{
  "id": 31191,
  "title": "Anyone else's models not like type 1?",
  "url": "/competitions/intel-mobileodt-cervical-cancer-screening/discussion/31191",
  "author_name": "",
  "post_date": "2017-04-05T18:51:42.138296Z",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<p>My CNN REALLY does not like to guess type one. Every now and then it gives ~40% to type one, but it usually ends up something like this:</p>\n\n<p>Type 1                    Type 2                     Type 3</p>\n\n<p>0.0782116726    0.3842832446    0.5375050306</p>\n\n<p>0.0670149922    0.3760157228    0.556969285</p>\n\n<p>0.0668902844    0.3754892647    0.5576205254</p>\n\n<p>0.0669433326    0.3758482039    0.5572084785</p>\n\n<p>0.0671149641    0.3761534691    0.5567315817</p>\n\n<p>0.0688785538    0.3773455322    0.5537759066</p>\n\n<p>0.0670827404    0.3761104643    0.5568067431</p>\n\n<p>0.0664959103    0.3751574457    0.5583466291</p>\n\n<p>0.0668536797    0.3756740689    0.557472229</p>\n\n<p>0.0681168213    0.3771105409    0.5547726154</p>",
  "messages": [
    {
      "id": "173035",
      "postDate": "04/05/2017 18:51:42",
      "content": "<p>My CNN REALLY does not like to guess type one. Every now and then it gives ~40% to type one, but it usually ends up something like this:</p>\n\n<p>Type 1                    Type 2                     Type 3</p>\n\n<p>0.0782116726    0.3842832446    0.5375050306</p>\n\n<p>0.0670149922    0.3760157228    0.556969285</p>\n\n<p>0.0668902844    0.3754892647    0.5576205254</p>\n\n<p>0.0669433326    0.3758482039    0.5572084785</p>\n\n<p>0.0671149641    0.3761534691    0.5567315817</p>\n\n<p>0.0688785538    0.3773455322    0.5537759066</p>\n\n<p>0.0670827404    0.3761104643    0.5568067431</p>\n\n<p>0.0664959103    0.3751574457    0.5583466291</p>\n\n<p>0.0668536797    0.3756740689    0.557472229</p>\n\n<p>0.0681168213    0.3771105409    0.5547726154</p>",
      "rawMarkdown": "My CNN REALLY does not like to guess type one. Every now and then it gives ~40% to type one, but it usually ends up something like this:\n\nType 1                    Type 2                     Type 3\n\n0.0782116726\t0.3842832446\t0.5375050306\n\n0.0670149922\t0.3760157228\t0.556969285\n\n0.0668902844\t0.3754892647\t0.5576205254\n\n0.0669433326\t0.3758482039\t0.5572084785\n\n0.0671149641\t0.3761534691\t0.5567315817\n\n0.0688785538\t0.3773455322\t0.5537759066\n\n0.0670827404\t0.3761104643\t0.5568067431\n\n0.0664959103\t0.3751574457\t0.5583466291\n\n0.0668536797\t0.3756740689\t0.557472229\n\n0.0681168213\t0.3771105409\t0.5547726154",
      "votes": null
    },
    {
      "id": "173039",
      "postDate": "04/05/2017 19:01:11",
      "content": "<p>Hum, you sure you're training it right? The scores are quite off of the training set distribution (should be closer to 0.169 0.527 0.304), so it doesn't even look like it's trying to give the best single answer to the problem.</p>",
      "rawMarkdown": "Hum, you sure you're training it right? The scores are quite off of the training set distribution (should be closer to 0.169 0.527 0.304), so it doesn't even look like it's trying to give the best single answer to the problem.",
      "votes": null
    },
    {
      "id": "173088",
      "postDate": "04/05/2017 22:33:14",
      "content": "<p>It isn't. This was from training on an evenly distributed set.</p>",
      "rawMarkdown": "It isn't. This was from training on an evenly distributed set.",
      "votes": null
    },
    {
      "id": "173089",
      "postDate": "04/05/2017 22:34:40",
      "content": "<p>Either the distribution got messed up somehow, or this is stuck at a terrible loss value.</p>",
      "rawMarkdown": "Either the distribution got messed up somehow, or this is stuck at a terrible loss value.",
      "votes": null
    },
    {
      "id": "173092",
      "postDate": "04/05/2017 22:41:18",
      "content": "<p>It seems the latter is correct, I just have no clue why...</p>",
      "rawMarkdown": "It seems the latter is correct, I just have no clue why...",
      "votes": null
    },
    {
      "id": "173101",
      "postDate": "04/05/2017 23:23:39",
      "content": "<p>After thorough examination, the only issue I could find is that I am resizing the images down to 64x64. Can this cause <em>serious</em> errors like this? I wonder what else it could be</p>",
      "rawMarkdown": "After thorough examination, the only issue I could find is that I am resizing the images down to 64x64. Can this cause *serious* errors like this? I wonder what else it could be",
      "votes": null
    },
    {
