Hydra Experiment Sweeper

hydra is a great library to configure model trainings and numerical experiments. Via its --multirun facility, you can launch multiple jobs at a time and evaluate multiple combinations of settings such as learning rates and seeds easily. However, the built-in sweepers only accept sweep ranges on the command line, which means that reproducing a set of experiments for a paper or similar relies on you to remember or write down the exact invocation of hydra that produced the results in the first place.

hydra-experiment-sweeper lets you configure experiments via YAML. This way colleagues and yourself a year from now can re-run the same jobs with exactly the same parameters.

defaults:
  - override hydra/sweeper: experiment

hydra:
  sweeper:
    overrides:
      learning_rate: 1e-5,1e-6,1e-7
      # Any other overrides such as seeds, model, dataset, etc.

Installation

pip install hydra-experiment-sweeper

How to Use

The sweeper accepts any override syntax that hydra accepts on the command line as well as YAML lists for long values such as seeds or paths.

defaults:
  - override hydra/sweeper: experiment

hydra:
  sweeper:
    max_batch_size: 2
    overrides:
      model: linreg,svm
      learning_rate: range(0.1, 0.3, 0.1)
      seed:
        - 123
        - 42

Running this experiment with python example/train.py --config-name experiment -m launches the configured jobs.

$ python example/train.py -m
[2021-10-15 17:57:18,241][HYDRA] ExperimentSweeper(max_batch_size=2, overrides={'model': 'linreg,svm', 'learning_rate': 'range(0.1, 0.3, 0.1)', 'seed': [123, 42]}) sweeping
[2021-10-15 17:57:18,242][HYDRA] Sweep output dir : multirun/2021-10-15/17-57-18
[2021-10-15 17:57:18,796][HYDRA] Launching 2 jobs locally
[2021-10-15 17:57:18,796][HYDRA] 	#0 : model=linreg learning_rate=0.1 seed=123
model:
  model_type: linear_regression
seed: 123
learning_rate: 0.1

[2021-10-15 17:57:18,888][HYDRA] 	#1 : model=linreg learning_rate=0.1 seed=42
model:
  model_type: linear_regression
seed: 42
learning_rate: 0.1

[2021-10-15 17:57:18,982][HYDRA] Launching 2 jobs locally
[2021-10-15 17:57:18,982][HYDRA] 	#2 : model=linreg learning_rate=0.2 seed=123
model:
  model_type: linear_regression
seed: 123
learning_rate: 0.2

[2021-10-15 17:57:19,079][HYDRA] 	#3 : model=linreg learning_rate=0.2 seed=42
model:
  model_type: linear_regression
seed: 42
learning_rate: 0.2

[2021-10-15 17:57:19,171][HYDRA] Launching 2 jobs locally
[2021-10-15 17:57:19,171][HYDRA] 	#4 : model=svm learning_rate=0.1 seed=123
model:
  model_type: support_vector_machine
seed: 123
learning_rate: 0.1

[2021-10-15 17:57:19,290][HYDRA] 	#5 : model=svm learning_rate=0.1 seed=42
model:
  model_type: support_vector_machine
seed: 42
learning_rate: 0.1

[2021-10-15 17:57:19,405][HYDRA] Launching 2 jobs locally
[2021-10-15 17:57:19,405][HYDRA] 	#6 : model=svm learning_rate=0.2 seed=123
model:
  model_type: support_vector_machine
seed: 123
learning_rate: 0.2

[2021-10-15 17:57:19,497][HYDRA] 	#7 : model=svm learning_rate=0.2 seed=42
model:
  model_type: support_vector_machine
seed: 42
learning_rate: 0.2

You can still override parameters on the command line. Let's say you want to run the same experiment but instead of sweeping over the configured learning rates, you want to try a different order of magnitude. python example/train.py --config-name experiment -m learning_rate=1e-6,5e-5,1e-5 works just as you would expect with the built-in sweeper and overwrites the configured override.