Awesome backbones for image classification

写在前面

• 若训练效果不佳，首先需要调整学习率和Batch size，这俩超参很大程度上影响收敛。其次，从关闭图像增强手段（尤其小数据集）开始，有的图像增强方法会污染数据，如

  如何去除增强？如efficientnetv2-b0配置文件中train_pipeline可更改为如下

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='RandomResizedCrop',
        size=192,
        efficientnet_style=True,
        interpolation='bicubic'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='ImageToTensor', keys=['img']),
    dict(type='ToTensor', keys=['gt_label']),
    dict(type='Collect', keys=['img', 'gt_label'])
]

  若你的数据集提前已经将shape更改为网络要求的尺寸，那么Resize操作也可以去除。

更新日志

2023.12.02

• 新增Issue中多人提及的输出Train Acc与Val loss

  • metrics_outputs.csv保存每周期train_loss, train_acc, train_precision, train_recall, train_f1-score, val_loss, val_acc, val_precision, val_recall, val_f1-score方便各位绘图
  • 终端由原先仅输出Val相关metrics升级为Train与Val都输出

  

2023.08.05

• 新增TinyViT(预训练权重不匹配)、DeiT3、EdgeNeXt、RevVisionTransformer

2023.03.07

• 新增MobileViT、DaViT、RepLKNet、BEiT、EVA、MixMIM、EfficientNetV2

2022.11.20

• 新增是否将测试集用作验证集选项，若不使用，从训练集按ratio划分验证集数量，随机从训练集某fold挑选作为验证集（类似k-fold但不是，可自己稍改达到k-fold目的），详见Training tutorial

2022.11.06

• 新增HorNet, EfficientFormer, SwinTransformer V2, MViT模型

测试环境

• Pytorch 1.7.1+
• Python 3.6+

资料

数据集	视频教程	人工智能技术探讨群
花卉数据集 提取码：0zat	点我跳转	1群：78174903 3群：584723646

快速开始

• 遵循环境搭建完成配置
• 下载MobileNetV3-Small权重至datas下
• Awesome-Backbones文件夹下终端输入

python tools/single_test.py datas/cat-dog.png models/mobilenet/mobilenet_v3_small.py --classes-map datas/imageNet1kAnnotation.txt

教程

• 环境搭建
• 数据集准备
• 配置文件解释
• 训练
• 模型评估
• 计算Flops&Params
• 添加新的模型组件
• 类别激活图可视化
• 学习率策略可视化
• 数据增强策略可视化

模型

• LeNet5
• AlexNet
• VGG
• DenseNet
• ResNet
• Wide-ResNet
• ResNeXt
• SEResNet
• SEResNeXt
• RegNet
• MobileNetV2
• MobileNetV3
• ShuffleNetV1
• ShuffleNetV2
• EfficientNet
• RepVGG
• Res2Net
• ConvNeXt
• HRNet
• ConvMixer
• CSPNet
• Swin-Transformer
• Vision-Transformer
• Transformer-in-Transformer
• MLP-Mixer
• DeiT
• Conformer
• T2T-ViT
• Twins
• PoolFormer
• VAN
• HorNet
• EfficientFormer
• Swin Transformer V2
• MViT V2
• MobileViT
• DaViT
• replknet
• BEiT
• EVA
• MixMIM
• EfficientNetV2

