训练数据集必须全部标注的机器方法（mmselfSup训练自己的数据集）

mmselfsup训练自己的数据集

       最近在做自监督学习的东西    
    
    
，使用无标签数据做预训练模型     


     


     


  ，做个分享吧




 




 




 
，写的不好
   

   

   

，请见谅    
    
    
  。

        mmselfsup地址：https://github.com/open-mmlab/mmselfsup

       相关文档：Welcome to MMSelfSup’s documentation! — MMSelfSup 0.10.1 documentation

一    
    
    
  、环境搭建

1.创建虚拟环境

conda create --name openmmlab python=3.8 -y

激活虚拟环境：

conda activate openmmlab

2.安装pytorch


     


     


     
、torchvision

根据自己的配置安装相应版本

pip install torch==1.7.1+cu101 torchvision==0.8.2+cu101 -f https://download.pytorch.org/whl/torch_stable.html

或手动下载     

     

     

 ，地址：https://download.pytorch.org/whl/torch_stable.html

3.下载I MMEngine 和 MMCV

pip install -U openmim mim install mmengine mim install mmcv>=2.0.0rc1

4.下载mmselfSup并编译

git clone https://github.com/open-mmlab/mmselfsup.git cd mmselfsup git checkout 1.x pip install -v -e .

二
 




 




 




、训练自监督模型（SimCLR）

1.构造数据集

数据集结构为datasetdir->｛namedirs｝->pics

 2.写模型自监督预训练的配置文件

新建一个名为 simclr_resnet50_1xb32-coslr-1e_tinyin200.py 的配置文件

新建位置自定 




  




  




  ，本人为：configs/selfsup/simclr下

写入

_base_ = [ ../_base_/models/simclr.py, # ../_base_/datasets/imagenet_mae.py, # removed ../_base_/schedules/lars_coslr-200e_in1k.py, ../_base_/default_runtime.py, ] # custom dataset dataset_type = mmcls.CustomDataset data_root = data/custom_dataset/ #数据集路径 file_client_args = dict(backend=disk) view_pipeline = [     dict(type=RandomResizedCrop, size=224, backend=pillow), dict(type=RandomFlip, prob=0.5),     dict(         type=RandomApply,   transforms=[             dict(                 type=ColorJitter,   brightness=0.8,                 contrast=0.8,   saturation=0.8,                 hue=0.2)  ],         prob=0.8),     dict(   type=RandomGrayscale,         prob=0.2,   keep_channels=True,         channel_weights=(0.114, 0.587, 0.2989)),   dict(type=RandomGaussianBlur, sigma_min=0.1, sigma_max=2.0, prob=0.5), ] train_pipeline = [     dict(type=LoadImageFromFile, file_client_args=file_client_args),   dict(type=MultiView, num_views=2, transforms=[view_pipeline]), dict(type=PackSelfSupInputs, meta_keys=[img_path]) ] train_dataloader = dict(     batch_size=32,    num_workers=4,     persistent_workers=True,   sampler=dict(type=DefaultSampler,shuffle=True), collate_fn=dict(type=default_collate),       dataset=dict(   type=dataset_type,         data_root=data_root,    # ann_file=meta/train.txt,         data_prefix=dict(img_path=./),   pipeline=train_pipeline)) # optimizer optimizer = dict(type=LARS, lr=0.3, momentum=0.9, weight_decay=1e-6) optim_wrapper = dict(     type=OptimWrapper,   optimizer=optimizer,     paramwise_cfg=dict(    custom_keys={             bn: dict(decay_mult=0, lars_exclude=True),   bias: dict(decay_mult=0, lars_exclude=True),   # bn layer in ResNet block downsample module   downsample.1: dict(decay_mult=0, lars_exclude=True), })) # runtime settings default_hooks = dict(     # only keeps the latest 3 checkpoints checkpoint=dict(type=CheckpointHook, interval=10, max_keep_ckpts=3)) 3.写数据预训练的配置文件(mae)

新建一个名为 selfsup_mae.py 的配置文件

新建位置自定

  


  


  


，本人为：configs/selfsup/_base_/datasets下

写入

_base_ = [     ../_base_/models/mae_vit-base-p16.py,    # ../_base_/datasets/imagenet_mae.py,  # removed     ../_base_/schedules/adamw_coslr-200e_in1k.py,     ../_base_/default_runtime.py, ] #<<<<<<<  modified <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<< # custom dataset dataset_type = mmcls.CustomDataset data_root = data/sbu/ #数据集地址 file_client_args = dict(backend=disk) train_pipeline = [     dict(type=LoadImageFromFile, file_client_args=file_client_args),     dict(         type=RandomResizedCrop,         size=224,         scale=(0.2, 1.0),         backend=pillow,         interpolation=bicubic),     dict(type=RandomFlip, prob=0.5),    dict(type=PackSelfSupInputs, meta_keys=[img_path]) ] # dataset 8 x 512 train_dataloader = dict(     batch_size=512,     num_workers=8,     persistent_workers=True,     sampler=dict(type=DefaultSampler, shuffle=True),    collate_fn=dict(type=default_collate),    dataset=dict(         type=dataset_type,         data_root=data_root,         # ann_file=meta/train.txt, # removed if you dont have the annotation file         data_prefix=dict(img_path=./),         pipeline=train_pipeline)) # optimizer wrapper optimizer = dict(     type=AdamW, lr=1.5e-4 * 4096 / 256, betas=(0.9, 0.95), weight_decay=0.05) optim_wrapper = dict(     type=OptimWrapper,     optimizer=optimizer,      paramwise_cfg=dict(         custom_keys={             ln: dict(decay_mult=0.0),             bias: dict(decay_mult=0.0),             pos_embed: dict(decay_mult=0.),             mask_token: dict(decay_mult=0.),             cls_token: dict(decay_mult=0.) })) # learning rate scheduler param_scheduler = [     dict(         type=LinearLR,        start_factor=1e-4,         by_epoch=True,         begin=0,         end=40,         convert_to_iter_based=True),     dict(         type=CosineAnnealingLR,         T_max=360,         by_epoch=True,         begin=40,         end=400,               convert_to_iter_based=True) ] # runtime settings # pre-train for 400 epochs train_cfg = dict(max_epochs=400) #训练次数 default_hooks = dict(     logger=dict(type=LoggerHook, interval=100),     # only keeps the latest 3 checkpoints     checkpoint=dict(type=CheckpointHook, interval=1, max_keep_ckpts=3)) # randomness randomness = dict(seed=0, diff_rank_seed=True) resume = True

4.启动训练程序

参数

训练

 结果