Version: Next

Nevergrad Sweeper plugin

PyPI PyPI - License PyPI - Python Version PyPI - Downloads Example application Plugin source

Nevergrad is a derivative-free optimization platform proposing a library of state-of-the art algorithms for hyperparameter search. This plugin provides a mechanism for Hydra applications to use Nevergrad algorithms for the optimization of experiments/applications parameters.


pip install hydra-nevergrad-sweeper --upgrade


Once installed, add hydra/sweeper=nevergrad to your command command. Alternatively, override hydra/sweeper in your config:

- hydra/sweeper: nevergrad

The default configuration is here.

Example of training using Nevergrad hyperparameter search

We include an example of how to use this plugin. The file example/ implements an example of how to perform minimization of a (dummy) function including a mixture of continuous and discrete parameters.

You can discover the Nevergrad sweeper parameters with:

$ python your_app hydra/sweeper=nevergrad --cfg hydra -p hydra.sweeper
# @package hydra.sweeper
_target_: hydra_plugins.hydra_nevergrad_sweeper.core.NevergradSweeper
optimizer: OnePlusOne
budget: 80
num_workers: 10
noisy: false
maximize: false
seed: null
parametrization: {}
version: 1

The function decorated with @hydra.main() returns a float which we want to minimize, the minimum is 0 and reached for:

db: mnist
lr: 0.12
dropout: 0.33

To run hyperparameter search and look for the best parameters for this function, clone the code and run the following command in the plugins/hydra_nevergrad_sweeper directory:

python example/ -m

You can also override the search space parametrization:

python example/ --multirun db=mnist,cifar batch_size=4,8,16 \
'lr=tag(log, interval(0.001, 1))' 'dropout=interval(0,1)'

The initialization of the sweep and the first 5 evaluations (out of 100) look like this:

[2020-10-08 20:13:53,592][HYDRA] NevergradSweeper(optimizer=OnePlusOne, budget=100, num_workers=10) minimization
[2020-10-08 20:13:53,593][HYDRA] with parametrization Dict(batch_size=Choice(choices=Tuple(4,8,16),weights=Array{(1,3)}),db=Choice(choices=Tuple(mnist,cifar),weights=Array{(1,2)}),dropout=Scalar{Cl(0,1,b)}[sigma=Log{exp=2.0}],lr=Log{exp=3.162277660168379,Cl(0.001,1,b)}):{'db': 'mnist', 'lr': 0.03162277660168379, 'dropout': 0.5, 'batch_size': 8}
[2020-10-08 20:13:53,593][HYDRA] Sweep output dir: multirun/2020-10-08/20-13-53
[2020-10-08 20:13:55,023][HYDRA] Launching 10 jobs locally
[2020-10-08 20:13:55,023][HYDRA] #0 : db=mnist lr=0.03162277660168379 dropout=0.5 batch_size=16
[2020-10-08 20:13:55,217][__main__][INFO] - dummy_training(dropout=0.500, lr=0.032, db=mnist, batch_size=16) = 13.258
[2020-10-08 20:13:55,218][HYDRA] #1 : db=cifar lr=0.018178519762066934 dropout=0.5061074452336254 batch_size=4
[2020-10-08 20:13:55,408][__main__][INFO] - dummy_training(dropout=0.506, lr=0.018, db=cifar, batch_size=4) = 0.278
[2020-10-08 20:13:55,409][HYDRA] #2 : db=cifar lr=0.10056825918734161 dropout=0.6399687427725211 batch_size=4
[2020-10-08 20:13:55,595][__main__][INFO] - dummy_training(dropout=0.640, lr=0.101, db=cifar, batch_size=4) = 0.329
[2020-10-08 20:13:55,596][HYDRA] #3 : db=mnist lr=0.06617542958182834 dropout=0.5059497416026679 batch_size=8
[2020-10-08 20:13:55,812][__main__][INFO] - dummy_training(dropout=0.506, lr=0.066, db=mnist, batch_size=8) = 5.230
[2020-10-08 20:13:55,813][HYDRA] #4 : db=mnist lr=0.16717013388679514 dropout=0.6519070394318255 batch_size=4
[2020-10-08 20:14:27,988][HYDRA] Best parameters: db=cifar lr=0.11961221693764439 dropout=0.37285878409770895 batch_size=4

and the final 2 evaluations look like this:

[HYDRA] #8 : db=mnist batch_size=4 lr=0.094 dropout=0.381
[__main__][INFO] -, lr=0.094, db=mnist, batch_size=4) = 1.077
[HYDRA] #9 : db=mnist batch_size=4 lr=0.094 dropout=0.381
[__main__][INFO] -, lr=0.094, db=mnist, batch_size=4) = 1.077
[HYDRA] Best parameters: db=mnist batch_size=4 lr=0.094 dropout=0.381

The run also creates an optimization_results.yaml file in your sweep folder with the parameters recommended by the optimizer:

best_evaluated_result: 0.381
batch_size: 4
db: mnist
dropout: 0.381
lr: 0.094
name: nevergrad

Defining the parameters

The plugin supports 2 types of parameters: Choices and Scalars. They can be defined either through config file or commandline override.

Defining through commandline override

Hydra provides a override parser that support rich syntax. More documentation can be found in (OverrideGrammer/Basic) and (OverrideGrammer/Extended). We recommend you go through them first before proceeding with this doc.


To override a field with choices:

'key=shuffle(range(1, 8))'

You can tag an override with ordered to indicate it's a TransitionChoice

`key=tag(ordered, choice(1,2,3))`


`key=interval(1,12)` # Interval are float by default
`key=int(interval(1,8))` # Scalar bounds cast to a int
`key=tag(log, interval(1,12))` # call ng.p.Log if tagged with log

Defining through config file


Choices are defined with with a list in a config file.

- mnist
- cifar


Scalars can be defined in a config files, with fields:

  • init: optional initial value
  • lower : optional lower bound
  • upper: optional upper bound
  • log: set to true for log distributed values
  • step: optional step size for looking for better parameters. In linear mode this is an additive step, in logarithmic mode it is multiplicative. 
  • integer: set to true for integers (favor floats over integers whenever possible)

Providing only lower and upper bound will set the initial value to the middle of the range, and the step to a sixth of the range. Note: unbounded scalars (scalars with no upper and/or lower bounds) can only be defined through a config file.

Last updated on by Rosario Scalise