{
  "id": 49114,
  "title": "My training accuracy is around 93% while the validation accuracy is 79% but in the LB it scores around 51.7%.. Any suggestion?",
  "url": "/competitions/sp-society-camera-model-identification/discussion/49114",
  "author_name": "",
  "post_date": "2018-02-07T01:56:34.418492100Z",
  "votes": 2,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Here is my approach.. I have taken 512x512 image crops from the center of each image.. then created generators which does the different transforms (jpeg compression 70,90, etc..) on 50 % of the images in each batch and finally I take 224x224 center crop from the transformed as well as unaltered images . Have used the same generator for both training and validation data. Used a resnet50 to train with \"imagenet\" weights as initial weights. The training accuracy is around 93% while the validation accuracy is differing from training at 79% . Also LB is at 51.7%. Any suggestions on the discrepancy ?</p>",
  "messages": [
    {
      "id": "278858",
      "postDate": "02/07/2018 01:56:34",
      "content": "<p>Here is my approach.. I have taken 512x512 image crops from the center of each image.. then created generators which does the different transforms (jpeg compression 70,90, etc..) on 50 % of the images in each batch and finally I take 224x224 center crop from the transformed as well as unaltered images . Have used the same generator for both training and validation data. Used a resnet50 to train with \"imagenet\" weights as initial weights. The training accuracy is around 93% while the validation accuracy is differing from training at 79% . Also LB is at 51.7%. Any suggestions on the discrepancy ?</p>",
      "rawMarkdown": "Here is my approach.. I have taken 512x512 image crops from the center of each image.. then created generators which does the different transforms (jpeg compression 70,90, etc..) on 50 % of the images in each batch and finally I take 224x224 center crop from the transformed as well as unaltered images . Have used the same generator for both training and validation data. Used a resnet50 to train with \"imagenet\" weights as initial weights. The training accuracy is around 93% while the validation accuracy is differing from training at 79% . Also LB is at 51.7%. Any suggestions on the discrepancy ?",
      "votes": null
    },
    {
      "id": "279075",
      "postDate": "02/07/2018 09:57:00",
      "content": "<p>I did a similar approach when I started. Try to perform 512x512 random crops in order to have more data for training. This helped me a lot.</p>",
      "rawMarkdown": "I did a similar approach when I started. Try to perform 512x512 random crops in order to have more data for training. This helped me a lot.",
      "votes": null
    },
    {
      "id": "279077",
      "postDate": "02/07/2018 10:07:47",
      "content": "<p>thanks IgorMuniz.. when you say 512x512 random crops you mean the ones to be fed to the neural network or the ones from the raw images before the  different transformation(50% of them)...</p>",
      "rawMarkdown": "thanks IgorMuniz.. when you say 512x512 random crops you mean the ones to be fed to the neural network or the ones from the raw images before the  different transformation(50% of them)...",
      "votes": null
    },
    {
      "id": "279093",
      "postDate": "02/07/2018 10:48:01",
      "content": "<p>Random crops from raw images. You can also do the transformations and then perform the random crops. The point is: do not just use center crops. This will decrease your training data and consequently the validation score. </p>",
      "rawMarkdown": "Random crops from raw images. You can also do the transformations and then perform the random crops. The point is: do not just use center crops. This will decrease your training data and consequently the validation score.",
      "votes": null
    },
    {
      "id": "279187",
      "postDate": "02/07/2018 15:06:13",
      "content": "<p>thanks Igor for the suggestion..</p>",
      "rawMarkdown": "thanks Igor for the suggestion..",
      "votes": null
    },
    {
      "id": "279504",
      "postDate": "02/08/2018 05:48:30",
      "content": "<p>I met the same situation in this competition, but I don't have good solution to solve it too. I only got a 0.842 LB score by chance one time. And I have never improved since then. That is what really puzzles me.</p>",
      "rawMarkdown": "I met the same situation in this competition, but I don't have good solution to solve it too. I only got a 0.842 LB score by chance one time. And I have never improved since then. That is what really puzzles me.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 279075,
      "author_name": "igormunizims",
      "author_url": "",
      "post_date": "02/07/2018 09:57:00",
      "content": "<p>I did a similar approach when I started. Try to perform 512x512 random crops in order to have more data for training. This helped me a lot.</p>",
      "votes": null,
      "replies": [
        {
          "id": 279077,
          "author_name": "santanuds",
          "author_url": "",
          "post_date": "02/07/2018 10:07:47",
          "content": "<p>thanks IgorMuniz.. when you say 512x512 random crops you mean the ones to be fed to the neural network or the ones from the raw images before the  different transformation(50% of them)...</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 279093,
          "author_name": "igormunizims",
          "author_url": "",
          "post_date": "02/07/2018 10:48:01",
          "content": "<p>Random crops from raw images. You can also do the transformations and then perform the random crops. The point is: do not just use center crops. This will decrease your training data and consequently the validation score. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 279187,
          "author_name": "santanuds",
          "author_url": "",
          "post_date": "02/07/2018 15:06:13",
          "content": "<p>thanks Igor for the suggestion..</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 279504,
      "author_name": "hzywish",
      "author_url": "",
      "post_date": "02/08/2018 05:48:30",
      "content": "<p>I met the same situation in this competition, but I don't have good solution to solve it too. I only got a 0.842 LB score by chance one time. And I have never improved since then. That is what really puzzles me.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "278858": "Here is my approach.. I have taken 512x512 image crops from the center of each image.. then created generators which does the different transforms (jpeg compression 70,90, etc..) on 50 % of the images in each batch and finally I take 224x224 center crop from the transformed as well as unaltered images . Have used the same generator for both training and validation data. Used a resnet50 to train with \"imagenet\" weights as initial weights. The training accuracy is around 93% while the validation accuracy is differing from training at 79% . Also LB is at 51.7%. Any suggestions on the discrepancy ?",
    "279075": "I did a similar approach when I started. Try to perform 512x512 random crops in order to have more data for training. This helped me a lot.",
    "279077": "thanks IgorMuniz.. when you say 512x512 random crops you mean the ones to be fed to the neural network or the ones from the raw images before the  different transformation(50% of them)...",
    "279093": "Random crops from raw images. You can also do the transformations and then perform the random crops. The point is: do not just use center crops. This will decrease your training data and consequently the validation score.",
    "279187": "thanks Igor for the suggestion..",
    "279504": "I met the same situation in this competition, but I don't have good solution to solve it too. I only got a 0.842 LB score by chance one time. And I have never improved since then. That is what really puzzles me."
  },
  "source": "meta"
}