{
  "id": 225704,
  "title": "Yolov5 img hyperparameter changes improve the score by 0.033",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/225704",
  "author_name": "Mostafa Ibrahim",
  "post_date": "2021-03-13T14:06:52.623000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I can't understand this to be honest. So if you train Yolov5 with the  default config on 512x512 images with the  --img flag as 512 on the train.py and detect.py command, u get a final score of 0.123. However, if you use  the same exact code, model, and 512 images but you change  the --img flag to 640 (although the  images are 512), u  get a final score of 0.156 (+0.033), this is a significant improvement. I original thought that the size on that --img flag must be exactly the same as the images, but I guess not. I have tried to look for an explanation of that  flag but I couldn't find one that gives more information than its just  the  image size. Can someone explain please? And does  the model actually  perform better or is this just some sort of overfitting?</p>",
  "messages": [
    {
      "id": 1236850,
      "postDate": "2021-03-13T14:06:52.623Z",
      "content": "<p>I can't understand this to be honest. So if you train Yolov5 with the  default config on 512x512 images with the  --img flag as 512 on the train.py and detect.py command, u get a final score of 0.123. However, if you use  the same exact code, model, and 512 images but you change  the --img flag to 640 (although the  images are 512), u  get a final score of 0.156 (+0.033), this is a significant improvement. I original thought that the size on that --img flag must be exactly the same as the images, but I guess not. I have tried to look for an explanation of that  flag but I couldn't find one that gives more information than its just  the  image size. Can someone explain please? And does  the model actually  perform better or is this just some sort of overfitting?</p>",
      "rawMarkdown": "I can't understand this to be honest. So if you train Yolov5 with the  default config on 512x512 images with the  --img flag as 512 on the train.py and detect.py command, u get a final score of 0.123. However, if you use  the same exact code, model, and 512 images but you change  the --img flag to 640 (although the  images are 512), u  get a final score of 0.156 (+0.033), this is a significant improvement. I original thought that the size on that --img flag must be exactly the same as the images, but I guess not. I have tried to look for an explanation of that  flag but I couldn't find one that gives more information than its just  the  image size. Can someone explain please? And does  the model actually  perform better or is this just some sort of overfitting?",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1236850": "I can't understand this to be honest. So if you train Yolov5 with the  default config on 512x512 images with the  --img flag as 512 on the train.py and detect.py command, u get a final score of 0.123. However, if you use  the same exact code, model, and 512 images but you change  the --img flag to 640 (although the  images are 512), u  get a final score of 0.156 (+0.033), this is a significant improvement. I original thought that the size on that --img flag must be exactly the same as the images, but I guess not. I have tried to look for an explanation of that  flag but I couldn't find one that gives more information than its just  the  image size. Can someone explain please? And does  the model actually  perform better or is this just some sort of overfitting?"
  }
}