For cifar10 ”DADA: Deep Adversarial Data Augmentation for Extremely Low Data Regime Classification“
Origin paper:https://arxiv.org/abs/1809.00981
Official Implementation(Theano): https://github.com/SchafferZhang/DADA
python3.5
pytorch 1.1.0
cuda8.0
torchvision
Keras
python3 train.py
Default Set:config.py
Best Acc
| Method | 400 | 600 | 800 | 1000 |
|---|---|---|---|---|
| DADA | 63.0 | 67.6 | 71.2 | 73.3 |
| DADA_augmented | 69.9 | 74.1 | 76.0 | 79.8 |
** one row represent one class (100 fixed noise)**
i remove weight_norm,because it cause bad performance,when i add weight_norm.