Skip to content

PyTorch dataset extended with map, cache etc. (tensorflow.data like)

License

Notifications You must be signed in to change notification settings

szymonmaszke/torchdatasets

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

Package renamed to torchdatasets!

  • Use map, apply, reduce or filter directly on Dataset objects
  • cache data in RAM/disk or via your own method (partial caching supported)
  • Full PyTorch's Dataset and IterableDataset support
  • General torchdatasets.maps like Flatten or Select
  • Extensible interface (your own cache methods, cache modifiers, maps etc.)
  • Useful torchdatasets.datasets classes designed for general tasks (e.g. file reading)
  • Support for torchvision datasets (e.g. ImageFolder, MNIST, CIFAR10) via td.datasets.WrapDataset
  • Minimal overhead (single call to super().__init__())
Version Docs Tests Coverage Style PyPI Python PyTorch Docker Roadmap
Version Documentation Tests Coverage codebeat PyPI Python PyTorch Docker Roadmap

๐Ÿ’ก Examples

Check documentation here: https://szymonmaszke.github.io/torchdatasets

General example

  • Create image dataset, convert it to Tensors, cache and concatenate with smoothed labels:
import torchdatasets as td
import torchvision

class Images(td.Dataset): # Different inheritance
    def __init__(self, path: str):
        super().__init__() # This is the only change
        self.files = [file for file in pathlib.Path(path).glob("*")]

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

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


images = Images("./data").map(torchvision.transforms.ToTensor()).cache()

You can concatenate above dataset with another (say labels) and iterate over them as per usual:

for data, label in images | labels:
    # Do whatever you want with your data
  • Cache first 1000 samples in memory, save the rest on disk in folder ./cache:
images = (
    ImageDataset.from_folder("./data").map(torchvision.transforms.ToTensor())
    # First 1000 samples in memory
    .cache(td.modifiers.UpToIndex(1000, td.cachers.Memory()))
    # Sample from 1000 to the end saved with Pickle on disk
    .cache(td.modifiers.FromIndex(1000, td.cachers.Pickle("./cache")))
    # You can define your own cachers, modifiers, see docs
)

To see what else you can do please check torchdatasets documentation

Integration with torchvision

Using torchdatasets you can easily split torchvision datasets and apply augmentation only to the training part of data without any troubles:

import torchvision

import torchdatasets as td

# Wrap torchvision dataset with WrapDataset
dataset = td.datasets.WrapDataset(torchvision.datasets.ImageFolder("./images"))

# Split dataset
train_dataset, validation_dataset, test_dataset = torch.utils.data.random_split(
    model_dataset,
    (int(0.6 * len(dataset)), int(0.2 * len(dataset)), int(0.2 * len(dataset))),
)

# Apply torchvision mappings ONLY to train dataset
train_dataset.map(
    td.maps.To(
        torchvision.transforms.Compose(
            [
                torchvision.transforms.RandomResizedCrop(224),
                torchvision.transforms.RandomHorizontalFlip(),
                torchvision.transforms.ToTensor(),
                torchvision.transforms.Normalize(
                    mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
                ),
            ]
        )
    ),
    # Apply this transformation to zeroth sample
    # First sample is the label
    0,
)

Please notice you can use td.datasets.WrapDataset with any existing torch.utils.data.Dataset instance to give it additional caching and mapping powers!

๐Ÿ”ง Installation

๐Ÿ pip

Latest release:

pip install --user torchdatasets

Nightly:

pip install --user torchdatasets-nightly

๐Ÿ‹ Docker

CPU standalone and various versions of GPU enabled images are available at dockerhub.

For CPU quickstart, issue:

docker pull szymonmaszke/torchdatasets:18.04

Nightly builds are also available, just prefix tag with nightly_. If you are going for GPU image make sure you have nvidia/docker installed and it's runtime set.

โ“ Contributing

If you find any issue or you think some functionality may be useful to others and fits this library, please open new Issue or create Pull Request.

To get an overview of thins one can do to help this project, see Roadmap