{
  "id": 54421,
  "title": "Fast image type clustering with imagecluster package",
  "url": "/competitions/data-science-bowl-2018/discussion/54421",
  "author_name": "",
  "post_date": "2018-04-13T02:18:59.173602700Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>At stage 1 I used <a href=\"https://github.com/elcorto/imagecluster\">https://github.com/elcorto/imagecluster</a> to help cluster nuclei stainings into different types. It's been useful even though the small size of stage 1 data can be easily clustered by hand). </p>\n\n<p>Now with the large amount of stage 2 test data, I found this package to be extremely useful for quickly showing how stage 2 data is distributed. (known issues: currently the clustering step ignores clusters with only one image. But it's easy to take the generated embeddings and do clustering yourself.)</p>",
  "messages": [
    {
      "id": "313259",
      "postDate": "04/13/2018 02:18:59",
      "content": "<p>At stage 1 I used <a href=\"https://github.com/elcorto/imagecluster\">https://github.com/elcorto/imagecluster</a> to help cluster nuclei stainings into different types. It's been useful even though the small size of stage 1 data can be easily clustered by hand). </p>\n\n<p>Now with the large amount of stage 2 test data, I found this package to be extremely useful for quickly showing how stage 2 data is distributed. (known issues: currently the clustering step ignores clusters with only one image. But it's easy to take the generated embeddings and do clustering yourself.)</p>",
      "rawMarkdown": "At stage 1 I used https://github.com/elcorto/imagecluster to help cluster nuclei stainings into different types. It's been useful even though the small size of stage 1 data can be easily clustered by hand). \n\nNow with the large amount of stage 2 test data, I found this package to be extremely useful for quickly showing how stage 2 data is distributed. (known issues: currently the clustering step ignores clusters with only one image. But it's easy to take the generated embeddings and do clustering yourself.)",
      "votes": null
    },
    {
      "id": "313273",
      "postDate": "04/13/2018 03:10:56",
      "content": "<p>Thanks for the link; it looks really useful. I starred it so I can use it next time. But isn't it a pretrained NN, and therefore External Data that needed to have been posted a week ago?</p>",
      "rawMarkdown": "Thanks for the link; it looks really useful. I starred it so I can use it next time. But isn't it a pretrained NN, and therefore External Data that needed to have been posted a week ago?",
      "votes": null
    },
    {
      "id": "313409",
      "postDate": "04/13/2018 08:07:15",
      "content": "<p>Good point! I didn't think of it as external data. But it use a very common model: vgg16 pretrained on ImageNet available in torchvision and keras.applications. Both torchvision and keras.applications models have been mentioned in the External Data thread, so I guess it should be fine?  </p>",
      "rawMarkdown": "Good point! I didn't think of it as external data. But it use a very common model: vgg16 pretrained on ImageNet available in torchvision and keras.applications. Both torchvision and keras.applications models have been mentioned in the External Data thread, so I guess it should be fine?",
      "votes": null
    },
    {
      "id": "313828",
      "postDate": "04/13/2018 22:20:00",
      "content": "<p>In that case I expect you're fine.</p>",
      "rawMarkdown": "In that case I expect you're fine.",
      "votes": null
    },
    {
      "id": "314004",
      "postDate": "04/14/2018 10:28:30",
      "content": "<p>Good Job</p>",
      "rawMarkdown": "Good Job",
      "votes": null
    },
    {
      "id": "314012",
      "postDate": "04/14/2018 11:11:46",
      "content": "<p>I should say... I just use the image's size(e.g. 256*256) to 'cluster' the images in stage 2 and it works pretty well, haha.\nAnyway, thanks for your code, now I could use both</p>",
      "rawMarkdown": "I should say... I just use the image's size(e.g. 256*256) to 'cluster' the images in stage 2 and it works pretty well, haha.\nAnyway, thanks for your code, now I could use both",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 313273,
      "author_name": "pointyointment",
      "author_url": "",
      "post_date": "04/13/2018 03:10:56",
      "content": "<p>Thanks for the link; it looks really useful. I starred it so I can use it next time. But isn't it a pretrained NN, and therefore External Data that needed to have been posted a week ago?</p>",
      "votes": null,
      "replies": [
        {
          "id": 313409,
          "author_name": "yueqiw",
          "author_url": "",
          "post_date": "04/13/2018 08:07:15",
          "content": "<p>Good point! I didn't think of it as external data. But it use a very common model: vgg16 pretrained on ImageNet available in torchvision and keras.applications. Both torchvision and keras.applications models have been mentioned in the External Data thread, so I guess it should be fine?  </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 313828,
          "author_name": "pointyointment",
          "author_url": "",
          "post_date": "04/13/2018 22:20:00",
          "content": "<p>In that case I expect you're fine.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 314004,
      "author_name": "kunwang",
      "author_url": "",
      "post_date": "04/14/2018 10:28:30",
      "content": "<p>Good Job</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 314012,
      "author_name": "atm584",
      "author_url": "",
      "post_date": "04/14/2018 11:11:46",
      "content": "<p>I should say... I just use the image's size(e.g. 256*256) to 'cluster' the images in stage 2 and it works pretty well, haha.\nAnyway, thanks for your code, now I could use both</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "313259": "At stage 1 I used https://github.com/elcorto/imagecluster to help cluster nuclei stainings into different types. It's been useful even though the small size of stage 1 data can be easily clustered by hand). \n\nNow with the large amount of stage 2 test data, I found this package to be extremely useful for quickly showing how stage 2 data is distributed. (known issues: currently the clustering step ignores clusters with only one image. But it's easy to take the generated embeddings and do clustering yourself.)",
    "313273": "Thanks for the link; it looks really useful. I starred it so I can use it next time. But isn't it a pretrained NN, and therefore External Data that needed to have been posted a week ago?",
    "313409": "Good point! I didn't think of it as external data. But it use a very common model: vgg16 pretrained on ImageNet available in torchvision and keras.applications. Both torchvision and keras.applications models have been mentioned in the External Data thread, so I guess it should be fine?",
    "313828": "In that case I expect you're fine.",
    "314004": "Good Job",
    "314012": "I should say... I just use the image's size(e.g. 256*256) to 'cluster' the images in stage 2 and it works pretty well, haha.\nAnyway, thanks for your code, now I could use both"
  },
  "source": "meta"
}