Skip to content

Latest commit

 

History

History
249 lines (175 loc) · 7.08 KB

README.md

File metadata and controls

249 lines (175 loc) · 7.08 KB
Raw

Randomness_Analysis for Data Augmentation

python pytorch lightning hydra

📌  Introduction

Randomness_Analysis for Data Augmentation [Paper] (Will be updated)


✅ This repo is about to data Augmentation containing cutmix, mixup, RandAugment ...

✅ You can do experiments with model ResNet50, ResNet101, WideResNet_28x10.

📌 How to use

1. Environments

# Create Conda Env
conda create -n myrand python=3.8

# Activate myrand
conda activate myrand

# cuda version
conda install cudatoolkit=11.3 -c pytorch

2. Data Preparation

Download Tiny-imagenet

# Make directory "data"
mkdir ./lightning-hydra-template/data

# unzip Tinyimagenet
unzip ./lightning-hydra-template/data/tiny-imagenet-200.zip
- Note that if you use Code related to CIFAR-100, then download the data automatically from torchvision.datasets

3. Setting Augmentations

If you want to use various Augmentations that we provide, you should change argument named "aug" at data_config and model_config. Note that these are options about augmentations.

  • [baseline, mixup, cutmix, randour_cutmix, cutmix_randour, randaug_mixup, randaug_cutmix, cutmix_randaug]

4. Training

- Note that these are options about {model_config_name}.
1. cifar100_resnet50
2. cifar100_resnet101
3. cifar100_wideresnet

4. tinyimagenet_resnet50
5. tinyimagenet_resnet101
6. tinyimagenet_wideresnet
- Note that these are options about {data_config_name}.
1. cifar100
2. tinyimagenet

command

# train on CPU
python train.py trainer=cpu model={model_config_name} data={data_config_name}

# train on 1 GPU
python train.py trainer=gpu model={model_config_name} data={data_config_name}

# train with DDP (Distributed Data Parallel) (4 GPUs)
python train.py trainer=ddp trainer.devices=4 model={model_config_name} data={data_config_name}

📌 Our results

1. Base Augmentation

TinyImagenet

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.6662 TinyImagenet Baseline
Wide_Resnet_28x10 0.6686 TinyImagenet Cutout
Wide_Resnet_28x10 0.6889 TinyImagenet mixup
Wide_Resnet_28x10 0.7057 TinyImagenet Cutmix
Wide_Resnet_28x10 0.687 TinyImagenet Randaugment

CIFAR-100

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.7943 Cifar100 Baseline
Wide_Resnet_28x10 0.8096 Cifar100 Cutout
Wide_Resnet_28x10 0.8232 Cifar100 mixup
Wide_Resnet_28x10 0.8237 Cifar100 Cutmix
Wide_Resnet_28x10 0.8089 Cifar100 Randaugment

2. RandAugment + mixup, cutmix

TinyImagenet

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.7041 TinyImagenet Randaugment → mixup
Wide_Resnet_28x10 0.716 TinyImagenet Randaugment → cutmix
Wide_Resnet_28x10 0.6889 TinyImagenet mixup
Wide_Resnet_28x10 0.7057 TinyImagenet cutmix
Resnet50 0.5871 TinyImagenet Randaugment → mixup
Resnet50 0.639 TinyImagenet Randaugment → cutmix
Resnet50 0.5298 TinyImagenet mixup
Resnet50 0.5768 TinyImagenet cutmix
Resnet101 0.5912 TinyImagenet Randaugment → mixup
Resnet101 0.6362 TinyImagenet Randaugment → cutmix
Resnet101 0.5789 TinyImagenet mixup
Resnet101 0.5285 TinyImagenet cutmix

CIFAR-100

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.8325 Cifar100 Randaugment + mixup
Wide_Resnet_28x10 0.8358 Cifar100 Randaugment + cutmix
Wide_Resnet_28x10 0.8232 Cifar100 mixup
Wide_Resnet_28x10 0.8237 Cifar100 cutmix
Resnet50 0.6485 Cifar100 Randaugment + mixup
Resnet50 0.6803 Cifar100 Randaugment + cutmix
Resnet50 0.6262 Cifar100 mixup
Resnet50 0.6453 Cifar100 cutmix
Resnet101 0.6431 Cifar100 Randaugment + mixup
Resnet101 0.6596 Cifar100 Randaugment + cutmix
Resnet101 0.619 Cifar100 mixup
Resnet101 0.6382 Cifar100 cutmix

3. RandAugment vs RandOur

TinyImagenet

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.7163 TinyImagenet Randaugment → cutmix
Wide_Resnet_28x10 0.7201 TinyImagenet Randours → cutmix
Resnet50 0.639 TinyImagenet Randaugment → cutmix
Resnet50 0.5989 TinyImagenet Randours → cutmix
Resnet101 0.6362 TinyImagenet Randaugment → cutmix
Resnet101 0.6 TinyImagenet Randours → cutmix

CIFAR-100

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.8358 Cifar100 Randaugment → cutmix
Wide_Resnet_28x10 0.8432 Cifar100 Randours → cutmix
Resnet50 0.685 Cifar100 Randaugment → cutmix
Resnet50 0.6536 Cifar100 Randours → cutmix
Resnet101 0.6596 Cifar100 Randaugment → cutmix
Resnet101 0.6517 Cifar100 Randours → cutmix

4. Randomness + Cutmix vs Cutmix + Randomness

TinyImagenet

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.7163 TinyImagenet Randaugment → cutmix
Wide_Resnet_28x10 0.7015 TinyImagenet cutmix → Randaugment
Resnet50 0.639 TinyImagenet Randaugment → cutmix
Resnet50 0.5807 TinyImagenet cutmix → Randaugment
Resnet101 0.6362 TinyImagenet Randaugment → cutmix
Resnet101 0.5835 TinyImagenet cutmix → Randaugment

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.7201 TinyImagenet Randours → cutmix
Wide_Resnet_28x10 0.7173 TinyImagenet cutmix → Randours
Resnet50 0.5989 TinyImagenet Randours → cutmix
Resnet50 0.5979 TinyImagenet cutmix → Randours
Resnet101 0.6 TinyImagenet Randours → cutmix
Resnet101 0.5967 TinyImagenet cutmix → Randours

CIFAR-100

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.8358 Cifar100 Randaugment → cutmix
Wide_Resnet_28x10 0.8253 Cifar100 cutmix → Randaugment
Resnet50 0.685 Cifar100 Randaugment → cutmix
Resnet50 0.6141 Cifar100 cutmix → Randaugment
Resnet101 0.6596 Cifar100 Randaugment → cutmix
Resnet101 0.6117 Cifar100 cutmix → Randaugment

Model Accuracy Dataset Augmentation
Wide_Resnet_28x10 0.843 Cifar100 Randours → cutmix
Wide_Resnet_28x10 0.84 Cifar100 cutmix → Randours
Resnet50 0.6536 Cifar100 Randours → cutmix
Resnet50 0.6516 Cifar100 cutmix → Randours
Resnet101 0.6517 Cifar100 Randours → cutmix
Resnet101 0.6438 Cifar100 cutmix → Randours