Cifar Classifier

Experiment

Changing the loss function for a given a known good convolutional neural network and evaluating it's performance.

Setup

For setting up a new AWS g2.x2 node running ubuntu 14.04, assuming CUDA already linked

virtualenv env
source env/bin/activate
pip install --upgrade pip
pip install keras
pip install Pillow
pip install matplotlib
python cifar_classifier.py

Main pieces of interest

• cifar_classifier.py : Contains all of the experiment code including loading data, the model, data generation

• plot*.py : Matplotlib wrapper + keras callback for generating accuracy and loss plots per epoch of training

Libraries used

• Keras
• Theano
• Matplotlib
• Pillow

Experiments ran on AWS g2.x2 node with K520 GPU with ~270 seconds per epoch.