A minimal framework for running hyperparameter tuning
Project description
hpt
This repository is under construction :construction:
A minimal hyperparameter tuning framework to help you train hundreds of models.
It's essentially a set of helpful wrappers over optuna.
Install
Install package from PyPI:
pip install hyperparameter-tuning
Getting started
from hpt.tuner import ObjectiveFunction, OptunaTuner
obj_func = ObjectiveFunction(
X_train, y_train, X_test, y_test,
hyperparameter_space=HYPERPARAM_SPACE_PATH, # path to YAML file
eval_metric="accuracy",
s_train=s_train,
s_val=s_test,
threshold=0.50,
)
tuner = OptunaTuner(
objective_function=obj_func,
direction="maximize", # NOTE: can pass other useful study kwargs here (e.g. storage)
)
# Then just run optimize as you would for an optuna.Study object
tuner.optimize(n_trials=20, n_jobs=4)
# Results are stored in tuner.results
tuner.results
# You can reconstruct the best predictor with:
clf = obj_func.reconstruct_model(obj_func.best_trial)
Defining a hyperparameter space
The hyperparameter space is provided either path to a YAML file, or as a dict
with the same structure.
Example hyperparameter spaces here and
here.
The YAML file must follow this structure:
# One or more top-level algorithms
DT:
# Full classpath of algorithm's constructor
classpath: sklearn.tree.DecisionTreeClassifier
# One or more key-word arguments to be passed to the constructor
kwargs:
# Kwargs may be sampled from a distribution
max_depth:
type: int # either 'int' or 'float'
range: [ 10, 100 ] # minimum and maximum values
log: True # (optionally) whether to use logarithmic scale
# Kwargs may be sampled from a fixed set of categories
criterion:
- 'gini'
- 'entropy'
# Kwargs may be a pre-defined value
min_samples_split: 4
# You may explore multiple algorithms at once
LR:
classpath: sklearn.linear_model.LogisticRegression
kwargs:
# An example of a float hyperparameter
C:
type: float
range: [ 0.01, 1.0 ]
log: True
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Close
Hashes for hyperparameter-tuning-0.3.1.tar.gz
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | 451e0b8a51e989d63913eba9bdc162224fe010401f38fdc051c5fdd343f23613 |
|
| MD5 | 1f922c9a889393ade8031bf5d65b76b8 |
|
| BLAKE2b-256 | c8c540907d79c62e8528a9bfa21ee1776f0b66d63a5e7da7287fbfca0ff92c97 |
Close
Hashes for hyperparameter_tuning-0.3.1-py3-none-any.whl
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 | df861403bbaad3263c826e369a215e76a2c8b038c85e856d8717343f639692da |
|
| MD5 | 9e64f4a4b4c46964389cf79a1438c928 |
|
| BLAKE2b-256 | ef0636f0d4a9188ec9654a1ab96a9c3c4c1618d0a2cb566f76120e5cf89f451c |
