{
  "id": 71267,
  "title": "Deeper is Beeter?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71267",
  "author_name": "Finlay",
  "post_date": "2018-11-12T04:32:27.958000",
  "votes": 5,
  "comment_count": 10,
  "views": 0,
  "content": "<p>As we all know, deeper model is more powerful.\nBut deeper model require more gpu space(and the batchsize is small), and it is slow.</p>",
  "messages": [
    {
      "id": 419501,
      "postDate": "2018-11-12T04:32:27.960Z",
      "content": "<p>As we all know, deeper model is more powerful.\nBut deeper model require more gpu space(and the batchsize is small), and it is slow.</p>",
      "rawMarkdown": "As we all know, deeper model is more powerful.\nBut deeper model require more gpu space(and the batchsize is small), and it is slow.",
      "votes": 5
    },
    {
      "id": 424406,
      "postDate": "2018-11-20T04:41:13.500Z",
      "content": "<p>So far this stands for me: NASnet &gt; SE-ResNeXt101 &gt; Xception &gt; SE-ResNeXt50 </p>",
      "rawMarkdown": "So far this stands for me: NASnet &gt; SE-ResNeXt101 &gt; Xception &gt; SE-ResNeXt50 ",
      "votes": 1,
      "replies": [
        {
          "id": 425324,
          "postDate": "2018-11-21T12:50:43.353Z",
          "content": "<p>@ Alexander Liao</p>\n\n<p>I was not able to find very good implementations for SE-ResNeXt. Can you point me to the reference code,  you are using or else did you build it on your own?</p>",
          "rawMarkdown": "@ Alexander Liao\n\nI was not able to find very good implementations for SE-ResNeXt. Can you point me to the reference code,  you are using or else did you build it on your own?"
        },
        {
          "id": 425406,
          "postDate": "2018-11-21T14:59:57.757Z",
          "content": "<p>search pytorch pretrained on github</p>",
          "rawMarkdown": "search pytorch pretrained on github"
        }
      ]
    },
    {
      "id": 419614,
      "postDate": "2018-11-12T09:45:11.337Z",
      "content": "<p>Yes, Depper is Better :-)</p>",
      "rawMarkdown": "Yes, Depper is Better :-)",
      "votes": 1,
      "replies": [
        {
          "id": 419628,
          "postDate": "2018-11-12T10:13:41.350Z",
          "content": "<p>But Deeper is slower, Hah..</p>",
          "rawMarkdown": "But Deeper is slower, Hah.."
        },
        {
          "id": 419631,
          "postDate": "2018-11-12T10:29:30.723Z",
          "content": "<p>So more hardware i guess :v</p>",
          "rawMarkdown": "So more hardware i guess :v"
        }
      ]
    },
    {
      "id": 420101,
      "postDate": "2018-11-13T04:01:33.903Z",
      "content": "<p>Did any actually get better performance with bigger models? I tried Mobilenet and InceptionV3 with similar optimizers and batch-size. Inception V3 performed worse for me. \nI do not understand what might be the reason for it. I was expecting InceptionV3 might try to overfit due to a larger capacity, but it saturated early in training accuracy too.\nCan someone help me to bring order to the universe again? :)</p>",
      "rawMarkdown": "Did any actually get better performance with bigger models? I tried Mobilenet and InceptionV3 with similar optimizers and batch-size. Inception V3 performed worse for me. \nI do not understand what might be the reason for it. I was expecting InceptionV3 might try to overfit due to a larger capacity, but it saturated early in training accuracy too.\nCan someone help me to bring order to the universe again? :)",
      "replies": [
        {
          "id": 421345,
          "postDate": "2018-11-14T23:37:59.983Z",
          "content": "<p>I try Resnet18/34/50, yes deeper is better.</p>",
          "rawMarkdown": "I try Resnet18/34/50, yes deeper is better.",
          "votes": 3
        },
        {
          "id": 421468,
          "postDate": "2018-11-15T03:39:06.590Z",
          "content": "<p>@ Finlay Liu \nWhat is the image size and sample/class for the above experiments ?</p>",
          "rawMarkdown": "@ Finlay Liu \nWhat is the image size and sample/class for the above experiments ?"