      "id": "173110",
      "postDate": "04/05/2017 23:32:52",
      "content": "<p>If you find just the cervix, isolate it and downsample to 64x64, you <em>might</em> be able to classify the type. If you take those sample images and directly downsample to 512x512 you have a very faint chance to figure out the type. With 64x64 whole images, you're feeding your classifier random noise.</p>",
      "rawMarkdown": "If you find just the cervix, isolate it and downsample to 64x64, you *might* be able to classify the type. If you take those sample images and directly downsample to 512x512 you have a very faint chance to figure out the type. With 64x64 whole images, you're feeding your classifier random noise.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 173039,
      "author_name": "mumech",
      "author_url": "",
      "post_date": "04/05/2017 19:01:11",
      "content": "<p>Hum, you sure you're training it right? The scores are quite off of the training set distribution (should be closer to 0.169 0.527 0.304), so it doesn't even look like it's trying to give the best single answer to the problem.</p>",
      "votes": null,
      "replies": [
        {
          "id": 173088,
          "author_name": "benchiislett",
          "author_url": "",
          "post_date": "04/05/2017 22:33:14",
          "content": "<p>It isn't. This was from training on an evenly distributed set.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 173089,
          "author_name": "mumech",
          "author_url": "",
          "post_date": "04/05/2017 22:34:40",
          "content": "<p>Either the distribution got messed up somehow, or this is stuck at a terrible loss value.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 173092,
          "author_name": "benchiislett",
          "author_url": "",
          "post_date": "04/05/2017 22:41:18",
          "content": "<p>It seems the latter is correct, I just have no clue why...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 173101,
          "author_name": "benchiislett",
          "author_url": "",
          "post_date": "04/05/2017 23:23:39",
          "content": "<p>After thorough examination, the only issue I could find is that I am resizing the images down to 64x64. Can this cause <em>serious</em> errors like this? I wonder what else it could be</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 173110,
          "author_name": "mumech",
          "author_url": "",
          "post_date": "04/05/2017 23:32:52",
          "content": "<p>If you find just the cervix, isolate it and downsample to 64x64, you <em>might</em> be able to classify the type. If you take those sample images and directly downsample to 512x512 you have a very faint chance to figure out the type. With 64x64 whole images, you're feeding your classifier random noise.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "173035": "My CNN REALLY does not like to guess type one. Every now and then it gives ~40% to type one, but it usually ends up something like this:\n\nType 1                    Type 2                     Type 3\n\n0.0782116726\t0.3842832446\t0.5375050306\n\n0.0670149922\t0.3760157228\t0.556969285\n\n0.0668902844\t0.3754892647\t0.5576205254\n\n0.0669433326\t0.3758482039\t0.5572084785\n\n0.0671149641\t0.3761534691\t0.5567315817\n\n0.0688785538\t0.3773455322\t0.5537759066\n\n0.0670827404\t0.3761104643\t0.5568067431\n\n0.0664959103\t0.3751574457\t0.5583466291\n\n0.0668536797\t0.3756740689\t0.557472229\n\n0.0681168213\t0.3771105409\t0.5547726154",
    "173039": "Hum, you sure you're training it right? The scores are quite off of the training set distribution (should be closer to 0.169 0.527 0.304), so it doesn't even look like it's trying to give the best single answer to the problem.",
    "173088": "It isn't. This was from training on an evenly distributed set.",
    "173089": "Either the distribution got messed up somehow, or this is stuck at a terrible loss value.",
    "173092": "It seems the latter is correct, I just have no clue why...",
    "173101": "After thorough examination, the only issue I could find is that I am resizing the images down to 64x64. Can this cause *serious* errors like this? I wonder what else it could be",
    "173110": "If you find just the cervix, isolate it and downsample to 64x64, you *might* be able to classify the type. If you take those sample images and directly downsample to 512x512 you have a very faint chance to figure out the type. With 64x64 whole images, you're feeding your classifier random noise."
  },
  "source": "meta"
}