DP 100 Azure Data Scientist Study notes - hyperparameters

April 20, 2024

Sampling methods

Grid sampling

Every possible combination, hyperparameter must be discrete

from azure.ai.ml.sweep import Choice

command_job_for_sweep = command_job(
    batch_size=Choice(values=[16, 32, 64]),
    learning_rate=Choice(values=[0.01, 0.1, 1.0]),
)

sweep_job = command_job_for_sweep.sweep(
    sampling_algorithm = "grid",
    ...
)

Random sampling

A mix of discrete and continuous values as shown in the following code example:

from azure.ai.ml.sweep import Normal, Uniform

command_job_for_sweep = command_job(
    batch_size=Choice(values=[16, 32, 64]),   
    learning_rate=Normal(mu=10, sigma=3),
)

sweep_job = command_job_for_sweep.sweep(
    sampling_algorithm = "random",
    ...
)

Sobol

Random sampling but reproduceable with a Seed:

from azure.ai.ml.sweep import RandomSamplingAlgorithm

sweep_job = command_job_for_sweep.sweep(
    sampling_algorithm = RandomSamplingAlgorithm(seed=123, rule="sobol"),
    ...
)

Bayesian sampling

Select combination based on previous select for improved perfomance

from azure.ai.ml.sweep import Uniform, Choice

command_job_for_sweep = job(
    batch_size=Choice(values=[16, 32, 64]),    
    learning_rate=Uniform(min_value=0.05, max_value=0.1),
)

sweep_job = command_job_for_sweep.sweep(
    sampling_algorithm = "bayesian",
    ...
)

Early Termination Policy

Bandit policy

Abandons a trial if the target performance metric underperforms the best trial so far by a specified margin.

from azure.ai.ml.sweep import BanditPolicy

sweep_job.early_termination = BanditPolicy(
    slack_amount = 0.2, 
    delay_evaluation = 5, 
    evaluation_interval = 1
)

Median stopping policy

Abandons trials where the target performance metric is worse than the median of the running averages for all trials.

from azure.ai.ml.sweep import MedianStoppingPolicy

sweep_job.early_termination = MedianStoppingPolicy(
    delay_evaluation = 5, 
    evaluation_interval = 1
)

Truncation selection policy

A truncation selection policy cancels the lowest performing X% of trials at each evaluation interval based on the truncation_percentage value you specify for X.

from azure.ai.ml.sweep import TruncationSelectionPolicy

sweep_job.early_termination = TruncationSelectionPolicy(
    evaluation_interval=1, 
    truncation_percentage=20, 
    delay_evaluation=4 
)

Sweep Job

The following example script trains a logistic regression model using a --regularization argument to set the regularization rate hyperparameter, and logs the accuracy metric with the name Accuracy:

import argparse
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
import mlflow

# get regularization hyperparameter
parser = argparse.ArgumentParser()
parser.add_argument('--regularization', type=float, dest='reg_rate', default=0.01)
args = parser.parse_args()
reg = args.reg_rate

# load the training dataset
data = pd.read_csv("data.csv")

# separate features and labels, and split for training/validatiom
X = data[['feature1','feature2','feature3','feature4']].values
y = data['label'].values
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.30)

# train a logistic regression model with the reg hyperparameter
model = LogisticRegression(C=1/reg, solver="liblinear").fit(X_train, y_train)

# calculate and log accuracy
y_hat = model.predict(X_test)
acc = np.average(y_hat == y_test)
mlflow.log_metric("Accuracy", acc)

Configure and run a sweep job

To prepare the sweep job, you must first create a base command job that specifies which script to run and defines the parameters used by the script:

from azure.ai.ml import command

# configure command job as base
job = command(
    code="./src",
    command="python train.py --regularization ${{inputs.reg_rate}}",
    inputs={
        "reg_rate": 0.01,
    },
    environment="AzureML-sklearn-0.24-ubuntu18.04-py37-cpu@latest",
    compute="aml-cluster",
    )

You can then override your input parameters with your search space:

from azure.ai.ml.sweep import Choice

command_job_for_sweep = job(
    reg_rate=Choice(values=[0.01, 0.1, 1]),
)

from azure.ai.ml import MLClient

# apply the sweep parameter to obtain the sweep_job
sweep_job = command_job_for_sweep.sweep(
    compute="aml-cluster",
    sampling_algorithm="grid",
    primary_metric="Accuracy",
    goal="Maximize",
)

# set the name of the sweep job experiment
sweep_job.experiment_name="sweep-example"

# define the limits for this sweep
sweep_job.set_limits(max_total_trials=4, max_concurrent_trials=2, timeout=7200)

# submit the sweep
returned_sweep_job = ml_client.create_or_update(sweep_job)