Enum VGGVariant
Defines the available VGG network architecture variants.
public enum VGGVariantFields
VGG11 = 0VGG-11: 11 weight layers (8 conv + 3 FC). The smallest VGG variant.
Architecture: [64] - [128] - [256, 256] - [512, 512] - [512, 512] - FC - FC - FC
For Beginners: VGG11 is the lightest VGG variant with approximately 133 million parameters. It's a good choice when you have limited computational resources or when training from scratch on smaller datasets. The brackets show the number of filters in each convolutional block, separated by max pooling layers.
VGG11_BN = 1VGG-11 with batch normalization after each convolutional layer.
For Beginners: This is VGG11 with batch normalization added. Batch normalization normalizes the inputs to each layer, which helps the network train faster and more stably. This variant typically achieves better accuracy than the original VGG11.
VGG13 = 2VGG-13: 13 weight layers (10 conv + 3 FC).
Architecture: [64, 64] - [128, 128] - [256, 256] - [512, 512] - [512, 512] - FC - FC - FC
For Beginners: VGG13 adds one more convolutional layer to each of the first two blocks compared to VGG11. This gives it more capacity to learn features at different scales. It has approximately 133 million parameters (similar to VGG11 due to the small conv layers).
VGG13_BN = 3VGG-13 with batch normalization after each convolutional layer.
For Beginners: VGG13 with batch normalization for improved training stability and accuracy.
VGG16 = 4VGG-16: 16 weight layers (13 conv + 3 FC). The most commonly used VGG variant.
Architecture: [64, 64] - [128, 128] - [256, 256, 256] - [512, 512, 512] - [512, 512, 512] - FC - FC - FC
For Beginners: VGG16 is the most popular VGG variant and is often used as a baseline for image classification tasks. It has approximately 138 million parameters. The deeper architecture allows it to learn more complex features, making it suitable for a wide range of image recognition tasks. It's commonly used for transfer learning where you take a pre-trained VGG16 and fine-tune it for your specific task.
VGG16_BN = 5VGG-16 with batch normalization after each convolutional layer.
For Beginners: VGG16 with batch normalization. This is often the recommended VGG variant for new projects because it combines the proven architecture of VGG16 with the training benefits of batch normalization. It typically achieves 1-2% better accuracy than the original VGG16.
VGG19 = 6VGG-19: 19 weight layers (16 conv + 3 FC). The deepest VGG variant.
Architecture: [64, 64] - [128, 128] - [256, 256, 256, 256] - [512, 512, 512, 512] - [512, 512, 512, 512] - FC - FC - FC
For Beginners: VGG19 is the deepest VGG variant with approximately 144 million parameters. It adds one more convolutional layer to each of the last three blocks compared to VGG16. While it can learn slightly more complex features, the additional depth provides diminishing returns and significantly increases training time and memory requirements. VGG16 is often preferred unless you have a specific need for the additional capacity.
VGG19_BN = 7VGG-19 with batch normalization after each convolutional layer.
For Beginners: VGG19 with batch normalization. Use this when you need maximum model capacity and have sufficient computational resources. The batch normalization helps mitigate some of the training difficulties associated with very deep networks.
Remarks
VGG networks are a family of deep convolutional neural networks developed by the Visual Geometry Group at Oxford University. They are characterized by their use of small (3x3) convolution filters stacked in increasing depth, which allows them to learn complex features while keeping the number of parameters manageable.
For Beginners: VGG networks are named after the Visual Geometry Group that created them. The number in the name (e.g., VGG16) refers to the total number of weight layers in the network. For example, VGG16 has 13 convolutional layers and 3 fully connected layers, totaling 16 weight layers. These networks were groundbreaking because they showed that network depth is critical for good performance. Despite being older architectures, they remain popular for transfer learning and as baselines.
Batch Normalization Variants: The "_BN" suffix indicates variants that include batch normalization layers after each convolutional layer. Batch normalization helps stabilize training and often allows for faster convergence and better final accuracy.