testbed/facebookresearch__hydra/website/docs/plugins/nevergrad_sweeper.md

metadata

id: nevergrad_sweeper
title: Nevergrad Sweeper plugin
sidebar_label: Nevergrad Sweeper plugin

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.

Installation

This plugin requires Hydra 1.0 (Release candidate)

$ pip install hydra-nevergrad-sweeper --pre

Usage

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

defaults:
  - 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/dummy_training.py implements an example of how to perform minimization of a (dummy) function including a mixture of continuous and discrete parameters.

This application has the following configuration:

defaults:
  - hydra/sweeper: nevergrad

hydra:
  sweeper:
    params:

      # configuration of the optimizer
      optim:
        # name of the Nevergrad optimizer to use. Here is a sample:
        #   - "OnePlusOne" extremely simple and robust, especially at low budget, but
        #     tends to converge early.
        #   - "CMA" very good algorithm, but may require a significant budget (> 120)
        #   - "TwoPointsDE": an algorithm good in a wide range of settings, for significant
        #     budgets (> 120).
        #   - "Shiwa" an algorithm aiming at identifying the best optimizer given your input
        #     definition (work in progress, it may still be ill-suited for low budget)
        # find out more within nevergrad's documentation:
        # https://github.com/facebookresearch/nevergrad/
        optimizer: OnePlusOne
        # total number of function evaluations to perform
        budget: 100
        # number of parallel workers for performing function evaluations
        num_workers: 10
        # maximize: true  # comment out for maximization

      # default parametrization of the search space
      parametrization:
        # either one or the other
        db:
          - mnist
          - cifar
        # a log-distributed positive scalar, evolving by factors of 2 on average
        lr:
          init: 0.02
          step: 2.0
          log: true
        # a linearly-distributed scalar between 0 and 1
        dropout:
          lower: 0.0
          upper: 1.0
        # an integer scalar going from 4 to 16
        # init and step parameters could also be provided,
        # by default init is set to the middle of the range
        # and step is set to a sixth of the range
        batch_size:
          lower: 4
          upper: 16
          integer: true

db: cifar
lr: 0.01
batch_size: 8
dropout: 0.6

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
batch_size=4

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/dummy_training.py -m

You can also override the search space parametrization:

python example/dummy_training.py -m db=mnist,cifar batch_size=4,8,16 lr=log:0.001:1 dropout=0:1

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

[HYDRA] NevergradSweeper(optimizer=OnePlusOne, budget=100, num_workers=10) minimization
[HYDRA] with parametrization Dict(batch_size=TransitionChoice(choices=Tuple(4,8,16),position=Scalar[sigma=Log{exp=1.2}],transitions=[1. 1.]),db=Choice(choices=Tuple(mnist,cifar),weights=Array{(2,)}),dropout=Scalar{Cl(0,1)}[sigma=Log{exp=1.2}],lr=Log{exp=3.162277660168379,Cl(0.001,1)}):{'db': 'cifar', 'batch_size': 8, 'lr': 0.03162277660168379, 'dropout': 0.5}
[HYDRA] Sweep output dir: multirun/2020-03-04/17-53-29
[HYDRA] Launching 10 jobs locally
[HYDRA] 	#0 : db=mnist batch_size=8 lr=0.032 dropout=0.5
[__main__][INFO] - dummy_training(dropout=0.500, lr=0.032, db=mnist, batch_size=8) = 5.258
[HYDRA] 	#1 : db=mnist batch_size=16 lr=0.035 dropout=0.714
[__main__][INFO] - dummy_training(dropout=0.714, lr=0.035, db=mnist, batch_size=16) = 13.469
[HYDRA] 	#2 : db=cifar batch_size=8 lr=0.053 dropout=0.408
[__main__][INFO] - dummy_training(dropout=0.408, lr=0.053, db=cifar, batch_size=8) = 4.145
[HYDRA] 	#3 : db=cifar batch_size=8 lr=0.012 dropout=0.305
[__main__][INFO] - dummy_training(dropout=0.305, lr=0.012, db=cifar, batch_size=8) = 4.133
[HYDRA] 	#4 : db=mnist batch_size=4 lr=0.030 dropout=0.204
[__main__][INFO] - dummy_training(dropout=0.204, lr=0.030, db=mnist, batch_size=4) = 1.216

and the final 2 evaluations look like this:

[HYDRA] 	#8 : db=mnist batch_size=4 lr=0.094 dropout=0.381
[__main__][INFO] - dummy_training(dropout=0.381, 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] - dummy_training(dropout=0.381, 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

best_evaluated_params:
  batch_size: 4
  db: mnist
  dropout: 0.381
  lr: 0.094

name: nevergrad:

Defining the parameters

The plugin can use 2 types of parameters:

Choices

Choices are defined with comma-separated values in the command-line (db=mnist,cifar or batch_size=4,8,12,16) or with a list in a config file. By default, values are processed as floats if all can be converted to it, but you can modify this behavior by adding colon-separated specifications int or str before the the list. (eg.: batch_size=int:4,8,12,16)

Note: sequences of increasing scalars are treated as a special case, easier to solve. Make sure to specify it this way when possible.

Scalars

Scalars can be defined:

• through a commandline override with :-separated values defining a range (eg: dropout=0:1). You can add specifications for log distributed values (eg.: lr=log:0.001:1) or integer values (eg.: batch_size=int:4:8) or a combination of both (eg.: batch_size=log:int:4:1024)

• through 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.