wz4515.zip contains the baseline and the improved methods for EE3-25 Deep Learning course.
Different optimisers are suggested for different networks, as suggested in the scripts:
Use Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-8, decay=0.0, amsgrad=False) for training all U-Net denoising network.
Use RMSprop(lr=0.0001, rho=0.9, epsilon=None, decay=0.0) for training descriptor networks with margin and margin hnm losses.
Use Adam(lr=0.00001, beta_1=0.9, beta_2=0.999, epsilon=1e-8, decay=0.0, amsgrad=False) for training descriptor networks with ratio and ratio hnm losses.
Use RMSprop(lr=0.001, rho=0.9, epsilon=None, decay=0.0) for training HardNet.
This repository is subject to updates as further work is carried out on the coursework.