mmdetection/docs/en/advanced_guides/customize_runtime.md

Customize Runtime Settings

Customize optimization settings

Optimization related configuration is now all managed by optim_wrapper which usually has three fields: optimizer, paramwise_cfg, clip_grad, refer to OptimWrapper for more detail. See the example below, where Adamw is used as an optimizer, the learning rate of the backbone is reduced by a factor of 10, and gradient clipping is added.

optim_wrapper = dict(
    type='OptimWrapper',
    # optimizer
    optimizer=dict(
        type='AdamW',
        lr=0.0001,
        weight_decay=0.05,
        eps=1e-8,
        betas=(0.9, 0.999)),

    # Parameter-level learning rate and weight decay settings
    paramwise_cfg=dict(
        custom_keys={
            'backbone': dict(lr_mult=0.1, decay_mult=1.0),
        },
        norm_decay_mult=0.0),

    # gradient clipping
    clip_grad=dict(max_norm=0.01, norm_type=2))

Customize optimizer supported by Pytorch

We already support to use all the optimizers implemented by PyTorch, and the only modification is to change the optimizer field in optim_wrapper field of config files. For example, if you want to use ADAM (note that the performance could drop a lot), the modification could be as the following.

optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='Adam', lr=0.0003, weight_decay=0.0001))

To modify the learning rate of the model, the users only need to modify the lr in optimizer. The users can directly set arguments following the API doc of PyTorch.

Customize self-implemented optimizer

1. Define a new optimizer

A customized optimizer could be defined as following.

Assume you want to add a optimizer named MyOptimizer, which has arguments a, b, and c. You need to create a new directory named mmdet/engine/optimizers. And then implement the new optimizer in a file, e.g., in mmdet/engine/optimizers/my_optimizer.py:

from mmdet.registry import OPTIMIZERS
from torch.optim import Optimizer


@OPTIMIZERS.register_module()
class MyOptimizer(Optimizer):

    def __init__(self, a, b, c)

2. Add the optimizer to registry

To find the above module defined above, this module should be imported into the main namespace at first. There are two options to achieve it.

• Modify mmdet/engine/optimizers/__init__.py to import it.

The newly defined module should be imported in mmdet/engine/optimizers/__init__.py so that the registry will find the new module and add it:

from .my_optimizer import MyOptimizer

• Use custom_imports in the config to manually import it

custom_imports = dict(imports=['mmdet.engine.optimizers.my_optimizer'], allow_failed_imports=False)

The module mmdet.engine.optimizers.my_optimizer will be imported at the beginning of the program and the class MyOptimizer is then automatically registered. Note that only the package containing the class MyOptimizer should be imported. mmdet.engine.optimizers.my_optimizer.MyOptimizer cannot be imported directly.

Actually users can use a totally different file directory structure using this importing method, as long as the module root can be located in PYTHONPATH.

3. Specify the optimizer in the config file

Then you can use MyOptimizer in optimizer field in optim_wrapper field of config files. In the configs, the optimizers are defined by the field optimizer like the following:

optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001))

To use your own optimizer, the field can be changed to

optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='MyOptimizer', a=a_value, b=b_value, c=c_value))

Customize optimizer wrapper constructor

Some models may have some parameter-specific settings for optimization, e.g. weight decay for BatchNorm layers. The users can do those fine-grained parameter tuning through customizing optimizer wrapper constructor.

from mmengine.optim import DefaultOptiWrapperConstructor

from mmdet.registry import OPTIM_WRAPPER_CONSTRUCTORS
from .my_optimizer import MyOptimizer


@OPTIM_WRAPPER_CONSTRUCTORS.register_module()
class MyOptimizerWrapperConstructor(DefaultOptimWrapperConstructor):

    def __init__(self,
                 optim_wrapper_cfg: dict,
                 paramwise_cfg: Optional[dict] = None):

    def __call__(self, model: nn.Module) -> OptimWrapper:

        return optim_wrapper

The default optimizer wrapper constructor is implemented here, which could also serve as a template for the new optimizer wrapper constructor.

