Hugging Face's logo

timm documentation

ResNet

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

ResNet

类别： CNN · 代码： timm/models/resnet.py · 架构： 81 · 预训练权重： 150（其中 144 个包含 ImageNet-1k 结果）

残差网络（ResNet）以及基于同一 ResNet 类构建的变体：ResNeXt 分组卷积、Squeeze-and-Excitation（SE）和 ECA 通道注意力、深层/分层 stem 以及平均池化下采样（“c”、“d”、“s” 和 “t” 变体）、抗锯齿下采样、Wide ResNet 和 ResNet-RS 配置。

论文

由预训练权重引用：

用法

请参阅快速入门了解如何加载预训练权重、进行图像预处理和微调，以及查阅特征提取了解如何将模型用作骨干网络。

>>> import timm
>>> model = timm.create_model('resnet50.a1_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)

模型 API 支持

功能ResNet
features_only=True 特征图✓（步长 2、4、8、16、32）
forward_intermediates()✓
prune_intermediate_layers()✓
任意输入尺寸，无需更改✓
set_input_size()–
dynamic_img_size=True–
set_grad_checkpointing()✓
逐层学习率衰减（group_matcher()）✓
reset_classifier()✓

✓ 支持，✓ (square) 仅支持方形尺寸，✗ 不支持，– 无需支持（模型本身已接受任意输入尺寸）。通过在每个功能上运行随机初始化的模型（ResNet 使用 resnet10t，尺寸为 176）进行验证，预期同一类的其他架构表现相同。详情请参阅 Model API 页面。

架构

架构参数（M）GMAC输入大小预训练权重
ecaresnet26t16.013.352561
ecaresnet50d25.584.352241
ecaresnet50d_pruned19.942.532241
ecaresnet50t25.575.642564
ecaresnet101d44.578.082241
ecaresnet101d_pruned24.883.482241
ecaresnet200d64.6920.002560
ecaresnet269d102.0941.533201
ecaresnetlight30.164.112241
ecaresnext26t_32x4d15.412.702240
ecaresnext50t_32x4d25.05-2240
resnet10t5.440.701761
resnet14t10.081.071761
resnet1811.691.822247
resnet18d11.712.062242
resnet2616.002.362241
resnet26d16.012.602241
resnet26t16.013.352561
resnet3421.803.672247
resnet34d21.823.912241
resnet5025.564.1122419
resnet50_gn25.564.142241
resnet50c25.584.352241
resnet50d25.584.352246
resnet50s25.685.472241
resnet50t25.574.322240
resnet10144.557.832247
resnet101c44.578.082241
resnet101d44.5710.552562
resnet101s44.679.192241
resnet15260.1911.562247
resnet152c60.2111.802241
resnet152d60.2115.412562
resnet152s60.3212.922241
resnet20064.6715.072240
resnet200d64.6920.002561
resnetaa34d21.824.432240
resnetaa5025.565.152241
resnetaa50d25.585.392243
resnetaa101d44.579.122242
resnetblur1811.692.342240
resnetblur5025.565.162241
resnetblur50d25.585.402240
resnetblur101d44.579.122240
resnetrs5035.692.291601
resnetrs10163.626.041921
resnetrs15286.6215.592561
resnetrs20093.2120.182561
resnetrs270129.8627.062561
resnetrs350163.9643.672881
resnetrs420191.8964.203201
resnext50_32x4d25.034.2622410
resnext50d_32x4d25.054.502241
resnext101_32x4d44.188.012243
resnext101_32x8d88.7916.482245
resnext101_32x16d194.0336.272243
resnext101_32x32d468.5387.292241
resnext101_64x4d83.4615.522243
senet154115.0920.772241
seresnet1811.781.822240
seresnet3421.963.672240
seresnet5028.094.112244
seresnet50t28.104.322240
seresnet10149.337.842240
seresnet15266.8211.572240
seresnet152d66.8415.422561
seresnet200d71.8620.012560
seresnet269d113.6726.592560
seresnetaa50d28.115.402240
seresnext26d_32x4d16.812.732241
seresnext26t_32x4d16.812.702241
seresnext50_32x4d27.564.262242
seresnext101_32x4d48.968.022241
seresnext101_32x8d93.5716.482241
seresnext101_64x4d88.2315.532241
seresnext101d_32x8d93.5916.722241
seresnextaa101d_32x8d93.5928.512884
seresnextaa201d_32x8d149.39101.113842
test_resnet0.470.101601
wide_resnet50_268.8811.432243
wide_resnet101_2126.8922.802242

