{
  "id": 71826,
  "title": "Some confusion about the pre-trained model",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/71826",
  "author_name": "Gary",
  "post_date": "2018-11-17T05:38:10.201000",
  "votes": 2,
  "comment_count": 7,
  "views": 0,
  "content": "<p>I have tried <strong>resnet50</strong> with all data <strong>(96*96)</strong> in pytorch, but I only got \n<strong>Local val: top1-acc = 0.820, top3-acc = 0.937, MAP@3 = 0.878. Public LB=0.929(without TTA).</strong> \nCan you try to help me find out the possible problems? Thank you.</p>",
  "messages": [
    {
      "id": 422938,
      "postDate": "2018-11-17T05:38:10.200Z",
      "content": "<p>I have tried <strong>resnet50</strong> with all data <strong>(96*96)</strong> in pytorch, but I only got \n<strong>Local val: top1-acc = 0.820, top3-acc = 0.937, MAP@3 = 0.878. Public LB=0.929(without TTA).</strong> \nCan you try to help me find out the possible problems? Thank you.</p>",
      "rawMarkdown": "I have tried **resnet50** with all data **(96*96)** in pytorch, but I only got \n**Local val: top1-acc = 0.820, top3-acc = 0.937, MAP@3 = 0.878. Public LB=0.929(without TTA).** \nCan you try to help me find out the possible problems? Thank you.",
      "votes": 2
    },
    {
      "id": 423256,
      "postDate": "2018-11-17T19:37:42.527Z",
      "content": "<p>I am wondering what is your learning rate schedule now?</p>",
      "rawMarkdown": "I am wondering what is your learning rate schedule now?"
    },
    {
      "id": 423094,
      "postDate": "2018-11-17T13:13:06.797Z",
      "content": "<blockquote>\n  <p>but I only got\n  Only?</p>\n</blockquote>",
      "rawMarkdown": "&gt;but I only got\nOnly?",
      "replies": [
        {
          "id": 423097,
          "postDate": "2018-11-17T13:22:11.437Z",
          "content": "<p>Oh, Sorry, If you don't like the way I speak. I was a little disappointed because it wasn't what I expected.</p>",
          "rawMarkdown": "Oh, Sorry, If you don't like the way I speak. I was a little disappointed because it wasn't what I expected."
        },
        {
          "id": 423250,
          "postDate": "2018-11-17T19:27:27.530Z",
          "content": "<p>It's ok, I didn't mean that. Just wanted to say that you're doing better than other 970 teams :)</p>",
          "rawMarkdown": "It's ok, I didn't mean that. Just wanted to say that you're doing better than other 970 teams :)",
          "votes": 1
        }
      ]
    },
    {
      "id": 422946,
      "postDate": "2018-11-17T05:47:15.323Z",
      "content": "<p>If you have enough computer resource you can try 128*128*3 image and Bigger batch_size\nDid you use the keras?</p>",
      "rawMarkdown": "If you have enough computer resource you can try 128*128*3 image and Bigger batch_size\nDid you use the keras?",
      "replies": [
        {
          "id": 422949,
          "postDate": "2018-11-17T05:53:21.320Z",
          "content": "<p>pytorch. But I think even 96*96 should be enough to get 0.93+ publick LB score without TTA.</p>",
          "rawMarkdown": "pytorch. But I think even 96*96 should be enough to get 0.93+ publick LB score without TTA."
        },
        {
          "id": 422972,
          "postDate": "2018-11-17T06:47:39.540Z",
          "content": "<p>you can try some hyperparameters and lr_schedule and snopshot\nI think those trick is useful to the public score below 0.930</p>",
          "rawMarkdown": "you can try some hyperparameters and lr_schedule and snopshot\nI think those trick is useful to the public score below 0.930"
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 423256,
      "author_name": "Strideradu",
      "author_url": "",
      "post_date": "2018-11-17T19:37:42.527000",
      "content": "<p>I am wondering what is your learning rate schedule now?</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 423094,
      "author_name": "Artyom Palvelev",
      "author_url": "",
      "post_date": "2018-11-17T13:13:06.797000",
      "content": "<blockquote>\n  <p>but I only got\n  Only?</p>\n</blockquote>",
      "votes": 0,
      "replies": [
        {
          "id": 423097,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-17T13:22:11.437000",
          "content": "<p>Oh, Sorry, If you don't like the way I speak. I was a little disappointed because it wasn't what I expected.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 423250,
          "author_name": "Artyom Palvelev",
          "author_url": "",
          "post_date": "2018-11-17T19:27:27.530000",
          "content": "<p>It's ok, I didn't mean that. Just wanted to say that you're doing better than other 970 teams :)</p>",
          "votes": 1,
          "replies": []
        }
      ]
    },
    {
      "id": 422946,
      "author_name": "TPloveYXT520",
      "author_url": "",
      "post_date": "2018-11-17T05:47:15.323000",
      "content": "<p>If you have enough computer resource you can try 128*128*3 image and Bigger batch_size\nDid you use the keras?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 422949,
          "author_name": "Gary",
          "author_url": "",
          "post_date": "2018-11-17T05:53:21.320000",
          "content": "<p>pytorch. But I think even 96*96 should be enough to get 0.93+ publick LB score without TTA.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 422972,
          "author_name": "TPloveYXT520",
          "author_url": "",
          "post_date": "2018-11-17T06:47:39.540000",
          "content": "<p>you can try some hyperparameters and lr_schedule and snopshot\nI think those trick is useful to the public score below 0.930</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "422938": "I have tried **resnet50** with all data **(96*96)** in pytorch, but I only got \n**Local val: top1-acc = 0.820, top3-acc = 0.937, MAP@3 = 0.878. Public LB=0.929(without TTA).** \nCan you try to help me find out the possible problems? Thank you.",
    "423256": "I am wondering what is your learning rate schedule now?",
    "423094": "&gt;but I only got\nOnly?",
    "422946": "If you have enough computer resource you can try 128*128*3 image and Bigger batch_size\nDid you use the keras?"
  }
}