SnapMix-tensorflow2

A Tensorflow2.x implementation of SnapMix as described in SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data

Features

• mixup
• cutmix
• snapmix
• resnet50,resnet101
• efficientB0~B7
• warmup
• cosinedecay lr scheduler
• step lr scheduler
• concat-max-and-average-pool
• custom dataset training

Installation

1. Clone project

git clone https://github.com/wangermeng2021/SnapMix-tensorflow2.git
cd SnapMix-tensorflow2

2. Install environment

• install tesnorflow ( skip this step if it's already installed)

• pip install -r requirements.txt
  

3. Download dataset

• Download cub dataset

  wget http://www.vision.caltech.edu/visipedia-data/CUB-200-2011/CUB_200_2011.tgz -P dataset/
  tar -xzf dataset/CUB_200_2011.tgz -C dataset/
  mv dataset/CUB_200_2011 dataset/cub
  

• Download cars dataset

  wget http://imagenet.stanford.edu/internal/car196/cars_train.tgz -P dataset/cars/
  wget http://imagenet.stanford.edu/internal/car196/cars_test.tgz -P dataset/cars/
  wget https://ai.stanford.edu/~jkrause/cars/car_devkit.tgz -P dataset/cars/
  tar -xzf  dataset/cars/cars_train.tgz -C dataset/cars
  tar -xzf  dataset/cars/cars_test.tgz -C dataset/cars
  tar -xzf  dataset/cars/car_devkit.tgz -C dataset/cars
  wget http://imagenet.stanford.edu/internal/car196/cars_test_annos_withlabels.mat -P dataset/cars/devkit/
  

Training

• For training on cub dataset,use:

  python train.py --dataset cub --dataset-dir dataset/cub --model ResNet50 --augment snapmix
  

• For training on Cars dataset,use:

  python train.py --dataset cars --dataset-dir dataset/cars --model ResNet50 --augment snapmix
  

• For training on your custom dataset,use:

  python train.py --dataset custom --dataset-dir your_dataset_root_directory  --model ResNet50  --augment snapmix
  

  you can try it on a toy dataset(No need to download dataset,it's already included in project:dataset/cat_dog):

  python train.py --dataset custom --dataset-dir dataset/cat_dog --model ResNet50  --augment snapmix
  

  your_dataset_root_directory:
  train
          class1_name
                 xxx.jpg
                 xxx.jpg
                 ...
         class2_name
                 xxx.jpg
                 xxx.jpg
                 ...
         ...
  valid
         class1_name
                 xxx.jpg
                 xxx.jpg
                 ...
         class2_name
                 xxx.jpg
                 xxx.jpg
                 ...
         ...

Evaluation results(RTX2080,epochs=300,batch_size=16):

model	cat_dog	cars	cub
ResNet50+cutmix	0.958
ResNet50+snapmix	0.979
EfficientNetB0+mixup	0.968
EfficientNetB0+cutmix	0.979
EfficientNetB0+snapmix	0.979
EfficientNetB3+cutmix	0.958
EfficientNetB3+snapmix	1.0

References

• https://github.com/Shaoli-Huang/SnapMix
• SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data