Computer vision models on TensorFlow 2.x

This is a collection of image classification, segmentation, detection, and pose estimation models. Many of them are pretrained on ImageNet-1K, CIFAR-10/100, SVHN, CUB-200-2011, Pascal VOC2012, ADE20K, Cityscapes, and COCO datasets and loaded automatically during use. All pretrained models require the same ordinary normalization. Scripts for training/evaluating/converting models are in the imgclsmob repo.

List of implemented models

• AlexNet ('One weird trick for parallelizing convolutional neural networks')
• ZFNet ('Visualizing and Understanding Convolutional Networks')
• VGG/BN-VGG ('Very Deep Convolutional Networks for Large-Scale Image Recognition')
• BN-Inception ('Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift')
• ResNet ('Deep Residual Learning for Image Recognition')
• PreResNet ('Identity Mappings in Deep Residual Networks')
• ResNeXt ('Aggregated Residual Transformations for Deep Neural Networks')
• SENet/SE-ResNet/SE-PreResNet/SE-ResNeXt ('Squeeze-and-Excitation Networks')
• ResNeSt(A) ('ResNeSt: Split-Attention Networks')
• IBN-ResNet/IBN-ResNeXt/IBN-DenseNet ('Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net')
• AirNet/AirNeXt ('Attention Inspiring Receptive-Fields Network for Learning Invariant Representations')
• BAM-ResNet ('BAM: Bottleneck Attention Module')
• CBAM-ResNet ('CBAM: Convolutional Block Attention Module')
• SCNet ('Improving Convolutional Networks with Self-Calibrated Convolutions')
• RegNet ('Designing Network Design Spaces')
• PyramidNet ('Deep Pyramidal Residual Networks')
• DiracNetV2 ('DiracNets: Training Very Deep Neural Networks Without Skip-Connections')
• DenseNet ('Densely Connected Convolutional Networks')
• PeleeNet ('Pelee: A Real-Time Object Detection System on Mobile Devices')
• WRN ('Wide Residual Networks')
• DRN-C/DRN-D ('Dilated Residual Networks')
• DPN ('Dual Path Networks')
• DarkNet Ref/Tiny/19 ('Darknet: Open source neural networks in c')
• DarkNet-53 ('YOLOv3: An Incremental Improvement')
• BagNet ('Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet')
• DLA ('Deep Layer Aggregation')
• DiCENet ('DiCENet: Dimension-wise Convolutions for Efficient Networks')
• HRNet ('Deep High-Resolution Representation Learning for Visual Recognition')
• VoVNet ('An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection')
• SelecSLS ('XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera')
• HarDNet ('HarDNet: A Low Memory Traffic Network')
• SqueezeNet/SqueezeResNet ('SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size')
• SqueezeNext ('SqueezeNext: Hardware-Aware Neural Network Design')
• ShuffleNet ('ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices')
• ShuffleNetV2 ('ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design')
• MENet ('Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications')
• MobileNet ('MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications')
• FD-MobileNet ('FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy')
• MobileNetV2 ('MobileNetV2: Inverted Residuals and Linear Bottlenecks')
• MobileNetV3 ('Searching for MobileNetV3')
• IGCV3 ('IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks')
• GhostNet ('GhostNet: More Features from Cheap Operations')
• MnasNet ('MnasNet: Platform-Aware Neural Architecture Search for Mobile')
• ProxylessNAS ('ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware')
• FBNet ('FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search')
• Xception ('Xception: Deep Learning with Depthwise Separable Convolutions')
• InceptionV3 ('Rethinking the Inception Architecture for Computer Vision')
• InceptionV4/InceptionResNetV1/InceptionResNetV2 ('Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning')
• PolyNet ('PolyNet: A Pursuit of Structural Diversity in Very Deep Networks')
• NASNet ('Learning Transferable Architectures for Scalable Image Recognition')
• PNASNet ('Progressive Neural Architecture Search')
• SPNASNet ('Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours')
• EfficientNet ('EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks')
• MixNet ('MixConv: Mixed Depthwise Convolutional Kernels')
• FCN-8s ('Fully Convolutional Networks for Semantic Segmentation')
• PSPNet ('Pyramid Scene Parsing Network')
• DeepLabv3 ('Rethinking Atrous Convolution for Semantic Image Segmentation')
• ICNet ('ICNet for Real-Time Semantic Segmentation on High-Resolution Images')
• Fast-SCNN ('Fast-SCNN: Fast Semantic Segmentation Network')
• CGNet ('CGNet: A Light-weight Context Guided Network for Semantic Segmentation')
• DABNet ('DABNet: Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation')
• BiSeNet ('BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation')
• DANet ('Dual Attention Network for Scene Segmentation')
• FPENet ('Feature Pyramid Encoding Network for Real-time Semantic Segmentation')
• LEDNet ('LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation')
• CenterNet ('Objects as Points')
• LFFD ('LFFD: A Light and Fast Face Detector for Edge Devices')
• AlphaPose ('RMPE: Regional Multi-person Pose Estimation')
• SimplePose ('Simple Baselines for Human Pose Estimation and Tracking')
• Lightweight OpenPose ('Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose')
• IBPPose ('Simple Pose: Rethinking and Improving a Bottom-up Approach for Multi-Person Pose Estimation')
• VOCA ('Capture, Learning, and Synthesis of 3D Speaking Styles')
• Neural Voice Puppetry Audio-to-Expression net ('Neural Voice Puppetry: Audio-driven Facial Reenactment')
• Jasper/JasperDR ('Jasper: An End-to-End Convolutional Neural Acoustic Model')
• QuartzNet ('QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions')

Installation

To use the models in your project, simply install the tf2cv package with tensorflow:

pip install tf2cv tensorflow>=2.0.0

To enable/disable different hardware supports, check out TensorFlow installation instructions.

Usage

Example of using a pretrained ResNet-18 model (with channels_first data format):

from tf2cv.model_provider import get_model as tf2cv_get_model
import tensorflow as tf

net = tf2cv_get_model("resnet18", pretrained=True, data_format="channels_last")
x = tf.random.normal((1, 224, 224, 3))
y_net = net(x)

Pretrained models (ImageNet-1K)

Some remarks:

• Top1/Top5 are the standard 1-crop Top-1/Top-5 errors (in percents) on the validation subset of the ImageNet-1K dataset.
• FLOPs/2 is the number of FLOPs divided by two to be similar to the number of MACs.
• Remark Converted from GL model means that the model was trained on MXNet/Gluon and then converted to TensorFlow.
• ResNet(A) is an average downsampled ResNet intended for use as an feature extractor in some pose estimation networks.
• ResNet(D) is a dilated ResNet intended for use as an feature extractor in some segmentation networks.
• Models with *-suffix use non-standard preprocessing (see the training log).

CIFAR-10

CIFAR-100

SVHN

CUB-200-2011

Pascal VOC20102

ADE20K

Cityscapes

COCO Semantic Segmentation

CelebAMask-HQ

COCO Keypoints Detection