Train CIFAR-10 using VGG-16-BN, which gives 93.39% test set accuracy. This is a PyTorch implementation.
CUDA_VISIBLE_DEVICES=0 ./vgg16_bn_tiny.py --base_lr 5e-2 --batch_size 256
--epochs 170 --weight_decay 1e-3
The model used here is a modified version of VGG-16-BN. Besides, we add a BN layer after each convolution layer. Dropout is used after fc6 and fc7. The network accepts a 33232 input, and the pool5 activation has shape 51211 since we down-sample 5 times.
conv1^2 (64) -> pool1 -> conv2^2 (128) -> pool2 -> conv3^3 (256) -> pool3 -> conv4^3 (512) -> pool4 -> conv5^3 (512) -> pool5 -> fc6 (512) -> fc7 (512) -> fc8 (10).