Deep Dive into running Hyper Parameter Optimization on AWS SageMaker#

Hyper Parameter Optimization (HPO) improves model quality by searching over hyperparameters, parameters not typically learned during the training process but rather values that control the learning process itself (e.g., model size/capacity). This search can significantly boost model quality relative to default settings and non-expert tuning; however, HPO can take a very long time on a non-accelerated platform. In this notebook, we containerize a RAPIDS workflow and run Bring-Your-Own-Container SageMaker HPO to show how we can overcome the computational complexity of model search.

We accelerate HPO in two key ways:

  • by scaling within a node (e.g., multi-GPU where each GPU brings a magnitude higher core count relative to CPUs), and

  • by scaling across nodes and running parallel trials on cloud instances.

By combining these two powers HPO experiments that feel unapproachable and may take multiple days on CPU instances can complete in just hours. For example, we find a 12x speedup in wall clock time (6 hours vs 3+ days) and a 4.5x reduction in cost when comparing between GPU and CPU EC2 Spot instances on 100 XGBoost HPO trials using 10 parallel workers on 10 years of the Airline Dataset (~63M flights) hosted in a S3 bucket. For additional details refer to the end of the notebook.

With all these powerful tools at our disposal, every data scientist should feel empowered to up-level their model before serving it to the world!


To get things rolling let’s make sure we can query our AWS SageMaker execution role and session as well as our account ID and AWS region.

!docker images
%pip install --upgrade boto3
Requirement already satisfied: boto3 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (1.34.119)
Requirement already satisfied: botocore<1.35.0,>=1.34.119 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from boto3) (1.34.119)
Requirement already satisfied: jmespath<2.0.0,>=0.7.1 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from boto3) (1.0.1)
Requirement already satisfied: s3transfer<0.11.0,>=0.10.0 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from boto3) (0.10.1)
Requirement already satisfied: python-dateutil<3.0.0,>=2.1 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from botocore<1.35.0,>=1.34.119->boto3) (2.9.0)
Requirement already satisfied: urllib3!=2.2.0,<3,>=1.25.4 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from botocore<1.35.0,>=1.34.119->boto3) (1.26.18)
Requirement already satisfied: six>=1.5 in /home/ec2-user/anaconda3/envs/rapids/lib/python3.10/site-packages (from python-dateutil<3.0.0,>=2.1->botocore<1.35.0,>=1.34.119->boto3) (1.16.0)
Note: you may need to restart the kernel to use updated packages.
import os

import sagemaker
from helper_functions import (
execution_role = sagemaker.get_execution_role()
session = sagemaker.Session()

account = !(aws sts get-caller-identity --query Account --output text)
region = !(aws configure get region)
account, region
(['561241433344'], ['us-west-2'])

Key Choices#

Let’s go ahead and choose the configuration options for our HPO run.

Below are two reference configurations showing a small and a large scale HPO (sized in terms of total experiments/compute).

The default values in the notebook are set for the small HPO configuration, however you are welcome to scale them up.

small HPO: 1_year, XGBoost, 3 CV folds, singleGPU, max_jobs = 10, max_parallel_jobs = 2

large HPO: 10_year, XGBoost, 10 CV folds, multiGPU, max_jobs = 100, max_parallel_jobs = 10


We offer free hosting for several demo datasets that you can try running HPO with, or alternatively you can bring your own dataset (BYOD).

By default we leverage the Airline dataset, which is a large public tracker of US domestic flight logs which we offer in various sizes (1 year, 3 year, and 10 year) and in Parquet (compressed column storage) format. The machine learning objective with this dataset is to predict whether flights will be more than 15 minutes late arriving to their destination (dataset link, additional details in Section 1.1).

As an alternative we also offer the NYC Taxi dataset which captures yellow cab trip details in Ney York in January 2020, stored in CSV format without any compression. The machine learning objective with this dataset is to predict whether a trip had an above average tip (>$2.20).

We host the demo datasets in public S3 demo buckets in both the us-east-1 (N. Virginia) or us-west-2 (Oregon) regions (i.e., sagemaker-rapids-hpo-us-east-1, and sagemaker-rapids-hpo-us-west-2). You should run the SageMaker HPO workflow in either of these two regions if you wish to leverage the demo datasets since SageMaker requires that the S3 dataset and the compute you’ll be renting are co-located.

Lastly, if you plan to use your own dataset refer to the BYOD checklist in the Appendix to help integrate into the workflow.




