Kaggle

Nolsigan's code for kaggle competitions using Tensorflow & keras!

Only deals with problems that can be solved with deep learning.

Cat vs Dog

• simple_vgg
  • Uses simplified version of VGG
  • After epoch 8, train loss still decreases but validation loss doesn't (overfitting)
  • Achieved 85.65% accuracy

MNIST

• simple_nn.ipynb

  • Uses simple neural network model with two 128-neurons hidden layers
  • Used extern data from official MNIST site, gaining 33% more data than Kaggle provides.
  • Loss decreases to zero after 100 epochs. ( Maybe this model is enough for this data? )
  • Achieved 99.714% accuracy

• conv2d.ipynb

  • Uses CNN model of tensorflow with two convolutional layers and max pooling, a 512-neurons hidden layer
  • Training speed is way slower than simple_nn. Local machine had hard time just training for 10 epochs.
  • Loss is 0.05 after 10 epochs, but still achieves pretty good Result.
  • Achieved 99.38% accuracy

Titanic

• gender.ipynb ( Kaggle Tutorial )

  • Uses simple classification using only gender property.
  • Achieved 77.65% accuracy

• simple_nn.ipynb

  • Uses simple neural network model with two 128-neurons hidden layers
  • Result is worse than simple gender model. Not suitable for neural network model.
  • Achieved 66.96% accuracy