Quick-Start¶
Installation¶
Python version 3.6 or above is necessary before installing Hypernets.
Conda¶
Install Hypernets with conda from the channel conda-forge:
conda install -c conda-forge hypernets
Pip¶
Install Hypernets with pip command:
pip install hypernets
Optional, to run Hypernets in JupyterLab notebooks, install Hypernets and JupyterLab with command:
pip install hypernets[notebook]
- Optional, to support experiment visualization base on web, install with command:
pip install hypernets[board]
Optional, to run Hypernets in distributed Dask cluster, install Hypernets with command:
pip install hypernets[dask]
Optional, to support simplified Chinese in feature generation, install jieba package before run Hypernets, or install Hypernets with command:
pip install hypernets[zhcn]
Optional, install all Hypernets components and dependencies with one command:
pip install hypernets[all]
``
Verify installation:
python -m hypernets.examples.smoke_testing
Getting started¶
In current version, we provide PlainModel (a plain HyperModel implementation), which can be used for hyper-parameter tuning with sklearn machine learning algorithms.
Basically, to search the best model only needs 4 steps:
- Step 1. Define
Search Space - Step 2. Select a
Searcher - Step 3. Select a
HyperModel - Step 4. Search and get the best model
Define Search space¶
Firstly, we define a search space for hyper-parameters of DecisionTreeClassifier and MLPClassifier:
from sklearn.neural_network import MLPClassifier
from sklearn.tree import DecisionTreeClassifier
from hypernets.core import get_random_state
from hypernets.core.ops import ModuleChoice, HyperInput, ModuleSpace
from hypernets.core.search_space import HyperSpace, Choice, Int
def my_search_space(enable_dt=True, enable_mlp=True):
space = HyperSpace()
with space.as_default():
hyper_input = HyperInput(name='input1')
estimators = []
if enable_dt:
estimators.append(dict(
cls=DecisionTreeClassifier,
criterion=Choice(["gini", "entropy"]),
splitter=Choice(["best", "random"]),
max_depth=Choice([None, 3, 5, 10, 20, 50]),
random_state=get_random_state(),
))
if enable_mlp:
estimators.append(dict(
cls=MLPClassifier,
max_iter=Int(500, 5000, step=500),
activation=Choice(['identity', 'logistic', 'tanh', 'relu']),
solver=Choice(['lbfgs', 'sgd', 'adam']),
learning_rate=Choice(['constant', 'invscaling', 'adaptive']),
random_state=get_random_state(),
))
modules = [ModuleSpace(name=f'{e["cls"].__name__}', **e) for e in estimators]
outputs = ModuleChoice(modules)(hyper_input)
space.set_inputs(hyper_input)
return space
Training with PlainModel¶
Turning and scoring with my_search_space for heart_disease_uci dataset.
def train():
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report
from hypernets.core.callbacks import SummaryCallback
from hypernets.examples.plain_model import PlainModel
from hypernets.searchers import make_searcher
from hypernets.tabular.datasets import dsutils
X = dsutils.load_heart_disease_uci()
y = X.pop('target')
X_train, X_eval, y_train, y_eval = \
train_test_split(X, y, test_size=0.3)
# make MCTS searcher
searcher = make_searcher('mcts', my_search_space, optimize_direction='max')
callbacks = [SummaryCallback()]
# create HyperModel and do 'search' action
hm = PlainModel(searcher=searcher, reward_metric='f1', callbacks=callbacks)
hm.search(X_train, y_train, X_eval, y_eval, )
# get best estimator
best = hm.get_best_trial()
estimator = hm.final_train(best.space_sample, X_train, y_train)
# scoring
y_pred = estimator.predict(X_eval)
print(classification_report(y_eval, y_pred))
if __name__ == '__main__':
train()
Run the example, we will get console output:
17:36:56 I hypernets.u.common.py 147 - 2 class detected, {0, 1}, so inferred as a [binary classification] task
17:36:56 I hypernets.c.meta_learner.py 22 - Initialize Meta Learner: dataset_id:e10ae1d61123d55062f7d3b64b79a6e1
17:36:56 I hypernets.c.callbacks.py 235 -
Trial No:1
--------------------------------------------------------------
(0) Module_ModuleChoice_1.hp_or: 0
(1) DecisionTreeClassifier.criterion: gini
(2) DecisionTreeClassifier.splitter: best
(3) DecisionTreeClassifier.max_depth: 5
--------------------------------------------------------------
...
precision recall f1-score support
0 0.68 0.67 0.67 42
1 0.72 0.73 0.73 49
accuracy 0.70 91
macro avg 0.70 0.70 0.70 91
weighted avg 0.70 0.70 0.70 91