-
For training
python main.py # argparser Default --print_freq 32 --save_dir ./save_model/ --save_every 10 --lr 0.1 --weight_decay 1e-4 --momentum 0.9 --Epoch 80 --batch_size 128 --test_batch_size 100 --cutout True --n_masks 1 --length 16 --normalize batchnorm
- Batch normalization
| resnet(from yeonsu repository) | SE + resnset (this code) + batchnorm | SEresnet + cutout + batch norm (this code) (cut length : 16) | |
|---|---|---|---|
| top - 1 error | 6.27 | 6.15 | 4.76 |
- Group normalization
| SE + resnet (this code) + groupnorm | SEresnet + cutout + groupnorm (this code) (cut length : 16) | |
|---|---|---|
| top - 1 error | 7.46 | 6.56 |
- Group normalization + weight standardization
| SE + resnet (this code) + groupnorm + weight standardization | SEresnet + cutout + groupnorm (this code) (cut length : 16) + weight standardization | |
|---|---|---|
| top - 1 error | 7.08 | 6.24 |
- Data augmentation
- 4pixels padded
- Randomly 32 x 32 crop
- Horizontal Flip
- Normalization with mean and standard deviation
- Option
- Cutout - Masking part to zero
- Cutmix https://github.com/LEEYEONSU/pytorch--CutMix
- Weight_initialization - kaiming_normal
- Optimizer
- SGD
- Learning_rate : 0.1
- Milestones [250, 375]
- gamma : 0.1
- Weight_decay : 0.4
- momentum : 0.9
- SGD
- group normalization
- weight standardization
- Shake-Shake regularization (in progress...)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent