This small project just run CNN for more practical experience and compare with different usage of CNN (and different features, but not the point).
-
Extract features from VGG19 at fc1 layer (4096), and train it with random forest
-
Fine tune VGG19 (using rmsprop first and then using SGD for stable descent)
-
Use external data like wiews, favorate ... may fail in testing stage (they directly gave photo without link).