s-themis / cnn-architectures

Keras implementation of influential CNN architectures.

cnn-architectures

Implementation of influential convolutional neural network architectures using the Keras functional API.

Architectures

AlexNet

Paper | Implementation
Year published: 2012
Depth: 8
Parameters: 62.4M
Reported training time: 5-6 days on two GTX 580 3GB GPUs
Key architectural points:

• Kernels of decreasing size 11x11, 5x5, 3x3
• ReLU activations
• Local response normalization
• Overlapping max pooling for downsampling
• Dropout on fully-connected layers

VGGNet (Configuration D - VGG16)

Paper | Implementation
Year published: 2014
Depth: 16
Parameters: 138.4M
Reported training time: 2-3 weeks on four Titan Black GPUs
Key architectural points:

• Homogeneous convolutional blocks with decreasing feature map size and increasing number of filters
• Homogeneous kernels of size 3x3
• ReLU activations
• No local response normalization
• Non-overlapping max pooling for downsampling on each block
• Dropout on fully-connected layers

ResNet (ResNet50 Variant)

Paper | Implementation
Year published: 2015
Depth: 50
Parameters: 25.6M
Reported training time: -
Key architectural points:

• Initial kernel size 7x7 followed by overlapping max-pooling for downsampling
• Homogenous stacks of residual blocks with decreasing feature map size and increasing number of filters
• Kernels of size 1x1, 3x3, 1x1 on each residual block to bottleneck number of filters on 3x3 layers
• ReLU activations
• Batch normalization before activations
• Downsampling with overlapping convolutions
• Global average pooling before single fully-connected layer
• No dropout