VGG

VGG Implementation

The repository includes the implementation of VGG11, VGG13, VGG16 and VGG19 in Tensorflow 2

VGG Architecrures

<center> <img src="img/vgg.png" align="center" width="800" height="600"/> </center>

image is taken from source

<center> <img src="img/vgg16.png" width="700" height="400"/> </center>

VGG16 Architecture source

Training on MNIST

<center> <img src="img/mnist.png" width="400" height="350"/> </center>

Requirement

python==3.7.0
numpy==1.18.1

How to use

Training & Prediction can be run as follows:
python train.py train
python train.py predict img.png

More information

• Please refer to the original paper of VGG here for more information.

Implementation Notes

• Note 1:
  Since VGG is somehow huge and painfully slow in training ,I decided to make number of filters variable. If you want to run it in your PC, you can reduce the number of filters into 32,16,8,4 or 2. (64 is by default). For example:
  model = vgg.VGG16((112, 112, 3), classes = 10, filters = 8)

• Note 2 :
  You can also make the size of images smaller, so that it can be ran faster and doesn't take too much memories.

Result for MNIST:

• Learning rate = 0.0001
• Batch size = 32
• Optimizer = Adam
• Fliters = 8
• epochs = 2

Name | Training Accuracy | Validation Accuracy | :---: | :---: | :---: VGG11 | 88.08% | 95.77% VGG13 | 93.99% | 94.85% VGG16 | 85.84% | 92.50% VGG19 | 93.66% | 93.75%