zoo

Model Zoo

All the models here are coded using design pattern for models. The models are based on their corresponding research paper and presented in two design patterns:

1. Idiomatic - procedural: for academic/educational audience.
2. Composable (end in _c.py) - object oriented programming: for production audience.

Model	Paper
Deep Convolutional Neural Networks
AlexNet	ImageNet Classification with Deep Convolutional Neural Networks, 2012
ZFNet	Visualizing and Understanding Convolutional Networks, 2013
VGG16	Very Deep Convolutional Networks for Large-Scale Image Recognition, 2014
VGG19	Very Deep Convolutional Networks for Large-Scale Image Recognition, 2014
Residual Convolutional Neural Networks
ResNet34	Deep Residual Learning for Image Recognition, 2015
ResNet50	Deep Residual Learning for Image Recognition, 2015
ResNet101	Deep Residual Learning for Image Recognition, 2015
ResNet152	Deep Residual Learning for Image Recognition, 2015
ResNet_cifar10	Deep Residual Learning for Image Recognition, 2015
ResNet50_v1.5	Deep Residual Learning for Image Recognition, 2015
ResNet50_v2.0	Identity Mappings in Deep Residual Networks, 2016
ResNet_cifar10_v2.0	Identity Mappings in Deep Residual Networks, 2016
SE-ResNet50	Squeeze-and-Excitation Networks, 2017
SE-ResNet101	Squeeze-and-Excitation Networks, 2017
SE-ResNet152	Squeeze-and-Excitation Networks, 2017
Wide Convolutional Neural Networks
Inception_v1	Going Deeper with Convolutions, 2015
Inception_v2	Going Deeper with Convolutions, 2015
Inception_v3	Rethinking the Inception Architecture for Computer Vision, 2015
ResNeXt50	Aggregated Residual Transformations for Deep Neural Networks, 2016
ResNeXt101	Aggregated Residual Transformations for Deep Neural Networks, 2016
ResNeXt152	Aggregated Residual Transformations for Deep Neural Networks, 2016
ResNeXt_cifar10	Aggregated Residual Transformations for Deep Neural Networks, 2016
WRN	Wide Residual Networks, 2016
Xception	Xception: Deep Learning with Depthwise Separable Convolutions, 2016
SE-ResNeXt50	Squeeze-and-Excitation Networks, 2017
SE-ResNeXt101	Squeeze-and-Excitation Networks, 2017
SE-ResNeXt152	Squeeze-and-Excitation Networks, 2017
Densely Connected Convolutional Neural Networks
DenseNet121	Densely Connected Convolutional Networks, 2016
DenseNet169	Densely Connected Convolutional Networks, 2016
DenseNet201	Densely Connected Convolutional Networks, 2016
Mobile Networks
MobileNet v1	MobileNets: Efficient Convolutional Neural Networks for Mobile VisionApplications, 2017
MobileNet v2	MobileNetV2: Inverted Residuals and Linear Bottlenecks, 2019
MobileNet v3 small	Searching for MobileNetV3, 2019
MobileNet v3 large	Searching for MobileNetV3, 2019
SqueezeNet	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size, 2016
SqueezeNet_bypass	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size, 2016
SqueezeNet_complex	SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size, 2016
ShuffleNet	ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices, 2017
One-Shot Classification Networks
Siamese Twin	Siamese Neural Networks for One-shot Image Recognition, 2015
AutoEncoders
Auto Encoder	no paper - general principles
SRCNN	Image Super-Resolution Using Deep Convolutional Networks
Generative Adversarial Networks
DCGAN	Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, 2016
Object Dectection
R-CNN	Rich feature hierarchies for accurate object detection and semantic segmentation, 2014
[Fast R-CNN](fast r-cnn)	Fast R-CNN, 2015
[Faster R-CNN](faster r-cnn)	Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks, 2015
Image Segmentation
U-Net	U-Net: Convolutional Networks for Biomedical Image Segmentation, 2015

Architecture Representation

The architecture representation of models consists of an overall macro-architecture and a micro-architecture.

Macro-Architecture

The macro architecture consists of a stem (entry) group, a collection of groups (middle), and a classifier group (exit). The number of groups is defined by the macro architecture. The macro architecture may optionally contain a pre-stem, which perform additional operations, such as data preprocessing, model aggregation, and prediction post-processing.

Micro-Architecture

The micro architecture consists of a collection of blocks, which collectively form a group. A block consists of an input and an output, and within the block is a set of layers connected by the block pattern. The number of blocks and the block pattern is defined by the meta-parameters (where parameters are the weights/biases learned, and hyper-parameters are the parameters used to train the model, but not part of the model).