FedRotatedMNIST#

class fl_sim.data_processing.FedRotatedMNIST(datadir: Path | str | None = None, num_rotations: int = 4, num_clients: int = 2400, transform: str | Callable | None = 'none', seed: int = 0)[source]#

Bases: FedVisionDataset

MNIST dataset with rotation augmentation.

The rotations are fixed and are multiples of 360 / num_rotations [Ghosh et al.[1]].

The original MNIST dataset https://pytorch.org/vision/stable/_modules/torchvision/datasets/mnist.html#MNIST contains 60,000 training images and 10,000 test images. Images are 28x28 grayscale images in 10 classes (0-9 handwritten digits).

Parameters:

  • datadir (str or pathlib.Path, optional) – Path to store the dataset. If not specified, the default path is used.

  • num_rotations (int, default 4) – Number of rotations to apply to the images in the dataset. Typical values are 2, 4.

  • num_clients (int, default 2400) – Number of clients to simulate. Typical values are 1200, 2400, 4800.

  • transform (str or callable, default 'none') – Transform (augmentation) to apply to the dataset. If ‘none’, no augmentation is applied, only the normalization transform is applied.

  • seed (int, default 0) – Random seed for reproducibility.

References

property candidate_models: Dict[str, Module]#

A set of candidate models.

property doi: List[str]#

DOI(s) related to the dataset.

download_if_needed() → None[source]#

Download data if needed.

evaluate(probs: Tensor, truths: Tensor) → Dict[str, float][source]#

Evaluation using predictions and ground truth.

Parameters:
Returns:

Evaluation results.

Return type:

Dict[str, float]

extra_repr_keys() → List[str][source]#

Extra keys for __repr__() and __str__().

get_dataloader(train_bs: int | None = None, test_bs: int | None = None, client_idx: int | None = None) → Tuple[DataLoader, DataLoader][source]#

Get local dataloader at client client_idx or get the global dataloader.

Parameters:

  • train_bs (int, optional) – Batch size for training dataloader. If None, use default batch size.

  • test_bs (int, optional) – Batch size for testing dataloader. If None, use default batch size.

  • client_idx (int, optional) – Index of the client to get dataloader. If None, get the dataloader containing all data. Usually used for centralized training.

Returns:

property label_map: dict#

Label map for the dataset.

property mirror: Dict[str, str]#

Mirror sites for downloading the dataset.

random_grid_view(nrow: int, ncol: int, save_path: Path | str | None = None) → None[source]#

Select randomly nrow x ncol images from the dataset and plot them in a grid.

Parameters:

  • nrow (int) – Number of rows in the grid.

  • ncol (int) – Number of columns in the grid.

  • save_path (Union[str, Path], optional) – Path to save the figure. If None, do not save the figure.

Return type:

None

property url: Dict[str, str]#

URLs for downloading the dataset.

view_image(client_idx: int, image_idx: int) → None[source]#

View a single image.

Parameters:

  • client_idx (int) – Index of the client on which the image is located.

  • image_idx (int) – Index of the image in the client.

Return type:

None