{
  "id": 78789,
  "title": "Batch size during testing",
  "url": "/competitions/quora-insincere-questions-classification/discussion/78789",
  "author_name": "mhiro2",
  "post_date": "2019-01-27T21:25:58.371000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I found that many public kernels use the same batch size during training and testing. In order to reduce validation and test time, we probably should use more larger batch size in validation and test such as 2048, 4096 and so on. (depend on your model)</p>\n\n<p>But the prediction speed is very fast compared to training (include backprop), the test batch size have little impact.</p>",
  "messages": [
    {
      "id": 462206,
      "postDate": "2019-01-27T21:25:58.370Z",
      "content": "<p>I found that many public kernels use the same batch size during training and testing. In order to reduce validation and test time, we probably should use more larger batch size in validation and test such as 2048, 4096 and so on. (depend on your model)</p>\n\n<p>But the prediction speed is very fast compared to training (include backprop), the test batch size have little impact.</p>",
      "rawMarkdown": "I found that many public kernels use the same batch size during training and testing. In order to reduce validation and test time, we probably should use more larger batch size in validation and test such as 2048, 4096 and so on. (depend on your model)\n\nBut the prediction speed is very fast compared to training (include backprop), the test batch size have little impact.",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "462206": "I found that many public kernels use the same batch size during training and testing. In order to reduce validation and test time, we probably should use more larger batch size in validation and test such as 2048, 4096 and so on. (depend on your model)\n\nBut the prediction speed is very fast compared to training (include backprop), the test batch size have little impact."
  }
}