zero.data.FnDataset

class zero.data.FnDataset(fn, args, transform=None)

A thin wrapper around a loader function and its arguments.

FnDataset allows to avoid implementing Dataset-classes (well, at least in simple cases). Below you can find the full tutorial and typical use cases, but here is a quick example:

Without FnDataset:

class ImagesList(Dataset):
    def __init__(self, filenames, transform):
        self.filenames = filenames
        self.transform = transform

    def __len__(self):
        return len(self.filenames)

    def __getitem__(self, index):
        return self.transform(Image.open(self.filenames[index]))

dataset = ImagesList(filenames, transform)

With FnDataset:

dataset = FnDataset(Image.open, filenames, transform)

Tutorial

With vanilla PyTorch, in order to create a dataset you have to inherit from torch.utils.data.Dataset and implement three methods:

• __init__

• __len__

• __getitem__

With FnDataset the only thing you may need to implement is the fn argument that will power __getitem__. The easiest way to learn FnDataset is to go through examples below.

A list of images:

dataset = FnDataset(Image.open, filenames)
# dataset[i] returns Image.open(filenames[i])

A list of images that are cached after the first load:

from functools import lru_cache
dataset = FnDataset(lru_cache(None)(Image.open), filenames)

pathlib.Path is very useful when you want to create a dataset that reads from files. For example:

images_dir = Path(...)
dataset = FnDataset(Image.open, images_dir.iterdir())

If there are many files, but you need only those with specific extensions, use pathlib.Path.glob:

dataset = FnDataset(Image.open, images_dir.glob(*.png))

If there are many files in many subfolders, but you need only those with specific extensions and that satisfy some condition, use pathlib.Path.rglob:

dataset = FnDataset(
    Image.open, (x for x in images_dir.rglob(*.png) if condition(x))
)

A segmentation dataset:

image_filenames = ...
gt_filenames = ...

def get(i):
    return Image.open(image_filenames[i]), Image.open(gt_filenames[i])

dataset = FnDataset(get, len(image_filenames))

A dummy dataset that demonstrates that FnDataset is a very general thing:

def f(x):
    return x * 10

def g(x):
    return x * 2

dataset = FnDataset(f, 3, g)
# dataset[i] returns g(f(i))
assert len(dataset) == 3
assert dataset[0] == 0
assert dataset[1] == 20
assert dataset[2] == 40

Methods

__init__	Initialize self.
__len__	Get the dataset size.
__getitem__	Get value by index.