电脑基础 · 2023年3月31日

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

mmselfSup训练自己的数据集

mmselfSup训练自己的数据集

 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 don't 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

mmselfSup训练自己的数据集

训练

mmselfSup训练自己的数据集

 结果

mmselfSup训练自己的数据集