Hyperparameter tuning with Amazon SageMaker for molecular property prediction

Contents

  1. Background

  2. Setup

  3. Code

  4. Tune

  5. Wrap-up

Background

This example notebook demonstrates a graph-based molecular property prediction model with automatic hyperparameter tuning. The implementation is based on DGL and PyTorch. To find the best hyperparameters, it leverages SageMaker to kick off multiple training jobs with different hyperparameter combinations. In this example, you use the Amazon SageMaker Python SDK to create a hyperparameter tuning job.

Setup

This notebook was created and tested on an ml.p3.2xlarge notebook instance.

Prerequisites * Before you start this tutorial, review the pytorch-gcn-tox21.ipynb example and ensure you have an account under your Amazon Elastic Container Registry (Amazon ECR) specified by {account}.dkr.ecr.{region}.amazonaws.com/sagemaker-dgl-pytorch-gcn-tox21:latest. * An S3 bucket and prefix exists that you want to use for training and model data. This should be within the same Region as the notebook instance, training, and hosting. * An IAM role ARN exists that you are going to use to give training and hosting access to your data. See the documentation for more details on creating these. Note that if a role is not associated with the current notebook instance, or more than one role is required for training or hosting, you should replace sagemaker.get_execution_role() with the appropriate full IAM role ARN strings.

[ ]:
import sagemaker

from sagemaker import get_execution_role
from sagemaker.session import Session

# Setup session
sess = sagemaker.Session()

# S3 bucket for saving code and model artifacts.
# Feel free to specify a different bucket here if you wish.
bucket = sess.default_bucket()

# Location to put your custom code.
custom_code_upload_location = "customcode"

# IAM execution role that gives Amazon SageMaker access to resources in your AWS account.
# Use the Amazon SageMaker Python SDK to get the role from the notebook environment.
role = get_execution_role()

Code

To run Docker containers with Amazon SageMaker, provide a Python script for the container to run. In this example, main.py provides all the code you need to train an Amazon SageMaker model.

[ ]:
!cat main.py

Tune

Similar to training a single training job in Amazon SageMaker, Define your training estimator passing in the code scripts, IAM role, (per job) hardware configuration, and any hyperparameters that you are not tuning.

You must have a Docker image in your Amazon Elastic Container Registry (Amazon ECR) following steps in pytorch-gcn-tox21.ipynb.

[ ]:
# Set target dgl-docker name
docker_name = "sagemaker-dgl-pytorch-gcn-tox21"

CODE_PATH = "main.py"
code_location = sess.upload_data(CODE_PATH, bucket=bucket, key_prefix=custom_code_upload_location)

account = sess.boto_session.client("sts").get_caller_identity()["Account"]
region = sess.boto_session.region_name
image = "{}.dkr.ecr.{}.amazonaws.com/{}:latest".format(account, region, docker_name)

estimator = sagemaker.estimator.Estimator(
    image,
    role,
    train_instance_count=1,
    train_instance_type="ml.p3.2xlarge",
    hyperparameters={"entrypoint": CODE_PATH},
    sagemaker_session=sess,
)

After you define your estimator, specify the hyperparameters that you want to tune and their possible values. Depending on the type of possible values, the hyperparameters can be divided into three classes:

  • Categorical: Its possible values form a discrete set and is represented by CategoricalParameter(list).

  • Continuous: It can take any real number within an interval [min, max] and is represented by ContinuousParameter(min, max).

  • Integer: It can take any integer value within an interval [min, max] and is represented by IntegerParameter(min, max).

Note that it’s almost always better to specify a value as the least restrictive type. For example, ContinuousParameter(0.01, 0.2) is less restrictive than CategoricalParameter([0.01, 0.1, 0.15, 0.2]).

[ ]:
from sagemaker.tuner import IntegerParameter, CategoricalParameter, ContinuousParameter

hyper_ranges = {
    "lr": ContinuousParameter(1e-4, 1e-2),
    "patience": IntegerParameter(5, 30),
    "n_hidden": CategoricalParameter([32, 64, 128]),
}

Next, specify the objective metric to tune and its definition. This includes the regular expression (regex) needed to extract that metric from the Amazon CloudWatch logs of the training job.

[ ]:
objective_name = "Validation_roc_auc"
metric_definitions = [{"Name": objective_name, "Regex": "Best validation score ([0-9\\.]+)"}]

Now, create a HyperparameterTuner object, which you pass:

  • The training estimator you created above

  • The hyperparameter ranges

  • Objective metric name and definition

  • Number of training jobs to run in total and how many training jobs should be run simultaneously. More parallel jobs will finish tuning sooner, but may sacrifice accuracy. We recommend you set the parallel jobs value to less than 10 percent of the total number of training jobs. It is set higher just for this example to keep it short.

  • Whether you should maximize or minimize the objective metric. You haven’t specified here since it defaults to ‘Maximize’, which is what you want for validation roc-auc)

You can also add a task_tag with value ‘DGL’ to help tracking the hyperparameter tuning task.

[ ]:
from sagemaker.tuner import HyperparameterTuner

task_tags = [{"Key": "ML Task", "Value": "DGL"}]
tuner = HyperparameterTuner(
    estimator,
    objective_name,
    hyper_ranges,
    metric_definitions,
    tags=task_tags,
    max_jobs=6,
    max_parallel_jobs=2,
)

Finally, start the tuning job by calling .fit().

[ ]:
tuner.fit(inputs={"training-code": code_location})

Check the hyperparameter tuning jobs status to make sure it started successfully and is InProgress.

[ ]:
import boto3

boto3.client("sagemaker").describe_hyper_parameter_tuning_job(
    HyperParameterTuningJobName=tuner.latest_tuning_job.job_name
)["HyperParameterTuningJobStatus"]

Wrap-up

After the hyperparameter tuning job is started, it runs in the background and you can close this notebook. When it’s finished, you can go to console to analyze the result.

For more information about Amazon SageMaker’s Hyperparameter Tuning, see the AWS documentation.