Skip to main content

Open source library for training and deploying models on Amazon SageMaker.

Project description


SageMaker Python SDK

Build Status CodeCov

SageMaker Python SDK is an open source library for training and deploying machine learning models on Amazon SageMaker.

With the SDK, you can train and deploy models using popular deep learning frameworks Apache MXNet and TensorFlow. You can also train and deploy models with Amazon algorithms, which are scalable implementations of core machine learning algorithms that are optimized for SageMaker and GPU training. If you have your own algorithms built into SageMaker compatible Docker containers, you can train and host models using these as well.

For detailed API reference please go to: Read the Docs

Installing the SageMaker Python SDK

The SageMaker Python SDK is built to PyPI and can be installed with pip as follows:

pip install sagemaker

You can install from source by cloning this repository and running a pip install command in the root directory of the repository:

git clone
cd sagemaker-python-sdk
pip install .

Supported Operating Systems

SageMaker Python SDK supports Unix/Linux and Mac.

Supported Python versions

SageMaker Python SDK is tested on: * Python 2.7 * Python 3.5


SageMaker Python SDK is licensed under the Apache 2.0 License. It is copyright 2018, Inc. or its affiliates. All Rights Reserved. The license is available at:

Running tests

SageMaker Python SDK has unit tests and integration tests.

Unit tests

tox is a prerequisite for running unit tests so you need to make sure you have it installed. To run the unit tests:

tox tests/unit

Integrations tests

To run the integration tests, the following prerequisites must be met

  1. Access to an AWS account to run the tests on
  2. AWS account credentials available to boto3 clients used in the tests
  3. The AWS account has an IAM role named SageMakerRole
  4. The libraries listed in the extras_require object in for test are installed. You can do this by running the following command: pip install --upgrade .[test]

You can run integ tests by issuing the following command:

pytest tests/integ

You can also filter by individual test function names (usable with any of the previous commands):

pytest -k 'test_i_care_about'

Building Sphinx docs

cd into the doc directory and run:

make html

You can edit the templates for any of the pages in the docs by editing the .rst files in the “doc” directory and then running “make html” again.

SageMaker Python SDK Overview

SageMaker Python SDK provides several high-level abstractions for working with Amazon SageMaker. These are:

  • Estimators: Encapsulate training on SageMaker.
  • Models: Encapsulate built ML models.
  • Predictors: Provide real-time inference and transformation using Python data-types against a SageMaker endpoint.
  • Session: Provides a collection of methods for working with SageMaker resources.

Estimator and Model implementations for MXNet, TensorFlow, Chainer, PyTorch, and Amazon ML algorithms are included. There’s also an Estimator that runs SageMaker compatible custom Docker containers, enabling you to run your own ML algorithms by using the SageMaker Python SDK.

The following sections of this document explain how to use the different estimators and models:

Using Estimators

Here is an end to end example of how to use a SageMaker Estimator:

from sagemaker.mxnet import MXNet

# Configure an MXNet Estimator (no training happens yet)
mxnet_estimator = MXNet('',

# Starts a SageMaker training job and waits until completion.'s3://my_bucket/my_training_data/')

# Deploys the model that was generated by fit() to a SageMaker endpoint
mxnet_predictor = mxnet_estimator.deploy(initial_instance_count=1, instance_type='ml.p2.xlarge')

# Serializes data and makes a prediction request to the SageMaker endpoint
response = mxnet_predictor.predict(data)

# Tears down the SageMaker endpoint

Training Metrics

The SageMaker Python SDK allows you to specify a name and a regular expression for metrics you want to track for training. A regular expression (regex) matches what is in the training algorithm logs, like a search function. Here is an example of how to define metrics:

# Configure an BYO Estimator with metric definitions (no training happens yet)
byo_estimator = Estimator(image_name=image_name,
                          role='SageMakerRole', train_instance_count=1,
                          metric_definitions=[{'Name': 'test:msd', 'Regex': '#quality_metric: host=\S+, test msd <loss>=(\S+)'},
                                              {'Name': 'test:ssd', 'Regex': '#quality_metric: host=\S+, test ssd <loss>=(\S+)'}])

All Amazon SageMaker algorithms come with built-in support for metrics. You can go to the AWS documentation for more details about built-in metrics of each Amazon SageMaker algorithm.

