mmengine/docs/en/common_usage/epoch_to_iter.md

EpochBasedTraining to IterBasedTraining

Epoch-based training and iteration-based training are two commonly used training way in MMEngine. For example, downstream repositories like MMDetection choose to train the model by epoch and MMSegmentation choose to train the model by iteration.

Many modules in MMEngine default to training models by epoch, such as ParamScheduler, LoggerHook, CheckPointHook, etc. Therefore, you need to adjust the configuration of these modules if you want to train by iteration. For example, a commonly used epoch based configuration is as follows:

param_scheduler = dict(
    type='MultiStepLR',
    milestones=[6, 8]
    by_epoch=True  # by_epoch is True by default
)

default_hooks = dict(
    logger=dict(type='LoggerHook', log_metric_by_epoch=True),  # log_metric_by_epoch is True by default
    checkpoint=dict(type='CheckpointHook', interval=2, by_epoch=True),  # by_epoch is True by default
)

train_cfg = dict(
    by_epoch=True,  # set by_epoch=True or type='EpochBasedTrainLoop'
    max_epochs=10,
    val_interval=2
)

log_processor = dict(
    by_epoch=True
)  # This is the default configuration, and just set it here for comparison.

runner = Runner(
    model=ResNet18(),
    work_dir='./work_dir',
    # Assuming train_dataloader is configured with an epoch-based sampler
    train_dataloader=train_dataloader_cfg,
    optim_wrapper=dict(optimizer=dict(type='SGD', lr=0.001, momentum=0.9)),
    param_scheduler=param_scheduler
    default_hooks=default_hooks,
    log_processor=log_processor,
    train_cfg=train_cfg,
    resume=True,
)

There are four steps to convert the above configuration to iteration based training:

1. Set by_epoch in train_cfg to False, and set max_iters to the total number of training iterations and val_interval to the interval between validation iterations.

   train_cfg = dict(
       by_epoch=False,
       max_iters=10000,
       val_interval=2000
     )
   

2. Set log_metric_by_epoch to False in logger and by_epoch to False in checkpoint.

   default_hooks = dict(
       logger=dict(type='LoggerHook', log_metric_by_epoch=False),
       checkpoint=dict(type='CheckpointHook', by_epoch=False, interval=2000),
   )
   

3. Set by_epoch in param_scheduler to False and convert any epoch-related parameters to iteration.

   param_scheduler = dict(
       type='MultiStepLR',
       milestones=[6000, 8000],
       by_epoch=False,
   )
   

   Alternatively, if you can ensure that the total number of iterations for IterBasedTraining and EpochBasedTraining is the same, simply set convert_to_iter_based to True.

   param_scheduler = dict(
       type='MultiStepLR',
       milestones=[6, 8]
       convert_to_iter_based=True
   )
   

4. Set by_epoch in log_processor to False.

   log_processor = dict(
       by_epoch=False
   )
   

Take training CIFAR10 as an example:

Step	Training by epoch	Training by iteration
Build model	import torch.nn.functional as F import torchvision from mmengine.model import BaseModel class MMResNet50(BaseModel): def __init__(self): super().__init__() self.resnet = torchvision.models.resnet50() def forward(self, imgs, labels, mode): x = self.resnet(imgs) if mode == 'loss': return {'loss': F.cross_entropy(x, labels)} elif mode == 'predict': return x, labels
Build dataloader	import torchvision.transforms as transforms from torch.utils.data import DataLoader norm_cfg = dict(mean=[0.491, 0.482, 0.447], std=[0.202, 0.199, 0.201]) train_dataloader = DataLoader( batch_size=32, shuffle=True, dataset=torchvision.datasets.CIFAR10( 'data/cifar10', train=True, download=True, transform=transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(**norm_cfg)]))) val_dataloader = DataLoader( batch_size=32, shuffle=False, dataset=torchvision.datasets.CIFAR10( 'data/cifar10', train=False, download=True, transform=transforms.Compose([ transforms.ToTensor(), transforms.Normalize(**norm_cfg)])))
Prepare metric	from mmengine.evaluator import BaseMetric class Accuracy(BaseMetric): def process(self, data_batch, data_samples): score, gt = data_samples # save the middle result of a batch to `self.results` self.results.append({ 'batch_size': len(gt), 'correct': (score.argmax(dim=1) == gt).sum().cpu(), }) def compute_metrics(self, results): total_correct = sum(item['correct'] for item in results) total_size = sum(item['batch_size'] for item in results) # return the dict containing the eval results # the key is the name of the metric name return dict(accuracy=100 * total_correct / total_size)
Configure default hooks	default_hooks = dict( logger=dict(type='LoggerHook', log_metric_by_epoch=True), checkpoint=dict(type='CheckpointHook', interval=2, by_epoch=True), )	default_hooks = dict( logger=dict(type='LoggerHook', log_metric_by_epoch=False), checkpoint=dict(type='CheckpointHook', by_epoch=False, interval=2000), )
Configure parameter scheduler	param_scheduler = dict( type='MultiStepLR', milestones=[6, 8] by_epoch=True )	param_scheduler = dict( type='MultiStepLR', milestones=[6000, 8000], by_epoch=False, )
Configure log_processor	# The default configuration of log_processor is used for epoch based training. # Defining it here additionally is for building runner with the same way. log_processor = dict(by_epoch=True)	log_processor = dict(by_epoch=False)
Configure train_cfg	train_cfg = dict( by_epoch=True, max_epochs=10, val_interval=2 )	train_cfg = dict( by_epoch=False, max_iters=10000, val_interval=2000 )
Build Runner	from torch.optim import SGD from mmengine.runner import Runner runner = Runner( model=MMResNet50(), work_dir='./work_dir', train_dataloader=train_dataloader, optim_wrapper=dict(optimizer=dict(type=SGD, lr=0.001, momentum=0.9)), train_cfg=train_cfg, log_processor=log_processor, default_hooks=default_hooks, val_dataloader=val_dataloader, val_cfg=dict(), val_evaluator=dict(type=Accuracy), ) runner.train()

If the base configuration file has configured a epoch/iteration based sampler for the train_dataloader, then it is necessary to change it to a specified type of sampler in the current configuration file, or set it to None. When the sampler in the dataloader is set to None, MMEngine will choose either the InfiniteSampler (when by_epoch is False) or the DefaultSampler (when by_epoch is True) according to the train_cfg parameter.

If `type` is configured for the `train_cfg` in the base configuration, you must overwrite the type to target type (EpochBasedTrainLoop or IterBasedTrainLoop) rather than simply set `by_epoch` to True/False.