dlibml / dnn

Deep Neural Network Architectures with dlib

dnn

Deep Neural Network Architectures

This repository contains the definitions for the following architectures, organized by task.

Contents

• Classification
  • AlexNet
  • SqueezeNet
  • VGGNet
  • GoogLeNet
  • ResNet
  • DenseNet
  • DarkNet
  • VoVNet
  • RepVGG
• Detection
  • YOLOv5

Classification

AlexNet

It contains the definition for the model that started it all.

Papers:

• ImageNet Classification with Deep Convolutional Neural Networks

SqueezeNet

In particular, it contains SqueezeNet-{v1.0,v1.1}.

Papers:

• SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size

VGGNet

In particular, it contains VGGNet-{11,13,16,19} variants with batch normalization.

Papers:

• Very Deep Convolutional Networks for Large-Scale Image Recognition

GoogLeNet

It contains the definition of the GoogLeNet, also known as InceptionV1.

Papers:

• Going Deeper with Convolutions

ResNet

In particular, it contains ResNet-{18,34,50,101,152}-B definitions, in contrast to dlib, which contains the A variants.

Papers:

• Deep Residual Learning for Image Recognition

DenseNet

In particular, it contains DenseNet-{121,169,201,264,161} definitions.

Papers:

• Densely Connected Convolutional Networks

DarkNet

In particular, it contains the backbones for DarkNet-19 (introduced in YOLOv1), DarkNet-53 (YOLOv3) and CSPDarknet-53 (YOLOv4).

Papers:

• You Only Look Once: Unified, Real-Time Object Detection
• YOLOv3: An Incremental Improvement
• CSPNet: A New Backbone that can Enhance Learning Capability of CNN
• YOLOv4: Optimal Speed and Accuracy of Object Detection

VoVNet

In particular, it contains implementations for VoVNetv2-{19slim,19,27slim,27,39,57,99}, which are very similar to VoVNetv1 (V2 have identiy mapping and effective Squeeze and Excitation on top of V1).

Papers:

• An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection
• CenterMask: Real-Time Anchor-Free Instance Segmentation

RepVGG

In particular, it contains implementations for RepVGG-{A0,A1,A2,B0,B1,B2,B3}.

Note that, at the moment, there is no way to convert from a trained RepVGG model into its inference counterpart. I will investigate how to do that soon.

Papers:

• RepVGG: Making VGG-style ConvNets Great Again

Detection

YOLOv5

In particular, it contains implementations for YOLOv5{n,s,m,l,x}, which match the ones in ultralytics/yolov5.