This repository comes with VGG implementation in TensorFlow. VGG took the 2nd place of the ILSVRC-2014 Competition.

Currently, the VGG models in this repo have been tested with CIFAR-10 and CIFAR-100 dataset. As an indivisual deep learner, it is hard to manage such a huge dataset, ImageNet. However, I will keep working on the ImageNet dataset, please wait for it.

VGG16 model example figure from Ref.

VGG: Visual Geometry Group @Oxford University

• scikit-images
• pickle
• tqdm
• numpy
• tensorflow-gpu (>1.7)

• From command line
  • Will download CIFAR-10 or CIFAR-100 dataset and pre-process of it, and run the training on VGG. It will produce the checkpoint file for performing inference later.

python vgg.py --model-type ['A'|'A-LRN'|'B'|'C'|'D'|'E'] --dataset ['cifar10'|'cifar100']

• From source code

import cifar10_utils
import cifar100_utils
from vgg import VGG

...
valid_set = (valid_features, valid_labels)
...

# model type, D is the most well known VGG16 without 1D conv layer
# check the bottom section to see what model types are supported
vggNet = VGG(dataset='cifar10', model_type='D', learning_rate=0.0001)
vggNet.train(epochs=10, 
              batch_size=128, 
              valid_set=valid_set, 
              save_model_path='./model')

• Environment
  • Floydhub GPU2 instance (1 x Tesla V100)

• A : 11 weight layers
• A-LRN : 11 weight layers with Local Response Normalization
• B : 13 weight layers
• C : 16 weight layers with 1D conv layers
• D : 16 weight layers
• E : 19 weight layers