Implementation of Swapout by chainer
git clone https://github.com/nutszebra/swapout.git
cd swapout
git submodule init
git submodule update
python main.py -g 0
All hyperparameters and network architecture are the same as in [1] except for data-augmentation.
-
Data augmentation
Train: Pictures are randomly resized in the range of [32, 36], then 32x32 patches are extracted randomly and are normalized locally. Horizontal flipping is applied with 0.5 probability.
Test: Pictures are resized to 32x32, then they are normalized locally. Single image test is used to calculate total accuracy. -
Stochastic inference
Implemented
| network | depth | k | total accuracy (%) |
|---|---|---|---|
| Swapout v2(20) Wx4[1] | 20 | 4 | 94.91 |
| Swapout v2(32) Wx4[1] | 32 | 4 | 95.24 |
| my implementation | 32 | 4 | 95.34 |
Swapout: Learning an ensemble of deep architectures [1]