Parallel Grid Pooling for Data Augmentation
This repository contains the code for the paper Parallel Grid Pooling for Data Augmentation.
Requirements
- Python 3.5+
- Chainer 4.0.0b2+
- CuPy 4.0.0b2+
- ChainerCV 0.9.0+
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 | #Params | Base | DConv | PGP |
|---|---|---|---|---|
| PreResNet-164 | 1.7M | 4.71 | 4.15 | 3.77 |
| All-CNN | 1.4M | 8.42 | 8.68 | 7.17 |
| WideResNet-28-10 | 36.5M | 3.44 | 3.88 | 3.13 |
| ResNeXt-29 (8x64d) | 34.4M | 3.86 | 3.87 | 3.22 |
| PyramidNet-164 (α=48) | 1.7M | 3.91 | 3.72 | 3.38 |
| DenseNet-BC-100 (k=12) | 0.8M | 4.60 | 4.35 | 4.11 |
Weight Transfer
Test errors (%) (Test-time data augmentation)
| Network | #Params | Base | PGP |
|---|---|---|---|
| PreResNet-164 | 1.7M | 4.71 | 4.56 |
| All-CNN | 1.4M | 8.42 | 9.03 |
| WideResNet-28-10 | 36.5M | 3.44 | 3.39 |
| ResNeXt-29 (8x64d) | 34.4M | 3.86 | 4.01 |
| PyramidNet-164 (α=48) | 1.7M | 3.91 | 3.82 |
| DenseNet-BC-100 (k=12) | 0.8M | 4.60 | 4.53 |
Test errors (%) (Training-time data augmentation)
| Network | #Params | Base | DConv | PGP |
|---|---|---|---|---|
| PreResNet-164 | 1.7M | 4.71 | 7.30 | 4.08 |
| All-CNN | 1.4M | 8.42 | 38.77 | 7.30 |
| WideResNet-28-10 | 36.5M | 3.44 | 7.90 | 3.30 |
| ResNeXt-29 (8x64d) | 34.4M | 3.86 | 16.91 | 3.36 |
| PyramidNet-164 (α=48) | 1.7M | 3.91 | 6.82 | 3.55 |
| DenseNet-BC-100 (k=12) | 0.8M | 4.60 | 7.03 | 4.36 |
Results on ImageNet and Pretrained Models
The error rates (%) shown are 224x224 1-crop test errors.
| Network | #Params | Top-1 error | Top-5 error | Model |
|---|---|---|---|---|
| ResNet-50 (Train: Base, Test: Base) | 25.6M | 23.69 | 7.00 | Download (91.1MB) |
| ResNet-50 (Train: DConv, Test: DConv) | 25.6M | 22.47 | 6.27 | Download (91.1MB) |
| ResNet-50 (Train: PGP, Test: PGP) | 25.6M | 22.40 | 6.30 | Download (91.1MB) |
| ResNet-50 (Train: Base, Test: PGP) | 25.6M | 23.32 | 6.85 | - |
| ResNet-50 (Train: DConv, Test: Base) | 25.6M | 31.44 | 11.40 | - |
| ResNet-50 (Train: PGP, Test: Base) | 25.6M | 23.01 | 6.66 | - |
| ResNet-101 (Train: Base, Test: Base) | 44.5M | 22.49 | 6.38 | Download (160MB) |
| ResNet-101 (Train: DConv, Test: DConv) | 44.5M | 21.26 | 5.61 | Download (160MB) |
| ResNet-101 (Train: PGP, Test: PGP) | 44.5M | 21.34 | 5.65 | Download (160MB) |
| ResNet-101 (Train: Base, Test: PGP) | 44.5M | 22.13 | 6.21 | - |
| ResNet-101 (Train: DConv, Test: Base) | 44.5M | 25.63 | 8.01 | - |
| ResNet-101 (Train: PGP, Test: Base) | 44.5M | 21.80 | 5.95 | - |
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,
}