michalwols/hp

A python library for managing Hyperparameters

hp: Hyperparameter Management Library

Defining, tuning and tracking hyperparameters in machine learning experiments can get messy. hp uses python classes to declaratively define your hyperparameters.

Benefits

1. Type annotated container for your parameters
2. Automatically generated command line interface
3. Handles saving and loading of parameters

Install

Install with pip

pip install hp

Tour

Define Your Parameters

import hp


class Params(hp.HyperParams):
  learning_rate: hp.Range(0.001, 0.1) = 0.03
  optimizer: hp.Choice(('SGD', 'Adam')) = 'SGD'

  batch_size = 32
  seed = 1

Command Line Parser

# parse from command line arguments
params = Params.from_command()

Environment Variables

params = Params.from_env(prefix='HP_')

Global Constants

params = Params.from_constants()  # load all CAP_CASE variables

YAML / JSON

hp.save(params, 'params.yaml')

params = Params.load('params.yaml')

Binding

params = hp.HyperParams()

@params.bind
def train(epochs=10):
  pass


train()  # use current param value (default to function default)

train(epochs=4)  # override params

trainer.batch_size = params.bind('batch_size')

params.bind(Optimizer, fields={'lr': 'learning_rate'})

Grid / Random Search

# grid search
for params in Params.grid():
  pass
 
# random samples without replacement
for params in Params.samples():
  pass

Change Tracking

@params.on_change
def log_changes(params, key, value):
  print(f"changing {key} from {params[key]} to {value}")
  
params.learning_rate = 0.001
# >> changing learning_rate from 0.03 to 0.001