{
  "id": 315734,
  "title": "Divides the image into 16 sections and soft-voting the predictions.",
  "url": "/competitions/sorghum-id-fgvc-9/discussion/315734",
  "author_name": "bobfromjapan",
  "post_date": "2022-03-29T14:06:16.887000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Although the data for this competition is large (1024x1024), it seemed to me that only a portion of the image would be sufficient to make predictions.<br>\nTherefore, I used the method shown in this notebook to segment the images and registered them in the Kaggle dataset.<br>\n<a href=\"https://www.kaggle.com/bobfromjapan/divide-image-into-16\" target=\"_blank\">https://www.kaggle.com/bobfromjapan/divide-image-into-16</a></p>\n<p>The train images totaled 355,088 and became very time consuming to train.<br>\nHowever, this approach could be effective. I trained the EfficientNetV2B0 with simple image Augmentation for about 20 epochs. And inferred the test image, which was also divided into 16 parts, and aggregate the results with soft voting, and obtained a score of 0.701! I did!<br>\nI'm looking to try training with TPU's next!</p>\n<p>Train dataset can be available here. Any comments are welcome!<br>\n<a href=\"https://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings\" target=\"_blank\">https://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings</a></p>",
  "messages": [
    {
      "id": 1738764,
      "postDate": "2022-03-29T14:06:16.887Z",
      "content": "<p>Although the data for this competition is large (1024x1024), it seemed to me that only a portion of the image would be sufficient to make predictions.<br>\nTherefore, I used the method shown in this notebook to segment the images and registered them in the Kaggle dataset.<br>\n<a href=\"https://www.kaggle.com/bobfromjapan/divide-image-into-16\" target=\"_blank\">https://www.kaggle.com/bobfromjapan/divide-image-into-16</a></p>\n<p>The train images totaled 355,088 and became very time consuming to train.<br>\nHowever, this approach could be effective. I trained the EfficientNetV2B0 with simple image Augmentation for about 20 epochs. And inferred the test image, which was also divided into 16 parts, and aggregate the results with soft voting, and obtained a score of 0.701! I did!<br>\nI'm looking to try training with TPU's next!</p>\n<p>Train dataset can be available here. Any comments are welcome!<br>\n<a href=\"https://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings\" target=\"_blank\">https://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings</a></p>",
      "rawMarkdown": "Although the data for this competition is large (1024x1024), it seemed to me that only a portion of the image would be sufficient to make predictions.\nTherefore, I used the method shown in this notebook to segment the images and registered them in the Kaggle dataset.\nhttps://www.kaggle.com/bobfromjapan/divide-image-into-16\n\nThe train images totaled 355,088 and became very time consuming to train.\nHowever, this approach could be effective. I trained the EfficientNetV2B0 with simple image Augmentation for about 20 epochs. And inferred the test image, which was also divided into 16 parts, and aggregate the results with soft voting, and obtained a score of 0.701! I did!\nI'm looking to try training with TPU's next!\n\nTrain dataset can be available here. Any comments are welcome!\nhttps://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1738764": "Although the data for this competition is large (1024x1024), it seemed to me that only a portion of the image would be sufficient to make predictions.\nTherefore, I used the method shown in this notebook to segment the images and registered them in the Kaggle dataset.\nhttps://www.kaggle.com/bobfromjapan/divide-image-into-16\n\nThe train images totaled 355,088 and became very time consuming to train.\nHowever, this approach could be effective. I trained the EfficientNetV2B0 with simple image Augmentation for about 20 epochs. And inferred the test image, which was also divided into 16 parts, and aggregate the results with soft voting, and obtained a score of 0.701! I did!\nI'm looking to try training with TPU's next!\n\nTrain dataset can be available here. Any comments are welcome!\nhttps://www.kaggle.com/datasets/bobfromjapan/sorghum-partitioned-4x4/settings"
  }
}