{
  "id": 71173,
  "title": "What optimizer do you prefer?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71173",
  "author_name": "beluga",
  "post_date": "2018-11-11T02:57:11.681000",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I haven't started to use the big guns mostly just played with smaller (64x64, 96x96) MobileNets. Adam seems to work better for me than RMSprop or SGD...</p>",
  "messages": [
    {
      "id": 418983,
      "postDate": "2018-11-11T02:57:11.680Z",
      "content": "<p>I haven't started to use the big guns mostly just played with smaller (64x64, 96x96) MobileNets. Adam seems to work better for me than RMSprop or SGD...</p>",
      "rawMarkdown": "I haven't started to use the big guns mostly just played with smaller (64x64, 96x96) MobileNets. Adam seems to work better for me than RMSprop or SGD...",
      "votes": 5
    },
    {
      "id": 419762,
      "postDate": "2018-11-12T14:21:16.543Z",
      "content": "<p>Tried to train the same model with Adam and SGD. Got approximately same results on CV and decided not to try on LB.</p>",
      "rawMarkdown": "Tried to train the same model with Adam and SGD. Got approximately same results on CV and decided not to try on LB."
    },
    {
      "id": 419426,
      "postDate": "2018-11-12T00:06:56.080Z",
      "content": "<p>In my case, SGD seems to find better optima than Adam does, still experimenting on the hyperparams though.. </p>",
      "rawMarkdown": "In my case, SGD seems to find better optima than Adam does, still experimenting on the hyperparams though.. "
    },
    {
      "id": 419054,
      "postDate": "2018-11-11T06:30:41.070Z",
      "content": "<p>I also have better scores with Adam than SGD. Although, correct me if I am wrong but I think SGD converges slowly and might require more epochs to get a better score. Also, anyone tried cyclicLR? If yes, does it help with the scores?</p>",
      "rawMarkdown": "I also have better scores with Adam than SGD. Although, correct me if I am wrong but I think SGD converges slowly and might require more epochs to get a better score. Also, anyone tried cyclicLR? If yes, does it help with the scores?"
    }
  ],
  "comments": [
    {
      "id": 419762,
      "author_name": "Oleg Panichev",
      "author_url": "",
      "post_date": "2018-11-12T14:21:16.543000",
      "content": "<p>Tried to train the same model with Adam and SGD. Got approximately same results on CV and decided not to try on LB.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 419426,
      "author_name": "Brian Lee",
      "author_url": "",
      "post_date": "2018-11-12T00:06:56.080000",
      "content": "<p>In my case, SGD seems to find better optima than Adam does, still experimenting on the hyperparams though.. </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 419054,
      "author_name": "Ankit Sati",
      "author_url": "",
      "post_date": "2018-11-11T06:30:41.070000",
      "content": "<p>I also have better scores with Adam than SGD. Although, correct me if I am wrong but I think SGD converges slowly and might require more epochs to get a better score. Also, anyone tried cyclicLR? If yes, does it help with the scores?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "418983": "I haven't started to use the big guns mostly just played with smaller (64x64, 96x96) MobileNets. Adam seems to work better for me than RMSprop or SGD...",
    "419762": "Tried to train the same model with Adam and SGD. Got approximately same results on CV and decided not to try on LB.",
    "419426": "In my case, SGD seems to find better optima than Adam does, still experimenting on the hyperparams though.. ",
    "419054": "I also have better scores with Adam than SGD. Although, correct me if I am wrong but I think SGD converges slowly and might require more epochs to get a better score. Also, anyone tried cyclicLR? If yes, does it help with the scores?"
  }
}