Object Classification using AlexNet and Asirra cats vs dogs

Collaborator

Hyunjae Gil: neural net construction, train parameter setting \ Chan Lee: dataset loading, neural net tuning, training on server, documentation

Introduction

Object Classification code with neural network model of AlexNet and Asirra cats vs dogs as dataset

ImageNet dataset has way far large file (need more than 300GB), so that Asirra dataset was used

Brief Explanation

assira.py for data loading, asirra_cnn.py for training, neural net construction in function_used.py

Load Data: assira.py

Load the Asirra Dogs vs. Cats data subset from disk and perform preprocessing for training AlexNet.\ Such as resizing, random crop for data augmentation, labeling

Construct Neural Net: function_used.py

keras has used for implementing neural network

Five convolution layers and three fully connected layers, Three pooling layers, two normalization layers and one dropout layer

Train & Val : asirra_cnn.py

1. Load augmented dataset
2. Set training hyper parameters

   learning_rate = 0.001
   training_epoch = 5
   batch_size = 64
   display_step = 20

3. Build Model, initialize a session and start training
4. Save the validation result with png file