tf-explorer

A command-line interface tool to explore TensorFlow checkpoints.

tf-explorer presents a hierarchical directory structure for variables in your checkpoint. You can list (ls), rename (mv), and print (cat) stored variables and scopes. And, of course, you can navigate (cd) through the scope hierarchy.

All changes to variables and scopes are kept in-memory until you're ready to commit them back to your checkpoint with commit. As a matter of precaution, I recommend keeping a backup of your original checkpoint before committing any changes to it.

Tutorial

The following example opens the latest checkpoint in /path/to/checkpoint_dir, prints the current value of the global_step variable, moves it into the foo/bar/ scope, and renames the variable to old_global_step.

  $ ./tf-explorer.sh /path/to/checkpoint_dir

  Checkpoint loaded from /path/to/checkpoint_dir/model.ckpt-106901.
  Type help or ? to list commands.

  > ls
  beta1_power
  beta2_power
  @conditioning/
  global_step
  @student/
  @teacher/
  > cat global_step
  106901
  > mv global_step foo/bar/old_global_step

The @ symbol indicates that the scope is "pure", i.e. it isn't both a variable name and a scope name. You cannot print out a pure scope since it doesn't store a value.

At this point, the changes to global_step are in-memory only. If you want to view which changes are in-memory but haven't been written to disk, use the mutations command:

  > mutations
  global_step -> foo/bar/old_global_step

If you're happy with these changes, write them back to disk:

  > commit

Otherwise, you can either continue exploring / making further changes or exit tf-explorer:

  > exit
  WARNING: there are pending mutations that have not been written to disk. Discard (y/N)? y

License

Apache 2.0