{
  "id": 543294,
  "title": "How to check if I'm overfitting",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/543294",
  "author_name": "TnkChaseMe",
  "post_date": "2024-10-29T17:33:50.447000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello everyone,<br>\nI utilized <a href=\"https://www.kaggle.com/cdeotte\" target=\"_blank\">@cdeotte</a> dataset fold 5 with a resolution of 512x512 for my training, validation, and testing processes. Specifically, I trained using Fold 0 and validated with Fold 1. My backbone architecture is EfficientNetB4, enhanced with batch normalization, dropout, and a ReLU activation function to adapt my classes to binary classification (Mel or Non-Mel).<br>\nI achieved an unexpectedly high F1 score of approximately 0.98 on my test set (comprising Folds 2, 3, and 4) and even incorporated the ISSIC 2019 dataset for additional testing. It’s quite remarkable that by applying focal loss and augmenting the minority class, I was able to attain such impressive results despite working with a relatively small dataset.<br>\nTo further assess whether my model is genuinely overfitting, what steps should I take?</p>",
  "messages": [
    {
      "id": 3031444,
      "postDate": "2024-10-29T17:33:50.447Z",
      "content": "<p>Hello everyone,<br>\nI utilized <a href=\"https://www.kaggle.com/cdeotte\" target=\"_blank\">@cdeotte</a> dataset fold 5 with a resolution of 512x512 for my training, validation, and testing processes. Specifically, I trained using Fold 0 and validated with Fold 1. My backbone architecture is EfficientNetB4, enhanced with batch normalization, dropout, and a ReLU activation function to adapt my classes to binary classification (Mel or Non-Mel).<br>\nI achieved an unexpectedly high F1 score of approximately 0.98 on my test set (comprising Folds 2, 3, and 4) and even incorporated the ISSIC 2019 dataset for additional testing. It’s quite remarkable that by applying focal loss and augmenting the minority class, I was able to attain such impressive results despite working with a relatively small dataset.<br>\nTo further assess whether my model is genuinely overfitting, what steps should I take?</p>",
      "rawMarkdown": "Hello everyone,\nI utilized @cdeotte dataset fold 5 with a resolution of 512x512 for my training, validation, and testing processes. Specifically, I trained using Fold 0 and validated with Fold 1. My backbone architecture is EfficientNetB4, enhanced with batch normalization, dropout, and a ReLU activation function to adapt my classes to binary classification (Mel or Non-Mel).\nI achieved an unexpectedly high F1 score of approximately 0.98 on my test set (comprising Folds 2, 3, and 4) and even incorporated the ISSIC 2019 dataset for additional testing. It’s quite remarkable that by applying focal loss and augmenting the minority class, I was able to attain such impressive results despite working with a relatively small dataset.\nTo further assess whether my model is genuinely overfitting, what steps should I take?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3031444": "Hello everyone,\nI utilized @cdeotte dataset fold 5 with a resolution of 512x512 for my training, validation, and testing processes. Specifically, I trained using Fold 0 and validated with Fold 1. My backbone architecture is EfficientNetB4, enhanced with batch normalization, dropout, and a ReLU activation function to adapt my classes to binary classification (Mel or Non-Mel).\nI achieved an unexpectedly high F1 score of approximately 0.98 on my test set (comprising Folds 2, 3, and 4) and even incorporated the ISSIC 2019 dataset for additional testing. It’s quite remarkable that by applying focal loss and augmenting the minority class, I was able to attain such impressive results despite working with a relatively small dataset.\nTo further assess whether my model is genuinely overfitting, what steps should I take?"
  }
}