GMACs 和参数量对应于每种架构的默认输入大小和分类器，数据来自基准测试结果。

预训练权重

Top-1 是 results-imagenet.csv 中在所列评估规模下的 ImageNet-1k 验证准确率，- 表示该权重没有 ImageNet-1k 结果（例如预训练权重或图文权重）。Classes 是预训练分类器的大小，对于没有分类器的权重则为 none。

权重按来源分组：预训练配置中给出的论文或来源，否则为 timm 以外的 Hub 组织，否则为预训练标签的第一部分（architecture.tag）。仅包含单个权重的来源列在 other 下。

ResNet 强势回归

预训练配置参考了 ResNet strikes back: An improved training procedure in timm。

权重输入大小Top-1（%）评估大小类许可证Hub
ecaresnet50t.a1_in1k224（测试 288）82.1102881000apache-2.0timm
ecaresnet50t.a2_in1k224（测试 288）81.6822881000apache-2.0timm
ecaresnet50t.a3_in1k160（测试 224）79.5482241000apache-2.0timm
resnet18.a1_in1k224（测试 288）73.1642881000apache-2.0timm
resnet18.a2_in1k224（测试 288）72.3642881000apache-2.0timm
resnet18.a3_in1k160（测试 224）68.2502241000apache-2.0timm
resnet18d.ra4_e3600_r224_in1k224（测试 288）76.0122881000apache-2.0timm
resnet34.a1_in1k224（测试 288）77.9302881000apache-2.0timm
resnet34.a2_in1k224（测试 288）77.1502881000apache-2.0timm
resnet34.a3_in1k160（测试 224）72.9962241000apache-2.0timm
resnet34.ra4_e3600_r224_in1k224（测试 288）78.9722881000apache-2.0timm
resnet50.a1_in1k224（测试 288）81.2402881000apache-2.0timm
resnet50.a1h_in1k176（测试 224）80.6542241000apache-2.0timm
resnet50.a2_in1k224（测试 288）80.7742881000apache-2.0timm
resnet50.a3_in1k160（测试 224）78.0522241000apache-2.0timm
resnet50.b1k_in1k224（测试 288）80.7082881000apache-2.0timm
resnet50.b2k_in1k224（测试 288）80.4662881000apache-2.0timm
resnet50.c1_in1k224（测试 288）80.9282881000apache-2.0timm
resnet50.c2_in1k224（测试 288）80.8622881000apache-2.0timm
resnet50.d_in1k224（测试 288）81.0122881000apache-2.0timm
resnet50d.a1_in1k224（测试 288）81.4662881000apache-2.0timm
resnet50d.a2_in1k224（测试 288）81.1642881000apache-2.0timm
resnet50d.a3_in1k160（测试 224）78.7422241000apache-2.0timm
resnet50d.ra4_e3600_r224_in1k224（测试 288）81.8402881000apache-2.0timm
resnet101.a1_in1k224（测试 288）82.3262881000apache-2.0timm
resnet101.a1h_in1k224（测试 288）82.7962881000apache-2.0timm
resnet101.a2_in1k224（测试 288）82.2182881000apache-2.0timm
resnet101.a3_in1k160（测试 224）79.8362241000apache-2.0timm
resnet152.a1_in1k224（测试 288）82.7422881000apache-2.0timm
resnet152.a1h_in1k224（测试 288）83.4942881000apache-2.0timm
resnet152.a2_in1k224（测试 288）82.6142881000apache-2.0timm
resnet152.a3_in1k160（测试 224）80.5482241000apache-2.0timm
resnetaa50.a1h_in1k224（测试 288）81.6182881000apache-2.0timm
resnext50_32x4d.a1_in1k224（测试 288）81.4802881000apache-2.0timm
resnext50_32x4d.a1h_in1k224（测试 288）82.0162881000apache-2.0timm
resnext50_32x4d.a2_in1k224（测试 288）81.3042881000apache-2.0timm
resnext50_32x4d.a3_in1k160（测试 224）79.2582241000apache-2.0timm
resnext101_64x4d.c1_in1k224（测试 288）83.1582881000apache-2.0timm
seresnet50.a1_in1k224（测试 288）81.1142881000apache-2.0timm
seresnet50.a2_in1k224（测试 288）81.0942881000apache-2.0timm
seresnet50.a3_in1k160（测试 224）77.0362241000apache-2.0timm
seresnext101_32x8d.ah_in1k224（测试 288）84.1922881000apache-2.0timm
seresnext101d_32x8d.ah_in1k224（测试 288）84.3802881000apache-2.0timm
seresnextaa101d_32x8d.ah_in1k224（测试 288）84.5802881000apache-2.0timm

