FlexibleRunner¶

class mmengine.runner.FlexibleRunner(model, *, work_dir='work_dirs', experiment_name=None, train_dataloader=None, optim_wrapper=None, param_scheduler=None, train_cfg=None, val_dataloader=None, val_evaluator=None, val_cfg=None, test_dataloader=None, test_evaluator=None, test_cfg=None, strategy=None, auto_scale_lr=None, default_hooks=None, custom_hooks=None, data_preprocessor=None, load_from=None, resume=False, launcher=None, env_cfg={'dist_cfg': {'backend': 'nccl'}}, log_processor=None, log_level='INFO', visualizer=None, default_scope='mmengine', randomness={'seed': None}, compile=False, cfg=None)[source]¶

A training helper for PyTorch.

Runner object can be built from config by runner = Runner.from_cfg(cfg) where the cfg usually contains training, validation, and test-related configurations to build corresponding components. We usually use the same config to launch training, testing, and validation tasks. However, only some of these components are necessary at the same time, e.g., testing a model does not need training or validation-related components.

To avoid repeatedly modifying config, the construction of Runner adopts lazy initialization to only initialize components when they are going to be used. Therefore, the model is always initialized at the beginning, and training, validation, and, testing related components are only initialized when calling runner.train(), runner.val(), and runner.test(), respectively.

Warning

This is an experimental feature, and its interface is subject to change.

Parameters:

model (torch.nn.Module or dict) – The model to be run. It can be a dict used for build a model.
work_dir (str) –
experiment_name (str | None) –
train_dataloader (DataLoader | Dict | None) –
optim_wrapper (OptimWrapper | Dict | None) –
param_scheduler (_ParamScheduler | Dict | List | None) –
train_cfg (Dict | None) –
val_dataloader (DataLoader | Dict | None) –
val_evaluator (Evaluator | Dict | List | None) –
val_cfg (Dict | None) –
test_dataloader (DataLoader | Dict | None) –
test_evaluator (Evaluator | Dict | List | None) –
test_cfg (Dict | None) –
strategy (BaseStrategy | Dict | None) –
auto_scale_lr (Dict | None) –
default_hooks (Dict[str, Hook | Dict] | None) –
custom_hooks (List[Hook | Dict] | None) –
data_preprocessor (Module | Dict | None) –
load_from (str | None) –
resume (str | bool) –
launcher (str | None) –
env_cfg (Dict) –
log_processor (Dict | None) –
log_level (str) –
visualizer (Visualizer | Dict | None) –
default_scope (str | None) –
randomness (Dict) –
compile (bool | Dict) –
cfg (Config) –

Kwargs:

