{
  "id": 206034,
  "title": "It looks like a CNN as a more layered and generalized nonlinear Fourier transform filter.",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/206034",
  "author_name": "",
  "post_date": "2020-12-22T21:33:57.070030700Z",
  "votes": -1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, All.</p>\n<p>I cut a rectangular image in half and make it a square to learn and infer.<br>\nAlthough it is cumbersome because you have to do it twice during inference.</p>\n<p>In addition to object detection, the recognition rate of the square is better for classification problems.<br>\nThere is another advantage of cutting a rectangle in half and entering it as a square.<br>\nThat rotation augmentation is easy and effective..</p>\n<p>I personally think that FFT2 is faster and better at a size of 2^n, and that a square input to CNN performs better, and this is closely related to each other.</p>\n<p>In other words, it looks like a CNN as a more layered and generalized nonlinear Fourier transform filter (Gabor filter).</p>\n<p>What do you think?</p>",
  "messages": [
    {
      "id": "1123056",
      "postDate": "12/22/2020 21:33:57",
      "content": "<p>Hi, All.</p>\n<p>I cut a rectangular image in half and make it a square to learn and infer.<br>\nAlthough it is cumbersome because you have to do it twice during inference.</p>\n<p>In addition to object detection, the recognition rate of the square is better for classification problems.<br>\nThere is another advantage of cutting a rectangle in half and entering it as a square.<br>\nThat rotation augmentation is easy and effective..</p>\n<p>I personally think that FFT2 is faster and better at a size of 2^n, and that a square input to CNN performs better, and this is closely related to each other.</p>\n<p>In other words, it looks like a CNN as a more layered and generalized nonlinear Fourier transform filter (Gabor filter).</p>\n<p>What do you think?</p>",
      "rawMarkdown": "Hi, All.\n\nI cut a rectangular image in half and make it a square to learn and infer.\nAlthough it is cumbersome because you have to do it twice during inference.\n\nIn addition to object detection, the recognition rate of the square is better for classification problems.\nThere is another advantage of cutting a rectangle in half and entering it as a square.\nThat rotation augmentation is easy and effective..\n\nI personally think that FFT2 is faster and better at a size of 2^n, and that a square input to CNN performs better, and this is closely related to each other.\n\nIn other words, it looks like a CNN as a more layered and generalized nonlinear Fourier transform filter (Gabor filter).\n\nWhat do you think?",
      "votes": null
    },
    {
      "id": "1123078",
      "postDate": "12/22/2020 22:05:13",
      "content": "<p>Sayyy whatttt, I dont understand.. The cnn kernels are trained, they can be anything technically. Also they can accept squares and rectangles</p>",
      "rawMarkdown": "Sayyy whatttt, I dont understand.. The cnn kernels are trained, they can be anything technically. Also they can accept squares and rectangles",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1123078,
      "author_name": "yannmajewski",
      "author_url": "",
      "post_date": "12/22/2020 22:05:13",
      "content": "<p>Sayyy whatttt, I dont understand.. The cnn kernels are trained, they can be anything technically. Also they can accept squares and rectangles</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1123056": "Hi, All.\n\nI cut a rectangular image in half and make it a square to learn and infer.\nAlthough it is cumbersome because you have to do it twice during inference.\n\nIn addition to object detection, the recognition rate of the square is better for classification problems.\nThere is another advantage of cutting a rectangle in half and entering it as a square.\nThat rotation augmentation is easy and effective..\n\nI personally think that FFT2 is faster and better at a size of 2^n, and that a square input to CNN performs better, and this is closely related to each other.\n\nIn other words, it looks like a CNN as a more layered and generalized nonlinear Fourier transform filter (Gabor filter).\n\nWhat do you think?",
    "1123078": "Sayyy whatttt, I dont understand.. The cnn kernels are trained, they can be anything technically. Also they can accept squares and rectangles"
  },
  "source": "meta"
}