{
  "id": 372628,
  "title": "playground code for loss and imabalance data",
  "url": "/competitions/rsna-breast-cancer-detection/discussion/372628",
  "author_name": "hengck23",
  "post_date": "2022-12-17T03:23:53.188000",
  "votes": 15,
  "comment_count": 1,
  "views": 0,
  "content": "<p><img src=\"https://i.ibb.co/f07N2y4/Selection-212.png\" alt=\"https://i.ibb.co/f07N2y4/Selection-212.png\"></p>\n<p>see attachment for code and results</p>\n<p>note:</p>\n<ul>\n<li>be careful to interpret results (you should ask: is kaggle dataset distribution same as the dummy data ?)</li>\n<li>run different seeds, different rate, different loss, different sampling, etc …</li>\n<li>try different models (weak, strong, very strong, over strong), how about freezing (fixed, random, etc) layers?</li>\n<li>try regularization loss</li>\n<li>try different data distribution and noise (separable and non separable)</li>\n<li>compute different metric (you can visualize : value of metric value = how predicted decision space looks like) </li>\n<li>try mode of failures </li>\n<li>we are free to use this code to create public kernel, etc</li>\n</ul>\n<p>have fun</p>",
  "messages": [
    {
      "id": 2067696,
      "postDate": "2022-12-17T03:23:53.190Z",
      "content": "<p><img src=\"https://i.ibb.co/f07N2y4/Selection-212.png\" alt=\"https://i.ibb.co/f07N2y4/Selection-212.png\"></p>\n<p>see attachment for code and results</p>\n<p>note:</p>\n<ul>\n<li>be careful to interpret results (you should ask: is kaggle dataset distribution same as the dummy data ?)</li>\n<li>run different seeds, different rate, different loss, different sampling, etc …</li>\n<li>try different models (weak, strong, very strong, over strong), how about freezing (fixed, random, etc) layers?</li>\n<li>try regularization loss</li>\n<li>try different data distribution and noise (separable and non separable)</li>\n<li>compute different metric (you can visualize : value of metric value = how predicted decision space looks like) </li>\n<li>try mode of failures </li>\n<li>we are free to use this code to create public kernel, etc</li>\n</ul>\n<p>have fun</p>",
      "rawMarkdown": "![https://i.ibb.co/f07N2y4/Selection-212.png](https://i.ibb.co/f07N2y4/Selection-212.png)\n\nsee attachment for code and results\n\nnote:\n- be careful to interpret results (you should ask: is kaggle dataset distribution same as the dummy data ?)\n- run different seeds, different rate, different loss, different sampling, etc ...\n- try different models (weak, strong, very strong, over strong), how about freezing (fixed, random, etc) layers?\n- try regularization loss\n- try different data distribution and noise (separable and non separable)\n- compute different metric (you can visualize : value of metric value = how predicted decision space looks like) \n- try mode of failures \n- we are free to use this code to create public kernel, etc\n\nhave fun\n ",
      "votes": 15
    },
    {
      "id": 2067697,
      "postDate": "2022-12-17T03:27:40.907Z",
      "content": "<p>i highlight here the effects of different learning rate</p>\n<p><a href=\"https://ibb.co/pZTfdqb\"><img src=\"https://i.ibb.co/J2TvHZm/bce-loss-oversample-pos-x-low-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000999\"></a><br>\n<a href=\"https://ibb.co/vmrYHX9\"><img src=\"https://i.ibb.co/J2bQK3Z/bce-loss-oversample-pos-x-low-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000300\"></a><br>\n<a href=\"https://ibb.co/k4NWKSm\"><img src=\"https://i.ibb.co/GWzb75H/bce-loss-oversample-pos-x-high-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000999\"></a><br>\n<a href=\"https://ibb.co/4TQQz5q\"><img src=\"https://i.ibb.co/2tbbxf9/bce-loss-oversample-pos-x-high-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000300\"></a></p>",
      "rawMarkdown": "i highlight here the effects of different learning rate\n\n<a href=\"https://ibb.co/pZTfdqb\"><img src=\"https://i.ibb.co/J2TvHZm/bce-loss-oversample-pos-x-low-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000999\" border=\"0\"></a>\n<a href=\"https://ibb.co/vmrYHX9\"><img src=\"https://i.ibb.co/J2bQK3Z/bce-loss-oversample-pos-x-low-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000300\" border=\"0\"></a>\n<a href=\"https://ibb.co/k4NWKSm\"><img src=\"https://i.ibb.co/GWzb75H/bce-loss-oversample-pos-x-high-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000999\" border=\"0\"></a>\n<a href=\"https://ibb.co/4TQQz5q\"><img src=\"https://i.ibb.co/2tbbxf9/bce-loss-oversample-pos-x-high-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000300\" border=\"0\"></a>\n",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 2067697,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2022-12-17T03:27:40.907000",
      "content": "<p>i highlight here the effects of different learning rate</p>\n<p><a href=\"https://ibb.co/pZTfdqb\"><img src=\"https://i.ibb.co/J2TvHZm/bce-loss-oversample-pos-x-low-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000999\"></a><br>\n<a href=\"https://ibb.co/vmrYHX9\"><img src=\"https://i.ibb.co/J2bQK3Z/bce-loss-oversample-pos-x-low-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000300\"></a><br>\n<a href=\"https://ibb.co/k4NWKSm\"><img src=\"https://i.ibb.co/GWzb75H/bce-loss-oversample-pos-x-high-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000999\"></a><br>\n<a href=\"https://ibb.co/4TQQz5q\"><img src=\"https://i.ibb.co/2tbbxf9/bce-loss-oversample-pos-x-high-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000300\"></a></p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2067696": "![https://i.ibb.co/f07N2y4/Selection-212.png](https://i.ibb.co/f07N2y4/Selection-212.png)\n\nsee attachment for code and results\n\nnote:\n- be careful to interpret results (you should ask: is kaggle dataset distribution same as the dummy data ?)\n- run different seeds, different rate, different loss, different sampling, etc ...\n- try different models (weak, strong, very strong, over strong), how about freezing (fixed, random, etc) layers?\n- try regularization loss\n- try different data distribution and noise (separable and non separable)\n- compute different metric (you can visualize : value of metric value = how predicted decision space looks like) \n- try mode of failures \n- we are free to use this code to create public kernel, etc\n\nhave fun\n ",
    "2067697": "i highlight here the effects of different learning rate\n\n<a href=\"https://ibb.co/pZTfdqb\"><img src=\"https://i.ibb.co/J2TvHZm/bce-loss-oversample-pos-x-low-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000999\" border=\"0\"></a>\n<a href=\"https://ibb.co/vmrYHX9\"><img src=\"https://i.ibb.co/J2bQK3Z/bce-loss-oversample-pos-x-low-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-low-learn-rate-000300\" border=\"0\"></a>\n<a href=\"https://ibb.co/k4NWKSm\"><img src=\"https://i.ibb.co/GWzb75H/bce-loss-oversample-pos-x-high-learn-rate-000999.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000999\" border=\"0\"></a>\n<a href=\"https://ibb.co/4TQQz5q\"><img src=\"https://i.ibb.co/2tbbxf9/bce-loss-oversample-pos-x-high-learn-rate-000300.png\" alt=\"bce-loss-oversample-pos-x-high-learn-rate-000300\" border=\"0\"></a>\n"
  }
}