{
  "id": 108660,
  "title": "Instance Segmentation->How to predict classes",
  "url": "/competitions/understanding_cloud_organization/discussion/108660",
  "author_name": "",
  "post_date": "2019-09-13T06:28:26.213370Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi\nI realized that we need to predict the mask for different categories (fish etc.) for each image. But Unet only outputs single class (that's what i am familiar with , semantic segmentation using carvana dataset).</p>\n\n<p>How can i extend it to multi class implementation (any other model , or replacing the layer at end? </p>",
  "messages": [
    {
      "id": "625505",
      "postDate": "09/13/2019 06:28:26",
      "content": "<p>Hi\nI realized that we need to predict the mask for different categories (fish etc.) for each image. But Unet only outputs single class (that's what i am familiar with , semantic segmentation using carvana dataset).</p>\n\n<p>How can i extend it to multi class implementation (any other model , or replacing the layer at end? </p>",
      "rawMarkdown": "Hi\nI realized that we need to predict the mask for different categories (fish etc.) for each image. But Unet only outputs single class (that's what i am familiar with , semantic segmentation using carvana dataset).\n\nHow can i extend it to multi class implementation (any other model , or replacing the layer at end?",
      "votes": null
    },
    {
      "id": "625586",
      "postDate": "09/13/2019 07:43:06",
      "content": "<p>In my opinion, I will build a classification model whose inputs are outputs of Unet</p>",
      "rawMarkdown": "In my opinion, I will build a classification model whose inputs are outputs of Unet",
      "votes": null
    },
    {
      "id": "627507",
      "postDate": "09/16/2019 04:39:36",
      "content": "<p>Check out     \"segmentation_models_pytorch\" on githhub specifically their unet version.</p>",
      "rawMarkdown": "Check out     \"segmentation_models_pytorch\" on githhub specifically their unet version.",
      "votes": null
    },
    {
      "id": "630284",
      "postDate": "09/20/2019 03:19:07",
      "content": "<p>You would change the decoder final conv layer output channels. <br>\n<code>final_conv = nn.Conv2d(previous_layer_output, final_channels, kernel_size=(1, 1))</code> </p>",
      "rawMarkdown": "You would change the decoder final conv layer output channels.  \n`final_conv = nn.Conv2d(previous_layer_output, final_channels, kernel_size=(1, 1))`",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 625586,
      "author_name": "thanhtinqn97",
      "author_url": "",
      "post_date": "09/13/2019 07:43:06",
      "content": "<p>In my opinion, I will build a classification model whose inputs are outputs of Unet</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 627507,
      "author_name": "petewills",
      "author_url": "",
      "post_date": "09/16/2019 04:39:36",
      "content": "<p>Check out     \"segmentation_models_pytorch\" on githhub specifically their unet version.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 630284,
      "author_name": "convexproblems",
      "author_url": "",
      "post_date": "09/20/2019 03:19:07",
      "content": "<p>You would change the decoder final conv layer output channels. <br>\n<code>final_conv = nn.Conv2d(previous_layer_output, final_channels, kernel_size=(1, 1))</code> </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "625505": "Hi\nI realized that we need to predict the mask for different categories (fish etc.) for each image. But Unet only outputs single class (that's what i am familiar with , semantic segmentation using carvana dataset).\n\nHow can i extend it to multi class implementation (any other model , or replacing the layer at end?",
    "625586": "In my opinion, I will build a classification model whose inputs are outputs of Unet",
    "627507": "Check out     \"segmentation_models_pytorch\" on githhub specifically their unet version.",
    "630284": "You would change the decoder final conv layer output channels.  \n`final_conv = nn.Conv2d(previous_layer_output, final_channels, kernel_size=(1, 1))`"
  },
  "source": "meta"
}