{
  "id": 153198,
  "title": " I can't balance the data correctly",
  "url": "/competitions/diabetic-retinopathy-detection/discussion/153198",
  "author_name": "Carlos de la Barrera Perez",
  "post_date": "2020-05-23T15:48:55.372000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I've am starting to try and get a good score with the data set.</p>\n\n<p>The first thing I realize is that the classes are ubalanced.</p>\n\n<p>So i tried this.</p>\n\n<p>Balance the data using 6000 images per class, filling class 1, 2, 3 and 4 with augmented images. If i do this my train improves like 1% per epoch, and then starts to overfit</p>\n\n<p>Balance the data using the scrips you can see in the jypyter books here, so my csv get oversampled, but if do this my score keeps in 20% every epoch</p>\n\n<p>I tried almost any solution posible: diferent learning rates, diferent optimizers, diferent number of layers and neurons but it always the same. Im pretty desesperated with this.</p>\n\n<p>Im using alex net now, but i also tried with Xception, Inception, VGG16, VGG19 and nothing.</p>\n\n<p>what can i do? i keep thinking that its fault of the unbalanced set and im not balancing it the right way. What can i do?</p>",
  "messages": [
    {
      "id": 858576,
      "postDate": "2020-05-23T15:48:55.373Z",
      "content": "<p>I've am starting to try and get a good score with the data set.</p>\n\n<p>The first thing I realize is that the classes are ubalanced.</p>\n\n<p>So i tried this.</p>\n\n<p>Balance the data using 6000 images per class, filling class 1, 2, 3 and 4 with augmented images. If i do this my train improves like 1% per epoch, and then starts to overfit</p>\n\n<p>Balance the data using the scrips you can see in the jypyter books here, so my csv get oversampled, but if do this my score keeps in 20% every epoch</p>\n\n<p>I tried almost any solution posible: diferent learning rates, diferent optimizers, diferent number of layers and neurons but it always the same. Im pretty desesperated with this.</p>\n\n<p>Im using alex net now, but i also tried with Xception, Inception, VGG16, VGG19 and nothing.</p>\n\n<p>what can i do? i keep thinking that its fault of the unbalanced set and im not balancing it the right way. What can i do?</p>",
      "rawMarkdown": "\nI've am starting to try and get a good score with the data set.\n\nThe first thing I realize is that the classes are ubalanced.\n\nSo i tried this.\n\nBalance the data using 6000 images per class, filling class 1, 2, 3 and 4 with augmented images. If i do this my train improves like 1% per epoch, and then starts to overfit\n\nBalance the data using the scrips you can see in the jypyter books here, so my csv get oversampled, but if do this my score keeps in 20% every epoch\n\nI tried almost any solution posible: diferent learning rates, diferent optimizers, diferent number of layers and neurons but it always the same. Im pretty desesperated with this.\n\nIm using alex net now, but i also tried with Xception, Inception, VGG16, VGG19 and nothing.\n\nwhat can i do? i keep thinking that its fault of the unbalanced set and im not balancing it the right way. What can i do?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "858576": "\nI've am starting to try and get a good score with the data set.\n\nThe first thing I realize is that the classes are ubalanced.\n\nSo i tried this.\n\nBalance the data using 6000 images per class, filling class 1, 2, 3 and 4 with augmented images. If i do this my train improves like 1% per epoch, and then starts to overfit\n\nBalance the data using the scrips you can see in the jypyter books here, so my csv get oversampled, but if do this my score keeps in 20% every epoch\n\nI tried almost any solution posible: diferent learning rates, diferent optimizers, diferent number of layers and neurons but it always the same. Im pretty desesperated with this.\n\nIm using alex net now, but i also tried with Xception, Inception, VGG16, VGG19 and nothing.\n\nwhat can i do? i keep thinking that its fault of the unbalanced set and im not balancing it the right way. What can i do?"
  }
}