pyramid (left) and the top-down pyramid (right). But, where U-net only copies the features and append them, FPN apply an a1x1 convolution layer before adding them. This allows the bottom-up pyramid called “backbone” to be pretty much whatever you want.",
      "votes": null
    },
    {
      "id": "642407",
      "postDate": "10/06/2019 02:54:22",
      "content": "<p>My best score is achieved by using the FPN with seresnet34 with 5 fold.</p>",
      "rawMarkdown": "My best score is achieved by using the FPN with seresnet34 with 5 fold.",
      "votes": null
    },
    {
      "id": "642426",
      "postDate": "10/06/2019 04:20:15",
      "content": "<p>I am using FPN for the task, but my score is only 0.630(backbone resnet50) first.</p>",
      "rawMarkdown": "I am using FPN for the task, but my score is only 0.630(backbone resnet50) first.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 642363,
      "author_name": "syumei",
      "author_url": "",
      "post_date": "10/06/2019 00:38:08",
      "content": "<p>I have never experimented FPN, but for better discussion, I leave the comment about the difference between them.</p>\n\n<p>The main difference is that there are multiple prediction layers: one for each upsampling layer. Like the U-Net, the FPN has laterals connection between the bottom-up pyramid (left) and the top-down pyramid (right). But, where U-net only copies the features and append them, FPN apply an a1x1 convolution layer before adding them. This allows the bottom-up pyramid called “backbone” to be pretty much whatever you want.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 642407,
      "author_name": "adish333",
      "author_url": "",
      "post_date": "10/06/2019 02:54:22",
      "content": "<p>My best score is achieved by using the FPN with seresnet34 with 5 fold.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 642426,
      "author_name": "bigyellower",
      "author_url": "",
      "post_date": "10/06/2019 04:20:15",
      "content": "<p>I am using FPN for the task, but my score is only 0.630(backbone resnet50) first.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "642048": "I believe most of us are experimenting with Unet models(+different encoders). Did anybody try `FPN` based models? How does it perform on LB?\nI have been training with `FPN` model and so far it looks promising on validation...will update the results once tested on LB",
    "642363": "I have never experimented FPN, but for better discussion, I leave the comment about the difference between them.\n\nThe main difference is that there are multiple prediction layers: one for each upsampling layer. Like the U-Net, the FPN has laterals connection between the bottom-up pyramid (left) and the top-down pyramid (right). But, where U-net only copies the features and append them, FPN apply an a1x1 convolution layer before adding them. This allows the bottom-up pyramid called “backbone” to be pretty much whatever you want.",
    "642407": "My best score is achieved by using the FPN with seresnet34 with 5 fold.",
    "642426": "I am using FPN for the task, but my score is only 0.630(backbone resnet50) first."
  },
  "source": "meta"
}