{
  "id": 75096,
  "title": "Techniques for model improvement suggestions",
  "url": "/competitions/humpback-whale-identification/discussion/75096",
  "author_name": "",
  "post_date": "2018-12-18T14:28:12.098506900Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have been trying to ensemble multiple renset architectures in my last couple of submissions but didn't get much improvements. Can someone share some augmentation techniques or training techniques which can help get some ideas for improving model performance. </p>\n\n<p>I have tried training models on small images first and then using bigger images to improve model but it didn't get any higher score than 0.693 on leaderboard. </p>\n\n<p>Some observations that I would like to share:</p>\n\n<ul>\n<li>Bigger images get better results</li>\n<li>Test time augmentation helps</li>\n<li>Smaller threshold helps if model is trained properly</li>\n<li>No independent model was able to beat max public score</li>\n</ul>\n\n<p>If anyone has some resources and would like to share it would be great way to learn more about how to handle such datasets for beginners like us.</p>",
  "messages": [
    {
      "id": "441308",
      "postDate": "12/18/2018 14:28:12",
      "content": "<p>I have been trying to ensemble multiple renset architectures in my last couple of submissions but didn't get much improvements. Can someone share some augmentation techniques or training techniques which can help get some ideas for improving model performance. </p>\n\n<p>I have tried training models on small images first and then using bigger images to improve model but it didn't get any higher score than 0.693 on leaderboard. </p>\n\n<p>Some observations that I would like to share:</p>\n\n<ul>\n<li>Bigger images get better results</li>\n<li>Test time augmentation helps</li>\n<li>Smaller threshold helps if model is trained properly</li>\n<li>No independent model was able to beat max public score</li>\n</ul>\n\n<p>If anyone has some resources and would like to share it would be great way to learn more about how to handle such datasets for beginners like us.</p>",
      "rawMarkdown": "I have been trying to ensemble multiple renset architectures in my last couple of submissions but didn't get much improvements. Can someone share some augmentation techniques or training techniques which can help get some ideas for improving model performance. \n\nI have tried training models on small images first and then using bigger images to improve model but it didn't get any higher score than 0.693 on leaderboard. \n\nSome observations that I would like to share:\n\n - Bigger images get better results\n - Test time augmentation helps\n - Smaller threshold helps if model is trained properly\n - No independent model was able to beat max public score\n\nIf anyone has some resources and would like to share it would be great way to learn more about how to handle such datasets for beginners like us.",
      "votes": null
    },
    {
      "id": "441450",
      "postDate": "12/18/2018 17:10:23",
      "content": "<p>f chollet - the inventor of keras produced a very good book on keras.  It has many good tips of resnet and image augmentation, the github code for the book is here:\n<a href=\"https://github.com/fchollet/deep-learning-with-python-notebooks\">https://github.com/fchollet/deep-learning-with-python-notebooks</a></p>\n\n<p>this kernel is a quite a good one but does not use resnet:</p>\n\n<p><a href=\"https://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn\">https://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn</a></p>",
      "rawMarkdown": "f chollet - the inventor of keras produced a very good book on keras.  It has many good tips of resnet and image augmentation, the github code for the book is here:\nhttps://github.com/fchollet/deep-learning-with-python-notebooks\n\nthis kernel is a quite a good one but does not use resnet:\n\nhttps://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 441450,
      "author_name": "richarde",
      "author_url": "",
      "post_date": "12/18/2018 17:10:23",
      "content": "<p>f chollet - the inventor of keras produced a very good book on keras.  It has many good tips of resnet and image augmentation, the github code for the book is here:\n<a href=\"https://github.com/fchollet/deep-learning-with-python-notebooks\">https://github.com/fchollet/deep-learning-with-python-notebooks</a></p>\n\n<p>this kernel is a quite a good one but does not use resnet:</p>\n\n<p><a href=\"https://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn\">https://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "441308": "I have been trying to ensemble multiple renset architectures in my last couple of submissions but didn't get much improvements. Can someone share some augmentation techniques or training techniques which can help get some ideas for improving model performance. \n\nI have tried training models on small images first and then using bigger images to improve model but it didn't get any higher score than 0.693 on leaderboard. \n\nSome observations that I would like to share:\n\n - Bigger images get better results\n - Test time augmentation helps\n - Smaller threshold helps if model is trained properly\n - No independent model was able to beat max public score\n\nIf anyone has some resources and would like to share it would be great way to learn more about how to handle such datasets for beginners like us.",
    "441450": "f chollet - the inventor of keras produced a very good book on keras.  It has many good tips of resnet and image augmentation, the github code for the book is here:\nhttps://github.com/fchollet/deep-learning-with-python-notebooks\n\nthis kernel is a quite a good one but does not use resnet:\n\nhttps://www.kaggle.com/gimunu/data-augmentation-with-keras-into-cnn"
  },
  "source": "meta"
}