{
  "id": 128436,
  "title": "MixMatch SSL with CutMix and MixMatch using fai",
  "url": "/competitions/bengaliai-cv19/discussion/128436",
  "author_name": "Morris",
  "post_date": "2020-01-31T10:40:59.284000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello, \nI basically adapted <a href=\"https://github.com/oguiza/fastai_extensions/blob/master/04a_MixMatch_extended.ipynb\">Oguiza's FastAI Mixmatch implementation</a> and some code by @machinelp  with np.random.rand()&lt;0.5 MixUp/CutMix regularization.\nThe idea being, that this Competitions challenge is all about regularization, hence all the CutOut/MixUp/GirdMask Magic.\nI only tested it once yesterday: \n- CV 96%\n- 10 epochs only\n- ResNext50\n- 80/20random split\n- 140k unlabeled and 20k labeled samples (+40k validation).\n- 224x224 1 channel</p>\n\n<p>Code: <a href=\"https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524\">https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524</a>\nUsage instructions at the bottom.</p>\n\n<p>Some Parts of the code are a bit iffy, due to hardcoded fp16, but can be adjusted easily.</p>\n\n<p>I'm thinking with more train epochs/other architectures like SE-resnext50/GridMask/fine-tune on fully labeled data this could be interesting</p>",
  "messages": [
    {
      "id": 733607,
      "postDate": "2020-01-31T10:40:59.283Z",
      "content": "<p>Hello, \nI basically adapted <a href=\"https://github.com/oguiza/fastai_extensions/blob/master/04a_MixMatch_extended.ipynb\">Oguiza's FastAI Mixmatch implementation</a> and some code by @machinelp  with np.random.rand()&lt;0.5 MixUp/CutMix regularization.\nThe idea being, that this Competitions challenge is all about regularization, hence all the CutOut/MixUp/GirdMask Magic.\nI only tested it once yesterday: \n- CV 96%\n- 10 epochs only\n- ResNext50\n- 80/20random split\n- 140k unlabeled and 20k labeled samples (+40k validation).\n- 224x224 1 channel</p>\n\n<p>Code: <a href=\"https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524\">https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524</a>\nUsage instructions at the bottom.</p>\n\n<p>Some Parts of the code are a bit iffy, due to hardcoded fp16, but can be adjusted easily.</p>\n\n<p>I'm thinking with more train epochs/other architectures like SE-resnext50/GridMask/fine-tune on fully labeled data this could be interesting</p>",
      "rawMarkdown": "Hello, \nI basically adapted [Oguiza's FastAI Mixmatch implementation](https://github.com/oguiza/fastai_extensions/blob/master/04a_MixMatch_extended.ipynb) and some code by @machinelp  with np.random.rand()&lt;0.5 MixUp/CutMix regularization.\nThe idea being, that this Competitions challenge is all about regularization, hence all the CutOut/MixUp/GirdMask Magic.\nI only tested it once yesterday: \n- CV 96%\n- 10 epochs only\n- ResNext50\n- 80/20random split\n- 140k unlabeled and 20k labeled samples (+40k validation).\n- 224x224 1 channel\n\nCode: https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524\nUsage instructions at the bottom.\n\nSome Parts of the code are a bit iffy, due to hardcoded fp16, but can be adjusted easily.\n\nI'm thinking with more train epochs/other architectures like SE-resnext50/GridMask/fine-tune on fully labeled data this could be interesting",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "733607": "Hello, \nI basically adapted [Oguiza's FastAI Mixmatch implementation](https://github.com/oguiza/fastai_extensions/blob/master/04a_MixMatch_extended.ipynb) and some code by @machinelp  with np.random.rand()&lt;0.5 MixUp/CutMix regularization.\nThe idea being, that this Competitions challenge is all about regularization, hence all the CutOut/MixUp/GirdMask Magic.\nI only tested it once yesterday: \n- CV 96%\n- 10 epochs only\n- ResNext50\n- 80/20random split\n- 140k unlabeled and 20k labeled samples (+40k validation).\n- 224x224 1 channel\n\nCode: https://gist.github.com/mlk1337/d6f62a76dabbc404c2fbe3da6033a524\nUsage instructions at the bottom.\n\nSome Parts of the code are a bit iffy, due to hardcoded fp16, but can be adjusted easily.\n\nI'm thinking with more train epochs/other architectures like SE-resnext50/GridMask/fine-tune on fully labeled data this could be interesting"
  }
}