MMselfSup训练自监督模型之mocov3
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
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 mocov3_resnet50_pretrain_8xb512-amp-coslr-1e-800e_in1k.py 的配置文件
新建位置自定，本人为：configs/selfsup/mocov3下
写入
_base_ = 'mocov3_resnet50_8xb512-amp-coslr-800e_in1k.py'
model = dict(base_momentum=0.996) # 0.99 for 100e and 300e, 0.996 for 800e
# optimizer
optimizer = dict(type='LARS', lr=9.6, weight_decay=1.5e-6)
optim_wrapper = dict(optimizer=optimizer)
# learning rate scheduler
param_scheduler = [
dict(
type='LinearLR',
start_factor=1e-4,
by_epoch=True,
begin=0,
end=10,
convert_to_iter_based=True),
dict(
type='CosineAnnealingLR',
T_max=790,
by_epoch=True,
begin=10,
end=800,
convert_to_iter_based=True)
]
# runtime settings
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=2000)
修改mocov3_resnet50_8xb512-amp-coslr-100e_in1k.py
_base_ = [
'../_base_/models/mocov3_resnet50.py',
#'../_base_/datasets/coco_MOCOV3.py',
'../_base_/schedules/lars_coslr-200e_in1k.py',
'../_base_/default_runtime.py',
]
#custom dataset
dataset_type = 'mmcls.CustomDataset'
data_root = '/macrosan/common/group/mjt/to_other/to_zym/mmselfsuptest'
#train_dataloader = dict(batch_size=16, num_workers=1)
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))
# <<<<<<<<<<<<<<<<<<<<<< End of Changed <<<<<<<<<<<<<<<<<<<<<<<<<<<
# optimizer
optimizer = dict(type='LARS', lr=9.6, weight_decay=1e-6, momentum=0.9)
optim_wrapper = dict(
type='AmpOptimWrapper',
loss_scale='dynamic',
optimizer=optimizer,
paramwise_cfg=dict(
custom_keys={
'bn': dict(decay_mult=0, lars_exclude=True),
'bias': dict(decay_mult=0, lars_exclude=True),
'downsample.1': dict(decay_mult=0, lars_exclude=True),
}),
)
# learning rate scheduler
param_scheduler = [
dict(
type='LinearLR',
start_factor=1e-4,
by_epoch=True,
begin=0,
end=10,
convert_to_iter_based=True),
dict(
type='CosineAnnealingLR',
T_max=90,
by_epoch=True,
begin=10,
end=100,
convert_to_iter_based=True)
]
# runtime settings
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=2000)
# runtime settings
# only keeps the latest 3 checkpoints
default_hooks = dict(checkpoint=dict(max_keep_ckpts=50))
3.训练
训练程序在tools/train.py
参数为mocov3_resnet50_pretrain_8xb512-amp-coslr-1e-800e_in1k.py
注：如果遇到不收敛（loss不降低）的情况
首先调整configs/selfsup/_base_/models/mocov3_resnet50.py文件中的温度系数参数
temperature = 0.1
若再不收敛可适当调整学习率
4.结果
