Skip to content

Code for SIIM-ISIC Melanoma image Classification contest on Kaggle

Notifications You must be signed in to change notification settings

wessam611/Melanoma-Classification

Repository files navigation

Image Classification Competition on Kaggle for detecting Melanoma skin Cancer (my best score: pu: 0.9235, pr: 0.9124) notebook

Experience.

  • It's my first competition on Kaggle to take seriously. I've learned some about the kaggle community, enjoyed the experience of joing a live competition on kaggle while reading other's ideas from the discussion and notebooks section.
  • I've gained new knowledge through the competition including (AUC metric, EfficientNets and crossEntropy label smoothing).
  • I've experimented more with Tensorflow Dataset module and TFRecords and feel more comfortable with it now.

Experiments.

  • I've tried using different models including (ResNets 'overfits', VGG, DenseNets, Inception 'overfits' and EfficientNets 'works best').
  • After using EfficientNets I built a Ensemble model using weighted sum of 5 EfficientNets predictions (weighted sums didn't improve the result much from just using the average prediction).
  • I've oversampled positive examples in the dataset in every 2048 records parallel to training (chech older versions of the notebook) making the distribution of the training data (50% pos, 50% neg) which resulted in overfitting (Same picture is used many times with heavy augmentations).
  • I've settled on using a more balanced Dataset which increased the accuracy and auc of the model while preventing overfitting.

Notes.

  • I've had different problems with Memory and overfitting, so I'd really appreciate any advice reagrding my code and this competition in general.
  • Follow me on Kaggle

About

Code for SIIM-ISIC Melanoma image Classification contest on Kaggle

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published