Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 94.86/5.14 | baseline/baseline.yaml | |
Mixup | 96.01/3.99 | mixup/200/mixup.yaml | α=1 |
Manifold Mixup | 96.10/3.90 | mixup/200/manifold_mixup01.yaml | α=2 layers=[0,1] |
Manifold Mixup | 96.01/3.99 | mixup/200/manifold_mixup012.yaml | α=2 layers=[0,1,2] |
Cutmix | 96.20/3.80 | mixup/200/cutmix.yaml | α=1 |
Manifold Cutmix | 96.06/3.94 | mixup/200/manifold_cutmix_a1.yaml | α=1 layers=[0,1,2] |
Manifold Cutmix | 95.70/4.30 | mixup/200/manifold_cutmix_a2.yaml | α=2 layers=[0,1,2] |
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 95.40/4.60 | baseline/baseline_600.yaml | |
Mixup | 96.59/3.41 | mixup/600/mixup.yaml | α=1 |
Manifold Mixup | 96.86/3.14 | mixup/600/manifold_mixup_a2.yaml | α=2 layers=[0,1,2] |
Manifold Mixup | 96.63/3.37 | mixup/600/manifold_mixup_a1.yaml | α=1 layers=[0,1,2] |
Cutmix | 96.76/3.24 | mixup/600/cutmix.yaml | α=1 |
Manifold Cutmix | 96.53/3.47 | mixup/600/manifold_cutmix_a1.yaml | α=1 layers=[0,1,2] |
Manifold Cutmix | 96.43/3.57 | mixup/600/manifold_cutmix_a2.yaml | α=2 layers=[0,1,2] |
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 95.59/4.41 | baseline/baseline_1200.yaml | |
Mixup | 96.85/3.15 | mixup/1200/mixup.yaml | α=1 |
Manifold Mixup | 97.19/2.81 | mixup/1200/manifold_mixup.yaml | α=2 layers=[0,1,2] |
- Images generated from conditional BigGAN.
% Generated Data | Acc/Error (%) | Config |
---|---|---|
100% | 66.23/33.77 | gan/gan_100.yaml |
50% | 92.34/7.66 | gan/gan_50.yaml |
25% | 93.67/6.33 | gan/gan_25.yaml |
10% | 94.33/5.67 | gan/gan_10.yaml |
0% | 94.86/5.14 | baseline/baseline.yaml |
- The standard augmentations are random horizonal flips, random translations and normalization. Addition augmentations are applied after the random flip and translation while normalization is always applied last. Cutout is applied after Autoaugment/RandAugment when used together.
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 94.86/5.14 | baseline/baseline.yaml | |
Cutout | 95.88/4.12 | augment/200/cutout.yaml | cutout=16x16 |
AutoAugment | 95.90/4.10 | augment/200/autoaugment.yaml | |
AutoAugment + Cutout | 96.31/3.69 | augment/200/autoaugment_cutout.yaml | cutout=16x16 |
RandAugment | 95.02/4.98 | augment/200/randaugment_n3m5.yaml | n=3 m=5 |
RandAugment | 94.93/5.07 | augment/200/randaugment_n3m4.yaml | n=3 m=4 |
RandAugment | 94.37/5.63 | augment/200/randaugment_n3m2.yaml | n=3 m=2 |
RandAugment | 95.65/4.35 | augment/200/randaugment_n2m5.yaml | n=2 m=5 |
RandAugment | 95.50/4.50 | augment/200/randaugment_n2m6.yaml | n=2 m=6 |
RandAugment + Cutout | 95.64/4.36 | augment/200/randaugment_cutout.yaml | n=2 m=5 cutout=16x16 |
RandAugment (w/ Cutout) | 95.63/4.37 | augment/200/randaugment_include_cutout.yaml | n=2 m=5 |
GridMask | 95.59/4.41 | augment/200/gridmask_8_32_r04.yaml | minD=8 maxD=32 r=0.4 |
GridMask | 95.88/4.12 | augment/200/gridmask_16_32_r04.yaml | minD=16 maxD=32 r=0.4 |
GridMask | 95.78/4.22 | augment/200/gridmask_16_40_r04.yaml | minD=16 maxD=40 r=0.4 |
GridMask | 95.80/4.20 | augment/200/gridmask_16_32_r05.yaml | minD=16 maxD=32 r=0.5 |
GridMask | 95.69/4.31 | augment/200/gridmask_16_32_r03.yaml | minD=16 maxD=32 r=0.3 |
