Dogs vs cats is a competition hosted on Kaggle, the aim is for the participants to write an algorithm for computers to classify images containing dogs or cats.

The dataset is too large to fit into memory all at once, so it is first wrote into a HDF5 file as described in @pyimagesearch Deep Learning for Computer Vision with Python book. Images are first scaled down (maintaining the aspect ratio), then a 256x256 patch from the centre is cropped. The training dataset from Kaggle is split into 3 sets: 20000 for training, 2500 for validation and 2500 for testing.

The AlexNet model is implemented with Tensorflow for this project. The model was trained for 200 epochs and it completed in ~8hours. Images are randomly augmented during training to reduce overfitting, my implementation is slow thus the long duration of training. An 88% accuracy is achieved on the testing set (this result is very different from the Keras implementation reported in the Deep Learning for Computer Vision with Python book, any ideas?)

utils: folder contatining Python files to preprocess images.
build_h5_dataset.ipynb: write the images and labels into a HDF5 file.
train_alexnet.ipynb: implementing AlexNet in Tensorflow and training