{
  "id": 73336,
  "title": "How to Identify ID for Test images?",
  "url": "/competitions/humpback-whale-identification/discussion/73336",
  "author_name": "",
  "post_date": "2018-12-02T05:10:33.834277700Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi All,</p>\n\n<p>Can anyone help me? On How to train a model to identify the test images containing this ID  \"new_whale w_23a388d w_9b5109b w_9c506f6 w_0369a5c\" . As you can see in the train data each image correspod to one of   new_whale or w_23a388d or w_9b5109b or w_9c506f6 or w_0369a5c</p>",
  "messages": [
    {
      "id": "431360",
      "postDate": "12/02/2018 05:10:33",
      "content": "<p>Hi All,</p>\n\n<p>Can anyone help me? On How to train a model to identify the test images containing this ID  \"new_whale w_23a388d w_9b5109b w_9c506f6 w_0369a5c\" . As you can see in the train data each image correspod to one of   new_whale or w_23a388d or w_9b5109b or w_9c506f6 or w_0369a5c</p>",
      "rawMarkdown": "Hi All,\n\nCan anyone help me? On How to train a model to identify the test images containing this ID\t\"new_whale w_23a388d w_9b5109b w_9c506f6 w_0369a5c\" . As you can see in the train data each image correspod to one of \tnew_whale or w_23a388d or w_9b5109b or w_9c506f6 or w_0369a5c",
      "votes": null
    },
    {
      "id": "431402",
      "postDate": "12/02/2018 07:31:31",
      "content": "<p>Test images don't have classes - you need to predict them yourself.\nYou can find test images in the folder \"../input/test\"</p>",
      "rawMarkdown": "Test images don't have classes - you need to predict them yourself.\nYou can find test images in the folder \"../input/test\"",
      "votes": null
    },
    {
      "id": "432297",
      "postDate": "12/03/2018 17:31:04",
      "content": "<p>Hi!</p>\n\n<p>It depends on how you handle the outputs of your model. I think that, for reaching a satisfactory response to your question, it is necessary to take into account two important things:</p>\n\n<ol>\n<li>As said before, you should configure your predictor output to provide as far as 5 possible candidates for the evaluated image (take into account that providing 5 candidates is not mandatory)</li>\n<li>You should think about how to handle the \"new_whale\" tag. It could be applied to an evaluation image that does not provide a sufficient likelihood to be any of the known tags, or you can train a model that handles the \"new_whale\" tag as any other. About this respect, I have no currently a clear idea on how to do it. However, you could surely look for hints in the competition kernels.</li>\n</ol>\n\n<p>Hope the answer has been useful. Regards!</p>",
      "rawMarkdown": "Hi!\n\nIt depends on how you handle the outputs of your model. I think that, for reaching a satisfactory response to your question, it is necessary to take into account two important things:\n\n1. As said before, you should configure your predictor output to provide as far as 5 possible candidates for the evaluated image (take into account that providing 5 candidates is not mandatory)\n2. You should think about how to handle the \"new_whale\" tag. It could be applied to an evaluation image that does not provide a sufficient likelihood to be any of the known tags, or you can train a model that handles the \"new_whale\" tag as any other. About this respect, I have no currently a clear idea on how to do it. However, you could surely look for hints in the competition kernels.\n\nHope the answer has been useful. Regards!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 431402,
      "author_name": "artgor",
      "author_url": "",
      "post_date": "12/02/2018 07:31:31",
      "content": "<p>Test images don't have classes - you need to predict them yourself.\nYou can find test images in the folder \"../input/test\"</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 432297,
      "author_name": "chelus",
      "author_url": "",
      "post_date": "12/03/2018 17:31:04",
      "content": "<p>Hi!</p>\n\n<p>It depends on how you handle the outputs of your model. I think that, for reaching a satisfactory response to your question, it is necessary to take into account two important things:</p>\n\n<ol>\n<li>As said before, you should configure your predictor output to provide as far as 5 possible candidates for the evaluated image (take into account that providing 5 candidates is not mandatory)</li>\n<li>You should think about how to handle the \"new_whale\" tag. It could be applied to an evaluation image that does not provide a sufficient likelihood to be any of the known tags, or you can train a model that handles the \"new_whale\" tag as any other. About this respect, I have no currently a clear idea on how to do it. However, you could surely look for hints in the competition kernels.</li>\n</ol>\n\n<p>Hope the answer has been useful. Regards!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "431360": "Hi All,\n\nCan anyone help me? On How to train a model to identify the test images containing this ID\t\"new_whale w_23a388d w_9b5109b w_9c506f6 w_0369a5c\" . As you can see in the train data each image correspod to one of \tnew_whale or w_23a388d or w_9b5109b or w_9c506f6 or w_0369a5c",
    "431402": "Test images don't have classes - you need to predict them yourself.\nYou can find test images in the folder \"../input/test\"",
    "432297": "Hi!\n\nIt depends on how you handle the outputs of your model. I think that, for reaching a satisfactory response to your question, it is necessary to take into account two important things:\n\n1. As said before, you should configure your predictor output to provide as far as 5 possible candidates for the evaluated image (take into account that providing 5 candidates is not mandatory)\n2. You should think about how to handle the \"new_whale\" tag. It could be applied to an evaluation image that does not provide a sufficient likelihood to be any of the known tags, or you can train a model that handles the \"new_whale\" tag as any other. About this respect, I have no currently a clear idea on how to do it. However, you could surely look for hints in the competition kernels.\n\nHope the answer has been useful. Regards!"
  },
  "source": "meta"
}