Optimization of Image Classification for CIFAR-10 dataset

Experiment

• Defined a CNN for classification of CIFAR-10 dataset
• Studied the influence of data augmentation
• Studied the influence of 2 different optimization algorithms - sgd and adam
• Studied the influence of 3 loss functions - cross-entropy, hinge loss, hinge squared loss

Setup

I used google cloud platform, which I personally felt very user-friendly. To setup google instance and GPU, I followed the instruction from standford cs231n. I used NVIDIA Tesla K80 GPU. See reports for details. http://cs231n.github.io/gce-tutorial/ http://cs231n.github.io/gce-tutorial-gpus/

Libraries

• Keras 1.20 (Note newer version will cause errors.)
• Tensorflow

Code

I used jupyter notebook to write up all the codes instead of terminal because it's very convenient for debugging. It is very easy to follow through the codes since I explained the purpose of each block of code. You could read the codes from its corresponding html file. A python file also generated using the original jupyter notebook file.

• cse543-finalproject.ipynb
• cse543-finalproject.html
• cse543-finalproject.py