{
  "id": 70762,
  "title": "possible speedup training",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/70762",
  "author_name": "hengck23",
  "post_date": "2018-11-07T04:38:13.803000",
  "votes": 13,
  "comment_count": 5,
  "views": 0,
  "content": "<p>training</p>\n\n<ol>\n<li><p>start with 64x64. use trained classifier to divide the train set into correct and wrong (e.g. based on threshold etc)</p></li>\n<li><p>use 64x64 classifier to initialise 128x128. train only on wrong images of 128x128\nnote. trainset.1 and trainset.2 should overlap a little</p></li>\n<li><p>repeat for 256x256, 320x320 , etc ,...</p></li>\n<li><p>finally train a fuse classifier that use features extracted from 64,128, ...  on all train mages</p></li>\n</ol>",
  "messages": [
    {
      "id": 416679,
      "postDate": "2018-11-07T04:38:13.803Z",
      "content": "<p>training</p>\n\n<ol>\n<li><p>start with 64x64. use trained classifier to divide the train set into correct and wrong (e.g. based on threshold etc)</p></li>\n<li><p>use 64x64 classifier to initialise 128x128. train only on wrong images of 128x128\nnote. trainset.1 and trainset.2 should overlap a little</p></li>\n<li><p>repeat for 256x256, 320x320 , etc ,...</p></li>\n<li><p>finally train a fuse classifier that use features extracted from 64,128, ...  on all train mages</p></li>\n</ol>",
      "rawMarkdown": "training\n\n1. start with 64x64. use trained classifier to divide the train set into correct and wrong (e.g. based on threshold etc)\n\n2. use 64x64 classifier to initialise 128x128. train only on wrong images of 128x128\n    note. trainset.1 and trainset.2 should overlap a little\n\n3. repeat for 256x256, 320x320 , etc ,...\n\n4. finally train a fuse classifier that use features extracted from 64,128, ...  on all train mages",
      "votes": 13
    },
    {
      "id": 416810,
      "postDate": "2018-11-07T10:00:31.793Z",
      "content": "<p>One better way, keep the input with same:</p>\n\n<ol>\n<li>Train model with 1k per class.</li>\n<li>Finetune the model with 10k per class.</li>\n<li>Finetune the model with 50k per class.</li>\n<li>Finetune the model with all data.</li>\n</ol>",
      "rawMarkdown": "One better way, keep the input with same:\n\n1. Train model with 1k per class.\n2. Finetune the model with 10k per class.\n3. Finetune the model with 50k per class.\n4. Finetune the model with all data.",
      "votes": 4,
      "replies": [
        {
          "id": 417569,
          "postDate": "2018-11-08T13:20:01.570Z",
          "content": "<p>For example like this?</p>\n\n<pre><code>data_amounts = [\"1k\", \"10k\", \"50k\"]\nepochs_per_amounts = [50, 30, 20]\ndata_train_per_amounts = [40000, 75000, 100000] # Data per epoch\nstart_lr_amounts = [0.1, 0.085, 0.065]\n</code></pre>",
          "rawMarkdown": "For example like this?\n\n    data_amounts = [\"1k\", \"10k\", \"50k\"]\n    epochs_per_amounts = [50, 30, 20]\n    data_train_per_amounts = [40000, 75000, 100000] # Data per epoch\n    start_lr_amounts = [0.1, 0.085, 0.065]"
        },
        {
          "id": 417886,
          "postDate": "2018-11-08T23:51:39.303Z",
          "content": "<p>Yes. like this.</p>",
          "rawMarkdown": "Yes. like this."
        }
      ]
    },
    {
      "id": 416710,
      "postDate": "2018-11-07T06:23:29.270Z",
      "content": "<p>Good idea.</p>",
      "rawMarkdown": "Good idea."
    },
    {
      "id": 423659,
      "postDate": "2018-11-18T20:01:01.910Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 416810,
      "author_name": "Finlay",
      "author_url": "",
      "post_date": "2018-11-07T10:00:31.793000",
      "content": "<p>One better way, keep the input with same:</p>\n\n<ol>\n<li>Train model with 1k per class.</li>\n<li>Finetune the model with 10k per class.</li>\n<li>Finetune the model with 50k per class.</li>\n<li>Finetune the model with all data.</li>\n</ol>",
      "votes": 4,
      "replies": [
        {
          "id": 417569,
          "author_name": "Mario Parreño Lara",
          "author_url": "",
          "post_date": "2018-11-08T13:20:01.570000",
          "content": "<p>For example like this?</p>\n\n<pre><code>data_amounts = [\"1k\", \"10k\", \"50k\"]\nepochs_per_amounts = [50, 30, 20]\ndata_train_per_amounts = [40000, 75000, 100000] # Data per epoch\nstart_lr_amounts = [0.1, 0.085, 0.065]\n</code></pre>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 417886,
          "author_name": "Finlay",
          "author_url": "",
          "post_date": "2018-11-08T23:51:39.303000",
          "content": "<p>Yes. like this.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 416710,
      "author_name": "good good study",
      "author_url": "",
      "post_date": "2018-11-07T06:23:29.270000",
      "content": "<p>Good idea.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 423659,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-11-18T20:01:01.910000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "416679": "training\n\n1. start with 64x64. use trained classifier to divide the train set into correct and wrong (e.g. based on threshold etc)\n\n2. use 64x64 classifier to initialise 128x128. train only on wrong images of 128x128\n    note. trainset.1 and trainset.2 should overlap a little\n\n3. repeat for 256x256, 320x320 , etc ,...\n\n4. finally train a fuse classifier that use features extracted from 64,128, ...  on all train mages",
    "416810": "One better way, keep the input with same:\n\n1. Train model with 1k per class.\n2. Finetune the model with 10k per class.\n3. Finetune the model with 50k per class.\n4. Finetune the model with all data.",
    "416710": "Good idea.",
    "423659": ""
  }
}