{
  "id": 133729,
  "title": "Accuracy Secret Sauce: 100% on Validation and 95% on Test",
  "url": "/competitions/bengaliai-cv19/discussion/133729",
  "author_name": "",
  "post_date": "2020-03-04T02:19:06.267393900Z",
  "votes": null,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello, I have run multiple models and integrated various augmentation techniques from notebooks and ended up having 100% accuracy on and cannot move the needle above 95% on Test data.</p>\n\n<p>It would be great if Pros could help on below questions that led them to 98%+:</p>\n\n<ol>\n<li>Single model or 3 models for individual classification?</li>\n<li>Any additional training data that is useful?</li>\n<li>Pseudo-labeling using heavy augmentation on training data? or with additional data?</li>\n<li>Did anyone above 98% implement metric learning? </li>\n<li>Any hints would be helpful without giving away the secret sauce.</li>\n</ol>",
  "messages": [
    {
      "id": "762996",
      "postDate": "03/04/2020 02:19:06",
      "content": "<p>Hello, I have run multiple models and integrated various augmentation techniques from notebooks and ended up having 100% accuracy on and cannot move the needle above 95% on Test data.</p>\n\n<p>It would be great if Pros could help on below questions that led them to 98%+:</p>\n\n<ol>\n<li>Single model or 3 models for individual classification?</li>\n<li>Any additional training data that is useful?</li>\n<li>Pseudo-labeling using heavy augmentation on training data? or with additional data?</li>\n<li>Did anyone above 98% implement metric learning? </li>\n<li>Any hints would be helpful without giving away the secret sauce.</li>\n</ol>",
      "rawMarkdown": "Hello, I have run multiple models and integrated various augmentation techniques from notebooks and ended up having 100% accuracy on and cannot move the needle above 95% on Test data.\n\nIt would be great if Pros could help on below questions that led them to 98%+:\n\n1. Single model or 3 models for individual classification?\n2. Any additional training data that is useful?\n3. Pseudo-labeling using heavy augmentation on training data? or with additional data?\n4. Did anyone above 98% implement metric learning? \n5. Any hints would be helpful without giving away the secret sauce.",
      "votes": null
    },
    {
      "id": "763142",
      "postDate": "03/04/2020 07:16:31",
      "content": "<p>Accuracy is not the competition metric, you should use the competition metric as I shown in a recent post.</p>",
      "rawMarkdown": "Accuracy is not the competition metric, you should use the competition metric as I shown in a recent post.",
      "votes": null
    },
    {
      "id": "763512",
      "postDate": "03/04/2020 15:06:03",
      "content": "<p>Thanks for your response. Can you please share the link to your post?</p>\n\n<p>Also, I'm having 100% accuracy with 98.5% ish recall on validation set. That leaves me with very little wiggle room to god knows what enhancements I can further bring in to increase my score. I doubt I'm overfitting the model at this point. Can you please advice on this? Also, how does measuring metric affect training of the model itself (in general)?</p>",
      "rawMarkdown": "Thanks for your response. Can you please share the link to your post?\n\nAlso, I'm having 100% accuracy with 98.5% ish recall on validation set. That leaves me with very little wiggle room to god knows what enhancements I can further bring in to increase my score. I doubt I'm overfitting the model at this point. Can you please advice on this? Also, how does measuring metric affect training of the model itself (in general)?",
      "votes": null
    },
    {
      "id": "763868",
      "postDate": "03/04/2020 23:10:08",
      "content": "<p>Where is that post?\nthank you!</p>",
      "rawMarkdown": "Where is that post?\nthank you!",
      "votes": null
    },
    {
      "id": "763877",
      "postDate": "03/04/2020 23:32:46",
      "content": "<p><a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/133219\">https://www.kaggle.com/c/bengaliai-cv19/discussion/133219</a></p>",
      "rawMarkdown": "https://www.kaggle.com/c/bengaliai-cv19/discussion/133219",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 763142,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "03/04/2020 07:16:31",
      "content": "<p>Accuracy is not the competition metric, you should use the competition metric as I shown in a recent post.</p>",
      "votes": null,
      "replies": [
        {
          "id": 763512,
          "author_name": "harshpatel1692",
          "author_url": "",
          "post_date": "03/04/2020 15:06:03",
          "content": "<p>Thanks for your response. Can you please share the link to your post?</p>\n\n<p>Also, I'm having 100% accuracy with 98.5% ish recall on validation set. That leaves me with very little wiggle room to god knows what enhancements I can further bring in to increase my score. I doubt I'm overfitting the model at this point. Can you please advice on this? Also, how does measuring metric affect training of the model itself (in general)?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 763868,
          "author_name": "joovasco",
          "author_url": "",
          "post_date": "03/04/2020 23:10:08",
          "content": "<p>Where is that post?\nthank you!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 763877,
          "author_name": "cpmpml",
          "author_url": "",
          "post_date": "03/04/2020 23:32:46",
          "content": "<p><a href=\"https://www.kaggle.com/c/bengaliai-cv19/discussion/133219\">https://www.kaggle.com/c/bengaliai-cv19/discussion/133219</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "762996": "Hello, I have run multiple models and integrated various augmentation techniques from notebooks and ended up having 100% accuracy on and cannot move the needle above 95% on Test data.\n\nIt would be great if Pros could help on below questions that led them to 98%+:\n\n1. Single model or 3 models for individual classification?\n2. Any additional training data that is useful?\n3. Pseudo-labeling using heavy augmentation on training data? or with additional data?\n4. Did anyone above 98% implement metric learning? \n5. Any hints would be helpful without giving away the secret sauce.",
    "763142": "Accuracy is not the competition metric, you should use the competition metric as I shown in a recent post.",
    "763512": "Thanks for your response. Can you please share the link to your post?\n\nAlso, I'm having 100% accuracy with 98.5% ish recall on validation set. That leaves me with very little wiggle room to god knows what enhancements I can further bring in to increase my score. I doubt I'm overfitting the model at this point. Can you please advice on this? Also, how does measuring metric affect training of the model itself (in general)?",
    "763868": "Where is that post?\nthank you!",
    "763877": "https://www.kaggle.com/c/bengaliai-cv19/discussion/133219"
  },
  "source": "meta"
}