{
  "id": 70970,
  "title": "Data Augmentation",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/70970",
  "author_name": "Mario Parreño Lara",
  "post_date": "2018-11-08T21:52:26.843000",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi guys</p>\n\n<p>I am a bit new in the world of Kaggle and this is my first competition, do not be too hard :) The fact is that I have trained a ResNet34 model using less maps (I do not have many resources), 64x64 images and 256 batch (occupies 7.7Gb of 8Gb in a gtx1080 using Pytorch). The case is that I have tried to apply Data Augmentation, something normal in problems of computer vision and the results are a little worse than without using anything (I only normalize dividing between 255). I am using the library of albumentations (<a href=\"https://github.com/albu/albumentations\">https://github.com/albu/albumentations</a>) and in particular I have tried to use together the horizontalFlip and shifRotation with small values ... Someone could guide / help me what techniques have been used in Data Augmentation.</p>\n\n<p>Thank you very much and happy Kaggle.</p>",
  "messages": [
    {
      "id": 417854,
      "postDate": "2018-11-08T21:52:26.843Z",
      "content": "<p>Hi guys</p>\n\n<p>I am a bit new in the world of Kaggle and this is my first competition, do not be too hard :) The fact is that I have trained a ResNet34 model using less maps (I do not have many resources), 64x64 images and 256 batch (occupies 7.7Gb of 8Gb in a gtx1080 using Pytorch). The case is that I have tried to apply Data Augmentation, something normal in problems of computer vision and the results are a little worse than without using anything (I only normalize dividing between 255). I am using the library of albumentations (<a href=\"https://github.com/albu/albumentations\">https://github.com/albu/albumentations</a>) and in particular I have tried to use together the horizontalFlip and shifRotation with small values ... Someone could guide / help me what techniques have been used in Data Augmentation.</p>\n\n<p>Thank you very much and happy Kaggle.</p>",
      "rawMarkdown": "Hi guys\n\nI am a bit new in the world of Kaggle and this is my first competition, do not be too hard :) The fact is that I have trained a ResNet34 model using less maps (I do not have many resources), 64x64 images and 256 batch (occupies 7.7Gb of 8Gb in a gtx1080 using Pytorch). The case is that I have tried to apply Data Augmentation, something normal in problems of computer vision and the results are a little worse than without using anything (I only normalize dividing between 255). I am using the library of albumentations (https://github.com/albu/albumentations) and in particular I have tried to use together the horizontalFlip and shifRotation with small values ... Someone could guide / help me what techniques have been used in Data Augmentation.\n\nThank you very much and happy Kaggle.",
      "votes": 1
    },
    {
      "id": 418250,
      "postDate": "2018-11-09T14:36:02.573Z",
      "content": "<p>I have not tried data augmentation yet for this contest because we have almost unlimited data (50 million images). Am I wrong in thinking that it is better to use more of the real data than use augmented data?</p>",
      "rawMarkdown": "I have not tried data augmentation yet for this contest because we have almost unlimited data (50 million images). Am I wrong in thinking that it is better to use more of the real data than use augmented data?"
    }
  ],
  "comments": [
    {
      "id": 418250,
      "author_name": "impulsecorp",
      "author_url": "",
      "post_date": "2018-11-09T14:36:02.573000",
      "content": "<p>I have not tried data augmentation yet for this contest because we have almost unlimited data (50 million images). Am I wrong in thinking that it is better to use more of the real data than use augmented data?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "417854": "Hi guys\n\nI am a bit new in the world of Kaggle and this is my first competition, do not be too hard :) The fact is that I have trained a ResNet34 model using less maps (I do not have many resources), 64x64 images and 256 batch (occupies 7.7Gb of 8Gb in a gtx1080 using Pytorch). The case is that I have tried to apply Data Augmentation, something normal in problems of computer vision and the results are a little worse than without using anything (I only normalize dividing between 255). I am using the library of albumentations (https://github.com/albu/albumentations) and in particular I have tried to use together the horizontalFlip and shifRotation with small values ... Someone could guide / help me what techniques have been used in Data Augmentation.\n\nThank you very much and happy Kaggle.",
    "418250": "I have not tried data augmentation yet for this contest because we have almost unlimited data (50 million images). Am I wrong in thinking that it is better to use more of the real data than use augmented data?"
  }
}