landerlini/hopaas_client

Hyperparameter Optimization as a Service (Client)

Hopaas Python Client: hopaas_client

Hopaas is a service to handle parameter optimization as a service, based on RESTful APIs.

hopaas_client provides a Python front-end to ease the access to the service, embedding the requests in objects and calls for a most Pythonic experience.

Install

hopaas_client is not released, yet. Still, it can be installed from source as :

pip install git+https://github.com/landerlini/hopaas_client.git

Configuration

The first time you run hopaas_client it will prompt you for the server address and port to be used, and the API token associated to your own account.

Sometimes (for example for batch jobs) this can be annoying. In order to configure hopaas_client manually you can create the file .hopaasrc

vim `python -c "import hopaas_client; import os.path; print(os.path.dirname(hopaas_client.__file__))"`/.hopaasrc

The configuration file should look like:

[server]
address = <http[s]://your-server.your-domain.com>
port = <80 or 443>

[auth]
api_token = <your API token>

Minimal example

The following code snippet uses hopaas via hopaas_client to optimize a BDT.

from sklearn.ensemble import GradientBoostingClassifier as GBDT
from sklearn.datasets import make_classification

# Load the hopaas frontend
import hopaas_client as hpc

# Create the dataset and split it into classification and validation samples
X, y = make_classification(n_samples=10000)
X_train, X_test = X[::2], X[1::2]
y_train, y_test = y[::2], y[1::2]

# Create the study
study = hpc.Study('Iris GBDT v7', properties=dict(
    n_estimators=hpc.suggestions.Int(2,20),
    learning_rate=0.1, 
    max_depth=hpc.suggestions.Int(2,5)
  ),
  sampler=hpc.samplers.TPESampler(),
  pruner=hpc.pruners.MedianPruner()
)

# Iterate over a given number of trials
for iTrial in range(20):
  # Requests a trial to the hopaas server
  with study.trial() as trial:
    # Instanciate a BDT with the parameters suggested by hopaas 
    bdt = GBDT(
        n_estimators=trial.n_estimators, 
        learning_rate=trial.learning_rate,
        max_depth=trial.max_depth
        )
    
    # Train the bdt
    bdt.fit(X_train, y_train)

    # Update the trial with the score of the BDT providing a loss 
    # that hopaas will try to minimize
    trial.loss = 1.-bdt.score(X_test, y_test)

Licence

hopaas_client is made available under MIT licence. The backend will be released under GPL 3 at a more advanced stage of the development.