fedflow.utils package

Fedflow utils

Some utils for fedflow.

class fedflow.utils.ModuleUtils

Bases: object

classmethod migrate_module(src: str, dst: str, dst_name: Optional[str] = None) → None

migrate a module from src to dst, and rename it to dst_name.

Parameters

• src – the module source dir.

• dst – target dir.

• dst_name – new module name, the module name will keep if this param is None.

Returns

classmethod import_module(name: str, path: Optional[str] = None)

Import the module dynamically

Parameters

• path – module path

• name – module name

Returns

module

classmethod exists_module(module: str) → bool

check if module is exists.

Parameters

module – the module to be checked.

Returns

a bool value.

classmethod get_name(module, obj) → Optional[str]

Get the name of obj in module.

Parameters

• module – the module to found obj.

• obj – the obj need to get name.

Returns

a string value represent the name of obj or None if not found.

Data utils

Some classes or methods for preprocessing datasets.

Trainers

Some classes or methods for training.

class fedflow.utils.trainer.SupervisedTrainer(model, optimizer, criterion, lr_scheduler=None, *, init_model_path=None, init_optim_path=None, dataset=None, batch_size=32, epoch=50, epoch_action=None, checkpoint_interval=10, device='cuda:0', console_out=None)

Bases: object

A trainer used for supervised training by pytorch.

After training of this trainer, there are 4 files will appeared in current dir:

• history.json: the data of history(contains loss and acc of train and validate).

• history.png: the chart of history

• parameter.pth: the parameters of model

• optimizer.pth: the parameters of optimizer

class History(train_loss, train_acc, val_loss, val_acc, lr)

Bases: tuple

Record history data during training

property lr

learning rate of every epoch

property train_acc

train accuracy of every epoch

property train_loss

train loss of every epoch

property val_acc

validate accuracy of every epoch

property val_loss

validate loss of every epoch

__init__(model, optimizer, criterion, lr_scheduler=None, *, init_model_path=None, init_optim_path=None, dataset=None, batch_size=32, epoch=50, epoch_action=None, checkpoint_interval=10, device='cuda:0', console_out=None)

Construct a trainer.

Parameters

• model – an instance of torch.nn.Module.

• optimizer – an instance of pytorch optimizer.

• criterion – loss function

• lr_scheduler – an instance of torch.optim.lr_scheduler._LRScheduler or torch.optim.lr_scheduler.ReduceLROnPlateau

• init_model_path – the init model parameters path.

• init_optim_path – the init optimizer parameters path.

• dataset – the datasets used for this trainer.

• batch_size – the batch size

• epoch – the epoch

• epoch_action –

  when every epoch finished, the epoch_action method will be called. In this method, you can update lr etc. The follow is an example of epoch_action:

  >>> class EpochAction(object):
  >>>     def __init__(self, optim):
  >>>         super(SupervisedTrainer, self).__init__()
  >>>         self.reduce = torch.optim.lr_scheduler.ReduceLROnPlateau(optim, mode="max")
  >>>
  >>>     # The complete method signature is:
  >>>     # def __call__(self, *, model, optimizer, criterion, lr_scheduler,
  >>>     #               train_loss, train_acc, val_loss, val_acc, lr):
  >>>     def __call__(self, *, val_acc, *args, **kwargs):
  >>>         self.reduce.step(val_acc)
  

• checkpoint_interval – the interval of save parameters, the trainer will not save parameters if this param if 0.

• device – the device used for training.

• console_out – redirect print.

mount_dataset(dataset, val_dataset=None, *, val_ratio=0.3, batch_size=32) → None

mount dataset to this trainer.

Parameters

• dataset – the complete dataset or train dataset.

• val_dataset – validate dataset, if it’s None, this method will split validate dataset from dataset.

• val_ratio – the ratio of validate dataset when split.

• batch_size – the batch size.

Returns

mount_dataloader(train_dataloader, val_dataloader) → None

Generally, this method is not recommended.

Only when the mount_dataset method unmet demand, you can directly mount a train_dataloader and a val_dataloader.

Parameters

• train_dataloader – dataloader used for training.

• val_dataloader – dataloader used for validating.

Returns

train() → dict

test(init_model_path, dataset=None, *, dataloader=None) → tuple

calculate the predict accuracy in dataset.

Parameters

• dataset – the dataset for predicting.

• dataloader – if dataloader if not None, the dataset param will be ignored.

Returns

a tuple (loss, correct, total)