work_dir (str, optional): The working directory to save checkpoints.: The logs will be saved in the subdirectory of work_dir named timestamp. Defaults to ‘work_dir’.
experiment_name (str, optional): Name of current experiment. If not: specified, timestamp will be used as experiment_name. Defaults to None.
train_dataloader (Dataloader or dict, optional): A dataloader object or: a dict to build a dataloader. If None is given, it means skipping training steps. Defaults to None. See build_dataloader() for more details.
optim_wrapper (OptimWrapper or dict, optional):: Computing gradient of model parameters. If specified, train_dataloader should also be specified. If automatic mixed precision or gradient accmulation training is required. The type of optim_wrapper should be AmpOptimizerWrapper. See build_optim_wrapper() for examples. Defaults to None.
param_scheduler (_ParamScheduler or dict or list, optional):: Parameter scheduler for updating optimizer parameters. If specified, optimizer should also be specified. Defaults to None. See build_param_scheduler() for examples.
train_cfg (dict, optional): A dict to build a training loop. If it does: not provide “type” key, it should contain “by_epoch” to decide which type of training loop EpochBasedTrainLoop or IterBasedTrainLoop should be used. If train_cfg specified, train_dataloader should also be specified. Defaults to None. See build_train_loop() for more details.
val_dataloader (Dataloader or dict, optional): A dataloader object or: a dict to build a dataloader. If None is given, it means skipping validation steps. Defaults to None. See build_dataloader() for more details.
val_evaluator (Evaluator or dict or list, optional): A evaluator object: used for computing metrics for validation. It can be a dict or a list of dict to build a evaluator. If specified, val_dataloader should also be specified. Defaults to None.
val_cfg (dict, optional): A dict to build a validation loop. If it does: not provide “type” key, ValLoop will be used by default. If val_cfg specified, val_dataloader should also be specified. If ValLoop is built with fp16=True`, runner.val() will be performed under fp16 precision.
test_dataloader (Dataloader or dict, optional): A dataloader object or: a dict to build a dataloader. If None is given, it means skipping test steps. Defaults to None. See build_dataloader() for more details. Defaults to None. See build_val_loop() for more details.
test_evaluator (Evaluator or dict or list, optional): A evaluator: object used for computing metrics for test steps. It can be a dict or a list of dict to build a evaluator. If specified, test_dataloader should also be specified. Defaults to None.
test_cfg (dict, optional): A dict to build a test loop. If it does: not provide “type” key, TestLoop will be used by default. If test_cfg specified, test_dataloader should also be specified. If ValLoop is built with fp16=True`, runner.val() will be performed under fp16 precision. Defaults to None. See build_test_loop() for more details.
strategy (BaseStrategy or dict, optional): A strategy object or a dict: to build a strategy. Defaults to None. If not specified, the strategy will be inferred automatically.
auto_scale_lr (dict, Optional): Config to scale the learning rate: automatically. It includes base_batch_size and enable. base_batch_size is the batch size that the optimizer lr is based on. enable is the switch to turn on and off the feature.
default_hooks (dict[str, dict] or dict[str, Hook], optional): Hooks to: execute default actions like updating model parameters and saving checkpoints. Default hooks are OptimizerHook, IterTimerHook, LoggerHook, ParamSchedulerHook and CheckpointHook. Defaults to None. See register_default_hooks() for more details.
custom_hooks (list[dict] or list[Hook], optional): Hooks to execute: custom actions like visualizing images processed by pipeline. Defaults to None.
data_preprocessor (dict, optional): The pre-process config of: BaseDataPreprocessor. If the model argument is a dict and doesn’t contain the key data_preprocessor, set the argument as the data_preprocessor of the model dict. Defaults to None.
load_from (str, optional): The checkpoint file to load from.: Defaults to None.
resume (bool): Whether to resume training. Defaults to False. If: resume is True and load_from is None, automatically to find latest checkpoint from work_dir. If not found, resuming does nothing.
launcher (str, optional): Way to launcher multi-process. Supported: launchers are ‘pytorch’, ‘mpi’, ‘slurm’ and ‘none’. If ‘none’ is provided, non-distributed environment will be launched. If launcher is None, the launcher will be inferred according some specified environments. Defaults to None.
env_cfg (dict): A dict used for setting environment. Defaults to: dict(dist_cfg=dict(backend=’nccl’)).
log_processor (dict, optional): A processor to format logs. Defaults to: None.
log_level (int or str): The log level of MMLogger handlers.: Defaults to ‘INFO’.
visualizer (Visualizer or dict, optional): A Visualizer object or a: dict build Visualizer object. Defaults to None. If not specified, default config will be used.
default_scope (str): Used to reset registries location.: Defaults to “mmengine”.
randomness (dict): Some settings to make the experiment as reproducible: as possible like seed and deterministic. Defaults to dict(seed=None). If seed is None, a random number will be generated and it will be broadcasted to all other processes if in distributed environment. If cudnn_benchmark is True in env_cfg but deterministic is True in randomness, the value of torch.backends.cudnn.benchmark will be False finally.
compile (bool or dict, optional): Whether to enable torch.compile.: Defaults to False.
cfg (dict or Configdict or Config, optional): Full config.: Defaults to None.

Note

Since PyTorch 2.0.0, you can enable torch.compile by passing in compile = True. If you want to control compile options, you can pass a dict, e.g. cfg.compile = dict(backend='eager'). Refer to PyTorch API Documentation for more valid options.

Examples

>>> from mmengine.runner import Runner
>>> cfg = dict(
>>>     model=dict(type='ToyModel'),
>>>     work_dir='path/of/work_dir',
>>>     train_dataloader=dict(
>>>     dataset=dict(type='ToyDataset'),
>>>     sampler=dict(type='DefaultSampler', shuffle=True),
>>>     batch_size=1,
>>>     num_workers=0),
>>>     val_dataloader=dict(
>>>         dataset=dict(type='ToyDataset'),
>>>         sampler=dict(type='DefaultSampler', shuffle=False),
>>>        batch_size=1,
>>>        num_workers=0),
>>>     test_dataloader=dict(
>>>         dataset=dict(type='ToyDataset'),
>>>         sampler=dict(type='DefaultSampler', shuffle=False),
>>>         batch_size=1,
>>>         num_workers=0),
>>>     auto_scale_lr=dict(base_batch_size=16, enable=False),
>>>     optim_wrapper=dict(type='OptimizerWrapper', optimizer=dict(
>>>         type='SGD', lr=0.01)),
>>>     param_scheduler=dict(type='MultiStepLR', milestones=[1, 2]),
>>>     val_evaluator=dict(type='ToyEvaluator'),
>>>     test_evaluator=dict(type='ToyEvaluator'),
>>>     train_cfg=dict(by_epoch=True, max_epochs=3, val_interval=1),
>>>     val_cfg=dict(),
>>>     test_cfg=dict(),
>>>     custom_hooks=[],
>>>     default_hooks=dict(
>>>         timer=dict(type='IterTimerHook'),
>>>         checkpoint=dict(type='CheckpointHook', interval=1),
>>>         logger=dict(type='LoggerHook'),
>>>         optimizer=dict(type='OptimizerHook', grad_clip=False),
>>>         param_scheduler=dict(type='ParamSchedulerHook')),
>>>     launcher='none',
>>>     env_cfg=dict(dist_cfg=dict(backend='nccl')),
>>>     log_processor=dict(window_size=20),
>>>     visualizer=dict(type='Visualizer',
>>>     vis_backends=[dict(type='LocalVisBackend',
>>>                        save_dir='temp_dir')])
>>>    )
>>> runner = Runner.from_cfg(cfg)
>>> runner.train()
>>> runner.test()

static build_dataloader(dataloader, seed=None, diff_rank_seed=False)[source]¶

Build dataloader.

The method builds three components:

Dataset
Sampler
Dataloader

An example of dataloader:

dataloader = dict(
    dataset=dict(type='ToyDataset'),
    sampler=dict(type='DefaultSampler', shuffle=True),
    batch_size=1,
    num_workers=9
)

Parameters:

dataloader (DataLoader or dict) – A Dataloader object or a dict to build Dataloader object. If dataloader is a Dataloader object, just returns itself.
seed (int, optional) – Random seed. Defaults to None.
diff_rank_seed (bool) – Whether or not set different seeds to different ranks. If True, the seed passed to sampler is set to None, in order to synchronize the seeds used in samplers across different ranks. Defaults to False.

Returns:

DataLoader build from dataloader_cfg.

Return type:

Dataloader

build_evaluator(evaluator)[source]¶

Build evaluator.

Examples of evaluator:

# evaluator could be a built Evaluator instance
evaluator = Evaluator(metrics=[ToyMetric()])

# evaluator can also be a list of dict
evaluator = [
    dict(type='ToyMetric1'),
    dict(type='ToyEvaluator2')
]

# evaluator can also be a list of built metric
evaluator = [ToyMetric1(), ToyMetric2()]

# evaluator can also be a dict with key metrics
evaluator = dict(metrics=ToyMetric())
# metric is a list
evaluator = dict(metrics=[ToyMetric()])

Parameters:: evaluator (Evaluator or dict or list) – An Evaluator object or a config dict or list of config dict used to build an Evaluator.
Returns:: Evaluator build from evaluator.
Return type:: Evaluator

build_log_processor(log_processor)[source]¶

Build test log_processor.

Examples of log_processor:

# LogProcessor will be used log_processor = dict()

# custom log_processor log_processor = dict(type=’CustomLogProcessor’)

Parameters:

log_processor (LogProcessor or dict) – A log processor or a dict
processor (to build log processor. If log_processor is a log) –
object –
itself. (just returns) –

Returns:

Log processor object build from log_processor_cfg.

Return type:

LogProcessor

build_message_hub(message_hub=None)[source]¶

Build a global asscessable MessageHub.

Parameters:: message_hub (dict, optional) – A dict to build MessageHub object. If not specified, default config will be used to build MessageHub object. Defaults to None.
Returns:: A MessageHub object build from message_hub.
Return type:: MessageHub

build_strategy(strategy=None, launcher='none', randomness=None, env_cfg={'dist_cfg': {'backend': 'nccl'}}, experiment_name=None, log_level=None)[source]¶

Build a strategy.

Parameters:

strategy (BaseStrategy, optional) – A strategy object or dict to build the strategy. Defaults to None.
launcher (str) –
randomness (dict | None) –
env_cfg (dict) –
experiment_name (str | None) –
log_level (str | None) –

Returns:

A strategy object.

Return type:

BaseStrategy

build_test_loop(loop)[source]¶

Build test loop.