# samples

storage type

time span

Airline Stats Small






Airline Stats Medium






Airline Stats Large






NYC Taxi





2020 January

Bring Your Own Dataset






# please choose dataset S3 bucket and directory
data_bucket = "sagemaker-rapids-hpo-" + region[0]
dataset_directory = "10_year"  # '1_year', '3_year', '10_year', 'NYC_taxi'

# please choose output bucket for trained model(s)
model_output_bucket = session.default_bucket()
s3_data_input = f"s3://{data_bucket}/{dataset_directory}"
s3_model_output = f"s3://{model_output_bucket}/trained-models"

best_hpo_model_local_save_directory = os.getcwd()


From a ML/algorithm perspective, we offer XGBoost, RandomForest and KMeans. You are free to switch between these algorithm choices and everything in the example will continue to work.

# please choose learning algorithm
algorithm_choice = "XGBoost"

assert algorithm_choice in ["XGBoost", "RandomForest", "KMeans"]

We can also optionally increase robustness via reshuffles of the train-test split (i.e., cross-validation folds). Typical values here are between 3 and 10 folds.

# please choose cross-validation folds
cv_folds = 10

assert cv_folds >= 1

ML Workflow Compute Choice#

We enable the option of running different code variations that unlock increasing amounts of parallelism in the compute workflow.

All of these code paths are available in the /workflows directory for your reference.

**Note that the single-CPU option will leverage multiple cores in the model training portion of the workflow; however, to unlock full parallelism in each stage of the workflow we use Dask.

# please choose code variant
ml_workflow_choice = "multiGPU"

assert ml_workflow_choice in ["singleCPU", "singleGPU", "multiCPU", "multiGPU"]

Search Ranges and Strategy#

One of the most important choices when running HPO is to choose the bounds of the hyperparameter search process. Below we’ve set the ranges of the hyperparameters to allow for interesting variation, you are of course welcome to revise these ranges based on domain knowledge especially if you plan to plug in your own dataset.

Note that we support additional algorithm specific parameters (refer to the parse_hyper_parameter_inputs function in, but for demo purposes have limited our choice to the three parameters that overlap between the XGBoost and RandomForest algorithms. For more details see the documentation for XGBoost parameters and RandomForest parameters. Since KMeans uses different parameters, we adjust accordingly.

# please choose HPO search ranges
hyperparameter_ranges = {
    "max_depth": sagemaker.parameter.IntegerParameter(5, 15),
    "n_estimators": sagemaker.parameter.IntegerParameter(100, 500),
    "max_features": sagemaker.parameter.ContinuousParameter(0.1, 1.0),
}  # see note above for adding additional parameters
if "XGBoost" in algorithm_choice:
    # number of trees parameter name difference b/w XGBoost and RandomForest
    hyperparameter_ranges["num_boost_round"] = hyperparameter_ranges.pop("n_estimators")
if "KMeans" in algorithm_choice:
    hyperparameter_ranges = {
        "n_clusters": sagemaker.parameter.IntegerParameter(2, 20),
        "max_iter": sagemaker.parameter.IntegerParameter(100, 500),

We can also choose between a Random and Bayesian search strategy for picking parameter combinations.

Random Search: Choose a random combination of values from within the ranges for each training job it launches. The choice of hyperparameters doesn’t depend on previous results so you can run the maximum number of concurrent workers without affecting the performance of the search.

Bayesian Search: Make a guess about which hyperparameter combinations are likely to get the best results. After testing the first set of hyperparameter values, hyperparameter tuning uses regression to choose the next set of hyperparameter values to test.

# please choose HPO search strategy
search_strategy = "Random"

assert search_strategy in ["Random", "Bayesian"]

Experiment Scale#

We also need to decide how may total experiments to run, and how many should run in parallel. Below we have a very conservative number of maximum jobs to run so that you don’t accidently spawn large computations when starting out, however for meaningful HPO searches this number should be much higher (e.g., in our experiments we often run 100 max_jobs). Note that you may need to request a quota limit increase for additional max_parallel_jobs parallel workers.

# please choose total number of HPO experiments[ we have set this number very low to allow for automated CI testing ]
max_jobs = 100
# please choose number of experiments that can run in parallel
max_parallel_jobs = 10

Let’s also set the max duration for an individual job to 24 hours so we don’t have run-away compute jobs taking too long.

max_duration_of_experiment_seconds = 60 * 60 * 24

Compute Platform#

Based on the dataset size and compute choice we will try to recommend an instance choice*, you are of course welcome to select alternate configurations.

e.g., For the 10_year dataset option, we suggest ml.p3.8xlarge instances (4 GPUs) and ml.m5.24xlarge CPU instances ( we will need upwards of 200GB CPU RAM during model training).

# we will recommend a compute instance type, feel free to modify
instance_type = recommend_instance_type(ml_workflow_choice, dataset_directory)
recommended instance type : ml.p3.8xlarge 
instance details          : 4x GPUs [ V100 ], 64GB GPU memory,  244GB CPU memory

In addition to choosing our instance type, we can also enable significant savings by leveraging AWS EC2 Spot Instances.

We highly recommend that you set this flag to True as it typically leads to 60-70% cost savings. Note, however that you may need to request a quota limit increase to enable Spot instances in SageMaker.

# please choose whether spot instances should be used
use_spot_instances_flag = True


s3 data input    =	s3://sagemaker-rapids-hpo-us-west-2/10_year
s3 model output  =	s3://sagemaker-us-west-2-561241433344/trained-models
compute          =	multiGPU
algorithm        =	XGBoost, 10 cv-fold
instance         =	ml.p3.8xlarge
spot instances   =	True
hpo strategy     =	Random
max_experiments  =	100
max_parallel     =	10
max runtime      =	86400 sec

1. ML Workflow


The default settings for this demo are built to utilize the Airline dataset (Carrier On-Time Performance 1987-2020, available from the Bureau of Transportation Statistics). Below are some additional details about this dataset, we plan to offer a companion notebook that does a deep dive on the data science behind this dataset. Note that if you are using an alternate dataset (e.g., NYC Taxi or BYOData) these details are not relevant.

The public dataset contains logs/features about flights in the United States (17 airlines) including:

  • Locations and distance ( Origin, Dest, Distance )

  • Airline / carrier ( Reporting_Airline )

  • Scheduled departure and arrival times ( CRSDepTime and CRSArrTime )

  • Actual departure and arrival times ( DpTime and ArrTime )

  • Difference between scheduled & actual times ( ArrDelay and DepDelay )

  • Binary encoded version of late, aka our target variable ( ArrDelay15 )

Using these features we will build a classifier model to predict whether a flight is going to be more than 15 minutes late on arrival as it prepares to depart.

Python ML Workflow#

To build a RAPIDS enabled SageMaker HPO we first need to build a SageMaker Estimator. An Estimator is a container image that captures all the software needed to run an HPO experiment. The container is augmented with entrypoint code that will be trggered at runtime by each worker. The entrypoint code enables us to write custom models and hook them up to data.

In order to work with SageMaker HPO, the entrypoint logic should parse hyperparameters (supplied by AWS SageMaker), load and split data, build and train a model, score/evaluate the trained model, and emit an output representing the final score for the given hyperparameter setting. We’ve already built multiple variations of this code.

If you would like to make changes by adding your custom model logic feel free to modify the and/or the specific workflow files in the workflows directory. You are also welcome to uncomment the cells below to load the read/review the code.

First, let’s switch our working directory to the location of the Estimator entrypoint and library code.

# %load
# %load workflows/

Build Estimator#

As we’ve already mentioned, the SageMaker Estimator represents the containerized software stack that AWS SageMaker will replicate to each worker node.

The first step to building our Estimator, is to augment a RAPIDS container with our ML Workflow code from above, and push this image to Amazon Elastic Cloud Registry so it is available to SageMaker.

Containerize and Push to ECR#

Now let’s turn to building our container so that it can integrate with the AWS SageMaker HPO API.

Our container can either be built on top of the latest RAPIDS [ nightly ] image as a starting layer or the RAPIDS stable image.

rapids_base_container = ""

Let’s also decide on the full name of our container.

image_base = "sagemaker-rapids-mnmg-100"
image_tag = rapids_base_container.split(":")[1]
ecr_fullname = (

Write Dockerfile#

We write out the Dockerfile to disk, and in a few cells execute the docker build command.

Let’s now write our selected RAPDIS image layer as the first FROM statement in the the Dockerfile.

with open("Dockerfile", "w") as dockerfile:
        f"FROM {rapids_base_container} \n\n"
        f'ENV AWS_DATASET_DIRECTORY="{dataset_directory}"\n'
        f'ENV AWS_ALGORITHM_CHOICE="{algorithm_choice}"\n'
        f'ENV AWS_ML_WORKFLOW_CHOICE="{ml_workflow_choice}"\n'
        f'ENV AWS_CV_FOLDS="{cv_folds}"\n'

Next let’s append write the remaining pieces of the Dockerfile, namely adding the sagemaker-training-toolkit, flask, dask-ml, and copying our python code.

