Skip to content

Latest commit

 

History

History
567 lines (470 loc) · 9.99 KB

param_scheduler.md

File metadata and controls

567 lines (470 loc) · 9.99 KB
Raw

Migrate parameter scheduler from MMCV to MMEngine

MMCV 1.x version uses LrUpdaterHook and MomentumUpdaterHook to adjust the learning rate and momentum. However, the design of LrUpdaterHook has been difficult to meet more abundant customization requirements due to the development of the training strategies. Hence, MMEngine proposes parameter schedulers (ParamScheduler).

The interface of the parameter scheduler is consistent with PyTroch's learning rate scheduler (LRScheduler). In addition, the parameter scheduler provides stronger functions. For details, please refer to Parameter Scheduler User Guide.

Learning rate scheduler (LrUpdater) migration

MMEngine uses LRScheduler instead of LrUpdaterHook. The field in the config file is changed from the original lr_config to param_scheduler. The learning rate config in MMCV corresponds to the parameter scheduler config in MMEngine as follows:

Learning rate warm-up migration

The learning rate warm-up can be achieved through the combination of schedulers by specifying the effective range begin and end. There are 3 learning rate warm-up methods in MMCV, namely 'constant', 'linear', 'exp'. The corresponding config in MMEngine should be modified as follows:

Constant warm-up

MMCV-1.x MMEngine 
lr_config = dict(
    warmup='constant',
    warmup_ratio=0.1,
    warmup_iters=500,
    warmup_by_epoch=False
)
 
param_scheduler = [
    dict(type='ConstantLR',
         factor=0.1,
         begin=0,
         end=500,
         by_epoch=False),
    dict(...) # the main learning rate scheduler
]

Linear warm-up

MMCV-1.x MMEngine 
lr_config = dict(
    warmup='linear',
    warmup_ratio=0.1,
    warmup_iters=500,
    warmup_by_epoch=False
)
 
param_scheduler = [
    dict(type='LinearLR',
         start_factor=0.1,
         begin=0,
         end=500,
         by_epoch=False),
    dict(...) # the main learning rate scheduler
]

Exponential warm-up

MMCV-1.x MMEngine 
lr_config = dict(
    warmup='exp',
    warmup_ratio=0.1,
    warmup_iters=500,
    warmup_by_epoch=False
)
 
param_scheduler = [
    dict(type='ExponentialLR',
         gamma=0.1,
         begin=0,
         end=500,
         by_epoch=False),
    dict(...) # the main learning rate scheduler
]

Fixed learning rate (FixedLrUpdaterHook) migration

MMCV-1.x MMEngine 
lr_config = dict(policy='fixed')
 
param_scheduler = [
    dict(type='ConstantLR', factor=1)
]

Step learning rate (StepLrUpdaterHook) migration

MMCV-1.x MMEngine 
lr_config = dict(
    policy='step',
    step=[8, 11],
    gamma=0.1,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='MultiStepLR',
         milestones=[8, 11],
         gamma=0.1,
         by_epoch=True)
]

Poly learning rate (PolyLrUpdaterHook) migration

MMCV-1.x MMEngine 
lr_config = dict(
    policy='poly',
    power=0.7,
    min_lr=0.001,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='PolyLR',
         power=0.7,
         eta_min=0.001,
         begin=0,
         end=num_epochs,
         by_epoch=True)
]

Exponential learning rate (ExpLrUpdaterHook) migration

MMCV-1.x MMEngine 
lr_config = dict(
    policy='exp',
    power=0.5,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='ExponentialLR',
         gamma=0.5,
         begin=0,
         end=num_epochs,
         by_epoch=True)
]

Cosine annealing learning rate (CosineAnnealingLrUpdaterHook) migration

MMCV-1.x MMEngine 
lr_config = dict(
    policy='CosineAnnealing',
    min_lr=0.5,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='CosineAnnealingLR',
         eta_min=0.5,
         T_max=num_epochs,
         begin=0,
         end=num_epochs,
         by_epoch=True)
]

