DUCK-Net-Pytorch

Unofficial implementation of DUCK-Net using Pytorch

Official Repo: https://github.com/RazvanDu/DUCK-Net

Paper: here

Requirement:

Pytorch

DUCK-Net for 2D image segmentation tasks: duck_net.py

official DuckNet model

from duck_net import DuckNet
model = DuckNet(in_channels=3, 
                out_channels=1, 
                depth=5, 
                init_features=17, # 34
                normalization='batch', 
                interpolation='nearest', 
                out_activation='sigmoid', 
                use_multiplier=False)
model.apply(init_weights_with_kaiming_uniform) # default init is xaiver uniform

Personal modified DuckNet model

# @ init_features=16: should be a power of 2 for better performance
# @ use_multiplier=True: for numerical stability
# @ normalization=None: reduce GPU memory usage
# @ out_activation=None: faster convergence when using Dice loss
from duck_net import DuckNet
model = DuckNet(in_channels=3, 
                out_channels=1, 
                depth=5, 
                init_features=16, 
                normalization=None, # 'batch', 'instance'
                interpolation='nearest', # 'bilinear', 'bicubic'
                out_activation=None, # 'sigmoid', 'relu', 'softmax'
                use_multiplier=True)

DUCK-Net for 3D medical image segmentation tasks: duck_net_3d.py (Performance not tested)

Personal modified DuckNet3D model

# Personal modified DuckNet3D model: 
# @ init_features=16: should be a power of 2 for better performance
# @ depth=4: reduce depth for faster training and less GPU memory usage
# @ use_multiplier=True: for numerical stability
# @ normalization=None: reduce GPU memory usage
# @ out_activation=None: faster convergence when using Dice loss
from duck_net_3d import DuckNet3D
model = DuckNet3D(in_channels=1, 
                out_channels=1, 
                depth=4, 
                init_features=16, 
                normalization=None, # 'batch', 'instance'
                interpolation='nearest', # 'trilinear'
                out_activation=None, # 'sigmoid', 'relu', 'softmax'
                use_multiplier=True)