%%writefile -a Dockerfile

# ensure printed output/log-messages retain correct order

# install a few more dependencies
RUN conda install --yes -n base \
        cupy \
        flask \
        protobuf \

# path where SageMaker looks for code when container runs in the cloud
ENV CLOUD_PATH="/opt/ml/code"

# copy our latest [local] code into the container 

Appending to Dockerfile

Lastly, let’s ensure that our Dockerfile correctly captured our base image selection.

!cat Dockerfile
FROM rapidsai/base:24.06a-cuda11.8-py3.10 


# ensure printed output/log-messages retain correct order

# install a few more dependencies
RUN conda install --yes -n base \
        cupy \
        flask \
        protobuf \

# path where SageMaker looks for code when container runs in the cloud
ENV CLOUD_PATH="/opt/ml/code"

# copy our latest [local] code into the container 


Build and Tag#

The build step will be dominated by the download of the RAPIDS image (base layer). If it’s already been downloaded the build will take less than 1 minute.

!docker pull $rapids_base_container
24.06a-cuda11.8-py3.10: Pulling from rapidsai/base

8493d397: Pulling fs layer 
7ee77381: Pulling fs layer 
37f007fd: Pulling fs layer 
774ad2ec: Pulling fs layer 
22adee62: Pulling fs layer 
68414a39: Pulling fs layer 
3710f323: Pulling fs layer 
8390d4e8: Pulling fs layer 
d0879975: Pulling fs layer 
1ab494af: Pulling fs layer 
763525ea: Pulling fs layer 
d4a79ee5: Pulling fs layer 
4ab83532: Pulling fs layer 
b700ef54: Pulling fs layer 
2adee62: Waiting fs layer 
5408789c: Pulling fs layer 
024a16c8: Pull complete  637B/637B7GBBExtracting  3.146MB/4.623MBDownloading  417.5MB/5.707GBExtracting  3.556GB/5.707GBExtracting   4.58GB/5.707GBExtracting  5.704GB/5.707GBDigest: sha256:e1995b699520fbe87a0196e3c24b6fecdd7e45797702e7dca49b4f44da1b23dd
Status: Downloaded newer image for rapidsai/base:24.06a-cuda11.8-py3.10
!docker images
REPOSITORY      TAG                      IMAGE ID       CREATED      SIZE
rapidsai/base   24.06a-cuda11.8-py3.10   a80bdce0d796   2 days ago   11.3GB
!docker build -t $ecr_fullname .
Sending build context to Docker daemon  455.2kB
Step 1/11 : FROM rapidsai/base:24.06a-cuda11.8-py3.10
 ---> a80bdce0d796
Step 2/11 : ENV AWS_DATASET_DIRECTORY="10_year"
 ---> Running in 6cedf114d041
Removing intermediate container 6cedf114d041
 ---> 6d41b5d4065d
 ---> Running in d419939bfbf0
Removing intermediate container d419939bfbf0
 ---> 8795a2713d5d
 ---> Running in 3cae981f7cab
Removing intermediate container 3cae981f7cab
 ---> 0678993d72a9
Step 5/11 : ENV AWS_CV_FOLDS="10"
 ---> Running in 150f81bcea62
Removing intermediate container 150f81bcea62
 ---> 7aaa519ee897
 ---> Running in 3a4a2d78589c
Removing intermediate container 3a4a2d78589c
 ---> 79a2ad0e4dcc
Step 7/11 : RUN conda install --yes -n base         cupy         flask         protobuf         sagemaker
 ---> Running in 0c9a24d11804
 - rapidsai-nightly
 - dask/label/dev
 - pytorch
 - conda-forge
 - nvidia
Platform: linux-64
Collecting package metadata (repodata.json): ...working... done
Solving environment: ...working... done

## Package Plan ##

  environment location: /opt/conda

  added / updated specs:
    - cupy
    - flask
    - protobuf
    - sagemaker

The following packages will be downloaded:

