Skip to content

flukeskywalker/highway-resnet-comparison

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Weighted Skip Connections are Not Harmful for Deep Nets

Comparison of 110-layer ResNets and HighwayNets on CIFAR-10, following Identity Mappings in Deep Residual Networks.

TL;DR:

The paper Identity Mappings in Deep Residual Networks has design mistakes leading to incorrect conclusions about training deep networks with gated skip connections. You should try gated/weighted skip connections yourself and see if they improve results on your problems.

See this accompanying blog post for details.

Requirement: The only requirements are pytorch and torchvision (for CIFAR10). Original results used pytorch 2.5.1.

pip install torch torchvision

To train the ResNet110 baseline:

python -u trainer.py --arch=resnet110 --amp --lr 0.1

To train Highway110:

python -u trainer.py --arch=highway110 --amp --wm 0.875

About

Weighted Skip Connections are Not Harmful for Deep Nets

Topics

Resources

License

Stars

Watchers

Forks

Languages