SoftwareGift / GroupNorm-reproduce

A collection of code in different frameworks that reproduces experiments in "Group Normalization"

A collection of code in different frameworks, that reproduces several experiment results in the paper:

• Group Normalization - Best Paper Honorable Mention at ECCV 2018.

There are a number of implementations of this paper online, however it's unsafe to trust any implementation (including your own) without reproducing results.

Official Code

• Object Detection with Mask R-CNN, in Caffe2.

Unofficial Code

• ImageNet Classification with ResNet50, in TensorFlow. Exactly reproduce.

• ImageNet Classification with ResNet50, in PyTorch. 0.1% different result due to fewer augmentations.

• ImageNet Classification with VGG16, in TensorFlow. Exactly reproduce.

• Object Detection with Mask R-CNN, in TensorFlow. Slighly different results due to different implementation of Mask R-CNN framework.

Not in the Paper

• Cifar10 Classification with ResNet, in TensorFlow.

Implementations of the GroupNorm operation:

• PyTorch: Implemented in C++, and available as a layer via torch.nn.GroupNorm.
• TensorFlow: Python implementation.
• Caffe2: Implemented in C++, available in Python via brew.spatial_gn.