Author Tianyi Zhao
This work is inspired by Kaggle Playground Prediction Competition: Dog Breed Identification (https://www.kaggle.com/c/dog-breed-identification)
Contents of Stanford Dogs dataset:
Number of categories: 120
Number of images: 20,580
Annotations: Class labels, Bounding boxes
Training set: Contains 12,000 images, 120 categories, 100 images for each category.
Test Set: Contains 8,580 images, 120 categories.
Use VGG19 which is pre-trained on ImageNet and transfer this model for our dog breed classification problem.
vgg_trainable.py: Construction of VGG19 network
Demo.py: Training VGG19
Test_Visualization.ipynb: Get visualization of our model.
Test.ipynb: Get result visualization on test set.
batch size = 64
Initial learning rate 0.001, decay step = 100, decay rate = 0.9
Activation function: ReLu
Dropout: 0.5