Customizing working directory pattern

Run output directory grouped by day:

hydra:
  run:
    dir: ./outputs/${now:%Y-%m-%d}/${now:%H-%M-%S}

Sweep sub directory contains the the job number and the override parameters for the job instance:

hydra:
  sweep:
    subdir: ${hydra.job.num}_${hydra.job.num}_${hydra.job.override_dirname}

Run output directory grouped by job name:

hydra:
  run:
    dir: outputs/${hydra.job.name}/${now:%Y-%m-%d_%H-%M-%S}

Run output directory can contain user configuration variables:

hydra:
  run:
    dir: outputs/${now:%Y-%m-%d_%H-%M-%S}/opt:${optimizer.type}

Run output directory can contain override parameters for the job

hydra:
  run:
    dir: output/${hydra.job.override_dirname}

Configuring hydra.job.override_dirname

hydra.job.override_dirname can be configured via hydra.job.config.override_dirname. You can exclude keys such as random_seed or change the separator used to construct override_dirname.

hydra:
  job:
    config:
      # configuration for the ${hydra.job.override_dirname} runtime variable
      override_dirname:
        kv_sep: '='
        item_sep: ','
        exclude_keys: []

Customizing outputs with substitution through the CLI

Outputs can also be configured through the CLI, like any other configuration.

python train.py model.nb_layers=3 hydra.run.dir=3_layers

This feature can become really powerful to write multiruns without boilerplate using substitution.

python train.py --multirun model.nb_layers=1,2,3,5 hydra.sweep.dir=multiruns/layers_effect hydra.sweep.subdir=\${model.nb_layers}

With bash, be careful to escape the $ symbol. Otherwise, bash will try to resolve the substitution, instead of passing it to Hydra.