Local Mode

The SageMaker Python SDK supports local mode, which allows you to create estimators and deploy them to your local environment. This is a great way to test your deep learning scripts before running them in SageMaker’s managed training or hosting environments.

We can take the example in Using Estimators , and use either local or local_gpu as the instance type.

from sagemaker.mxnet import MXNet

# Configure an MXNet Estimator (no training happens yet)
mxnet_estimator = MXNet('',

# In Local Mode, fit will pull the MXNet container Docker image and run it locally's3://my_bucket/my_training_data/')

# Alternatively, you can train using data in your local file system. This is only supported in Local mode.'file:///tmp/my_training_data')

# Deploys the model that was generated by fit() to local endpoint in a container
mxnet_predictor = mxnet_estimator.deploy(initial_instance_count=1, instance_type='local')

# Serializes data and makes a prediction request to the local endpoint
response = mxnet_predictor.predict(data)

# Tears down the endpoint container

If you have an existing model and want to deploy it locally, don’t specify a sagemaker_session argument to the MXNetModel constructor. The correct session is generated when you call model.deploy().

Here is an end-to-end example:

import numpy
from sagemaker.mxnet import MXNetModel

model_location = 's3://mybucket/my_model.tar.gz'
code_location = 's3://mybucket/sourcedir.tar.gz'
s3_model = MXNetModel(model_data=model_location, role='SageMakerRole',
                      entry_point='', source_dir=code_location)

predictor = s3_model.deploy(initial_instance_count=1, instance_type='local')
data = numpy.zeros(shape=(1, 1, 28, 28))

# Tear down the endpoint container

If you don’t want to deploy your model locally, you can also choose to perform a Local Batch Transform Job. This is useful if you want to test your container before creating a Sagemaker Batch Transform Job. Note that the performance will not match Batch Transform Jobs hosted on SageMaker but it is still a useful tool to ensure you have everything right or if you are not dealing with huge amounts of data.

Here is an end-to-end example:

from sagemaker.mxnet import MXNet

mxnet_estimator = MXNet('',
transformer = mxnet_estimator.transformer(1, 'local', assemble_with='Line', max_payload=1)
transformer.transform('s3://my/transform/data, content_type='text/csv', split_type='Line')

For detailed examples of running Docker in local mode, see:

A few important notes:

  • Only one local mode endpoint can be running at a time.
  • If you are using S3 data as input, it is pulled from S3 to your local environment. Ensure you have sufficient space to store the data locally.
  • If you run into problems it often due to different Docker containers conflicting. Killing these containers and re-running often solves your problems.
  • Local Mode requires Docker Compose and nvidia-docker2 for local_gpu.
  • Distributed training is not yet supported for local_gpu.

Incremental Training

Incremental training allows you to bring a pre-trained model into a SageMaker training job and use it as a starting point for a new model. There are several situations where you might want to do this:

  • You want to perform additional training on a model to improve its fit on your data set.
  • You want to import a pre-trained model and fit it to your data.
  • You want to resume a training job that you previously stopped.

To use incremental training with SageMaker algorithms, you need model artifacts compressed into a tar.gz file. These artifacts are passed to a training job via an input channel configured with the pre-defined settings Amazon SageMaker algorithms require.

To use model files with a SageMaker estimator, you can use the following parameters:

  • model_uri: points to the location of a model tarball, either in S3 or locally. Specifying a local path only works in local mode.
  • model_channel_name: name of the channel SageMaker will use to download the tarball specified in model_uri. Defaults to ‘model’.

This is converted into an input channel with the specifications mentioned above once you call fit() on the predictor. In bring-your-own cases, model_channel_name can be overriden if you require to change the name of the channel while using the same settings.

If your bring-your-own case requires different settings, you can create your own s3_input object with the settings you require.

Here’s an example of how to use incremental training:

# Configure an estimator
estimator = sagemaker.estimator.Estimator(training_image,

# Start a SageMaker training job and waits until completion.'s3://my_bucket/my_training_data/')

# Create a new estimator using the previous' model artifacts
incr_estimator = sagemaker.estimator.Estimator(training_image,

# Start a SageMaker training job using the original model for incremental training's3://my_bucket/my_training_data/')

Currently, the following algorithms support incremental training:

  • Image Classification
  • Object Detection
  • Semantics Segmentation

MXNet SageMaker Estimators

By using MXNet SageMaker Estimators, you can train and host MXNet models on Amazon SageMaker.

Supported versions of MXNet: 1.2.1, 1.1.0, 1.0.0, 0.12.1.

We recommend that you use the latest supported version, because that’s where we focus most of our development efforts.

For more information, see MXNet SageMaker Estimators and Models.

TensorFlow SageMaker Estimators

By using TensorFlow SageMaker Estimators, you can train and host TensorFlow models on Amazon SageMaker.

Supported versions of TensorFlow: 1.4.1, 1.5.0, 1.6.0, 1.7.0, 1.8.0, 1.9.0, 1.10.0, 1.11.0.

We recommend that you use the latest supported version, because that’s where we focus most of our development efforts.

For more information, see TensorFlow SageMaker Estimators and Models.

Chainer SageMaker Estimators

By using Chainer SageMaker Estimators, you can train and host Chainer models on Amazon SageMaker.

Supported versions of Chainer: 4.0.0, 4.1.0, 5.0.0.

We recommend that you use the latest supported version, because that’s where we focus most of our development efforts.

For more information about Chainer, see

For more information about Chainer SageMaker Estimators, see Chainer SageMaker Estimators and Models.

PyTorch SageMaker Estimators

With PyTorch SageMaker Estimators, you can train and host PyTorch models on Amazon SageMaker.

Supported versions of PyTorch: 0.4.0, (“Preview”).

We recommend that you use the latest supported version, because that’s where we focus most of our development efforts.

You can try the “Preview” version of PyTorch by specifying '' for framework_version when creating your PyTorch estimator. This will ensure you’re using the latest version of torch-nightly.

For more information about PyTorch, see

For more information about PyTorch SageMaker Estimators, see PyTorch SageMaker Estimators and Models.

AWS SageMaker Estimators

Amazon SageMaker provides several built-in machine learning algorithms that you can use to solve a variety of problems.

The full list of algorithms is available at:

The SageMaker Python SDK includes estimator wrappers for the AWS K-means, Principal Components Analysis (PCA), Linear Learner, Factorization Machines, Latent Dirichlet Allocation (LDA), Neural Topic Model (NTM), Random Cut Forest, k-nearest neighbors (k-NN), Object2Vec, and IP Insights algorithms.

For more information, see AWS SageMaker Estimators and Models.

BYO Docker Containers with SageMaker Estimators

To use a Docker image that you created and use the SageMaker SDK for training, the easiest way is to use the dedicated Estimator class. You can create an instance of the Estimator class with desired Docker image and use it as described in previous sections.

Please refer to the full example in the examples repo:

git clone

The example notebook is is located here: advanced_functionality/scikit_bring_your_own/scikit_bring_your_own.ipynb

SageMaker Automatic Model Tuning

All of the estimators can be used with SageMaker Automatic Model Tuning, which performs hyperparameter tuning jobs. A hyperparameter tuning job finds the best version of a model by running many training jobs on your dataset using the algorithm with different values of hyperparameters within ranges that you specify. It then chooses the hyperparameter values that result in a model that performs the best, as measured by a metric that you choose. If you’re not using an Amazon SageMaker built-in algorithm, then the metric is defined by a regular expression (regex) you provide. The hyperparameter tuning job parses the training job’s logs to find metrics that match the regex you defined. For more information about SageMaker Automatic Model Tuning, see AWS documentation.

The SageMaker Python SDK contains a HyperparameterTuner class for creating and interacting with hyperparameter training jobs. Here is a basic example of how to use it:

from sagemaker.tuner import HyperparameterTuner, ContinuousParameter

# Configure HyperparameterTuner
my_tuner = HyperparameterTuner(estimator=my_estimator,  # previously-configured Estimator object
                               hyperparameter_ranges={'learning-rate': ContinuousParameter(0.05, 0.06)},
                               metric_definitions=[{'Name': 'validation-accuracy', 'Regex': 'validation-accuracy=(\d\.\d+)'}],

# Start hyperparameter tuning job{'train': 's3://my_bucket/my_training_data', 'test': 's3://my_bucket_my_testing_data'})

# Deploy best model
my_predictor = my_tuner.deploy(initial_instance_count=1, instance_type='ml.m4.xlarge')

# Make a prediction against the SageMaker endpoint
response = my_predictor.predict(my_prediction_data)

# Tear down the SageMaker endpoint

This example shows a hyperparameter tuning job that creates up to 100 training jobs, running up to 10 training jobs at a time. Each training job’s learning rate is a value between 0.05 and 0.06, but this value will differ between training jobs. You can read more about how these values are chosen in the AWS documentation.

A hyperparameter range can be one of three types: continuous, integer, or categorical. The SageMaker Python SDK provides corresponding classes for defining these different types. You can define up to 20 hyperparameters to search over, but each value of a categorical hyperparameter range counts against that limit.

If you are using an Amazon SageMaker built-in algorithm, you don’t need to pass in anything for metric_definitions. In addition, the fit() call uses a list of RecordSet objects instead of a dictionary:

# Create RecordSet object for each data channel
train_records = RecordSet(...)
test_records = RecordSet(...)

# Start hyperparameter tuning job[train_records, test_records])

To help attach a previously-started hyperparameter tuning job to a HyperparameterTuner instance, fit() adds the module path of the class used to create the tuner to the list of static hyperparameters by default. If the algorithm you are using cannot handle unknown hyperparameters (for example, an Amazon SageMaker built-in algorithm that does not have a custom estimator in the Python SDK), set include_cls_metadata to False when you call fit, so that it does not add the module path as a static hyperparameter:{'train': 's3://my_bucket/my_training_data', 'test': 's3://my_bucket_my_testing_data'},

There is also an analytics object associated with each HyperparameterTuner instance that contains useful information about the hyperparameter tuning job. For example, the dataframe method gets a pandas dataframe summarizing the associated training jobs:

# Retrieve analytics object
my_tuner_analytics =

# Look at summary of associated training jobs
my_dataframe = my_tuner_analytics.dataframe()

For more detailed examples of running hyperparameter tuning jobs, see:

For more detailed explanations of the classes that this library provides for automatic model tuning, see:

SageMaker Batch Transform

After you train a model, you can use Amazon SageMaker Batch Transform to perform inferences with the model. Batch Transform manages all necessary compute resources, including launching instances to deploy endpoints and deleting them afterward. You can read more about SageMaker Batch Transform in the AWS documentation.

If you trained the model using a SageMaker Python SDK estimator, you can invoke the estimator’s transformer() method to create a transform job for a model based on the training job:

transformer = estimator.transformer(instance_count=1, instance_type='ml.m4.xlarge')

Alternatively, if you already have a SageMaker model, you can create an instance of the Transformer class by calling its constructor:

transformer = Transformer(model_name='my-previously-trained-model',

For a full list of the possible options to configure by using either of these methods, see the API docs for Estimator or Transformer.

After you create a Transformer object, you can invoke transform() to start a batch transform job with the S3 location of your data. You can also specify other attributes of your data, such as the content type.


For more details about what can be specified here, see API docs.

Secure Training and Inference with VPC

Amazon SageMaker allows you to control network traffic to and from model container instances using Amazon Virtual Private Cloud (VPC). You can configure SageMaker to use your own private VPC in order to further protect and monitor traffic.

For more information about Amazon SageMaker VPC features, and guidelines for configuring your VPC, see the following documentation:

You can also reference or reuse the example VPC created for integration tests: tests/integ/

To train a model using your own VPC, set the optional parameters subnets and security_group_ids on an Estimator:

from sagemaker.mxnet import MXNet

# Configure an MXNet Estimator with subnets and security groups from your VPC
mxnet_vpc_estimator = MXNet('',
                            subnets=['subnet-1', 'subnet-2'],

# SageMaker Training Job will set VpcConfig and container instances will run in your VPC's3://my_bucket/my_training_data/')

When you create a Predictor from the Estimator using deploy(), the same VPC configurations will be set on the SageMaker Model:

# Creates a SageMaker Model and Endpoint using the same VpcConfig
# Endpoint container instances will run in your VPC
mxnet_vpc_predictor = mxnet_vpc_estimator.deploy(initial_instance_count=1,

# You can also set ``vpc_config_override`` to use a different VpcConfig
other_vpc_config = {'Subnets': ['subnet-3', 'subnet-4'],
                    'SecurityGroupIds': ['sg-2']}
mxnet_predictor_other_vpc = mxnet_vpc_estimator.deploy(initial_instance_count=1,

# Setting ``vpc_config_override=None`` will disable VpcConfig
mxnet_predictor_no_vpc = mxnet_vpc_estimator.deploy(initial_instance_count=1,

Likewise, when you create Transformer from the Estimator using transformer(), the same VPC configurations will be set on the SageMaker Model:

# Creates a SageMaker Model using the same VpcConfig
mxnet_vpc_transformer = mxnet_vpc_estimator.transformer(instance_count=1,

# Transform Job container instances will run in your VPC


I want to train a SageMaker Estimator with local data, how do I do this?

Upload the data to S3 before training. You can use the AWS Command Line Tool (the aws cli) to achieve this.

If you don’t have the aws cli, you can install it using pip:

pip install awscli --upgrade --user

If you don’t have pip or want to learn more about installing the aws cli, see the official Amazon aws cli installation guide.

After you install the AWS cli, you can upload a directory of files to S3 with the following command:

aws s3 cp /tmp/foo/ s3://bucket/path

For more information about using the aws cli for manipulating S3 resources, see AWS cli command reference.

How do I make predictions against an existing endpoint?

Create a Predictor object and provide it with your endpoint name, then call its predict() method with your input.

You can use either the generic RealTimePredictor class, which by default does not perform any serialization/deserialization transformations on your input, but can be configured to do so through constructor arguments:

Or you can use the TensorFlow / MXNet specific predictor classes, which have default serialization/deserialization logic:

Example code using the TensorFlow predictor:

from sagemaker.tensorflow import TensorFlowPredictor

predictor = TensorFlowPredictor('myexistingendpoint')
result = predictor.predict(['my request body'])

BYO Model

You can also create an endpoint from an existing model rather than training one. That is, you can bring your own model:

First, package the files for the trained model into a .tar.gz file, and upload the archive to S3.

Next, create a Model object that corresponds to the framework that you are using: MXNetModel or TensorFlowModel.

Example code using MXNetModel:

from sagemaker.mxnet.model import MXNetModel

sagemaker_model = MXNetModel(model_data='s3://path/to/model.tar.gz',

After that, invoke the deploy() method on the Model:

predictor = sagemaker_model.deploy(initial_instance_count=1,

This returns a predictor the same way an Estimator does when deploy() is called. You can now get inferences just like with any other model deployed on Amazon SageMaker.

A full example is available in the Amazon SageMaker examples repository.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filename, size & hash SHA256 hash help File type Python version Upload date
sagemaker-1.15.1.tar.gz (212.7 kB) Copy SHA256 hash SHA256 Source None

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN SignalFx SignalFx Supporter DigiCert DigiCert EV certificate StatusPage StatusPage Status page