Skip to content
A TensorFlow implementation of Group Normalization on the task of image classification
Branch: master
Clone or download
shaohua0116 Merge pull request #8 from iyah4888/patch-1
Fixed the common errors by 'None' batch size and multiple uses
Latest commit 551511c Oct 9, 2018
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
datasets add tiny imagenet Mar 27, 2018
figure
.gitignore
LICENSE add training codes and the model Mar 25, 2018
README.md edit RAEDME Apr 16, 2018
__init__.py
download.py
input_ops.py
model.py
ops.py Fixed the common error by 'None' batch size and multiple uses Oct 9, 2018
trainer.py
util.py

README.md

Group Normalization

As part of the implementation series of Joseph Lim's group at USC, our motivation is to accelerate (or sometimes delay) research in the AI community by promoting open-source projects. To this end, we implement state-of-the-art research papers, and publicly share them with concise reports. Please visit our group github site for other projects.

This project is implemented by Shao-Hua Sun and the codes have been reviewed by Te-Lin Wu before being published.

Descriptions

This project includes a Tensorflow implementation of Group Normalizations proposed in the paper Group Normalization by Wu et al. Batch Normalization (BN) has been widely employed in the trainings of deep neural networks to alleviate the internal covariate shift [1].Specifically, BN aims to transform the inputs of each layer in such a way that they have a mean output activation of zero and standard deviation of one. While BN demonstrates it effectiveness in a variety of fields including computer vision, natural language processing, speech processing, robotics, etc., BN's performance substantially decrease when the training batch size become smaller, which limits the gain of utilizing BN in a task requiring small batches constrained by memory consumption.

Motivated by this phenomenon, the Group Normalization (GN) technique is proposed. Instead of normalizing along the batch dimension, GN divides the channels into groups and computes within each group the mean and variance. Therefore, GN’s computation is independent of batch sizes, and so does its accuracy. The experiment section of the paper demonstrates the effectiveness of GN in a wide range of visual tasks, which include image classification (ImageNet), object detection and segmentation (COCO), and video classification (Kinect). This repository is simply a toy repository for those who want to quickly test GN and compare it against BN.

The illustration from the original GN paper. Each cube represent a 4D tensor of feature maps. Note that the spatial dimension are combined as a single dimension for visualization. N denotes the batch axis, C denotes the channel axis, and H, W denote the spatial axes. The values in blue are normalized by the same mean and variance, computed by aggregating the values of these pixels.

Based on the implementation of this repository, GN is around 20% slower than BN on datasets such as CIFAR-10 and SVHN, which is probably because of the extra reshape and transpose operations. However, when the network goes deeper and the number of channels increase, GN gets even slower due to a larger group size. The model uses GN is around 4 times slower than the one uses BN when being trained ImageNet. This is not reported in the original GN paper.

*This code is still being developed and subject to change.

Prerequisites

Usage

Datasets

Train models on MNIST, Fashion MNIST, SVHN, CIFAR-10 datasets:

  • Download datasets with:
$ python download.py --dataset MNIST Fashion SVHN CIFAR10

Train models on Tiny ImageNet

  • Download the dataset from the webpage.
  • Move the downloaded file (named ) to datasets/tiny_imagenet and unzip it.

Train models on ImageNet

  • The ImageNet dataset is located in the Downloads section of the website. Please specify the path to the downloaded dataset by changing the variable __IMAGENET_IMG_PATH__ in datasets/ImageNet.py. Also, please provide a list of file names for trainings in the directory __IMAGENET_LIST_PATH__ with the file name train_list.txt. By default, the train_list.txt includes all the training images in ImageNet dataset.

Train models with downloaded datasets:

Specify the type of normalization you want to use by --norm_type batch or --norm_type group and specify the batch size with --batch_size BATCH_SIZE.

$ python trainer.py --dataset MNIST --learning_rate 1e-3
$ python trainer.py --dataset Fashion --prefix test
$ python trainer.py --dataset SVHN --batch_size 128
$ python trainer.py --dataset CIFAR10 

Train and test your own datasets:

  • Create a directory
$ mkdir datasets/YOUR_DATASET
  • Store your data as an h5py file datasets/YOUR_DATASET/data.hy and each data point contains
    • 'image': has shape [h, w, c], where c is the number of channels (grayscale images: 1, color images: 3)
    • 'label': represented as an one-hot vector
  • Maintain a list datasets/YOUR_DATASET/id.txt listing ids of all data points
  • Modify trainer.py including args, data_info, etc.
  • Finally, train and test models:
$ python trainer.py --dataset YOUR_DATASET
$ python evaler.py --dataset YOUR_DATASET

Results

CIFAR-10

Color Batch Size
Orange 1
Blue 2
Sky blue 4
Red 8
Green 16
Pink 32
  • Loss

  • Accuracy

SVHN

Color Batch Size
Pink 1
Blue 2
Sky blue 4
Green 8
Red 16
Orange 32
  • Loss

  • Accuracy

ImageNet

The trainings are ongoing...

Color Norm Type
Orange Group Normalization
Blue Batch Normalization

Conclusion

The Group Normalization divides the channels into groups and computes within each group the mean and variance, and therefore its performance independent of training batch sizes, which is verified with this implementation. However, the performance of Batch Normalization does not vary a lot with different batch sizes on smaller image datasets including CIFAR-10, SVHN, etc. The ImageNet experiments are ongoing and the results will be updated later.

Related works

Author

Shao-Hua Sun / @shaohua0116 @ Joseph Lim's research lab @ USC

You can’t perform that action at this time.