{
  "id": 455426,
  "title": "Help needed!     Inference is timing out.",
  "url": "/competitions/UBC-OCEAN/discussion/455426",
  "author_name": "Huang Jin Feng",
  "post_date": "2023-11-14T16:55:32.762000",
  "votes": 5,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Is there any way to speed up the process of cutting PNG images into patches, other than the maximum sampling number threshold?</p>\n<p>I used parallel processing with three cores, and because my strategy is weakly supervised multiple instance learning, I don't need to store a large number of cut patches under the temp space.   Instead, I convert the patches from a whole-slide image directly into feature vectors in memory and then save them as .pt files in the temp space.</p>\n<p>Setting random sampling has a big impact on multiple instance learning…</p>",
  "messages": [
    {
      "id": 2524985,
      "postDate": "2023-11-14T16:55:32.763Z",
      "content": "<p>Is there any way to speed up the process of cutting PNG images into patches, other than the maximum sampling number threshold?</p>\n<p>I used parallel processing with three cores, and because my strategy is weakly supervised multiple instance learning, I don't need to store a large number of cut patches under the temp space.   Instead, I convert the patches from a whole-slide image directly into feature vectors in memory and then save them as .pt files in the temp space.</p>\n<p>Setting random sampling has a big impact on multiple instance learning…</p>",
      "rawMarkdown": "Is there any way to speed up the process of cutting PNG images into patches, other than the maximum sampling number threshold?\n\nI used parallel processing with three cores, and because my strategy is weakly supervised multiple instance learning, I don't need to store a large number of cut patches under the temp space.   Instead, I convert the patches from a whole-slide image directly into feature vectors in memory and then save them as .pt files in the temp space.\n\nSetting random sampling has a big impact on multiple instance learning...",
      "votes": 5
    },
    {
      "id": 2525267,
      "postDate": "2023-11-15T01:24:33.017Z",
      "content": "<p>Have you looked at Mayo competition 2nd place winning solution used a strategy called smart tiles to create patches. Might be useful</p>",
      "rawMarkdown": "Have you looked at Mayo competition 2nd place winning solution used a strategy called smart tiles to create patches. Might be useful\n",
      "votes": 3,
      "replies": [
        {
          "id": 2526103,
          "postDate": "2023-11-15T15:28:35.180Z",
          "content": "<p>Sure, thank you, I'll check it out right away.</p>",
          "rawMarkdown": "\nSure, thank you, I'll check it out right away.",
          "votes": 1
        }
      ]
    },
    {
      "id": 2526097,
      "postDate": "2023-11-15T15:22:28.540Z",
      "content": "<p>I created an unsupervised approach to select tiles.  This might reduce the number of tiles that goes into memory. <a href=\"https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne\" target=\"_blank\">https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne</a></p>",
      "rawMarkdown": "I created an unsupervised approach to select tiles.  This might reduce the number of tiles that goes into memory. https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne",
      "votes": 1,
      "replies": [
        {
          "id": 2526106,
          "postDate": "2023-11-15T15:30:36.937Z",
          "content": "<p>Wow！ Could you tell me which method is faster, this way or cutting patches directly?</p>",
          "rawMarkdown": "Wow！ Could you tell me which method is faster, this way or cutting patches directly?",
          "votes": 1,
          "replies": [
            {
              "id": 2526275,
              "postDate": "2023-11-15T17:26:24.797Z",
              "content": "<p>Working with pre-extracted patches is likely to be more efficient. This is because there are already datasets available that contain these extracted patches, eliminating the need for you to perform this extraction process. Additionally, extracting features simplifies the complexity, though it should be noted that this might result in some information loss.</p>",
              "rawMarkdown": "Working with pre-extracted patches is likely to be more efficient. This is because there are already datasets available that contain these extracted patches, eliminating the need for you to perform this extraction process. Additionally, extracting features simplifies the complexity, though it should be noted that this might result in some information loss."
            }
          ]
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 2525267,
      "author_name": "samu2505",
      "author_url": "",
      "post_date": "2023-11-15T01:24:33.017000",
      "content": "<p>Have you looked at Mayo competition 2nd place winning solution used a strategy called smart tiles to create patches. Might be useful</p>",
      "votes": 3,
      "replies": [
        {
          "id": 2526103,
          "author_name": "Huang Jin Feng",
          "author_url": "",
          "post_date": "2023-11-15T15:28:35.180000",
          "content": "<p>Sure, thank you, I'll check it out right away.</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 2526097,
      "author_name": "dhinesh",
      "author_url": "",
      "post_date": "2023-11-15T15:22:28.540000",
      "content": "<p>I created an unsupervised approach to select tiles.  This might reduce the number of tiles that goes into memory. <a href=\"https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne\" target=\"_blank\">https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne</a></p>",
      "votes": 1,
      "replies": [
        {
          "id": 2526106,
          "author_name": "Huang Jin Feng",
          "author_url": "",
          "post_date": "2023-11-15T15:30:36.937000",
          "content": "<p>Wow！ Could you tell me which method is faster, this way or cutting patches directly?</p>",
          "votes": 1,
          "replies": [
            {
              "id": 2526275,
              "author_name": "dhinesh",
              "author_url": "",
              "post_date": "2023-11-15T17:26:24.797000",
              "content": "<p>Working with pre-extracted patches is likely to be more efficient. This is because there are already datasets available that contain these extracted patches, eliminating the need for you to perform this extraction process. Additionally, extracting features simplifies the complexity, though it should be noted that this might result in some information loss.</p>",
              "votes": 0,
              "replies": []
            }
          ]
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2524985": "Is there any way to speed up the process of cutting PNG images into patches, other than the maximum sampling number threshold?\n\nI used parallel processing with three cores, and because my strategy is weakly supervised multiple instance learning, I don't need to store a large number of cut patches under the temp space.   Instead, I convert the patches from a whole-slide image directly into feature vectors in memory and then save them as .pt files in the temp space.\n\nSetting random sampling has a big impact on multiple instance learning...",
    "2525267": "Have you looked at Mayo competition 2nd place winning solution used a strategy called smart tiles to create patches. Might be useful\n",
    "2526097": "I created an unsupervised approach to select tiles.  This might reduce the number of tiles that goes into memory. https://www.kaggle.com/code/dhinkris/robust-feature-extraction-resnet-k-means-t-sne"
  }
}