{
  "id": 163902,
  "title": "one idea to train efficient-b7 on large 512x512 image",
  "url": "/competitions/alaska2-image-steganalysis/discussion/163902",
  "author_name": "hengck23",
  "post_date": "2020-07-03T22:46:33.941000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>pretraining:\n- crop 512x512 into smaller size, e.g. 128x128 or 64x64\n- if you are using 128x128, then your feature map size before the fc layer is e.g. 4x4.\n- if can get the label of the crop or each of the 4x4 region by counting the number of changed pixels (or changed DCT coefficients) in the  cover-stegano pair\n- for me , i pretrained the deep network to label each of the 4x4 feature block (instead of the  whole crop), but i think either way will work.</p>\n\n<p>finetuning (lower learning rate)\n- now train with full 512x512 images. freeze the bottom layers to feed in more images. \n- finally here, we use image level label.</p>\n\n<p>you may repeat the two steps for better results. This is the same trick for training deep net for large segmentation problem. </p>",
  "messages": [
    {
      "id": 914458,
      "postDate": "2020-07-03T22:46:33.940Z",
      "content": "<p>pretraining:\n- crop 512x512 into smaller size, e.g. 128x128 or 64x64\n- if you are using 128x128, then your feature map size before the fc layer is e.g. 4x4.\n- if can get the label of the crop or each of the 4x4 region by counting the number of changed pixels (or changed DCT coefficients) in the  cover-stegano pair\n- for me , i pretrained the deep network to label each of the 4x4 feature block (instead of the  whole crop), but i think either way will work.</p>\n\n<p>finetuning (lower learning rate)\n- now train with full 512x512 images. freeze the bottom layers to feed in more images. \n- finally here, we use image level label.</p>\n\n<p>you may repeat the two steps for better results. This is the same trick for training deep net for large segmentation problem. </p>",
      "rawMarkdown": "pretraining:\n- crop 512x512 into smaller size, e.g. 128x128 or 64x64\n- if you are using 128x128, then your feature map size before the fc layer is e.g. 4x4.\n- if can get the label of the crop or each of the 4x4 region by counting the number of changed pixels (or changed DCT coefficients) in the  cover-stegano pair\n- for me , i pretrained the deep network to label each of the 4x4 feature block (instead of the  whole crop), but i think either way will work.\n\nfinetuning (lower learning rate)\n- now train with full 512x512 images. freeze the bottom layers to feed in more images. \n- finally here, we use image level label.\n\nyou may repeat the two steps for better results. This is the same trick for training deep net for large segmentation problem. ",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "914458": "pretraining:\n- crop 512x512 into smaller size, e.g. 128x128 or 64x64\n- if you are using 128x128, then your feature map size before the fc layer is e.g. 4x4.\n- if can get the label of the crop or each of the 4x4 region by counting the number of changed pixels (or changed DCT coefficients) in the  cover-stegano pair\n- for me , i pretrained the deep network to label each of the 4x4 feature block (instead of the  whole crop), but i think either way will work.\n\nfinetuning (lower learning rate)\n- now train with full 512x512 images. freeze the bottom layers to feed in more images. \n- finally here, we use image level label.\n\nyou may repeat the two steps for better results. This is the same trick for training deep net for large segmentation problem. "
  }
}