Additional settings

Tricks not implemented by the optimizer should be implemented through optimizer wrapper constructor (e.g., set parameter-wise learning rates) or hooks. We list some common settings that could stabilize the training or accelerate the training. Feel free to create PR, issue for more settings.

• Use gradient clip to stabilize training: Some models need gradient clip to clip the gradients to stabilize the training process. An example is as below:

  optim_wrapper = dict(
      _delete_=True, clip_grad=dict(max_norm=35, norm_type=2))

If your config inherits the base config which already sets the optim_wrapper, you might need _delete_=True to override the unnecessary settings. See the config documentation for more details.

• Use momentum schedule to accelerate model convergence: We support momentum scheduler to modify model's momentum according to learning rate, which could make the model converge in a faster way. Momentum scheduler is usually used with LR scheduler, for example, the following config is used in 3D detection to accelerate convergence. For more details, please refer to the implementation of CosineAnnealingLR and CosineAnnealingMomentum.

  param_scheduler = [
      # learning rate scheduler
      # During the first 8 epochs, learning rate increases from 0 to lr * 10
      # during the next 12 epochs, learning rate decreases from lr * 10 to lr * 1e-4
      dict(
          type='CosineAnnealingLR',
          T_max=8,
          eta_min=lr * 10,
          begin=0,
          end=8,
          by_epoch=True,
          convert_to_iter_based=True),
      dict(
          type='CosineAnnealingLR',
          T_max=12,
          eta_min=lr * 1e-4,
          begin=8,
          end=20,
          by_epoch=True,
          convert_to_iter_based=True),
      # momentum scheduler
      # During the first 8 epochs, momentum increases from 0 to 0.85 / 0.95
      # during the next 12 epochs, momentum increases from 0.85 / 0.95 to 1
      dict(
          type='CosineAnnealingMomentum',
          T_max=8,
          eta_min=0.85 / 0.95,
          begin=0,
          end=8,
          by_epoch=True,
          convert_to_iter_based=True),
      dict(
          type='CosineAnnealingMomentum',
          T_max=12,
          eta_min=1,
          begin=8,
          end=20,
          by_epoch=True,
          convert_to_iter_based=True)
  ]

Customize training schedules

By default we use step learning rate with 1x schedule, this calls MultiStepLR in MMEngine. We support many other learning rate schedule here, such as CosineAnnealingLR and PolyLR schedule. Here are some examples

• Poly schedule:

  param_scheduler = [
      dict(
          type='PolyLR',
          power=0.9,
          eta_min=1e-4,
          begin=0,
          end=8,
          by_epoch=True)]

• ConsineAnnealing schedule:

  param_scheduler = [
      dict(
          type='CosineAnnealingLR',
          T_max=8,
          eta_min=lr * 1e-5,
          begin=0,
          end=8,
          by_epoch=True)]

Customize train loop

By default, EpochBasedTrainLoop is used in train_cfg and validation is done after every train epoch, as follows.

train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=12, val_begin=1, val_interval=1)

Actually, both IterBasedTrainLoop and EpochBasedTrainLoop support dynamical interval, see the following example.

# Before 365001th iteration, we do evaluation every 5000 iterations.
# After 365000th iteration, we do evaluation every 368750 iterations,
# which means that we do evaluation at the end of training.

interval = 5000
max_iters = 368750
dynamic_intervals = [(max_iters // interval * interval + 1, max_iters)]
train_cfg = dict(
    type='IterBasedTrainLoop',
    max_iters=max_iters,
    val_interval=interval,
    dynamic_intervals=dynamic_intervals)

Customize hooks

Customize self-implemented hooks

1. Implement a new hook

MMEngine provides many useful hooks, but there are some occasions when the users might need to implement a new hook. MMDetection supports customized hooks in training in v3.0 . Thus the users could implement a hook directly in mmdet or their mmdet-based codebases and use the hook by only modifying the config in training. Here we give an example of creating a new hook in mmdet and using it in training.

