To help you get started, we've collected the most common ways that torchvision is being used within popular public projects.

kakaobrain / torchgpipe / benchmarks / resnet101-accuracy / main.py View on Github

def dataloaders(batch_size: int, num_workers: int = 32) -> Tuple[DataLoader, DataLoader]:
    num_workers = num_workers if batch_size <= 4096 else num_workers // 2

    post_transforms = torchvision.transforms.Compose([
        torchvision.transforms.ToTensor(),
        torchvision.transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
    ])

    train_dataset = torchvision.datasets.ImageNet(
        root='imagenet',
        split='train',
        transform=torchvision.transforms.Compose([
            torchvision.transforms.RandomResizedCrop(224, scale=(0.08, 1.0)),
            torchvision.transforms.RandomHorizontalFlip(),
            post_transforms,
        ])
    )
    test_dataset = torchvision.datasets.ImageNet(
        root='imagenet',
        split='val',
        transform=torchvision.transforms.Compose([

View more ways to use torchvision

torchvision

The torchvision package consists of popular datasets, model architectures, and common image transformations for computer vision.

Installation

We recommend Anaconda as Python package management system. Please refer to pytorch.org for the detail of PyTorch (torch) installation. The following is the corresponding torchvision versions and supported Python versions.

Anaconda:

conda install torchvision -c pytorch

pip:

pip install torchvision

From source:

python setup.py install
# or, for OSX
# MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install

We don't officially support building from source using pip, but if you do, you'll need to use the --no-build-isolation flag. In case building TorchVision from source fails, install the nightly version of PyTorch following the linked guide on the contributing page and retry the install.

By default, GPU support is built if CUDA is found and torch.cuda.is_available() is true. It's possible to force building GPU support by setting FORCE_CUDA=1 environment variable, which is useful when building a docker image.

Image Backend

Torchvision currently supports the following image backends:

• Pillow (default)
• Pillow-SIMD - a much faster drop-in replacement for Pillow with SIMD. If installed will be used as the default.
• accimage - if installed can be activated by calling torchvision.set_image_backend('accimage')
• libpng - can be installed via conda conda install libpng or any of the package managers for debian-based and RHEL-based Linux distributions.
• libjpeg - can be installed via conda conda install jpeg or any of the package managers for debian-based and RHEL-based Linux distributions. libjpeg-turbo can be used as well.

Notes: libpng and libjpeg must be available at compilation time in order to be available. Make sure that it is available on the standard library locations, otherwise, add the include and library paths in the environment variables TORCHVISION_INCLUDE and TORCHVISION_LIBRARY, respectively.

Video Backend

Torchvision currently supports the following video backends:

• pyav (default) - Pythonic binding for ffmpeg libraries.
• video_reader - This needs ffmpeg to be installed and torchvision to be built from source. There shouldn't be any conflicting version of ffmpeg installed. Currently, this is only supported on Linux.

conda install -c conda-forge ffmpeg
python setup.py install

Using the models on C++

TorchVision provides an example project for how to use the models on C++ using JIT Script.

Installation From source:

mkdir build
cd build
# Add -DWITH_CUDA=on support for the CUDA if needed
cmake ..
make
make install

Once installed, the library can be accessed in cmake (after properly configuring CMAKE_PREFIX_PATH) via the TorchVision::TorchVision target:

find_package(TorchVision REQUIRED)
target_link_libraries(my-target PUBLIC TorchVision::TorchVision)

The TorchVision package will also automatically look for the Torch package and add it as a dependency to my-target, so make sure that it is also available to cmake via the CMAKE_PREFIX_PATH.

For an example setup, take a look at examples/cpp/hello_world.

Python linking is disabled by default when compiling TorchVision with CMake, this allows you to run models without any Python dependency. In some special cases where TorchVision's operators are used from Python code, you may need to link to Python. This can be done by passing -DUSE_PYTHON=on to CMake.

TorchVision Operators

In order to get the torchvision operators registered with torch (eg. for the JIT), all you need to do is to ensure that you #include in your project.

Documentation

You can find the API documentation on the pytorch website:

Contributing

See the CONTRIBUTING file for how to help out.

Disclaimer on Datasets

This is a utility library that downloads and prepares public datasets. We do not host or distribute these datasets, vouch for their quality or fairness, or claim that you have license to use the dataset. It is your responsibility to determine whether you have permission to use the dataset under the dataset's license.

If you're a dataset owner and wish to update any part of it (description, citation, etc.), or do not want your dataset to be included in this library, please get in touch through a GitHub issue. Thanks for your contribution to the ML community!

Pre-trained Model License

The pre-trained models provided in this library may have their own licenses or terms and conditions derived from the dataset used for training. It is your responsibility to determine whether you have permission to use the models for your use case.

More specifically, SWAG models are released under the CC-BY-NC 4.0 license. See SWAG LICENSE for additional details.

Citing TorchVision

If you find TorchVision useful in your work, please consider citing the following BibTeX entry:

@software{torchvision2016,
    title        = {TorchVision: PyTorch's Computer Vision library},
    author       = {TorchVision maintainers and contributors},
    year         = 2016,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/pytorch/vision}}
}