Home

Awesome

Collective Residual Networks

This repository contains the code and trained models of "Sharing Residual Units Through Collective Tensor Factorization in Deep Neural Networks".

Implementation

Augmentation

MethodSettings
Random MirrorTrue
Random Crop8% - 100%
Aspect Ratio3/4 - 4/3
Random HSL[20,40,50]

Note: We did not use PCA Lighting and any other advanced augmentation methods.

Normalization

The augmented input images are substrated by mean RGB = [ 124, 117, 104 ], and then multiplied by 0.0167.

Results

ImageNet-1k

Single crop validation error (center 224x224 crop from resized image with shorter side=256):

ModelSettingModel SizeTop-1
CRU-Net-56 @x1432x4d98MB21.9%
CRU-Net-56 @x14136x1d98MB21.7%
CRU-Net-116 @x28x1432x4d168MB20.6%
CRU-Net-116, wider @x28x1464x4d318MB20.3%

We also trained a tiny CRU-Net-56 with less than half the size of ResNet-50.

Single crop validation error (center 224x224 crop from resized image with shorter side=256):

ModelSettingModel SizeTop-1
CRU-Net-56,tiny @x1432x4d48MB22.9%

Places365-Standard

10-crop validation accuracy (averaging softmax scores of 10 224x224 crops from resized image with shorter side=256):

ModelSettingModel SizeTop-1
CRU-Net-116 @x28x1432x4d163MB56.6%

Trained Models

ModelSettingDatasetLink
CRU-Net-56,tiny @x1432x4dImageNet-1kGoogleDrive
CRU-Net-56 @x1432x4dImageNet-1kGoogleDrive
CRU-Net-56 @x14136x1dImageNet-1kGoogleDrive
CRU-Net-116 @x28x1432x4dImageNet-1kGoogleDrive
CRU-Net-116 @x28x1432x4dPlaces365-StandardGoogleDrive