from mmengine.hooks import Hook
from mmdet.registry import HOOKS


@HOOKS.register_module()
class MyHook(Hook):

    def __init__(self, a, b):

    def before_run(self, runner) -> None:

    def after_run(self, runner) -> None:

    def before_train(self, runner) -> None:

    def after_train(self, runner) -> None:

    def before_train_epoch(self, runner) -> None:

    def after_train_epoch(self, runner) -> None:

    def before_train_iter(self,
                          runner,
                          batch_idx: int,
                          data_batch: DATA_BATCH = None) -> None:

    def after_train_iter(self,
                         runner,
                         batch_idx: int,
                         data_batch: DATA_BATCH = None,
                         outputs: Optional[dict] = None) -> None:

Depending on the functionality of the hook, the users need to specify what the hook will do at each stage of the training in before_run, after_run, before_train, after_train , before_train_epoch, after_train_epoch, before_train_iter, and after_train_iter. There are more points where hooks can be inserted, refer to base hook class for more detail.

2. Register the new hook

Then we need to make MyHook imported. Assuming the file is in mmdet/engine/hooks/my_hook.py there are two ways to do that:

• Modify mmdet/engine/hooks/__init__.py to import it.

The newly defined module should be imported in mmdet/engine/hooks/__init__.py so that the registry will find the new module and add it:

from .my_hook import MyHook

• Use custom_imports in the config to manually import it

custom_imports = dict(imports=['mmdet.engine.hooks.my_hook'], allow_failed_imports=False)

3. Modify the config

custom_hooks = [
    dict(type='MyHook', a=a_value, b=b_value)
]

You can also set the priority of the hook by adding key priority to 'NORMAL' or 'HIGHEST' as below

custom_hooks = [
    dict(type='MyHook', a=a_value, b=b_value, priority='NORMAL')
]

By default the hook's priority is set as NORMAL during registration.

Use hooks implemented in MMDetection

If the hook is already implemented in MMDectection, you can directly modify the config to use the hook as below

Example: NumClassCheckHook

We implement a customized hook named NumClassCheckHook to check whether the num_classes in head matches the length of classes in the metainfo of dataset.

We set it in default_runtime.py.

custom_hooks = [dict(type='NumClassCheckHook')]

Modify default runtime hooks

There are some common hooks that are registered through default_hooks, they are

• IterTimerHook: A hook that logs 'data_time' for loading data and 'time' for a model train step.
• LoggerHook: A hook that Collect logs from different components of Runner and write them to terminal, JSON file, tensorboard and wandb .etc.
• ParamSchedulerHook: A hook to update some hyper-parameters in optimizer, e.g., learning rate and momentum.
• CheckpointHook: A hook that saves checkpoints periodically.
• DistSamplerSeedHook: A hook that sets the seed for sampler and batch_sampler.
• DetVisualizationHook: A hook used to visualize validation and testing process prediction results.

IterTimerHook, ParamSchedulerHook and DistSamplerSeedHook are simple and no need to be modified usually, so here we reveals how what we can do with LoggerHook, CheckpointHook and DetVisualizationHook.

CheckpointHook

Except saving checkpoints periodically, CheckpointHook provides other options such as max_keep_ckpts, save_optimizer and etc. The users could set max_keep_ckpts to only save small number of checkpoints or decide whether to store state dict of optimizer by save_optimizer. More details of the arguments are here

default_hooks = dict(
    checkpoint=dict(
        type='CheckpointHook',
        interval=1,
        max_keep_ckpts=3,
        save_optimizer=True))

LoggerHook

The LoggerHook enables to set intervals. And the detail usages can be found in the docstring.

default_hooks = dict(logger=dict(type='LoggerHook', interval=50))

DetVisualizationHook

DetVisualizationHook use DetLocalVisualizer to visualize prediction results, and DetLocalVisualizer current supports different backends, e.g., TensorboardVisBackend and WandbVisBackend (see docstring for more detail). The users could add multi backbends to do visualization, as follows.

default_hooks = dict(
    visualization=dict(type='DetVisualizationHook', draw=True))

vis_backends = [dict(type='LocalVisBackend'),
                dict(type='TensorboardVisBackend')]
visualizer = dict(
    type='DetLocalVisualizer', vis_backends=vis_backends, name='visualizer')