Skip to main content
Version: Next

Re-run a job from previous config

Β Example (Click Here)

This is an experimental feature. Please read through this page to understand what is supported.

We use the example app linked above for demonstration. To save the configs for re-run, first use the experimental Hydra Callback for saving the job info:

_target_: hydra.experimental.callbacks.PickleJobInfoCallback
Example function
@hydra.main(version_base=None, config_path=".", config_name="config")
def my_app(cfg: DictConfig) -> None:"output_dir={HydraConfig.get().runtime.output_dir}")"{}")

Run the example app:

$ python
[2022-03-16 14:51:30,905][hydra.experimental.pickle_job_info_callback][INFO] - Saving job configs in /Users/jieru/workspace/hydra/examples/experimental/outputs/2022-03-16/14-51-30/.hydra/config.pickle
[2022-03-16 14:51:30,906][__main__][INFO] - Output_dir=/Users/jieru/workspace/hydra/examples/experimental/outputs/2022-03-16/14-51-30
[2022-03-16 14:51:30,906][__main__][INFO] -
[2022-03-16 14:51:30,906][hydra.experimental.pickle_job_info_callback][INFO] - Saving job_return in /Users/jieru/workspace/hydra/examples/experimental/outputs/2022-03-16/14-51-30/.hydra/job_return.pickle

The Callback saves config.pickle in .hydra sub dir, this is what we will use for rerun.

Now rerun the app

$ OUTPUT_DIR=/Users/jieru/workspace/hydra/examples/experimental/outputs/2022-03-16/14-51-30/.hydra/
$ python --experimental-rerun $OUTPUT_DIR/config.pickle
/Users/jieru/workspace/hydra/hydra/ UserWarning: Experimental rerun CLI option.
warnings.warn(msg, UserWarning)
[2022-03-16 14:59:21,666][__main__][INFO] - Output_dir=/Users/jieru/workspace/hydra/examples/experimental/outputs/2022-03-16/14-51-30
[2022-03-16 14:59:21,666][__main__][INFO] -

You will notice my_app.log is updated with the logging from the second run, but Callbacks are not called this time. Read on to learn more.

Important Notes​

This is an experimental feature. Please reach out if you have any question.

  • Only single run is supported.
  • --experimental-rerun cannot be used with other command-line options or overrides. They will simply be ignored.
  • Rerun passes in a cfg_passthrough directly to your application, this means except for logging, no other hydra.main functions are called (such as change working dir, or calling callbacks.)
  • The configs are preserved and reconstructed to the best efforts. Meaning we can only guarantee that the cfg object itself passed in by hydra.main stays the same across runs. However, configs are resolved lazily. Meaning we cannot guarantee your application will behave the same if your application resolves configs during run time. In the following example, cfg.time_now will resolve to different value every run.
time_now: ${now:%H-%M-%S}

Example function
@hydra.main(version_base=None, config_path=".", config_name="config")
def my_app(cfg: DictConfig) -> None:
val = cfg.time_now
# the rest of the application