{
  "id": 198341,
  "title": "Let's merge the tiles and do inference. ",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/198341",
  "author_name": "",
  "post_date": "2020-11-20T18:55:09.252863Z",
  "votes": 3,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Most of us are using <a href=\"https://www.kaggle.com/iafoss\" target=\"_blank\">@iafoss</a> 256x256 png images for training.</p>\n<p><strong>But, what about the inference with such images?</strong></p>\n<p><em>Note: I am calling such PNG image, a tile.</em></p>\n<p>Idea  👇:<br>\nModel will predict the mask-tile corresponding to each train-tile. <br>\nThen we have to merge these mask-tiles into one big mask. This big mask will be the prediction corresponding to the one test tiff image. We have 5 such test tiff images in the test folder.</p>\n<p>Implementation of this idea is here 👇 :<br>\n<strong><a href=\"https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference\" target=\"_blank\">https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference</a></strong></p>\n<p>Current version is giving OOM when trying to convert that big mask (Prediction for one test tiff image) to RLE.</p>",
  "messages": [
    {
      "id": "1085211",
      "postDate": "11/20/2020 18:55:09",
      "content": "<p>Most of us are using <a href=\"https://www.kaggle.com/iafoss\" target=\"_blank\">@iafoss</a> 256x256 png images for training.</p>\n<p><strong>But, what about the inference with such images?</strong></p>\n<p><em>Note: I am calling such PNG image, a tile.</em></p>\n<p>Idea  👇:<br>\nModel will predict the mask-tile corresponding to each train-tile. <br>\nThen we have to merge these mask-tiles into one big mask. This big mask will be the prediction corresponding to the one test tiff image. We have 5 such test tiff images in the test folder.</p>\n<p>Implementation of this idea is here 👇 :<br>\n<strong><a href=\"https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference\" target=\"_blank\">https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference</a></strong></p>\n<p>Current version is giving OOM when trying to convert that big mask (Prediction for one test tiff image) to RLE.</p>",
      "rawMarkdown": "Most of us are using @iafoss 256x256 png images for training.\n\n**But, what about the inference with such images?**\n\n*Note: I am calling such PNG image, a tile.*\n\nIdea  👇:\nModel will predict the mask-tile corresponding to each train-tile. \nThen we have to merge these mask-tiles into one big mask. This big mask will be the prediction corresponding to the one test tiff image. We have 5 such test tiff images in the test folder.\n\nImplementation of this idea is here 👇 :\n**https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference**\n\nCurrent version is giving OOM when trying to convert that big mask (Prediction for one test tiff image) to RLE.",
      "votes": null
    },
    {
      "id": "1094349",
      "postDate": "11/28/2020 14:30:11",
      "content": "<p>Hi,I have the same question and I'm trying it for several days.If I use CPU version(16 GB RAM), the time is longer than 9 hours.</p>",
      "rawMarkdown": "Hi,I have the same question and I'm trying it for several days.If I use CPU version(16 GB RAM), the time is longer than 9 hours.",
      "votes": null
    },
    {
      "id": "1094601",
      "postDate": "11/28/2020 18:54:26",
      "content": "<p>What is taking your time?</p>",
      "rawMarkdown": "What is taking your time?",
      "votes": null
    },
    {
      "id": "1095296",
      "postDate": "11/29/2020 13:06:27",
      "content": "<p>I'm trying to reduce the memory in testing stage.</p>",
      "rawMarkdown": "I'm trying to reduce the memory in testing stage.",
      "votes": null
    },
    {
      "id": "1095991",
      "postDate": "11/30/2020 06:12:11",
      "content": "<p>Use the lib like pyvips or rasterio, to only read some part of the image at once, instead of reading and loading the whole image at once. <br>\nRefer to this NB:<br>\n<a href=\"https://www.kaggle.com/leighplt/pytorch-fcn-resnet50\" target=\"_blank\">https://www.kaggle.com/leighplt/pytorch-fcn-resnet50</a></p>",
      "rawMarkdown": "Use the lib like pyvips or rasterio, to only read some part of the image at once, instead of reading and loading the whole image at once. \nRefer to this NB:\nhttps://www.kaggle.com/leighplt/pytorch-fcn-resnet50",
      "votes": null
    },
    {
      "id": "1096048",
      "postDate": "11/30/2020 07:20:39",
      "content": "<p>Thanks a lot.I will try it.</p>",
      "rawMarkdown": "Thanks a lot.I will try it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1094349,
      "author_name": "xinggang1221",
      "author_url": "",
      "post_date": "11/28/2020 14:30:11",
      "content": "<p>Hi,I have the same question and I'm trying it for several days.If I use CPU version(16 GB RAM), the time is longer than 9 hours.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1094601,
          "author_name": "joshi98kishan",
          "author_url": "",
          "post_date": "11/28/2020 18:54:26",
          "content": "<p>What is taking your time?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1095296,
          "author_name": "xinggang1221",
          "author_url": "",
          "post_date": "11/29/2020 13:06:27",
          "content": "<p>I'm trying to reduce the memory in testing stage.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1095991,
          "author_name": "joshi98kishan",
          "author_url": "",
          "post_date": "11/30/2020 06:12:11",
          "content": "<p>Use the lib like pyvips or rasterio, to only read some part of the image at once, instead of reading and loading the whole image at once. <br>\nRefer to this NB:<br>\n<a href=\"https://www.kaggle.com/leighplt/pytorch-fcn-resnet50\" target=\"_blank\">https://www.kaggle.com/leighplt/pytorch-fcn-resnet50</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1096048,
          "author_name": "xinggang1221",
          "author_url": "",
          "post_date": "11/30/2020 07:20:39",
          "content": "<p>Thanks a lot.I will try it.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1085211": "Most of us are using @iafoss 256x256 png images for training.\n\n**But, what about the inference with such images?**\n\n*Note: I am calling such PNG image, a tile.*\n\nIdea  👇:\nModel will predict the mask-tile corresponding to each train-tile. \nThen we have to merge these mask-tiles into one big mask. This big mask will be the prediction corresponding to the one test tiff image. We have 5 such test tiff images in the test folder.\n\nImplementation of this idea is here 👇 :\n**https://www.kaggle.com/joshi98kishan/hubmap-keras-pipeline-training-inference**\n\nCurrent version is giving OOM when trying to convert that big mask (Prediction for one test tiff image) to RLE.",
    "1094349": "Hi,I have the same question and I'm trying it for several days.If I use CPU version(16 GB RAM), the time is longer than 9 hours.",
    "1094601": "What is taking your time?",
    "1095296": "I'm trying to reduce the memory in testing stage.",
    "1095991": "Use the lib like pyvips or rasterio, to only read some part of the image at once, instead of reading and loading the whole image at once. \nRefer to this NB:\nhttps://www.kaggle.com/leighplt/pytorch-fcn-resnet50",
    "1096048": "Thanks a lot.I will try it."
  },
  "source": "meta"
}