Home

Awesome

Parallel Grid Pooling for Data Augmentation

This repository contains the code for the paper Parallel Grid Pooling for Data Augmentation.

Requirements

Training

To train PreResNet-164 on CIFAR-10 dataset with single-GPU:

$ python train.py --dataset cifar10 --model PreResNet164 --gpus 0

To train ResNet-50 on ImageNet dataset with multi-GPU:

$ python train_imagenet.py --model ResNet50_fb --gpus 0,1,2,3,4,5,6,7

Results on CIFAR-10

Test errors (%)

Network#ParamsBaseDConvPGP
PreResNet-1641.7M4.714.153.77
All-CNN1.4M8.428.687.17
WideResNet-28-1036.5M3.443.883.13
ResNeXt-29 (8x64d)34.4M3.863.873.22
PyramidNet-164 (α=48)1.7M3.913.723.38
DenseNet-BC-100 (k=12)0.8M4.604.354.11

Weight Transfer

Test errors (%) (Test-time data augmentation)

Network#ParamsBasePGP
PreResNet-1641.7M4.714.56
All-CNN1.4M8.429.03
WideResNet-28-1036.5M3.443.39
ResNeXt-29 (8x64d)34.4M3.864.01
PyramidNet-164 (α=48)1.7M3.913.82
DenseNet-BC-100 (k=12)0.8M4.604.53

Test errors (%) (Training-time data augmentation)

Network#ParamsBaseDConvPGP
PreResNet-1641.7M4.717.304.08
All-CNN1.4M8.4238.777.30
WideResNet-28-1036.5M3.447.903.30
ResNeXt-29 (8x64d)34.4M3.8616.913.36
PyramidNet-164 (α=48)1.7M3.916.823.55
DenseNet-BC-100 (k=12)0.8M4.607.034.36

Results on ImageNet and Pretrained Models

The error rates (%) shown are 224x224 1-crop test errors.

Network#ParamsTop-1 errorTop-5 errorModel
ResNet-50 (Train: Base, Test: Base)25.6M23.697.00Download (91.1MB)
ResNet-50 (Train: DConv, Test: DConv)25.6M22.476.27Download (91.1MB)
ResNet-50 (Train: PGP, Test: PGP)25.6M22.406.30Download (91.1MB)
ResNet-50 (Train: Base, Test: PGP)25.6M23.326.85-
ResNet-50 (Train: DConv, Test: Base)25.6M31.4411.40-
ResNet-50 (Train: PGP, Test: Base)25.6M23.016.66-
ResNet-101 (Train: Base, Test: Base)  44.5M22.496.38Download (160MB)
ResNet-101 (Train: DConv, Test: DConv)44.5M21.265.61Download (160MB)
ResNet-101 (Train: PGP, Test: PGP)44.5M21.345.65Download (160MB)
ResNet-101 (Train: Base, Test: PGP)44.5M22.136.21-
ResNet-101 (Train: DConv, Test: Base)44.5M25.638.01-
ResNet-101 (Train: PGP, Test: Base)44.5M21.805.95-

Citation

@article{Takeki18Parallel,
    title = {Parallel Grid Pooling for Data Augmentation},
    author = {Takeki, Akito and Ikami, Daiki and Irie, Go and Aizawa, Kiyoharu},
    journal = {arXiv preprint arXiv:1803.11370},
    year = 2018,
}