{
  "id": 34674,
  "title": "Anyone try loading in grayscale?",
  "url": "/competitions/intel-mobileodt-cervical-cancer-screening/discussion/34674",
  "author_name": "",
  "post_date": "2017-06-14T00:14:10.862818600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Since colour change a lot, anyone get correct LB training and testing using grayscale?</p>",
  "messages": [
    {
      "id": "192540",
      "postDate": "06/14/2017 00:14:10",
      "content": "<p>Since colour change a lot, anyone get correct LB training and testing using grayscale?</p>",
      "rawMarkdown": "Since colour change a lot, anyone get correct LB training and testing using grayscale?",
      "votes": null
    },
    {
      "id": "192668",
      "postDate": "06/14/2017 10:40:37",
      "content": "<p>Overfitting problems when using simple CNN model and 128/grayscale images sets.</p>",
      "rawMarkdown": "Overfitting problems when using simple CNN model and 128/grayscale images sets.",
      "votes": null
    },
    {
      "id": "192720",
      "postDate": "06/14/2017 14:55:13",
      "content": "<p>@Pierre Tisseur,  I am also interested in how good grayscale is compared to color in this competition. </p>\n\n<p>I considered grayscale as a means to save computing time but soon discarded it. The reason: In my understanding <strong>when you convert to grayscale you just manually choose a linear combination of the three colour channels</strong> and go on from that point. </p>\n\n<p>If, on the other hand  you feed your first layer with the 3 color channel dimensions you are allowing your filters to try many more possible combinations of this pixel color values, combined with their neighbouring pixels. As a consequence you will not achieve better results with grayscale, <strong>unless computing resources are too limited</strong> for handling all channels from the beginning.</p>\n\n<p>About <strong>overfitting</strong>, it is always an issue but in this competition with such a small number of images... it overfits blazing fast!!! (using 128 color images here) I think a Convnet has no problem in memorizing this small image set, color or grayscale. :-)</p>",
      "rawMarkdown": "Pierre Tisseur,  I am also interested in how good grayscale is compared to color in this competition. \n\nI considered grayscale as a means to save computing time but soon discarded it. The reason: In my understanding **when you convert to grayscale you just manually choose a linear combination of the three colour channels** and go on from that point. \n\nIf, on the other hand  you feed your first layer with the 3 color channel dimensions you are allowing your filters to try many more possible combinations of this pixel color values, combined with their neighbouring pixels. As a consequence you will not achieve better results with grayscale, **unless computing resources are too limited** for handling all channels from the beginning.\n\nAbout **overfitting**, it is always an issue but in this competition with such a small number of images... it overfits blazing fast!!! (using 128 color images here) I think a Convnet has no problem in memorizing this small image set, color or grayscale. :-)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 192668,
      "author_name": "pierretisseur",
      "author_url": "",
      "post_date": "06/14/2017 10:40:37",
      "content": "<p>Overfitting problems when using simple CNN model and 128/grayscale images sets.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 192720,
      "author_name": "miguelpm",
      "author_url": "",
      "post_date": "06/14/2017 14:55:13",
      "content": "<p>@Pierre Tisseur,  I am also interested in how good grayscale is compared to color in this competition. </p>\n\n<p>I considered grayscale as a means to save computing time but soon discarded it. The reason: In my understanding <strong>when you convert to grayscale you just manually choose a linear combination of the three colour channels</strong> and go on from that point. </p>\n\n<p>If, on the other hand  you feed your first layer with the 3 color channel dimensions you are allowing your filters to try many more possible combinations of this pixel color values, combined with their neighbouring pixels. As a consequence you will not achieve better results with grayscale, <strong>unless computing resources are too limited</strong> for handling all channels from the beginning.</p>\n\n<p>About <strong>overfitting</strong>, it is always an issue but in this competition with such a small number of images... it overfits blazing fast!!! (using 128 color images here) I think a Convnet has no problem in memorizing this small image set, color or grayscale. :-)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "192540": "Since colour change a lot, anyone get correct LB training and testing using grayscale?",
    "192668": "Overfitting problems when using simple CNN model and 128/grayscale images sets.",
    "192720": "Pierre Tisseur,  I am also interested in how good grayscale is compared to color in this competition. \n\nI considered grayscale as a means to save computing time but soon discarded it. The reason: In my understanding **when you convert to grayscale you just manually choose a linear combination of the three colour channels** and go on from that point. \n\nIf, on the other hand  you feed your first layer with the 3 color channel dimensions you are allowing your filters to try many more possible combinations of this pixel color values, combined with their neighbouring pixels. As a consequence you will not achieve better results with grayscale, **unless computing resources are too limited** for handling all channels from the beginning.\n\nAbout **overfitting**, it is always an issue but in this competition with such a small number of images... it overfits blazing fast!!! (using 128 color images here) I think a Convnet has no problem in memorizing this small image set, color or grayscale. :-)"
  },
  "source": "meta"
}