模型摘要
包含的模型架构来自广泛的来源。以下列出了来源,包括论文、我重写/适配的原始实现(“参考代码”),以及我直接利用的PyTorch实现(“代码”)。
大多数包含的模型都有预训练的权重。这些权重是
- 来自其原始来源
- 由我移植自不同框架(例如Tensorflow模型)的原始实现
- 使用所提供的训练脚本从头开始训练
预训练权重的验证结果在这里
在timm
中模型的更多精彩视图(带有漂亮图片)可以在paperswithcode找到。
大规模迁移ResNetV2 (BiT)
- 实现:resnetv2.py
- 论文:
大规模迁移 (BiT):通用视觉表示学习
- https://arxiv.org/abs/1912.11370 - 参考代码:https://github.com/google-research/big_transfer
跨阶段部分网络
- 实现:cspnet.py
- 论文:
CSPNet:一种能增强CNN学习能力的全新骨架
- https://arxiv.org/abs/1911.11929 - 参考实现:https://github.com/WongKinYiu/CrossStagePartialNetworks
DenseNet
- 实现:[densenet.py](https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/densenet.py "densenet.py")
- 论文:[Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993 "https://arxiv.org/abs/1608.06993")
- 代码:[https://github.com/pytorch/vision/tree/master/torchvision/models](https://github.com/pytorch/vision/tree/master/torchvision/models)
DLA
- 实现:[dla.py](https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/dla.py "dla.py")
- 论文:[https://arxiv.org/abs/1707.06484](https://arxiv.org/abs/1707.06484)
- 代码:[https://github.com/ucbdrive/dla](https://github.com/ucbdrive/dla)
Dual-Path Networks
- 实现:[dpn.py](https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/dpn.py "dpn.py")
- 论文:[Dual Path Networks](https://arxiv.org/abs/1707.01629 "https://arxiv.org/abs/1707.01629")
- 我的 PyTorch 代码:[https://github.com/rwightman/pytorch-dpn-pretrained](https://github.com/rwightman/pytorch-dpn-pretrained)
- 参考代码:[https://github.com/cypw/DPNs](https://github.com/cypw/DPNs)
GPU-Efficient Networks
- 实现:byobnet.py
- 论文:
用于GPU高效网络的神经网络架构设计
- https://arxiv.org/abs/2006.14090 - 参考代码:https://github.com/idstcv/GPU-Efficient-Networks
HRNet
- 实现:hrnet.py
- 论文:
用于视觉识别的高分辨率表征学习
- https://arxiv.org/abs/1908.07919 - 代码:https://github.com/HRNet/HRNet-Image-Classification
Inception-V3
- 实现:inception_v3.py
- 论文:
重新思考用于计算机视觉的Inception架构
- https://arxiv.org/abs/1512.00567 - 代码:[https://github.com/pytorch/vision/tree/master/torchvision/models](https://github.com/pytorch/vision/tree/master/torchvision/models)
Inception-V4
- 实现:inception_v4.py
- 论文:
Inception-v4, Inception-ResNet和残差连接对学习的影响
- https://arxiv.org/abs/1602.07261 - 代码:https://github.com/Cadene/pretrained-models.pytorch
- 参考代码:https://github.com/tensorflow/models/tree/master/research/slim/nets
Inception-ResNet-V2
- 实现: inception_resnet_v2.py
- 论文:
Inception-v4, Inception-ResNet和残差连接对学习的影响
- https://arxiv.org/abs/1602.07261 - 代码:https://github.com/Cadene/pretrained-models.pytorch
- 参考代码:https://github.com/tensorflow/models/tree/master/research/slim/nets
NASNet-A
- 实现: nasnet.py
- 论文:
可扩展图像识别的可迁移架构学习
- https://arxiv.org/abs/1707.07012 - 代码:https://github.com/Cadene/pretrained-models.pytorch
- 参考代码: https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet
PNasNet-5
- 实现: pnasnet.py
- 论文:
渐进式神经网络架构搜索
- https://arxiv.org/abs/1712.00559 - 代码:https://github.com/Cadene/pretrained-models.pytorch
- 参考代码: https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet
EfficientNet
- 实现:efficientnet.py
- 论文
- EfficientNet NoisyStudent (B0-B7, L2) - https://arxiv.org/abs/1911.04252
- EfficientNet AdvProp (B0-B8) - https://arxiv.org/abs/1911.09665
- EfficientNet (B0-B7) - https://arxiv.org/abs/1905.11946
- EfficientNet-EdgeTPU (S, M, L) - https://ai.googleblog.com/2019/08/efficientnet-edgetpu-creating.html
- MixNet - https://arxiv.org/abs/1907.09595
- MNASNet B1, A1 (Squeeze-Excite), and Small - https://arxiv.org/abs/1807.11626
- MobileNet-V2 - https://arxiv.org/abs/1801.04381
- FBNet-C - https://arxiv.org/abs/1812.03443
- Single-Path NAS - https://arxiv.org/abs/1904.02877
- 我的PyTorch代码:https://github.com/rwightman/gen-efficientnet-pytorch
- 参考代码:https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet
MobileNet-V3
- 实现:mobilenetv3.py
- 论文:
Searching for MobileNetV3
- https://arxiv.org/abs/1905.02244 - 参考代码:https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
RegNet
- 实现:regnet.py
- 论文:
设计网络设计空间
- https://arxiv.org/abs/2003.13678 - 参考代码:https://github.com/facebookresearch/pycls/blob/master/pycls/models/regnet.py
RepVGG
- 实现:byobnet.py
- 论文:
让VGG风格的卷积神经网络重焕新生
- https://arxiv.org/abs/2101.03697 - 参考代码:https://github.com/DingXiaoH/RepVGG
ResNet, ResNeXt
实现:resnet.py
ResNet (V1B)
- 论文:
深度残差学习用于图像识别
- https://arxiv.org/abs/1512.03385 - 代码:[https://github.com/pytorch/vision/tree/master/torchvision/models](https://github.com/pytorch/vision/tree/master/torchvision/models)
- 论文:
ResNeXt
- 论文:
深度神经网络中的聚合残差变换
- https://arxiv.org/abs/1611.05431 - 代码:[https://github.com/pytorch/vision/tree/master/torchvision/models](https://github.com/pytorch/vision/tree/master/torchvision/models)
- 论文:
‘技巧包’/ Gluon C, D, E, S ResNet变体
Instagram预训练/ ImageNet调优的ResNeXt101
- 论文:
探索弱监督预训练的极限
- https://arxiv.org/abs/1805.00932 - 权重:https://pytorch.ac.cn/hub/facebookresearch_WSL-Images_resnext(注意:CC BY-NC 4.0许可,非商业友好型)
- 论文:
半监督(SSL)/半弱监督(SWSL)的ResNet和ResNeXts
- 论文:
用于图像分类的十亿规模半监督学习
- https://arxiv.org/abs/1905.00546 - 权重:https://github.com/facebookresearch/semi-supervised-ImageNet1K-models(注意:CC BY-NC 4.0许可,非商业友好型)
- 论文:
感知与激励网络(Squeeze-and-Excitation Networks)
- 论文:
压缩与激励网络
- https://arxiv.org/abs/1709.01507 - 代码:添加到ResNet基础版本,这将是未来的当前版本,旧的
senet.py
将被弃用
- 论文:
ECAResNet (ECA-Net)
- 论文:
ECA-Net:用于深度CNN的效率通道注意力
- https://arxiv.org/abs/1910.03151v4 - 代码:添加到ResNet基础版本,ECA模块由@VRandme贡献,参考https://github.com/BangguWu/ECANet
- 论文:
Res2Net
- 实现:res2net.py
- 论文:
Res2Net: A New Multi-scale Backbone Architecture
- https://arxiv.org/abs/1904.01169 - 代码:https://github.com/gasvn/Res2Net
ResNeSt
- 实现:resnest.py
- 论文:
ResNeSt: Split-Attention Networks
- https://arxiv.org/abs/2004.08955 - 代码:https://github.com/zhanghang1989/ResNeSt
ReXNet
- 实现:rexnet.py
- 论文:
ReXNet: Diminishing Representational Bottleneck on CNN
- https://arxiv.org/abs/2007.00992 - 代码:https://github.com/clovaai/rexnet
选择性核网络
- 实现:sknet.py
- 论文:
选择性核网络
- https://arxiv.org/abs/1903.06586 - 代码:https://github.com/implus/SKNet,https://github.com/clovaai/assembled-cnn
SelecSLS
- 实现:selecsls.py
- 论文:
XNect: Real-time Multi-Person 3D Motion Capture with a Single RGB Camera
- https://arxiv.org/abs/1907.00837 - 代码:https://github.com/mehtadushy/SelecSLS-Pytorch
squeeze-and-excitation-networks
实现:senet.py 备注:我将废弃这个版本的网络,新的版本包含在
resnet.py
中。论文:
压缩与激励网络
- https://arxiv.org/abs/1709.01507
TResNet
- 实现:tresnet.py
- 论文:《TResNet:针对GPU的高性能架构》 - https://arxiv.org/abs/2003.13630
- 代码:https://github.com/mrT23/TResNet
VGG
- 实现:vgg.py
- 论文:《非常深的卷积神经网络用于大规模图像识别》 - https://arxiv.org/pdf/1409.1556.pdf
- 参考代码:https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py
视觉Transformer
- 实现:vision_transformer.py
- 论文:《一张图片等于16x16词:大规模图像识别中的变换器》 - https://arxiv.org/abs/2010.11929
- 参考代码和预训练权重:https://github.com/google-research/vision_transformer
VovNet V2 和 V1
- 实现:vovnet.py
- 论文:《CenterMask:基于锚点的实时实例分割》- https://arxiv.org/abs/1911.06667
- 参考代码:https://github.com/youngwanLEE/vovnet-detectron2
Xception
- 实现:xception.py
- 论文:《Xception:使用深度可分离卷积的深度学习》- https://arxiv.org/abs/1610.02357
- 代码:https://github.com/Cadene/pretrained-models.pytorch
Xception (修改后,对齐,Gluon)
- 实现:gluon_xception.py
- 论文:《带有膨胀可分离卷积的编码器-解码器用于语义图像分割》- https://arxiv.org/abs/1802.02611
- 参考代码:https://github.com/dmlc/gluon-cv/tree/master/gluoncv/model_zoo,https://github.com/jfzhang95/pytorch-deeplab-xception/
Xception (修改后的对齐,TF)
- 实现:aligned_xception.py
- 论文:《带有膨胀可分离卷积的编码器-解码器用于语义图像分割》- https://arxiv.org/abs/1802.02611
- 参考代码:https://github.com/tensorflow/models/tree/master/research/deeplab