FlatCosineAnnealingLrUpdaterHook migration

The learning rate strategy combined by multiple phases like FlatCosineAnnealing originally needs to be achieved by rewriting a Hook. But in MMEngine, it can be achieved with combining two parameter scheduler configs:

MMCV-1.x MMEngine 
lr_config = dict(
    policy='FlatCosineAnnealing',
    start_percent=0.5,
    min_lr=0.005,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='ConstantLR', factor=1, begin=0, end=num_epochs * 0.75)
    dict(type='CosineAnnealingLR',
         eta_min=0.005,
         begin=num_epochs * 0.75,
         end=num_epochs,
         T_max=num_epochs * 0.25,
         by_epoch=True)
]

CosineRestartLrUpdaterHook migration

MMCV-1.x MMEngine 
lr_config = dict(policy='CosineRestart',
                 periods=[5, 10, 15],
                 restart_weights=[1, 0.7, 0.3],
                 min_lr=0.001,
                 by_epoch=True)
 
param_scheduler = [
    dict(type='CosineRestartLR',
         periods=[5, 10, 15],
         restart_weights=[1, 0.7, 0.3],
         eta_min=0.001,
         by_epoch=True)
]

OneCycleLrUpdaterHook migration

MMCV-1.x MMEngine 
lr_config = dict(policy='OneCycle',
                 max_lr=0.02,
                 total_steps=90000,
                 pct_start=0.3,
                 anneal_strategy='cos',
                 div_factor=25,
                 final_div_factor=1e4,
                 three_phase=True,
                 by_epoch=False)
 
param_scheduler = [
    dict(type='OneCycleLR',
         eta_max=0.02,
         total_steps=90000,
         pct_start=0.3,
         anneal_strategy='cos',
         div_factor=25,
         final_div_factor=1e4,
         three_phase=True,
         by_epoch=False)
]

Notice: by_epoch defaults to False in MMCV. It now defaults to True in MMEngine.

LinearAnnealingLrUpdaterHook migration

MMCV-1.x MMEngine 
lr_config = dict(
    policy='LinearAnnealing',
    min_lr_ratio=0.01,
    by_epoch=True
)
 
param_scheduler = [
    dict(type='LinearLR',
         start_factor=1,
         end_factor=0.01,
         begin=0,
         end=num_epochs,
         by_epoch=True)
]

MomentumUpdater migration

MMCV uses momentum_config field and MomentumUpdateHook to adjust momentum. The momentum in MMEngine is also controlled by the parameter scheduler. Users can simply change the LR of the learning rate scheduler to Momentum to use the same strategy to adjust the momentum. The momentum scheduler shares the same param_scheduler field in the config with the learning rate scheduler:

MMCV-1.x MMEngine 
lr_config = dict(...)
momentum_config = dict(
    policy='CosineAnnealing',
    min_momentum=0.1,
    by_epoch=True
)
 
param_scheduler = [
    # config of learning rate schedulers
    dict(...),
    # config of momentum schedulers
    dict(type='CosineAnnealingMomentum',
         eta_min=0.1,
         T_max=num_epochs,
         begin=0,
         end=num_epochs,
         by_epoch=True)
]

Migrate parameter update frequency related config

If you want to update the parameter rate based on iteration while using the epoch-based training loop and setting the effective range (begin, end) or period (T_max) and other variables according to epoch in MMCV, you need to set by_epoch to False.

However, in MMEngine, the by_epoch in the config still needs to be set to True. Instead, you need to add convert_to_iter_based=True in the config to build a parameter scheduler which updates by iteration, see Parameter Scheduler Tutorial for more details.

Take the migration of CosineAnnealing as an example:

MMCV-1.x MMEngine 
lr_config = dict(
    policy='CosineAnnealing',
    min_lr=0.5,
    by_epoch=False
)
 
param_scheduler = [
    dict(
        type='CosineAnnealingLR',
        eta_min=0.5,
        T_max=num_epochs,
        by_epoch=True,  # Notice, by_epoch need to be set to True
        convert_to_iter_based=True  # convert to an iter-based scheduler
    )
]

You may also want to read parameter scheduler tutorial or parameter scheduler API documentations.