{
  "id": 38654,
  "title": "terrible accuracy with transfer learning score of 17+",
  "url": "/competitions/dogs-vs-cats-redux-kernels-edition/discussion/38654",
  "author_name": "",
  "post_date": "2017-08-28T14:06:13.896741100Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>i used VGG16 for transfer learning and trained for 20 epochs got a training accuracy of 87 and validation accuracy of 87% and when i submitted my score was 17+ i could not able to find any reason why my score is so bad\ni am attaching my submission file can someone please help me</p>",
  "messages": [
    {
      "id": "216876",
      "postDate": "08/28/2017 14:06:13",
      "content": "<p>i used VGG16 for transfer learning and trained for 20 epochs got a training accuracy of 87 and validation accuracy of 87% and when i submitted my score was 17+ i could not able to find any reason why my score is so bad\ni am attaching my submission file can someone please help me</p>",
      "rawMarkdown": "i used VGG16 for transfer learning and trained for 20 epochs got a training accuracy of 87 and validation accuracy of 87% and when i submitted my score was 17+ i could not able to find any reason why my score is so bad\ni am attaching my submission file can someone please help me",
      "votes": null
    },
    {
      "id": "220431",
      "postDate": "09/12/2017 04:55:35",
      "content": "<p>Try to fine tune few of the lower conv layers.\nI did fine tuning of the lower conv layers other than just the top layer and my validation accuracy shot to 99.99%.\nI will very soon publish the kernel with my codebase.</p>",
      "rawMarkdown": "Try to fine tune few of the lower conv layers.\nI did fine tuning of the lower conv layers other than just the top layer and my validation accuracy shot to 99.99%.\nI will very soon publish the kernel with my codebase.",
      "votes": null
    },
    {
      "id": "220708",
      "postDate": "09/12/2017 21:37:26",
      "content": "<p>I believe you get a score of around 17 if you are returning whether it is a dog or not (0 vs. 1) values. In the Evaluation section of the Overview section, it says that the scoring is based on log loss. You would need to grab the probabilities of the predictions for whether it is a dog. Additionally, you would want to clip the values with min and max values that are not (0, 1) i.e. 0.05, 0.999 using is_dog = is_dog.clip(min=0.05, max=0.95) or is_dog = np.clip(predictions, min=0.05, max=0.95)).</p>",
      "rawMarkdown": "I believe you get a score of around 17 if you are returning whether it is a dog or not (0 vs. 1) values. In the Evaluation section of the Overview section, it says that the scoring is based on log loss. You would need to grab the probabilities of the predictions for whether it is a dog. Additionally, you would want to clip the values with min and max values that are not (0, 1) i.e. 0.05, 0.999 using is_dog = is_dog.clip(min=0.05, max=0.95) or is_dog = np.clip(predictions, min=0.05, max=0.95)).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 220431,
      "author_name": "abhayizm",
      "author_url": "",
      "post_date": "09/12/2017 04:55:35",
      "content": "<p>Try to fine tune few of the lower conv layers.\nI did fine tuning of the lower conv layers other than just the top layer and my validation accuracy shot to 99.99%.\nI will very soon publish the kernel with my codebase.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 220708,
      "author_name": "kmatsuda",
      "author_url": "",
      "post_date": "09/12/2017 21:37:26",
      "content": "<p>I believe you get a score of around 17 if you are returning whether it is a dog or not (0 vs. 1) values. In the Evaluation section of the Overview section, it says that the scoring is based on log loss. You would need to grab the probabilities of the predictions for whether it is a dog. Additionally, you would want to clip the values with min and max values that are not (0, 1) i.e. 0.05, 0.999 using is_dog = is_dog.clip(min=0.05, max=0.95) or is_dog = np.clip(predictions, min=0.05, max=0.95)).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "216876": "i used VGG16 for transfer learning and trained for 20 epochs got a training accuracy of 87 and validation accuracy of 87% and when i submitted my score was 17+ i could not able to find any reason why my score is so bad\ni am attaching my submission file can someone please help me",
    "220431": "Try to fine tune few of the lower conv layers.\nI did fine tuning of the lower conv layers other than just the top layer and my validation accuracy shot to 99.99%.\nI will very soon publish the kernel with my codebase.",
    "220708": "I believe you get a score of around 17 if you are returning whether it is a dog or not (0 vs. 1) values. In the Evaluation section of the Overview section, it says that the scoring is based on log loss. You would need to grab the probabilities of the predictions for whether it is a dog. Additionally, you would want to clip the values with min and max values that are not (0, 1) i.e. 0.05, 0.999 using is_dog = is_dog.clip(min=0.05, max=0.95) or is_dog = np.clip(predictions, min=0.05, max=0.95))."
  },
  "source": "meta"
}