Skip to main content
Version: Next

Multi-run

Sometimes you want to run the same application with multiple different configurations.
E.g. running a performance test on each of the databases with each of the schemas.

You can multirun a Hydra application via either commandline or configuration:

Configure hydra.mode (new in Hydra 1.2)​

You can configure hydra.mode in any supported way. The legal values are RUN and MULTIRUN. The following shows how to override from the command-line and sweep over all combinations of the dbs and schemas. Setting hydra.mode=MULTIRUN in your input config would make your application multi-run by default.

$ python my_app.py hydra.mode=MULTIRUN db=mysql,postgresql schema=warehouse,support,school
[2021-01-20 17:25:03,317][HYDRA] Launching 6 jobs locally
[2021-01-20 17:25:03,318][HYDRA] #0 : db=mysql schema=warehouse
[2021-01-20 17:25:03,458][HYDRA] #1 : db=mysql schema=support
[2021-01-20 17:25:03,602][HYDRA] #2 : db=mysql schema=school
[2021-01-20 17:25:03,755][HYDRA] #3 : db=postgresql schema=warehouse
[2021-01-20 17:25:03,895][HYDRA] #4 : db=postgresql schema=support
[2021-01-20 17:25:04,040][HYDRA] #5 : db=postgresql schema=school

The printed configurations have been omitted for brevity.

--multirun (-m) from the command-line​

You can achieve the above from command-line as well:

python my_app.py --multirun db=mysql,postgresql schema=warehouse,support,school

or

python my_app.py -m db=mysql,postgresql schema=warehouse,support,school

You can access hydra.mode at runtime to determine whether the application is in RUN or MULTIRUN mode. Check here on how to access Hydra config at run time.

If conflicts arise (eg, hydra.mode=RUN and the application was run with --multirun), Hydra will determine the value of hydra.mode at run time. The following table shows what runtime hydra.mode value you'd get with different input configs and commandline combinations.

No multirun commandline flag--multirun ( -m)
hydra.mode=RUNRunMode.RUNRunMode.MULTIRUN (with UserWarning)
hydra.mode=MULTIRUNRunMode.MULTIRUNRunMode.MULTIRUN
hydra.mode=None (default)RunMode.RUNRunMode.MULTIRUN
important

Hydra composes configs lazily at job launching time. If you change code or configs after launching a job/sweep, the final composed configs might be impacted.

Sweeping via hydra.sweeper.params​

Β Example (Click Here)

You can also define sweeping in the input configs by overriding hydra.sweeper.params. Using the above example, the same multirun could be achieved via the following config.

hydra:
sweeper:
params:
db: mysql,postgresql
schema: warehouse,support,school

The syntax are consistent for both input configs and commandline overrides. If a sweep is specified in both an input config and at the command line, then the commandline sweep will take precedence over the sweep defined in the input config. If we run the same application with the above input config and a new commandline override:

$ python my_app.py -m db=mysql
[2021-01-20 17:25:03,317][HYDRA] Launching 3 jobs locally
[2021-01-20 17:25:03,318][HYDRA] #0 : db=mysql schema=warehouse
[2021-01-20 17:25:03,458][HYDRA] #1 : db=mysql schema=support
[2021-01-20 17:25:03,602][HYDRA] #2 : db=mysql schema=school
info

The above configuration methods only apply to Hydra's default BasicSweeper for now. For other sweepers, please checkout the corresponding documentations.

Additional sweep types​

Hydra supports other kinds of sweeps, e.g:

x=range(1,10)                  # 1-9
schema=glob(*) # warehouse,support,school
schema=glob(*,exclude=w*) # support,school

See the Extended Override syntax for details.

Sweeper​

The default sweeping logic is built into Hydra. Additional sweepers are available as plugins. For example, the Ax Sweeper can automatically find the best parameter combination!

Launcher​

By default, Hydra runs your multi-run jobs locally and serially. Other launchers are available as plugins for launching in parallel and on different clusters. For example, the JobLib Launcher can execute the different parameter combinations in parallel on your local machine using multi-processing.