    package                    |            build
    blinker-1.8.2              |     pyhd8ed1ab_0          14 KB  conda-forge
    boto3-1.34.119             |     pyhd8ed1ab_0          79 KB  conda-forge
    botocore-1.34.119          |pyge310_1234567_0         6.8 MB  conda-forge
    dill-0.3.8                 |     pyhd8ed1ab_0          86 KB  conda-forge
    flask-3.0.3                |     pyhd8ed1ab_0          79 KB  conda-forge
    google-pasta-0.2.0         |     pyh8c360ce_0          42 KB  conda-forge
    itsdangerous-2.2.0         |     pyhd8ed1ab_0          19 KB  conda-forge
    jmespath-1.0.1             |     pyhd8ed1ab_0          21 KB  conda-forge
    multiprocess-0.70.16       |  py310h2372a71_0         238 KB  conda-forge
    openssl-3.3.1              |       h4ab18f5_0         2.8 MB  conda-forge
    pathos-0.3.2               |     pyhd8ed1ab_1          52 KB  conda-forge
    pox-0.3.4                  |     pyhd8ed1ab_0          26 KB  conda-forge
    ppft-               |     pyhd8ed1ab_0          33 KB  conda-forge
    protobuf-4.25.3            |  py310ha8c1f0e_0         325 KB  conda-forge
    protobuf3-to-dict-0.1.5    |  py310hff52083_8          14 KB  conda-forge
    s3transfer-0.10.1          |     pyhd8ed1ab_0          61 KB  conda-forge
    sagemaker-2.75.1           |     pyhd8ed1ab_0         377 KB  conda-forge
    smdebug-rulesconfig-1.0.1  |     pyhd3deb0d_1          20 KB  conda-forge
    werkzeug-3.0.3             |     pyhd8ed1ab_0         237 KB  conda-forge
                                           Total:        11.2 MB

The following NEW packages will be INSTALLED:

  blinker            conda-forge/noarch::blinker-1.8.2-pyhd8ed1ab_0 
  boto3              conda-forge/noarch::boto3-1.34.119-pyhd8ed1ab_0 
  botocore           conda-forge/noarch::botocore-1.34.119-pyge310_1234567_0 
  dill               conda-forge/noarch::dill-0.3.8-pyhd8ed1ab_0 
  flask              conda-forge/noarch::flask-3.0.3-pyhd8ed1ab_0 
  google-pasta       conda-forge/noarch::google-pasta-0.2.0-pyh8c360ce_0 
  itsdangerous       conda-forge/noarch::itsdangerous-2.2.0-pyhd8ed1ab_0 
  jmespath           conda-forge/noarch::jmespath-1.0.1-pyhd8ed1ab_0 
  multiprocess       conda-forge/linux-64::multiprocess-0.70.16-py310h2372a71_0 
  pathos             conda-forge/noarch::pathos-0.3.2-pyhd8ed1ab_1 
  pox                conda-forge/noarch::pox-0.3.4-pyhd8ed1ab_0 
  ppft               conda-forge/noarch::ppft- 
  protobuf           conda-forge/linux-64::protobuf-4.25.3-py310ha8c1f0e_0 
  protobuf3-to-dict  conda-forge/linux-64::protobuf3-to-dict-0.1.5-py310hff52083_8 
  s3transfer         conda-forge/noarch::s3transfer-0.10.1-pyhd8ed1ab_0 
  sagemaker          conda-forge/noarch::sagemaker-2.75.1-pyhd8ed1ab_0 
  smdebug-rulesconf~ conda-forge/noarch::smdebug-rulesconfig-1.0.1-pyhd3deb0d_1 
  werkzeug           conda-forge/noarch::werkzeug-3.0.3-pyhd8ed1ab_0 

The following packages will be UPDATED:

  openssl                                  3.3.0-h4ab18f5_3 --> 3.3.1-h4ab18f5_0 

Downloading and Extracting Packages: ...working... done
Preparing transaction: ...working... done
Verifying transaction: ...working... done
Executing transaction: ...working... done
Removing intermediate container 0c9a24d11804
 ---> 14a9c592c437
Step 8/11 : ENV CLOUD_PATH="/opt/ml/code"
 ---> Running in bce7a9615822
Removing intermediate container bce7a9615822
 ---> ae3d6adb9f11
Step 9/11 : COPY . $CLOUD_PATH
 ---> 6fdd122a0b80
 ---> Running in f259ffd20090
Removing intermediate container f259ffd20090
 ---> b7c9af873fa8
Step 11/11 : ENTRYPOINT ["./"]
 ---> Running in 22cd3b9875fb
Removing intermediate container 22cd3b9875fb
 ---> 5eb1ea40d589
Successfully built 5eb1ea40d589
Successfully tagged
CPU times: user 977 ms, sys: 240 ms, total: 1.22 s
Wall time: 1min 15s
!docker images
REPOSITORY                                                               TAG                      IMAGE ID       CREATED         SIZE   24.06a-cuda11.8-py3.10   5eb1ea40d589   2 minutes ago   12GB
rapidsai/base                                                            24.06a-cuda11.8-py3.10   a80bdce0d796   2 days ago      11.3GB

Publish to Elastic Cloud Registry (ECR)#

Now that we’ve built and tagged our container its time to push it to Amazon’s container registry (ECR). Once in ECR, AWS SageMaker will be able to leverage our image to build Estimators and run experiments.

Docker Login to ECR

docker_login_str = !(aws ecr get-login --region {region[0]} --no-include-email)
WARNING! Using --password via the CLI is insecure. Use --password-stdin.
WARNING! Your password will be stored unencrypted in /home/ec2-user/.docker/config.json.
Configure a credential helper to remove this warning. See

Login Succeeded

Create ECR repository [ if it doesn’t already exist]

repository_query = !(aws ecr describe-repositories --repository-names $image_base)
if repository_query[0] == "":
    !(aws ecr create-repository --repository-name $image_base)

Let’s now actually push the container to ECR

Note the first push to ECR may take some time (hopefully less than 10 minutes).

!docker push $ecr_fullname
The push refers to repository []

53ae5466: Preparing 
7b5ba20f: Preparing 
bdce5066: Preparing 
923ec1b3: Preparing 
3fcfb3d4: Preparing 
bf18a086: Preparing 
f3ff1008: Preparing 
b6fb91b8: Preparing 
7bf1eb99: Preparing 
264186e1: Preparing 
7d7711e0: Preparing 
ee96f292: Preparing 
e2a80b3f: Preparing 
0a873d7a: Preparing 
bcc60d01: Preparing 
1dcee623: Preparing 
9a46b795: Preparing 
f18a086: Waiting g 
9a46b795: Layer already exists 24.06a-cuda11.8-py3.10: digest: sha256:ad469ed961b5a7cee6ff7d71069ead5d2bde9906097298075d320d3d5c305d29 size: 4299

Create Estimator#

Having built our container [ +custom logic] and pushed it to ECR, we can finally compile all of efforts into an Estimator instance.

!docker images
REPOSITORY                                                               TAG                      IMAGE ID       CREATED         SIZE   24.06a-cuda11.8-py3.10   5eb1ea40d589   4 minutes ago   12GB
rapidsai/base                                                            24.06a-cuda11.8-py3.10   a80bdce0d796   2 days ago      11.3GB
# 'volume_size' - EBS volume size in GB, default = 30
estimator_params = {
    "image_uri": ecr_fullname,
    "role": execution_role,
    "instance_type": instance_type,
    "instance_count": 2,
    "input_mode": "File",
    "output_path": s3_model_output,
    "use_spot_instances": use_spot_instances_flag,
    "max_run": max_duration_of_experiment_seconds,  # 24 hours
    "sagemaker_session": session,

if use_spot_instances_flag:
    estimator_params.update({"max_wait": max_duration_of_experiment_seconds + 1})
estimator = sagemaker.estimator.Estimator(**estimator_params)

Test Estimator#

Now we are ready to test by asking SageMaker to run the BYOContainer logic inside our Estimator. This is a useful step if you’ve made changes to your custom logic and are interested in making sure everything works before launching a large HPO search.

Note: This verification step will use the default hyperparameter values declared in our custom train code, as SageMaker HPO will not be orchestrating a search for this single run.

s3 data input    =	s3://sagemaker-rapids-hpo-us-west-2/10_year
s3 model output  =	s3://sagemaker-us-west-2-561241433344/trained-models
compute          =	multiGPU
algorithm        =	XGBoost, 10 cv-fold
instance         =	ml.p3.8xlarge
spot instances   =	True
hpo strategy     =	Random
max_experiments  =	100
max_parallel     =	10
max runtime      =	86400 sec
job_name = new_job_name_from_config(
generated job name : air-mGPU-XGB-10cv-31d03d8b015bfc, job_name=job_name.lower())

Run HPO#

With a working SageMaker Estimator in hand, the hardest part is behind us. In the key choices section we already defined our search strategy and hyperparameter ranges, so all that remains is to choose a metric to evaluate performance on. For more documentation check out the AWS SageMaker Hyperparameter Tuner documentation.

Define Metric#

We only focus on a single metric, which we call ‘final-score’, that captures the accuracy of our model on the test data unseen during training. You are of course welcome to add aditional metrics, see AWS SageMaker documentation on Metrics. When defining a metric we provide a regular expression (i.e., string parsing rule) to extract the key metric from the output of each Estimator/worker.

metric_definitions = [{"Name": "final-score", "Regex": "final-score: (.*);"}]
objective_metric_name = "final-score"

Define Tuner#

Finally we put all of the elements we’ve been building up together into a HyperparameterTuner declaration.

hpo = sagemaker.tuner.HyperparameterTuner(

Run HPO#

s3 data input    =	s3://sagemaker-rapids-hpo-us-west-2/10_year
s3 model output  =	s3://sagemaker-us-west-2-561241433344/trained-models
compute          =	multiGPU
algorithm        =	XGBoost, 10 cv-fold
instance         =	ml.p3.8xlarge
spot instances   =	True
hpo strategy     =	Random
max_experiments  =	100
max_parallel     =	10
max runtime      =	86400 sec

Let’s be sure we take a moment to confirm before launching all of our HPO experiments. Depending on your configuration options running this cell can kick off a massive amount of computation!

Once this process begins, we recommend that you use the SageMaker UI to keep track of the health of the HPO process and the individual workers.

# tuning_job_name = new_job_name_from_config(dataset_directory, region, ml_workflow_choice,
#                                            algorithm_choice, cv_folds,
# #                                            instance_type)
# inputs=s3_data_input,
#          job_name=tuning_job_name,
#          wait=True,
#          logs='All')

# hpo.wait()  # block until the .fit call above is completed

Results and Summary#

Once your job is complete there are multiple ways to analyze the results. Below we display the performance of the best job, as well printing each HPO trial/job as a row of a dataframe.

tuning_job_name = "air-mGPU-XGB-10cv-527fd372fa4d8d"
hpo_results = summarize_hpo_results(tuning_job_name)
INFO:botocore.credentials:Found credentials from IAM Role: BaseNotebookInstanceEc2InstanceRole
best score: 0.9203665256500244
best params: {'max_depth': '7', 'max_features': '0.29751893065195945', 'num_boost_round': '346'}
best job-name: air-mGPU-XGB-10cv-527fd372fa4d8d-042-ed1ff13b
INFO:botocore.credentials:Found credentials from IAM Role: BaseNotebookInstanceEc2InstanceRole
max_depth max_features num_boost_round TrainingJobName TrainingJobStatus FinalObjectiveValue TrainingStartTime TrainingEndTime TrainingElapsedTimeSeconds
0 5.0 0.715196 116.0 air-mGPU-XGB-10cv-527fd372fa4d8d-100-c04c691b Completed 0.920362 2023-01-23 21:01:38+00:00 2023-01-23 21:06:21+00:00 283.0
1 12.0 0.855974 243.0 air-mGPU-XGB-10cv-527fd372fa4d8d-099-97d44628 Completed 0.920355 2023-01-23 21:17:56+00:00 2023-01-23 21:22:34+00:00 278.0
2 11.0 0.549247 395.0 air-mGPU-XGB-10cv-527fd372fa4d8d-098-e74f483f Completed 0.920356 2023-01-23 20:56:06+00:00 2023-01-23 21:00:44+00:00 278.0
3 7.0 0.882803 179.0 air-mGPU-XGB-10cv-527fd372fa4d8d-097-50755cd6 Completed 0.920356 2023-01-23 20:54:35+00:00 2023-01-23 20:59:13+00:00 278.0
4 8.0 0.416939 267.0 air-mGPU-XGB-10cv-527fd372fa4d8d-096-5c95eb2f Completed 0.920355 2023-01-23 20:51:24+00:00 2023-01-23 20:56:02+00:00 278.0
... ... ... ... ... ... ... ... ... ...
95 5.0 0.426204 330.0 air-mGPU-XGB-10cv-527fd372fa4d8d-005-7b755a81 Completed 0.920355 2023-01-23 18:48:35+00:00 2023-01-23 18:53:48+00:00 313.0
96 5.0 0.283752 256.0 air-mGPU-XGB-10cv-527fd372fa4d8d-004-e4d086fb Completed 0.920356 2023-01-23 18:48:34+00:00 2023-01-23 18:53:47+00:00 313.0
97 5.0 0.137874 377.0 air-mGPU-XGB-10cv-527fd372fa4d8d-003-89cd8506 Completed 0.920355 2023-01-23 18:48:31+00:00 2023-01-23 18:53:44+00:00 313.0
98 15.0 0.934718 365.0 air-mGPU-XGB-10cv-527fd372fa4d8d-002-caf8f6c3 Completed 0.920360 2023-01-23 18:48:25+00:00 2023-01-23 18:53:48+00:00 323.0
99 15.0 0.356588 460.0 air-mGPU-XGB-10cv-527fd372fa4d8d-001-e8a6d247 Completed 0.920356 2023-01-23 18:48:29+00:00 2023-01-23 18:53:47+00:00 318.0

100 rows × 9 columns

For a more in depth look at the HPO process we invite you to check out the HPO_Analyze_TuningJob_Results.ipynb notebook which shows how we can explore interesting things like the impact of each individual hyperparameter on the performance metric.

Getting the best Model#

Next let’s download the best trained model from our HPO runs.

local_filename, s3_path_to_best_model = download_best_model(
INFO:botocore.credentials:Found credentials from IAM Role: BaseNotebookInstanceEc2InstanceRole
Successfully downloaded best model
> filename: /home/ec2-user/SageMaker/cloud-ml-examples/aws/best_model.tar.gz
> local directory : /home/ec2-user/SageMaker/cloud-ml-examples/aws

full S3 path : s3://sagemaker-us-west-2-561241433344/trained-models/air-mGPU-XGB-10cv-527fd372fa4d8d-042-ed1ff13b/output/model.tar.gz

Model Serving#

With your best model in hand, you can now move on to serving this model on SageMaker.

In the example below we show you how to build a RealTimePredictor using the best model found during the HPO search. We will add a lightweight Flask server to our RAPIDS Estimator (a.k.a., container) which will handle the incoming requests and pass them along to the trained model for inference. If you are curious about how this works under the hood check out the Use Your Own Inference Server documentation and reference the code in

If you are interested in additional serving options (e.g., large batch with batch-transform), we plan to add a companion notebook that will provide additional details.

GPU serving#

endpoint_model = sagemaker.model.Model(
    image_uri=ecr_fullname, role=execution_role, model_data=s3_path_to_best_model

Kick off an instance for prediction [ recommend ‘ml.g4dn.2xlarge’ ]


        initial_instance_count=1, instance_type="ml.g4dn.2xlarge"
    )  #'ml.p3.2xlarge'
INFO:sagemaker:Creating model with name: rapids-sagemaker-mnmg-100-2023-01-23-22-24-22-008
INFO:sagemaker:Creating endpoint-config with name rapids-sagemaker-mnmg-100-2023-01-23-22-24-22-498
INFO:sagemaker:Creating endpoint with name rapids-sagemaker-mnmg-100-2023-01-23-22-24-22-498

Perform the prediction and return the result(s).

Below we’ve compiled examples to sanity test the trained model performance on the Airline dataset.

The first example is from a 2019 flight that departed nine minutes early,

The second example is from a 2018 flight that was more than two hours late to depart.

When we run these samples we expect to see b’[0.0, 1.0] as the printed result.

We encourage you to modify the queries below especially if you plug in your own dataset.

    predictor = sagemaker.predictor.Predictor(
        endpoint_name=str(endpoint_model.endpoint_name), sagemaker_session=session

    if dataset_directory in ["1_year", "3_year", "10_year"]:
        on_time_example = [
        ]  # 9 minutes early departure
        late_example = [
        example_payload = str(list([on_time_example, late_example]))
        example_payload = ""  # fill in a sample payload

    result = predictor.predict(example_payload)
b'[0.0, 1.0]'

Once we are finished with the serving example, we should be sure to clean up and delete the endpoint.


#     predictor.delete_endpoint()


We’ve now successfully built a RAPIDS ML workflow, containerized it (as a SageMaker Estimator), and launched a set of HPO experiments to find the best hyperparamters for our model.

If you are curious to go further, we invite you to plug in your own dataset and tweak the configuration settings to find your champion model!

HPO Experiment Details

As mentioned in the introduction we find a 12X speedup in wall clock time and a 4.5x reduction in cost when comparing between GPU and CPU instances on 100 HPO trials using 10 parallel workers on 10 years of the Airline Dataset (~63M flights). In these experiments we used the XGBoost algorithm with the multi-GPU vs multi-CPU Dask cluster and 10 cross validaiton folds. Below we offer a table with additional details.

In the case of the CPU runs, 12 jobs were stopped since they exceeded the 24 hour limit we set. CPU Job Summary Image

In the case of the GPU runs, no jobs were stopped. GPU Job Summary Image

Note that in both cases 1 job failed because a spot instance was terminated. But 1 failed job out of 100 is a minimal tradeoff for the significant cost savings.


Bring Your Own Dataset Checklist#

If you plan to use your own dataset (BYOD) here is a checklist to help you integrate into the workflow:

  • [ ] Dataset should be in either CSV or Parquet format.

  • [ ] Dataset is already pre-processed (and all feature-engineering is done).

  • [ ] Dataset is uploaded to S3 and data_bucket and dataset_directory have been set to the location of your data.

  • [ ] Dataset feature and target columns have been enumerated in /

Rapids References#

SageMaker References#

Spot Instances docs, and blog