tensorplay.hub

How to implement an entrypoint?

Here is a code snippet specifies an entrypoint for alexnet model if we expand the implementation in tensorplay/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.

  dependencies = ['tensorplay']
  from tensorplay.vision.models.alexnet import alexnet as _alexnet

  # alexnet is the name of entrypoint
  def alexnet(pretrained=False, **kwargs):
      """ # This docstring shows up in hub.help()
      Alexnet model
      pretrained (bool): kwargs, load pretrained weights into the model
      """
      # Call the model, load pretrained weights
      model = _alexnet(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 tensorplay.hub.list().

• Pretrained weights can either be stored locally in the GitHub repo, or loadable by tensorplay.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 tensorplay.vision.models.alexnet.alexnet 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 = tensorplay.load(checkpoint)
      model.load_state_dict(state_dict)

      # For checkpoint saved elsewhere
      checkpoint = 'https://download.tensorplay.cn/models/alexnet-owt-7be5be79.pth'
      model.load_state_dict(tensorplay.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.

Running a loaded model:

Note that *args and **kwargs in tensorplay.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.

Loading from a local directory

Besides downloading from GitHub, tensorplay.hub.load() accepts source='local' to load an entrypoint from a directory on disk that contains a hubconf.py, which is handy while developing a hub repository.

Where are my downloaded models saved?

The locations are used in the order of

• Calling hub.set_dir(<PATH_TO_HUB_DIR>)

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

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

• ~/.cache/tensorplay/hub

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.

For GitHub sources the cache records a .commit pin: the exact commit the snapshot was fetched at. On subsequent loads the ref’s current commit is checked (one API call; set GITHUB_TOKEN to raise the rate limit) — an unchanged ref reuses the cache, a moved ref refreshes it, and when the API is unreachable the cache is used as-is. Pass skip_validation=True to skip the check entirely and never contact the API on a cache hit.

Downloads are resumable: an interrupted transfer leaves a <file>.part file that the next attempt continues from, and a completed download writes a <file>.sha256 sidecar used to verify the file on later cache hits.

Trust and validation

Loading a repository executes its hubconf.py, so the code must be treated as untrusted. By default the ref (branch/tag) is checked against the branches and tags of the repository owner (set GITHUB_TOKEN to raise the API rate limit), and first-time downloads from unknown owners prompt for acknowledgement. Pass trust_repo=True to skip the prompt, or skip_validation=True to skip the ref check (needed e.g. for refs that only exist on a fork).

MEGA backend

Weights can also be served from the MEGA hub (mega.tensorplay.cn, via the optional megatensors package). MEGA repos are artifact repositories: no hubconf.py is executed, downloads are revision-resolved, resumable, and integrity-checked by the SDK.

import tensorplay

# full snapshot into the hub cache
tensorplay.hub.snapshot_download("org/model", source="mega", revision="main")

# state dict from a MEGA repo (revision-pinned)
state_dict = tensorplay.hub.load_state_dict("org/model", device="cpu")

# or straight to a ready-to-run model
model = tensorplay.hub.load_model("org/model", model_class=MyNet)

trust_policy accepts a megatensors TrustPolicy on load_state_dict and load_model: the loader then verifies the signature provenance of every served artifact before reading tensors. Unsigned or wrongly signed artifacts are rejected (or warn, with strict=False). This is MEGA-only; GitHub sources are governed by trust_repo instead.

Known limitations:

The 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. Loading two repos in the same python process is fine as long as their sub-package names do not collide; separate processes are always safe.