Hugging Face's logo

timm documentation

BYOBNet（GENet、RepVGG、MobileOne 等）

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

BYOBNet（GENet、RepVGG、MobileOne 等）

类别： CNN · 代码： timm/models/byobnet.py · 架构： 53 · 预训练权重： 57（其中 41 个包含 ImageNet-1k 结果）

Bring-Your-Own-Blocks 网络是一种灵活的架构，它基于 dataclass 配置，将 blocks 堆叠成 stages。它实现了 GPU-Efficient Networks（gernet_*）、RepVGG、MobileOne、RegNetZ、CLIP 的 ResNet 图像塔（*_clip）以及多种 ResNet 风格的 timm 变体。

论文

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('repvgg_a2.rvgg_in1k', pretrained=True)
>>> data_config = timm.data.resolve_model_data_config(model)

模型 API 支持

功能ByobNet
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（ByobNet 使用 mobileone_s1，尺寸为 224）进行验证，预期同一类的其他架构表现相同。有关详细信息，请参阅 模型 API 页面。

架构

架构参数（M）GMAC输入大小预训练权重
bat_resnext26ts10.732.532561
eca_resnet33ts19.684.762561
eca_resnext26ts10.302.432561
gcresnet33ts19.884.762561
gcresnet50t25.905.422561
gcresnext26ts10.482.432561
gcresnext50ts15.673.752561
gernet_l31.084.572561
gernet_m21.143.022241
gernet_s8.170.752241
mobileone_s02.080.282241
mobileone_s14.760.832241
mobileone_s27.811.302241
mobileone_s310.081.902241
mobileone_s414.842.982241
regnetz_b169.721.452241
regnetz_b16_evos9.741.432240
regnetz_c1613.462.512561
regnetz_c16_evos13.492.482561
regnetz_d823.373.972561
regnetz_d8_evos23.464.502561
regnetz_d3227.585.982561
regnetz_e857.709.912561
repvgg_a08.311.362241
repvgg_a112.792.362241
repvgg_a225.505.122241
repvgg_b014.343.062241
repvgg_b151.8311.822241
repvgg_b1g436.137.312241
repvgg_b280.3218.382241
repvgg_b2g455.7811.332241
repvgg_b3110.9626.212241
repvgg_b3g475.6316.062241
repvgg_d2se120.3966.993201
resnet32ts17.964.632561
resnet33ts19.684.762561
resnet50_clip38.326.142243
resnet50_clip_gap23.535.392243
resnet50_mlp26.657.052560
resnet50x4_clip87.1421.352881
resnet50x4_clip_gap65.6219.572881
resnet50x16_clip167.3374.903841
resnet50x16_clip_gap136.2070.323841
resnet50x64_clip420.38265.024481
resnet50x64_clip_gap365.03253.964481
resnet51q35.706.382561
resnet61q36.857.802561
resnet101_clip56.269.812242
resnet101_clip_gap42.529.112242
resnext26ts10.302.432561
seresnet33ts19.784.762561
seresnext26ts10.392.432561
test_byobnet0.460.031601

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。

rvgg 标签

权重输入大小Top-1（%）评估大小类许可证Hub
repvgg_a0.rvgg_in1k22472.4262241000mittimm
repvgg_a1.rvgg_in1k22474.4742241000mittimm
repvgg_a2.rvgg_in1k22476.4882241000mittimm
repvgg_b0.rvgg_in1k22475.1582241000mittimm
repvgg_b1.rvgg_in1k22478.3702241000mittimm
repvgg_b1g4.rvgg_in1k22477.6082241000mittimm
repvgg_b2.rvgg_in1k22478.7882241000mittimm
repvgg_b2g4.rvgg_in1k22479.3962241000mittimm
repvgg_b3.rvgg_in1k22480.5282241000mittimm
repvgg_b3g4.rvgg_in1k22480.2242241000mittimm
repvgg_d2se.rvgg_in1k32083.5503201000mittimm