Examples of loop:

# `TestLoop` will be used
loop = dict()

# custom test loop
loop = dict(type='CustomTestLoop')

Parameters:: loop (BaseLoop or dict) – A test loop or a dict to build test loop. If loop is a test loop object, just returns itself.
Returns:: Test loop object build from loop_cfg.
Return type:: BaseLoop

build_train_loop(loop)[source]¶

Build training loop.

Examples of loop:

# `EpochBasedTrainLoop` will be used
loop = dict(by_epoch=True, max_epochs=3)

# `IterBasedTrainLoop` will be used
loop = dict(by_epoch=False, max_epochs=3)

# custom training loop
loop = dict(type='CustomTrainLoop', max_epochs=3)

Parameters:: loop (BaseLoop or dict) – A training loop or a dict to build training loop. If loop is a training loop object, just returns itself.
Returns:: Training loop object build from loop.
Return type:: BaseLoop

build_val_loop(loop)[source]¶

Build validation loop.

Examples of loop:

# ValLoop will be used loop = dict()

# custom validation loop loop = dict(type=’CustomValLoop’)

Parameters:: loop (BaseLoop or dict) – A validation loop or a dict to build validation loop. If loop is a validation loop object, just returns itself.
Returns:: Validation loop object build from loop.
Return type:: BaseLoop

build_visualizer(visualizer=None)[source]¶

Build a global asscessable Visualizer.

Parameters:: visualizer (Visualizer or dict, optional) – A Visualizer object or a dict to build Visualizer object. If visualizer is a Visualizer object, just returns itself. If not specified, default config will be used to build Visualizer object. Defaults to None.
Returns:: A Visualizer object build from visualizer.
Return type:: Visualizer

call_hook(fn_name, **kwargs)[source]¶

Call all hooks.

Parameters:

fn_name (str) – The function name in each hook to be called, such as “before_train_epoch”.
**kwargs – Keyword arguments passed to hook.

Return type:

None

property deterministic¶

Whether cudnn to select deterministic algorithms.

Type:: int

property distributed¶

Whether current environment is distributed.

Type:: bool

dump_config()[source]¶

Dump config to work_dir.

Return type:: None

property epoch¶

Current epoch.

Type:: int

property experiment_name¶

Name of experiment.

Type:: str

classmethod from_cfg(cfg)[source]¶

Build a runner from config.

Parameters:: cfg (ConfigType) – A config used for building runner. Keys of cfg can see __init__().
Returns:: A runner build from cfg.
Return type:: Runner

property hooks¶

A list of registered hooks.

Type:: list[Hook]

property iter¶

Current iteration.

Type:: int

load_checkpoint(filename, map_location='cpu', strict=False, revise_keys=[('^module.', '')])[source]¶

Load checkpoint from given filename.

Parameters:

filename (str) – Accept local filepath, URL, torchvision://xxx, open-mmlab://xxx.
map_location (str or callable) – A string or a callable function to specifying how to remap storage locations. Defaults to ‘cpu’.
strict (bool) – strict (bool): Whether to allow different params for the model and checkpoint.
revise_keys (list) – A list of customized keywords to modify the state_dict in checkpoint. Each item is a (pattern, replacement) pair of the regular expression operations. Defaults to strip the prefix ‘module.’ by [(r’^module.’, ‘’)].

load_or_resume()[source]¶: load or resume checkpoint.

property max_epochs¶

Total epochs to train model.

Type:: int

property max_iters¶

Total iterations to train model.

Type:: int

property model_name¶

Name of the model, usually the module class name.

Type:: str

property rank¶

Rank of current process.

Type:: int

register_custom_hooks(hooks)[source]¶

Register custom hooks into hook list.

Parameters:: hooks (list[Hook | dict]) – List of hooks or configs to be registered.
Return type:: None

register_default_hooks(hooks=None)[source]¶

Register default hooks into hook list.

hooks will be registered into runner to execute some default actions like updating model parameters or saving checkpoints.

Default hooks and their priorities:

Hooks	Priority
RuntimeInfoHook	VERY_HIGH (10)
IterTimerHook	NORMAL (50)
DistSamplerSeedHook	NORMAL (50)
LoggerHook	BELOW_NORMAL (60)
ParamSchedulerHook	LOW (70)
CheckpointHook	VERY_LOW (90)

If hooks is None, above hooks will be registered by default:

default_hooks = dict(
    runtime_info=dict(type='RuntimeInfoHook'),
    timer=dict(type='IterTimerHook'),
    sampler_seed=dict(type='DistSamplerSeedHook'),
    logger=dict(type='LoggerHook'),
    param_scheduler=dict(type='ParamSchedulerHook'),
    checkpoint=dict(type='CheckpointHook', interval=1),
)

If not None, hooks will be merged into default_hooks. If there are None value in default_hooks, the corresponding item will be popped from default_hooks:

hooks = dict(timer=None)

The final registered default hooks will be RuntimeInfoHook, DistSamplerSeedHook, LoggerHook, ParamSchedulerHook and CheckpointHook.

