{
  "id": 47765,
  "title": "Non image-processing techniques",
  "url": "/competitions/sp-society-camera-model-identification/discussion/47765",
  "author_name": "tarobxl",
  "post_date": "2018-01-18T16:29:40.718000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I am trying to see if non image-processing techniques can boost up the prediction accuracy in this competition. Actually one simple post-processing step helped me to go from 0.200 to 0.205 but I think that step could bring more values when the original prediction is better.</p>\n\n<p>It would be cool if any of you could share your output + prediction probability for each camera. Of course the score for that output should be bellow the medal range in order to not screw up the LB. </p>\n\n<p>Here is the code I used at the end of the awesome the1owl's kernel.</p>\n\n<pre><code>prob = etr.predict_proba(xtest)\n\nfor i in range(10):\n\n    c = \"prob\" + str(i)\n\n    test[c] = prob[:,i]\n\ntest.to_csv(\"test_prob.csv\", index=False)\n</code></pre>",
  "messages": [
    {
      "id": 270630,
      "postDate": "2018-01-18T16:29:40.720Z",
      "content": "<p>I am trying to see if non image-processing techniques can boost up the prediction accuracy in this competition. Actually one simple post-processing step helped me to go from 0.200 to 0.205 but I think that step could bring more values when the original prediction is better.</p>\n\n<p>It would be cool if any of you could share your output + prediction probability for each camera. Of course the score for that output should be bellow the medal range in order to not screw up the LB. </p>\n\n<p>Here is the code I used at the end of the awesome the1owl's kernel.</p>\n\n<pre><code>prob = etr.predict_proba(xtest)\n\nfor i in range(10):\n\n    c = \"prob\" + str(i)\n\n    test[c] = prob[:,i]\n\ntest.to_csv(\"test_prob.csv\", index=False)\n</code></pre>",
      "rawMarkdown": "I am trying to see if non image-processing techniques can boost up the prediction accuracy in this competition. Actually one simple post-processing step helped me to go from 0.200 to 0.205 but I think that step could bring more values when the original prediction is better.\n\nIt would be cool if any of you could share your output + prediction probability for each camera. Of course the score for that output should be bellow the medal range in order to not screw up the LB. \n\nHere is the code I used at the end of the awesome the1owl's kernel.\n\n\n    prob = etr.predict_proba(xtest)\n\n    for i in range(10):\n\n        c = \"prob\" + str(i)\n\n        test[c] = prob[:,i]\n\n    test.to_csv(\"test_prob.csv\", index=False)\n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "270630": "I am trying to see if non image-processing techniques can boost up the prediction accuracy in this competition. Actually one simple post-processing step helped me to go from 0.200 to 0.205 but I think that step could bring more values when the original prediction is better.\n\nIt would be cool if any of you could share your output + prediction probability for each camera. Of course the score for that output should be bellow the medal range in order to not screw up the LB. \n\nHere is the code I used at the end of the awesome the1owl's kernel.\n\n\n    prob = etr.predict_proba(xtest)\n\n    for i in range(10):\n\n        c = \"prob\" + str(i)\n\n        test[c] = prob[:,i]\n\n    test.to_csv(\"test_prob.csv\", index=False)\n"
  }
}