openai 标签

权重输入大小Top-1（%）评估大小类许可证Hub
resnet50_clip.openai224--1024mittimm
resnet50_clip_gap.openai224--nonemittimm
resnet50x4_clip.openai288--640mittimm
resnet50x4_clip_gap.openai288--nonemittimm
resnet50x16_clip.openai384--768mittimm
resnet50x16_clip_gap.openai384--nonemittimm
resnet50x64_clip.openai448--1024mittimm
resnet50x64_clip_gap.openai448--nonemittimm
resnet101_clip.openai224--512mittimm
resnet101_clip_gap.openai224--nonemittimm

ra2 标签

权重输入大小Top-1（%）评估大小类许可证Hub
eca_resnet33ts.ra2_in1k256（测试 288）80.7162881000apache-2.0timm
gcresnet33ts.ra2_in1k256（测试 288）80.5882881000apache-2.0timm
gcresnet50t.ra2_in1k256（测试 288）81.4582881000apache-2.0timm
resnet32ts.ra2_in1k256（测试 288）79.3882881000apache-2.0timm
resnet33ts.ra2_in1k256（测试 288）79.7102881000apache-2.0timm
resnet51q.ra2_in1k256（测试 288）82.3562881000apache-2.0timm
resnet61q.ra2_in1k256（测试 288）82.5322881000apache-2.0timm
resnext26ts.ra2_in1k256（测试 288）77.1862881000apache-2.0timm
seresnet33ts.ra2_in1k256（测试 288）80.8142881000apache-2.0timm

ch 标签

权重输入大小Top-1（%）评估大小类许可证Hub
bat_resnext26ts.ch_in1k25678.2802561000apache-2.0timm
eca_resnext26ts.ch_in1k256（测试 288）78.0202881000apache-2.0timm
gcresnext26ts.ch_in1k256（测试 288）78.4362881000apache-2.0timm
gcresnext50ts.ch_in1k256（测试 288）81.2422881000apache-2.0timm
regnetz_c16_evos.ch_in1k256（测试 320）82.6603201000apache-2.0timm
regnetz_d8_evos.ch_in1k256（测试 320）84.1383201000apache-2.0timm
seresnext26ts.ch_in1k256（测试 288）78.2922881000apache-2.0timm

apple 标签

权重输入大小Top-1（%）评估大小类许可证Hub
mobileone_s0.apple_in1k22471.3982241000mobileone-licensetimm
mobileone_s1.apple_in1k22475.7602241000mobileone-licensetimm
mobileone_s2.apple_in1k22477.5042241000mobileone-licensetimm
mobileone_s3.apple_in1k22478.0022241000mobileone-licensetimm
mobileone_s4.apple_in1k22479.4462241000mobileone-licensetimm

ra3 标签

权重输入大小Top-1（%）评估大小类许可证Hub
regnetz_b16.ra3_in1k224（测试 288）80.7402881000apache-2.0timm
regnetz_c16.ra3_in1k256（测试 320）82.6623201000apache-2.0timm
regnetz_d8.ra3_in1k256（测试 320）84.0443201000apache-2.0timm
regnetz_d32.ra3_in1k256（测试 320）84.0283201000apache-2.0timm
regnetz_e8.ra3_in1k256（测试 320）85.0163201000apache-2.0timm

yfcc15m 标签

权重输入大小Top-1（%）评估大小类许可证Hub
resnet50_clip.yfcc15m224--1024mittimm
resnet50_clip_gap.yfcc15m224--nonemittimm
resnet101_clip.yfcc15m224--512mittimm
resnet101_clip_gap.yfcc15m224--nonemittimm

idstcv 标签

权重输入大小Top-1（%）评估大小类许可证Hub
gernet_l.idstcv_in1k25681.3542561000apache-2.0timm
gernet_m.idstcv_in1k22480.7022241000apache-2.0timm
gernet_s.idstcv_in1k22476.8742241000apache-2.0timm

cc12m 标签

权重输入大小Top-1（%）评估大小类许可证Hub
resnet50_clip.cc12m224--1024mittimm
resnet50_clip_gap.cc12m224--nonemittimm

其他

权重输入大小Top-1（%）评估大小类许可证Hub
test_byobnet.r160_in1k16045.8581601000apache-2.0timm
Update on GitHub

←对齐的 Xception ConvMixer→