huggingface/pytorch-image-models

来源：github.com/huggingface/pytorch-image-models

权重输入大小Top-1（%）评估大小类许可证Hub
ecaresnet26t.ra2_in1k256（测试 320）79.9003201000apache-2.0timm
ecaresnet269d.ra2_in1k320（测试 352）84.9683521000apache-2.0timm
resnet10t.c3_in1k176（测试 224）68.3542241000apache-2.0timm
resnet14t.c3_in1k176（测试 224）72.2422241000apache-2.0timm
resnet18d.ra2_in1k224（测试 288）73.7902881000apache-2.0timm
resnet26.bt_in1k224（测试 288）76.3842881000apache-2.0timm
resnet26d.bt_in1k224（测试 288）77.4222881000apache-2.0timm
resnet26t.ra2_in1k256（测试 320）78.3303201000apache-2.0timm
resnet34.bt_in1k224（测试 288）76.5062881000apache-2.0timm
resnet34d.ra2_in1k224（测试 288）78.4402881000apache-2.0timm
resnet50.bt_in1k224（测试 288）79.6522881000apache-2.0timm
resnet50.ra_in1k224（测试 288）79.8642881000apache-2.0timm
resnet50.ram_in1k224（测试 288）79.9822881000apache-2.0timm
resnet50_gn.a1h_in1k224（测试 288）81.2162881000apache-2.0timm
resnet50d.ra2_in1k224（测试 288）81.3542881000apache-2.0timm
resnet101d.ra2_in1k256（测试 320）83.0203201000apache-2.0timm
resnet152d.ra2_in1k256（测试 320）83.6843201000apache-2.0timm
resnet200d.ra2_in1k256（测试 320）83.9763201000apache-2.0timm
resnetaa50d.d_in12k224（测试 288）--11821apache-2.0timm
resnetaa50d.sw_in12k224（测试 288）--11821apache-2.0timm
resnetaa50d.sw_in12k_ft_in1k224（测试 288）82.6562881000apache-2.0timm
resnetaa101d.sw_in12k224（测试 288）--11821apache-2.0timm
resnetaa101d.sw_in12k_ft_in1k224（测试 288）84.1142881000apache-2.0timm
resnetblur50.bt_in1k224（测试 288）80.2482881000apache-2.0timm
resnext50_32x4d.ra_in1k224（测试 288）80.7202881000apache-2.0timm
resnext50d_32x4d.bt_in1k224（测试 288）80.6762881000apache-2.0timm
seresnet50.ra2_in1k224（测试 288）81.2882881000apache-2.0timm
seresnet152d.ra2_in1k256（测试 320）84.3443201000apache-2.0timm
seresnext26d_32x4d.bt_in1k224（测试 288）78.8162881000apache-2.0timm
seresnext26t_32x4d.bt_in1k224（测试 288）78.7362881000apache-2.0timm
seresnext50_32x4d.racm_in1k224（测试 288）82.1862881000apache-2.0timm
seresnextaa101d_32x8d.sw_in12k224（测试 288）--11821apache-2.0timm
seresnextaa101d_32x8d.sw_in12k_ft_in1k224（测试 288）86.5242881000apache-2.0timm
seresnextaa101d_32x8d.sw_in12k_ft_in1k_288288（测试 320）86.7223201000apache-2.0timm
wide_resnet50_2.racm_in1k224（测试 288）82.2602881000apache-2.0timm

cv.gluon.ai

来源：cv.gluon.ai/model_zoo/classification.html

权重输入大小Top-1（%）评估大小类许可证Hub
resnet18.gluon_in1k22470.8442241000apache-2.0timm
resnet34.gluon_in1k22474.5742241000apache-2.0timm
resnet50.gluon_in1k22477.5802241000apache-2.0timm
resnet50c.gluon_in1k22478.0162241000apache-2.0timm
resnet50d.gluon_in1k22479.0802241000apache-2.0timm
resnet50s.gluon_in1k22478.7102241000apache-2.0timm
resnet101.gluon_in1k22479.3062241000apache-2.0timm
resnet101c.gluon_in1k22479.5422241000apache-2.0timm
resnet101d.gluon_in1k22480.4662241000apache-2.0timm
resnet101s.gluon_in1k22480.2782241000apache-2.0timm
resnet152.gluon_in1k22479.7122241000apache-2.0timm
resnet152c.gluon_in1k22479.9042241000apache-2.0timm
resnet152d.gluon_in1k22480.4762241000apache-2.0timm
resnet152s.gluon_in1k22481.0482241000apache-2.0timm
resnext50_32x4d.gluon_in1k22479.3542241000apache-2.0timm
resnext101_32x4d.gluon_in1k22480.3362241000apache-2.0timm
resnext101_64x4d.gluon_in1k22480.6402241000apache-2.0timm
senet154.gluon_in1k22481.2582241000apache-2.0timm
seresnext50_32x4d.gluon_in1k22479.9162241000apache-2.0timm
seresnext101_32x4d.gluon_in1k22480.8862241000apache-2.0timm
seresnext101_64x4d.gluon_in1k22480.8842241000apache-2.0timm

pytorch/vision

来源：github.com/pytorch/vision

权重输入大小Top-1（%）评估大小类许可证Hub
resnet18.tv_in1k22469.7522241000bsd-3-clausetimm
resnet34.tv_in1k22473.2982241000bsd-3-clausetimm
resnet50.tv2_in1k176（测试 224）80.8462241000bsd-3-clausetimm
resnet50.tv_in1k22476.1562241000bsd-3-clausetimm
resnet101.tv2_in1k176（测试 224）81.9062241000bsd-3-clausetimm
resnet101.tv_in1k22477.4042241000bsd-3-clausetimm
resnet152.tv2_in1k176（测试 224）82.2942241000bsd-3-clausetimm
resnet152.tv_in1k22478.3302241000bsd-3-clausetimm
resnext50_32x4d.tv2_in1k176（测试 224）81.1902241000bsd-3-clausetimm
resnext50_32x4d.tv_in1k22477.6322241000bsd-3-clausetimm
resnext101_32x8d.tv2_in1k176（测试 224）82.8322241000bsd-3-clausetimm
resnext101_32x8d.tv_in1k22479.3082241000bsd-3-clausetimm
resnext101_64x4d.tv_in1k22482.9782241000bsd-3-clausetimm
wide_resnet50_2.tv2_in1k176（测试 224）81.6122241000bsd-3-clausetimm
wide_resnet50_2.tv_in1k22478.4862241000bsd-3-clausetimm
wide_resnet101_2.tv2_in1k176（测试 224）82.5142241000bsd-3-clausetimm
wide_resnet101_2.tv_in1k22478.8622241000bsd-3-clausetimm

facebookresearch/semi-supervised-ImageNet1K-models

来源：github.com/facebookresearch/semi-supervised-ImageNet1K-models

权重输入大小Top-1（%）评估大小类许可证Hub
resnet18.fb_ssl_yfcc100m_ft_in1k22472.6362241000cc-by-nc-4.0timm
resnet18.fb_swsl_ig1b_ft_in1k22473.2862241000cc-by-nc-4.0timm
resnet50.fb_ssl_yfcc100m_ft_in1k22479.2522241000cc-by-nc-4.0timm
resnet50.fb_swsl_ig1b_ft_in1k22481.1462241000cc-by-nc-4.0timm
resnext50_32x4d.fb_ssl_yfcc100m_ft_in1k22480.3162241000cc-by-nc-4.0timm
resnext50_32x4d.fb_swsl_ig1b_ft_in1k22482.1622241000cc-by-nc-4.0timm
resnext101_32x4d.fb_ssl_yfcc100m_ft_in1k22480.9362241000cc-by-nc-4.0timm
resnext101_32x4d.fb_swsl_ig1b_ft_in1k22483.2682241000cc-by-nc-4.0timm
resnext101_32x8d.fb_ssl_yfcc100m_ft_in1k22481.6202241000cc-by-nc-4.0timm
resnext101_32x8d.fb_swsl_ig1b_ft_in1k22484.2762241000cc-by-nc-4.0timm
resnext101_32x16d.fb_ssl_yfcc100m_ft_in1k22481.8562241000cc-by-nc-4.0timm
resnext101_32x16d.fb_swsl_ig1b_ft_in1k22483.3402241000cc-by-nc-4.0timm

tf 标签

权重输入大小Top-1（%）评估大小类许可证Hub
resnetrs50.tf_in1k160（测试 224）79.9342241000apache-2.0timm
resnetrs101.tf_in1k192（测试 288）82.2882881000apache-2.0timm
resnetrs152.tf_in1k256（测试 320）83.7103201000apache-2.0timm
resnetrs200.tf_in1k256（测试 320）84.4483201000apache-2.0timm
resnetrs270.tf_in1k256（测试 352）84.4383521000apache-2.0timm
resnetrs350.tf_in1k288（测试 384）84.7103841000apache-2.0timm
resnetrs420.tf_in1k320（测试 416）85.0104161000apache-2.0timm

miil 标签

权重输入大小Top-1（%）评估大小类许可证Hub
ecaresnet50d.miil_in1k224（测试 288）81.6482881000apache-2.0timm
ecaresnet50d_pruned.miil_in1k224（测试 288）80.8082881000apache-2.0timm
ecaresnet101d.miil_in1k224（测试 288）83.0062881000apache-2.0timm
ecaresnet101d_pruned.miil_in1k224（测试 288）82.0162881000apache-2.0timm
ecaresnetlight.miil_in1k224（测试 288）81.4302881000apache-2.0timm

facebookresearch/WSL-Images

来源：github.com/facebookresearch/WSL-Images

权重输入大小Top-1（%）评估大小类许可证Hub
resnext101_32x8d.fb_wsl_ig1b_ft_in1k22482.7202241000cc-by-nc-4.0timm
resnext101_32x16d.fb_wsl_ig1b_ft_in1k22484.1662241000cc-by-nc-4.0timm
resnext101_32x32d.fb_wsl_ig1b_ft_in1k22485.0902241000cc-by-nc-4.0timm

sw 标签

权重输入大小Top-1（%）评估大小类许可证Hub
seresnextaa201d_32x8d.sw_in12k320（测试 384）--11821apache-2.0timm
seresnextaa201d_32x8d.sw_in12k_ft_in1k_38438487.3043841000apache-2.0timm

其他

权重输入大小Top-1（%）评估大小类许可证Hub
ecaresnet50t.ra2_in1k256（测试 320）82.3703201000apache-2.0timm
resnet50.am_in1k22479.0042241000apache-2.0timm
resnet50.bamboo224--none-timm
test_resnet.r160_in1k16041.6021601000apache-2.0timm
Update on GitHub

←ResNeSt ResNetV2 (BiT)→