torch.hub

Pytorch Hub is a pre-trained model repository designed to facilitate research reproducibility.

Publishing models

Pytorch Hub supports publishing pre-trained models(model definitions and pre-trained weights) to a GitHub repository by adding a simple hubconf.py file;

hubconf.py can have multiple entrypoints. Each entrypoint is defined as a python function (example: a pre-trained model you want to publish).

def entrypoint_name(*args, **kwargs):
    # args & kwargs are optional, for models which take positional/keyword arguments.
    ...

How to implement an entrypoint?

Here is a code snippet specifies an entrypoint for resnet18 model if we expand the implementation in pytorch/vision/hubconf.py. In most case importing the right function in hubconf.py is sufficient. Here we just want to use the expanded version as an example to show how it works. You can see the full script in pytorch/vision repo

dependencies = ['torch']
from torchvision.models.resnet import resnet18 as _resnet18

# resnet18 is the name of entrypoint
def resnet18(pretrained=False, **kwargs):
    """ # This docstring shows up in hub.help()
    Resnet18 model
    pretrained (bool): kwargs, load pretrained weights into the model
    """
    # Call the model, load pretrained weights
    model = _resnet18(pretrained=pretrained, **kwargs)
    return model

• dependencies variable is a list of package names required to load the model. Note this might be slightly different from dependencies required for training a model.

• args and kwargs are passed along to the real callable function.

• Docstring of the function works as a help message. It explains what does the model do and what are the allowed positional/keyword arguments. It’s highly recommended to add a few examples here.

• Entrypoint function can either return a model(nn.module), or auxiliary tools to make the user workflow smoother, e.g. tokenizers.

• Callables prefixed with underscore are considered as helper functions which won’t show up in torch.hub.list().

• Pretrained weights can either be stored locally in the GitHub repo, or loadable by torch.hub.load_state_dict_from_url(). If less than 2GB, it’s recommended to attach it to a project release and use the url from the release. In the example above torchvision.models.resnet.resnet18 handles pretrained, alternatively you can put the following logic in the entrypoint definition.

if pretrained:
    # For checkpoint saved in local GitHub repo, e.g. <RELATIVE_PATH_TO_CHECKPOINT>=weights/save.pth
    dirname = os.path.dirname(__file__)
    checkpoint = os.path.join(dirname, <RELATIVE_PATH_TO_CHECKPOINT>)
    state_dict = torch.load(checkpoint)
    model.load_state_dict(state_dict)

    # For checkpoint saved elsewhere
    checkpoint = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'
    model.load_state_dict(torch.hub.load_state_dict_from_url(checkpoint, progress=False))

Important Notice

• The published models should be at least in a branch/tag. It can’t be a random commit.

Loading models from Hub

Pytorch Hub provides convenient APIs to explore all available models in hub through torch.hub.list(), show docstring and examples through torch.hub.help() and load the pre-trained models using torch.hub.load().

torch.hub.list(github, force_reload=False, skip_validation=False, trust_repo=None, verbose=True)

List all callable entrypoints available in the repo specified by github.

Parameters

• github (str) – a string with format “repo_owner/repo_name[:ref]” with an optional ref (tag or branch). If ref is not specified, the default branch is assumed to be main if it exists, and otherwise master. Example: ‘pytorch/vision:0.10’

• force_reload (bool, optional) – whether to discard the existing cache and force a fresh download. Default is False.

• skip_validation (bool, optional) – if False, torchhub will check that the branch or commit specified by the github argument properly belongs to the repo owner. This will make requests to the GitHub API; you can specify a non-default GitHub token by setting the GITHUB_TOKEN environment variable. Default is False.

• trust_repo (bool, str or None) –

  "check", True, False or None. This parameter was introduced in v1.12 and helps ensuring that users only run code from repos that they trust.

  • If False, a prompt will ask the user whether the repo should be trusted.

  • If True, the repo will be added to the trusted list and loaded without requiring explicit confirmation.

  • If "check", the repo will be checked against the list of trusted repos in the cache. If it is not present in that list, the behaviour will fall back onto the trust_repo=False option.

  • If None: this will raise a warning, inviting the user to set trust_repo to either False, True or "check". This is only present for backward compatibility and will be removed in v2.0.

  Default is None and will eventually change to "check" in v2.0.

• verbose (bool, optional) – If False, mute messages about hitting local caches. Note that the message about first download cannot be muted. Default is True.

Returns

The available callables entrypoint

Return type

list

Example

>>> entrypoints = torch.hub.list('pytorch/vision', force_reload=True)

torch.hub.help(github, model, force_reload=False, skip_validation=False, trust_repo=None)

Show the docstring of entrypoint model.

Parameters

• github (str) – a string with format <repo_owner/repo_name[:ref]> with an optional ref (a tag or a branch). If ref is not specified, the default branch is assumed to be main if it exists, and otherwise master. Example: ‘pytorch/vision:0.10’

• model (str) – a string of entrypoint name defined in repo’s hubconf.py

• force_reload (bool, optional) – whether to discard the existing cache and force a fresh download. Default is False.

• skip_validation (bool, optional) – if False, torchhub will check that the ref specified by the github argument properly belongs to the repo owner. This will make requests to the GitHub API; you can specify a non-default GitHub token by setting the GITHUB_TOKEN environment variable. Default is False.

• trust_repo (bool, str or None) –

  "check", True, False or None. This parameter was introduced in v1.12 and helps ensuring that users only run code from repos that they trust.

  • If False, a prompt will ask the user whether the repo should be trusted.

  • If True, the repo will be added to the trusted list and loaded without requiring explicit confirmation.

  • If "check", the repo will be checked against the list of trusted repos in the cache. If it is not present in that list, the behaviour will fall back onto the trust_repo=False option.

  • If None: this will raise a warning, inviting the user to set trust_repo to either False, True or "check". This is only present for backward compatibility and will be removed in v2.0.

  Default is None and will eventually change to "check" in v2.0.

Example

>>> print(torch.hub.help('pytorch/vision', 'resnet18', force_reload=True))

torch.hub.load(repo_or_dir, model, *args, source='github', trust_repo=None, force_reload=False, verbose=True, skip_validation=False, **kwargs)

Load a model from a github repo or a local directory.

Note: Loading a model is the typical use case, but this can also be used to for loading other objects such as tokenizers, loss functions, etc.

If source is ‘github’, repo_or_dir is expected to be of the form repo_owner/repo_name[:ref] with an optional ref (a tag or a branch).

If source is ‘local’, repo_or_dir is expected to be a path to a local directory.

Parameters

• repo_or_dir (str) – If source is ‘github’, this should correspond to a github repo with format repo_owner/repo_name[:ref] with an optional ref (tag or branch), for example ‘pytorch/vision:0.10’. If ref is not specified, the default branch is assumed to be main if it exists, and otherwise master. If source is ‘local’ then it should be a path to a local directory.

• model (str) – the name of a callable (entrypoint) defined in the repo/dir’s hubconf.py.

• *args (optional) – the corresponding args for callable model.

• source (str, optional) – ‘github’ or ‘local’. Specifies how repo_or_dir is to be interpreted. Default is ‘github’.

• trust_repo (bool, str or None) –

  "check", True, False or None. This parameter was introduced in v1.12 and helps ensuring that users only run code from repos that they trust.

  • If False, a prompt will ask the user whether the repo should be trusted.

  • If True, the repo will be added to the trusted list and loaded without requiring explicit confirmation.

  • If "check", the repo will be checked against the list of trusted repos in the cache. If it is not present in that list, the behaviour will fall back onto the trust_repo=False option.

  • If None: this will raise a warning, inviting the user to set trust_repo to either False, True or "check". This is only present for backward compatibility and will be removed in v2.0.

  Default is None and will eventually change to "check" in v2.0.

• force_reload (bool, optional) – whether to force a fresh download of the github repo unconditionally. Does not have any effect if source = 'local'. Default is False.

• verbose (bool, optional) – If False, mute messages about hitting local caches. Note that the message about first download cannot be muted. Does not have any effect if source = 'local'. Default is True.

• skip_validation (bool, optional) – if False, torchhub will check that the branch or commit specified by the github argument properly belongs to the repo owner. This will make requests to the GitHub API; you can specify a non-default GitHub token by setting the GITHUB_TOKEN environment variable. Default is False.

• **kwargs (optional) – the corresponding kwargs for callable model.

Returns

The output of the model callable when called with the given *args and **kwargs.

Example

>>> # from a github repo
>>> repo = 'pytorch/vision'
>>> model = torch.hub.load(repo, 'resnet50', weights='ResNet50_Weights.IMAGENET1K_V1')
>>> # from a local directory
>>> path = '/some/local/path/pytorch/vision'
>>> model = torch.hub.load(path, 'resnet50', weights='ResNet50_Weights.DEFAULT')

torch.hub.download_url_to_file(url, dst, hash_prefix=None, progress=True)

Download object at the given URL to a local path.

Parameters

• url (str) – URL of the object to download

• dst (str) – Full path where object will be saved, e.g. /tmp/temporary_file

• hash_prefix (str, optional) – If not None, the SHA256 downloaded file should start with hash_prefix. Default: None

• progress (bool, optional) – whether or not to display a progress bar to stderr Default: True

Example

>>> torch.hub.download_url_to_file('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth', '/tmp/temporary_file')

torch.hub.load_state_dict_from_url(url, model_dir=None, map_location=None, progress=True, check_hash=False, file_name=None, weights_only=False)

Loads the Torch serialized object at the given URL.

If downloaded file is a zip file, it will be automatically decompressed.

If the object is already present in model_dir, it’s deserialized and returned. The default value of model_dir is <hub_dir>/checkpoints where hub_dir is the directory returned by get_dir().

Parameters

• url (str) – URL of the object to download

• model_dir (str, optional) – directory in which to save the object

• map_location (optional) – a function or a dict specifying how to remap storage locations (see torch.load)

• progress (bool, optional) – whether or not to display a progress bar to stderr. Default: True

• check_hash (bool, optional) – If True, the filename part of the URL should follow the naming convention filename-<sha256>.ext where <sha256> is the first eight or more digits of the SHA256 hash of the contents of the file. The hash is used to ensure unique names and to verify the contents of the file. Default: False

• file_name (str, optional) – name for the downloaded file. Filename from url will be used if not set.

• weights_only (bool, optional) – If True, only weights will be loaded and no complex pickled objects. Recommended for untrusted sources. See load() for more details.

Return type

Dict[str, Any]

Example

>>> state_dict = torch.hub.load_state_dict_from_url('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth')

Running a loaded model:

Note that *args and **kwargs in torch.hub.load() are used to instantiate a model. After you have loaded a model, how can you find out what you can do with the model? A suggested workflow is

• dir(model) to see all available methods of the model.

• help(model.foo) to check what arguments model.foo takes to run

To help users explore without referring to documentation back and forth, we strongly recommend repo owners make function help messages clear and succinct. It’s also helpful to include a minimal working example.

Where are my downloaded models saved?

The locations are used in the order of

• Calling hub.set_dir(<PATH_TO_HUB_DIR>)

• $TORCH_HOME/hub, if environment variable TORCH_HOME is set.

• $XDG_CACHE_HOME/torch/hub, if environment variable XDG_CACHE_HOME is set.

• ~/.cache/torch/hub

torch.hub.get_dir()

Get the Torch Hub cache directory used for storing downloaded models & weights.

If set_dir() is not called, default path is $TORCH_HOME/hub where environment variable $TORCH_HOME defaults to $XDG_CACHE_HOME/torch. $XDG_CACHE_HOME follows the X Design Group specification of the Linux filesystem layout, with a default value ~/.cache if the environment variable is not set.

torch.hub.set_dir(d)

Optionally set the Torch Hub directory used to save downloaded models & weights.

Parameters

d (str) – path to a local folder to save downloaded models & weights.

Caching logic

By default, we don’t clean up files after loading it. Hub uses the cache by default if it already exists in the directory returned by get_dir().

Users can force a reload by calling hub.load(..., force_reload=True). This will delete the existing GitHub folder and downloaded weights, reinitialize a fresh download. This is useful when updates are published to the same branch, users can keep up with the latest release.

Known limitations:

Torch hub works by importing the package as if it was installed. There are some side effects introduced by importing in Python. For example, you can see new items in Python caches sys.modules and sys.path_importer_cache which is normal Python behavior. This also means that you may have import errors when importing different models from different repos, if the repos have the same sub-package names (typically, a model subpackage). A workaround for these kinds of import errors is to remove the offending sub-package from the sys.modules dict; more details can be found in this GitHub issue.

A known limitation that is worth mentioning here: users CANNOT load two different branches of the same repo in the same python process. It’s just like installing two packages with the same name in Python, which is not good. Cache might join the party and give you surprises if you actually try that. Of course it’s totally fine to load them in separate processes.