Analytical-Learning-Theory
所属分类:人工智能/神经网络/深度学习
开发工具:Python
文件大小:28KB
下载次数:0
上传日期:2019-03-06 22:33:43
上 传 者:
sh-1993
说明: 使用PyTorch在CIFAR10上达到97.39%
(97.39% on CIFAR10 with PyTorch)
文件列表:
affine_transforms.py (16715, 2019-03-07)
cifar10 (0, 2019-03-07)
cifar10\resnext (0, 2019-03-07)
cifar10\resnext\LICENSE (1068, 2019-03-07)
cifar10\resnext\logger.py (2279, 2019-03-07)
cifar10\resnext\main.py (18110, 2019-03-07)
cifar10\resnext\models (0, 2019-03-07)
cifar10\resnext\models\__init__.py (995, 2019-03-07)
cifar10\resnext\models\caffe_cifar.py (1750, 2019-03-07)
cifar10\resnext\models\densenet.py (3629, 2019-03-07)
cifar10\resnext\models\imagenet_resnet.py (5564, 2019-03-07)
cifar10\resnext\models\preresnet.py (4727, 2019-03-07)
cifar10\resnext\models\res_utils.py (976, 2019-03-07)
cifar10\resnext\models\resnet.py (4962, 2019-03-07)
cifar10\resnext\models\resnext.py (4339, 2019-03-07)
cifar10\resnext\models\wide_resnet.py (4999, 2019-03-07)
cifar10\resnext\test.py (199, 2019-03-07)
cifar10\resnext\utils.py (3919, 2019-03-07)
helpers.py (4359, 2019-03-07)
load_data.py (5373, 2019-03-07)
lr_scheduler.py (4312, 2019-03-07)
plots.py (2070, 2019-03-07)
This repository contains a PyTorch implementation code for reproducing the results in our paper:
**[Generalization in Machine Learning via Analytical Learning Theory](https://arxiv.org/pdf/1802.07426.pdf)** \
*Kenji Kawaguchi, Yoshua Bengio, Vikas Verma, and Leslie Pack Kaelbling*
#### Test error (\%) with WideResNet28_10 and different regularization methods
| Regularization Method | CIFAR-10 | CIFAR-100 | SVHN |
|:----------:|:--------------:|:--------------:|:------:|
| Standard | 3.79 ± 0.07 | 19.85 ± 0.14 | 2.47 ± 0.04|
| Single-cutout | 3.19 ± 0.09 | 18.13 ± 0.28 | 2.23 ± 0.03 |
| Dual-cutout | 2.61 ± 0.04 | 17.54 ± 0.09 | 2.06 ± 0.06|
* Dual-cutout is proposed in our paper based on a new learning theory.
### How to run DualCutout
```
python cifar10/resnext/main.py --dualcutout --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size *** --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1 --alpha 0.1 --cutsize 16
```
Add the --temp_dir and --home_dir as appropriate in the above commands. For Cifar10 and Cifar100, we used --cutsize 16, and for SVHN, we used --cutsize 20.
### How to run Single Cutout
```
python cifar10/resnext/main.py --singlecutout --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size *** --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1 --alpha 0.1 --cutsize 16
```
### How to run baseline
```
python cifar10/resnext/main.py --dataset cifar10 --arch wrn28_10 \
--epochs 300 --batch_size *** --learning_rate 0.1 --data_aug 1 --decay 0.0005 --schedule 150 225 \
--gamma 0.1 0.1
```
This code has been tested with
python 2.7.9
torch 0.3.1
torchvision 0.2.0
近期下载者:
相关文件:
收藏者: