A PyTorch implementation of Very Deep Convolutional Networks For Large-Scale Image Recognition
git clone https://github.com/andregaio/vgg.git
cd vgg
conda create -n vgg python=3.8
conda activate vgg
pip install -r requirements.txt
| Name | Accuracy |
|---|---|
| VGG-A | 90.368 |
| VGG-A-LRN | 90.813 |
| VGG-B | 91.604 |
| VGG-C | 91.822 |
| VGG-D | 92.247 |
| VGG-E | 91.901 |
python train.py --model vgg_A --dataset cifar10
python eval.py --model vgg_A --weights weights/checkpoint_00070.pt --dataset cifar10
python infer.py --model vgg_A --weights weights/checkpoint_00070.pt --image assets/cat.png
This implementation is not designed to be a complete replica of the original - the main differences are:
- Batchnorm layers have been added prior to each activation
- Learning rate modified to 10e-3
- Accuracy has been used to evaluate classification performance
- Has been trained on CIFAR10
- Input resolution and FC layer sizes have been changed to 32x32 match dataset
- Automatic Mixed Precision (AMP) training with gradient scaling and autocasting
- Kaiming initialisation
- RGB colour shift has not been used
- Dropout set to 0.3 for all except vgg_C which was kept at 0.5
- Learning rate dropped 2 times during training process