Awesome

RICAP: Data Augmentation using Random Image Cropping and Patching for Deep CNNs

PyTorch implementation of data augmentation method RICAP for deep CNNs proposed by "Data Augmentation using Random Image Cropping and Patching for Deep CNNs."

Prerequisites

• Python 3.5
• PyTorch 1.0
• GPU (recommended)

Datasets

• CIFAR-10/100: automatically downloaded by PyTorch scripts to data folder
• ImageNet: manually downloaded from ImageNet (ILSVRC2012 version) and moved to train and val folders in your dataroot path (e.g., ./imagenet/)

Results

CIFAR

ImageNet

• Details are in our paper.

How to Train

Our script occupies all available GPUs. Please set environment CUDA_VISIBLE_DEVICES.

CIFAR-10 and WideResNet28-10

with RICAP

python main.py --dataset cifar10 --model WideResNetDropout --depth 28 --params 10 --beta_of_ricap 0.3 --postfix ricap0.3

without RICAP

python main.py --dataset cifar10 --model WideResNetDropout --depth 28 --params 10

We trained these models on a single GPU (GeForce GTX 1080).

CIFAR-100 and WideResNet28-10

with RICAP

python main.py --dataset cifar100 --model WideResNetDropout --depth 28 --params 10 --beta_of_ricap 0.3 --postfix ricap0.3

without RICAP

python main.py --dataset cifar100 --model WideResNetDropout --depth 28 --params 10

We trained these models on a single GPU (GeForce GTX 1080).

ImageNet and WideResNetBottleneck50-2 for 100 epochs

with RICAP

python main.py --dataset ImageNet --dataroot [your imagenet folder path(like ./imagenet)] --model WideResNetBottleneck --depth 50 --epoch 100 --adlr 30,60,90 --droplr 0.1 --wd 1e-4 --batch 256 --params 2 --beta_of_ricap 0.3 --postfix ricap0.3

without RICAP

python main.py --dataset ImageNet --dataroot [your imagenet folder path(like ./imagenet)] --model WideResNetBottleneck --depth 50 --epoch 100 --adlr 30,60,90 --droplr 0.1 --wd 1e-4 --batch 256 --params 2

We trained these models on four GPUs (GeForce GTX 1080).

References

@inproceedings{RICAP2018ACML,
  title = {RICAP: Random Image Cropping and Patching Data Augmentation for Deep CNNs},
  author = {Takahashi, Ryo and Matsubara, Takashi and Uehara, Kuniaki},
  booktitle = {Asian Conference on Machine Learning (ACML)},
  url={http://proceedings.mlr.press/v95/takahashi18a.html},
  year = {2018}
}

@article{RICAP2018arXiv,
  title={Data Augmentation using Random Image Cropping and Patching for Deep CNNs},
  author={Takahashi, Ryo and Matsubara, Takashi and Uehara, Kuniaki},
  journal={arXiv},
  url={https://arxiv.org/abs/1811.09030},
  year={2018}
}