预训练权重

名称	权重	名称	权重	名称	权重
LeNet5	None	AlexNet	None	VGG	VGG-11 VGG-13 VGG-16 VGG-19 VGG-11-BN VGG-13-BN VGG-16-BN VGG-19-BN
ResNet	ResNet-18 ResNet-34 ResNet-50 ResNet-101 ResNet-152	ResNetV1C	ResNetV1C-50 ResNetV1C-101 ResNetV1C-152	ResNetV1D	ResNetV1D-50 ResNetV1D-101 ResNetV1D-152
ResNeXt	ResNeXt-50 ResNeXt-101 ResNeXt-152	SEResNet	SEResNet-50 SEResNet-101	SEResNeXt	None
RegNet	RegNetX-400MF RegNetX-800MF RegNetX-1.6GF RegNetX-3.2GF RegNetX-4.0GF RegNetX-6.4GF RegNetX-8.0GF RegNetX-12GF	MobileNetV2	MobileNetV2	MobileNetV3	MobileNetV3-Small MobileNetV3-Large
ShuffleNetV1	ShuffleNetV1	ShuffleNetV2	ShuffleNetV2	EfficientNet	EfficientNet-B0 EfficientNet-B1 EfficientNet-B2 EfficientNet-B3 EfficientNet-B4 EfficientNet-B5 EfficientNet-B6 EfficientNet-B7 EfficientNet-B8
RepVGG	RepVGG-A0 RepVGG-A1 RepVGG-A2 RepVGG-B0 RepVGG-B1 RepVGG-A1 RepVGG-B1g2 RepVGG-B1g4 RepVGG-B2 RepVGG-B2g4 RepVGG-B2g4 RepVGG-B3 RepVGG-B3g4 RepVGG-D2se	Res2Net	Res2Net-50-14w-8s Res2Net-50-26w-8s Res2Net-101-26w-4s	ConvNeXt	ConvNeXt-Tiny ConvNeXt-Small ConvNeXt-Base ConvNeXt-Large ConvNeXt-XLarge
HRNet	HRNet-W18 HRNet-W30 HRNet-W32 HRNet-W40 HRNet-W44 HRNet-W48 HRNet-W64	ConvMixer	ConvMixer-768/32 ConvMixer-1024/20 ConvMixer-1536/20	CSPNet	CSPDarkNet50 CSPResNet50 CSPResNeXt50
Swin Transformer	tiny-224 small-224 base-224 large-224 base-384 large-384	Vision Transformer	vit_base_p16_224 vit_base_p32_224 vit_large_p16_224 vit_base_p16_384 vit_base_p32_384 vit_large_p16_384	Transformer in Transformer	TNT-small
MLP Mixer	base_p16 large_p16	Deit	DeiT-tiny DeiT-tiny distilled DeiT-small DeiT-small distilled DeiT-base DeiT-base distilled DeiT-base 384px DeiT-base distilled 384px	Conformer	Conformer-tiny-p16 Conformer-small-p32 Conformer-small-p16 Conformer-base-p16
T2T-ViT	T2T-ViT_t-14 T2T-ViT_t-19 T2T-ViT_t-24	Twins	PCPVT-small PCPVT-base PCPVT-large SVT-small SVT-base SVT-large	PoolFormer	PoolFormer-S12 PoolFormer-S24 PoolFormer-S36 PoolFormer-M36 PoolFormer-M48
DenseNet	DenseNet121 DenseNet161 DenseNet169 DenseNet201	Visual Attention Network(VAN)	VAN-Tiny VAN-Small VAN-Base VAN-Large	Wide-ResNet	WRN-50 WRN-101
HorNet	HorNet-Tiny HorNet-Tiny-GF HorNet-Small HorNet-Small-GF HorNet-Base HorNet-Base-GF HorNet-Large HorNet-Large-GF HorNet-Large-GF384	EfficientFormer	efficientformer-l1 efficientformer-l3 efficientformer-l7	Swin Transformer v2	tiny-256 window 8 tiny-256 window 16 small-256 window 8 small-256 window 16 base-256 window 8 base-256 window 16 large-256 window 16 large-384 window 24
MViTv2	MViTv2-Tiny MViTv2-Small MViTv2-Base MViTv2-Large	MobileVit	MobileViT-XXSmall MobileViT-XSmall MobileViT-Small	DaViT	DaViT-T DaViT-S DaViT-B
RepLKNet	RepLKNet-31B-224 RepLKNet-31B-384 RepLKNet-31L-384 RepLKNet-XL	BEiT	BEiT-base	EVA	EVA-G-p14-224 EVA-G-p14-336 EVA-G-p14-560 EVA-G-p16-224 EVA-L-p14-224 EVA-L-p14-196 EVA-L-p14-336
MixMIM	mixmim-base	EfficientNetV2	EfficientNetV2-b0 EfficientNetV2-b1 EfficientNetV2-b2 EfficientNetV2-b3 EfficientNetV2-s EfficientNetV2-m EfficientNetV2-l EfficientNetV2-xl	DeiT3	deit3_small_p16 deit3_small_p16_384 deit3_base_p16 deit3_base_p16_384 deit3_medium_p16 deit3_large_p16 deit3_large_p16_384 deit3_huge_p16
EdgeNeXt	edgenext-base edgenext-small edgenext-X-small edgenext-XX-small	RevVisionTransformer	revvit-small revvit-base

我维护的其他项目

• 图片数据不够？我做了一款图像增强软件
• 一键转换与编辑图像标注文件软件，极大提高效率

参考

@repo{2020mmclassification,
    title={OpenMMLab's Image Classification Toolbox and Benchmark},
    author={MMClassification Contributors},
    howpublished = {\url{https://github.com/open-mmlab/mmclassification}},
    year={2020}
}