Parameters:: hooks (dict[str, Hook or dict], optional) – Default hooks or configs to be registered.
Return type:: None

register_hook(hook, priority=None)[source]¶

Register a hook into the hook list.

The hook will be inserted into a priority queue, with the specified priority (See Priority for details of priorities). For hooks with the same priority, they will be triggered in the same order as they are registered.

Priority of hook will be decided with the following priority:

priority argument. If priority is given, it will be priority of hook.
If hook argument is a dict and priority in it, the priority will be the value of hook['priority'].
If hook argument is a dict but priority not in it or hook is an instance of hook, the priority will be hook.priority.

Parameters:

hook (Hook or dict) – The hook to be registered.
priority (int or str or Priority, optional) – Hook priority. Lower value means higher priority.

Return type:

None

register_hooks(default_hooks=None, custom_hooks=None)[source]¶

Register default hooks and custom hooks into hook list.

Parameters:

default_hooks (dict[str, dict] or dict[str, Hook], optional) – Hooks to execute default actions like updating model parameters and saving checkpoints. Defaults to None.
custom_hooks (list[dict] or list[Hook], optional) – Hooks to execute custom actions like visualizing images processed by pipeline. Defaults to None.

Return type:

None

resume(filename, resume_optimizer=True, resume_param_scheduler=True, map_location='default')[source]¶

Resume model from checkpoint.

Parameters:

filename (str) – Accept local filepath, URL, torchvision://xxx, open-mmlab://xxx.
resume_optimizer (bool) – Whether to resume optimizer state. Defaults to True.
resume_param_scheduler (bool) – Whether to resume param scheduler state. Defaults to True.
map_location (str or callable) – A string or a callable function to specifying how to remap storage locations. Defaults to ‘default’.

Return type:

None

save_checkpoint(out_dir, filename, file_client_args=None, save_optimizer=True, save_param_scheduler=True, meta=None, by_epoch=True, backend_args=None)[source]¶

Save checkpoints.

CheckpointHook invokes this method to save checkpoints periodically.

Parameters:

out_dir (str) – The directory that checkpoints are saved.
filename (str) – The checkpoint filename.
file_client_args (dict, optional) – Arguments to instantiate a FileClient. See mmengine.fileio.FileClient for details. Defaults to None. It will be deprecated in future. Please use backend_args instead.
save_optimizer (bool) – Whether to save the optimizer to the checkpoint. Defaults to True.
save_param_scheduler (bool) – Whether to save the param_scheduler to the checkpoint. Defaults to True.
meta (dict, optional) – The meta information to be saved in the checkpoint. Defaults to None.
by_epoch (bool) – Whether the scheduled momentum is updated by epochs. Defaults to True.
backend_args (dict, optional) – Arguments to instantiate the prefix of uri corresponding backend. Defaults to None.

property seed¶

A number to set random modules.

Type:: int

test()[source]¶

Launch test.

Returns:: A dict of metrics on testing set.
Return type:: dict

property test_dataloader¶: The data loader for testing.

property test_evaluator¶

An evaluator for testing.

Type:: Evaluator

property test_loop¶

A loop to run testing.

Type:: BaseLoop

property timestamp¶

Timestamp when creating experiment.

Type:: str

train()[source]¶

Launch training.

Returns:: The model after training.
Return type:: nn.Module

property train_dataloader¶: The data loader for training.

property train_loop¶

A loop to run training.

Type:: BaseLoop

val()[source]¶

Launch validation.

Returns:: A dict of metrics on validation set.
Return type:: dict

property val_begin¶

The epoch/iteration to start running validation during training.

Type:: int

property val_dataloader¶: The data loader for validation.

property val_evaluator¶

An evaluator for validation.

Type:: Evaluator

property val_interval¶

Interval to run validation during training.

Type:: int

property val_loop¶

A loop to run validation.

Type:: BaseLoop

property work_dir¶

The working directory to save checkpoints and logs.

Type:: str

property world_size¶

Number of processes participating in the job.

Type:: int