{
  "id": 170701,
  "title": "Hosts & kaggle Team: Is Sub  Using Test data Trained Model Allowed",
  "url": "/competitions/global-wheat-detection/discussion/170701",
  "author_name": "Jaideep",
  "post_date": "2020-07-28T16:20:44.835000",
  "votes": -1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi Host/Kaggle team\nCan you please clarify how pseudo labelling is allowed there are two ways\n1) Take only 10 image ids of Test data,predict the labels and include that test data small subset in the training\n2) Since inference time is very less , so take 69% of test data that is visible ,perform training on that and then use that test data trained model  to make final predictions for submission ( I think many people are doing this )  if yes then what is purpose of Test data in that case</p>",
  "messages": [
    {
      "id": 949424,
      "postDate": "2020-07-28T16:20:44.837Z",
      "content": "<p>Hi Host/Kaggle team\nCan you please clarify how pseudo labelling is allowed there are two ways\n1) Take only 10 image ids of Test data,predict the labels and include that test data small subset in the training\n2) Since inference time is very less , so take 69% of test data that is visible ,perform training on that and then use that test data trained model  to make final predictions for submission ( I think many people are doing this )  if yes then what is purpose of Test data in that case</p>",
      "rawMarkdown": "Hi Host/Kaggle team\nCan you please clarify how pseudo labelling is allowed there are two ways\n1) Take only 10 image ids of Test data,predict the labels and include that test data small subset in the training\n2) Since inference time is very less , so take 69% of test data that is visible ,perform training on that and then use that test data trained model  to make final predictions for submission ( I think many people are doing this )  if yes then what is purpose of Test data in that case\n\n",
      "votes": -1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "949424": "Hi Host/Kaggle team\nCan you please clarify how pseudo labelling is allowed there are two ways\n1) Take only 10 image ids of Test data,predict the labels and include that test data small subset in the training\n2) Since inference time is very less , so take 69% of test data that is visible ,perform training on that and then use that test data trained model  to make final predictions for submission ( I think many people are doing this )  if yes then what is purpose of Test data in that case\n\n"
  }
}