        },
        {
          "id": 421481,
          "postDate": "2018-11-15T03:55:18.433Z",
          "content": "<p>input 224/256/299, i use all the dataset.</p>",
          "rawMarkdown": "input 224/256/299, i use all the dataset."
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 424406,
      "author_name": "Peiyuan Liao",
      "author_url": "",
      "post_date": "2018-11-20T04:41:13.500000",
      "content": "<p>So far this stands for me: NASnet &gt; SE-ResNeXt101 &gt; Xception &gt; SE-ResNeXt50 </p>",
      "votes": 1,
      "replies": [
        {
          "id": 425324,
          "author_name": "remidi",
          "author_url": "",
          "post_date": "2018-11-21T12:50:43.353000",
          "content": "<p>@ Alexander Liao</p>\n\n<p>I was not able to find very good implementations for SE-ResNeXt. Can you point me to the reference code,  you are using or else did you build it on your own?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 425406,
          "author_name": "Strideradu",
          "author_url": "",
          "post_date": "2018-11-21T14:59:57.757000",
          "content": "<p>search pytorch pretrained on github</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 419614,
      "author_name": "Lukasz Grad",
      "author_url": "",
      "post_date": "2018-11-12T09:45:11.337000",
      "content": "<p>Yes, Depper is Better :-)</p>",
      "votes": 1,
      "replies": [
        {
          "id": 419628,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-12T10:13:41.350000",
          "content": "<p>But Deeper is slower, Hah..</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 419631,
          "author_name": "Hoàng Tùng Lâm",
          "author_url": "",
          "post_date": "2018-11-12T10:29:30.723000",
          "content": "<p>So more hardware i guess :v</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 420101,
      "author_name": "ankdesh",
      "author_url": "",
      "post_date": "2018-11-13T04:01:33.903000",
      "content": "<p>Did any actually get better performance with bigger models? I tried Mobilenet and InceptionV3 with similar optimizers and batch-size. Inception V3 performed worse for me. \nI do not understand what might be the reason for it. I was expecting InceptionV3 might try to overfit due to a larger capacity, but it saturated early in training accuracy too.\nCan someone help me to bring order to the universe again? :)</p>",
      "votes": 0,
      "replies": [
        {
          "id": 421345,
          "author_name": "Finlay",
          "author_url": "",
          "post_date": "2018-11-14T23:37:59.983000",
          "content": "<p>I try Resnet18/34/50, yes deeper is better.</p>",
          "votes": 3,
          "replies": []
        },
        {
          "id": 421468,
          "author_name": "remidi",
          "author_url": "",
          "post_date": "2018-11-15T03:39:06.590000",
          "content": "<p>@ Finlay Liu \nWhat is the image size and sample/class for the above experiments ?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 421481,
          "author_name": "Finlay",
          "author_url": "",
          "post_date": "2018-11-15T03:55:18.433000",
          "content": "<p>input 224/256/299, i use all the dataset.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "419501": "As we all know, deeper model is more powerful.\nBut deeper model require more gpu space(and the batchsize is small), and it is slow.",
    "424406": "So far this stands for me: NASnet &gt; SE-ResNeXt101 &gt; Xception &gt; SE-ResNeXt50 ",
    "419614": "Yes, Depper is Better :-)",
    "420101": "Did any actually get better performance with bigger models? I tried Mobilenet and InceptionV3 with similar optimizers and batch-size. Inception V3 performed worse for me. \nI do not understand what might be the reason for it. I was expecting InceptionV3 might try to overfit due to a larger capacity, but it saturated early in training accuracy too.\nCan someone help me to bring order to the universe again? :)"
  }
}