Autoaugment + GridMask | 96.18/3.82 | augment/200/autoaugment_gridmask_16_32_r04.yaml | minD=16 maxD=32 r=0.4 |
Autoaugment + GridMask | 95.95/4.05 | augment/200/autoaugment_gridmask_16_32_r03.yaml | minD=16 maxD=32 r=0.3 |
AugMix | 95.63/4.37 | augment/200/augmix_w3_d3_s3.yaml | width=3 depth=3 severity=3 |
AugMix | 95.53/4.47 | augment/200/augmix_w3_d3_s5.yaml | width=3 depth=3 severity=5 |
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 95.40/4.60 | baseline/baseline_600.yaml | |
AutoAugment + Cutout | 97.09/2.91 | augment/600/autoaugment_cutout.yaml | cutout=16x16 |
RandAugment + Cutout | 96.5/3.5 | augment/600/randaugment_cutout.yaml | n=2 m=5 cutout=16x16 |
RandAugment (w/ Cutout) | 96.58/3.42 | augment/600/randaugment_include_cutout.yaml | n=2 m=5 |
Autoaugment + GridMask | 97.02/2.98 | augment/600/autoaugment_gridmask_16_32_r04.yaml | minD=16 maxD=32 r=0.4 |
Autoaugment + GridMask | 96.81/3.19 | augment/600/autoaugment_gridmask_16_32_r03.yaml | minD=16 maxD=32 r=0.3 |
AugMix | 96.33/3.67 | augment/600/augmix_w3_d3_s3.yaml | width=3 depth=3 severity=3 |
- AutoAugment shows better results on CIFAR compared to RandAugment which appears to make the augmentations too strong for the size of the model without proper hyperparameter tuning. The hyperparameters however offer better flexiblity which should benefit a wider variety of datasets compared to AutoAugment which is tuned specifically on a target dataset.
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 94.86/5.14 | baseline/baseline.yaml | |
Label Smoothing | 94.69/5.31 | other/smoothing_01.yaml | ε=0.1 |
Method | Acc/Error (%) | Config | Parameters |
---|---|---|---|
Baseline | 95.06/4.94 | baseline/baseline_600_cos.yaml | |
AutoAugment + Cutout | 97.06/2.94 | combination/autoaugment_cutout.yaml | cutout=16x16 |
AutoAugment + Cutout + Label Smoothing | - | combination/autoaugment_cutout_smoothing.yaml | cutout=16x16 ε=1 |
AutoAugment + Cutout + Mixup | 97.25/2.75 | combination/autoaugment_cutout_mixup.yaml | cutout=16x16 α=1 |
AutoAugment + Cutout + Mixup + Label Smoothing | 97.26/2.74 | combination/aa_cutout_mixup_smoothing.yaml | cutout=16x16 α=1 ε=0.1 |
AutoAugment + Cutout + Manifold Mixup | 97.05/2.95 | combination/autoaugment_cutout_manifold.yaml | cutout=16x16 α=2 layers=[0,1,2] |
AutoAugment + Cutout + Manifold Mixup + Label Smoothing | - | combination/aa_cutout_manifold_smoothing.yaml | cutout=16x16 α=2 layers=[0,1,2] ε=0.1 |
AutoAugment + Cutmix | - | combination/autoaugment_cutmix.yaml | α=1 |
- DropBlock, StochDepth
- Truncated z sampling in GAN
- CIFAR100
- Different architectures
- Architecture modifications (Shake-Shake, ShakeDrop, etc)
- Further investigate AugMix and GridMask
- mixup: Beyond Empirical Risk Minimization
- Manifold Mixup: Better Representations by Interpolating Hidden States
- CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features
- Pretrained BigGAN
- Improved Regularization of Convolutional Neural Networks with Cutout
- AutoAugment: Learning Augmentation Policies from Data
- RandAugment: Practical automated data augmentation with a reduced search space
- GridMask Data Augmentation
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty