• Stars
    star
    110
  • Rank 316,770 (Top 7 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created about 4 years ago
  • Updated about 4 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

graftr: an interactive shell to view and edit PyTorch checkpoints.

graftr

PyPI Release License

An interactive shell to view and edit PyTorch checkpoints.

graftr cast

graftr can be used to remove, rename, and move around the layers and parameters of your saved model. It's also a handy tool to peek into the structure of pre-trained PyTorch models that you can find online (e.g. Transformer, DCGAN, etc.).

The screencast above shows an example of taking a pre-trained Densenet and preparing it for integration into a larger model. We remove the final classification layer and move the feature extractor into its own densenet module.

Install

pip install graftr

Documentation

graftr presents a hierarchical directory structure for state_dicts and parameters in your checkpoint. You can list (ls), move/rename (mv), and print (cat) parameters. And, of course, you can navigate (cd) through the hierarchy. It also supports standard shell beahvior like command history, up-arrow, tab-completion, etc.

All changes are kept in-memory until you're ready to write them back to your checkpoint with save.

Supported commands

  • cd - change working directory.
  • pwd - print working directory.
  • ls - list directory contents.
  • cat - print the contents of a value or directory.
  • cp - copy value or directory.
  • mv - move/rename value or directory.
  • rm - remove value or directory.
  • parameters - print the number of model parameters under a directory.
  • shape - print tensor shape.
  • device - get or set the device of a tensor or group of tensors.
  • save - write back changes to disk.
  • where - print the location on disk where changes will be saved.
  • exit - exits the shell.

FAQ

Should this be a mountable user-mode filesystem instead?

Maybe? Some operations (e.g. shape, parameters, device) don't map easily onto standard filesystem operations. On the other hand, it would be interesting to insert/extract tensors by copying NumPy files in and out of the virtual filesystem.