Dog vs Cats

My experiments with the Kaggle competition Dogs vs Cats Redux. Project is structured as follows:

• dataset.py: handles preprocessing and queueing of input data
• bottleneck.py: handles preprocessing and queueing of Inception v4 bottlenecks
• tfutil.py: a collection of helper functions for building, training and evaluating TensorFlow networks
• DogsVsCats_Intro.ipynb: introduction, control and basic fully connected network (~66% on test set)
• DogsVsCats_Conv.ipynb: convolutional network
• DogsVsCats_Inception.ipynb: transfer learning with Google Inception v4
  • inception_v4.py, inception_utils.py: code to build Inception v4 network with TF-Slim, from TF-Slim models page

Setup

To work with this project, the data must be obtained from Kaggle and placed in data/raw. For transfer learning the Inception v4 checkpoint must also be downloaded and extracted into the project root.