Use MXNet with the SageMaker Python SDK

With the SageMaker Python SDK, you can train and host MXNet models on Amazon SageMaker.

For information about supported versions of MXNet, see the AWS documentation. We recommend that you use the latest supported version because that’s where we focus our development efforts.

For general information about using the SageMaker Python SDK, see Using the SageMaker Python SDK.

Train a Model with MXNet

To train an MXNet model by using the SageMaker Python SDK:

  1. Prepare a training script

  2. Create a sagemaker.mxnet.MXNet Estimator

  3. Call the estimator’s fit method

Prepare an MXNet Training Script

The training script is very similar to a training script you might run outside of Amazon SageMaker, but you can access useful properties about the training environment through various environment variables, including the following:

  • SM_MODEL_DIR: A string that represents the path where the training job writes the model artifacts to. After training, artifacts in this directory are uploaded to Amazon S3 for model hosting.

  • SM_NUM_GPUS: An integer representing the number of GPUs available to the host.

  • SM_CHANNEL_XXXX: A string that represents the path to the directory that contains the input data for the specified channel. For example, if you specify two input channels in the MXNet estimator’s fit call, named ‘train’ and ‘test’, the environment variables SM_CHANNEL_TRAIN and SM_CHANNEL_TEST are set.

  • SM_HPS: A JSON dump of the hyperparameters preserving JSON types (boolean, integer, etc.)

For the exhaustive list of available environment variables, see the SageMaker Containers documentation.

A typical training script loads data from the input channels, configures training with hyperparameters, trains a model, and saves a model to model_dir so that it can be deployed for inference later. Hyperparameters are passed to your script as arguments and can be retrieved with an argparse.ArgumentParser instance. For example, a training script might start with the following:

import argparse
import os
import json

if __name__ =='__main__':

    parser = argparse.ArgumentParser()

    # hyperparameters sent by the client are passed as command-line arguments to the script.
    parser.add_argument('--epochs', type=int, default=10)
    parser.add_argument('--batch-size', type=int, default=100)
    parser.add_argument('--learning-rate', type=float, default=0.1)

    # an alternative way to load hyperparameters via SM_HPS environment variable.
    parser.add_argument('--sm-hps', type=json.loads, default=os.environ['SM_HPS'])

    # input data and model directories
    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])
    parser.add_argument('--train', type=str, default=os.environ['SM_CHANNEL_TRAIN'])
    parser.add_argument('--test', type=str, default=os.environ['SM_CHANNEL_TEST'])

    args, _ = parser.parse_known_args()

    # ... load from args.train and args.test, train a model, write model to args.model_dir.

Because Amazon SageMaker imports your training script, you should put your training code in a main guard (if __name__=='__main__':) if you are using the same script to host your model, so that Amazon SageMaker does not inadvertently run your training code at the wrong point in execution.

Note that Amazon SageMaker doesn’t support argparse actions. If you want to use, for example, boolean hyperparameters, you need to specify type as bool in your script and provide an explicit True or False value for this hyperparameter when instantiating your MXNet estimator.

For more on training environment variables, please visit SageMaker Containers.


If you want to use MXNet 1.2 or lower, see an older version of this page.

Save a Checkpoint

It is good practice to save the best model after each training epoch, so that you can resume a training job if it gets interrupted. This is particularly important if you are using Managed Spot training.

To save MXNet model checkpoints, do the following in your training script:

  • Set the CHECKPOINTS_DIR environment variable and enable checkpoints.

    CHECKPOINTS_DIR = '/opt/ml/checkpoints'
    checkpoints_enabled = os.path.exists(CHECKPOINTS_DIR)
  • Make sure you are emitting a validation metric to test the model. For information, see Evaluation Metric API.

  • After each training epoch, test whether the current model performs the best with respect to the validation metric, and if it does, save that model to CHECKPOINTS_DIR.

    if checkpoints_enabled and current_host == hosts[0]:
           if val_acc > best_accuracy:
               best_accuracy = val_acc
     'Saving the model, params and optimizer state')
               net.export(CHECKPOINTS_DIR + "/%.4f-cifar10"%(best_accuracy), epoch)
               trainer.save_states(CHECKPOINTS_DIR + '/%.4f-cifar10-%d.states'%(best_accuracy, epoch))

For a complete example of an MXNet training script that impelements checkpointing, see

Save the Model

There is a default save method that can be imported when training on SageMaker:

from sagemaker_mxnet_training.training_utils import save

if __name__ == '__main__':
    # arg parsing and training (shown above) goes here

    save(args.model_dir, model)

The default serialization system generates three files:

  • model-shapes.json: A JSON list, containing a serialization of the Module data_shapes property. Each object in the list contains the serialization of one DataShape in the returned Module. Each object has a name property, containing the DataShape name and a shape property, which is a list of that dimensions for the shape of that DataShape. For example:

    {"name":"images", "shape":[100, 1, 28, 28]},
    {"name":"labels", "shape":[100, 1]}
  • model-symbol.json: The MXNet Module Symbol serialization, produced by invoking save on the symbol property of the Module being saved.

  • modle.params: The MXNet Module parameters, produced by invoking save_params on the Module being saved.

Use third-party libraries

When running your training script on Amazon SageMaker, it has access to some pre-installed third-party libraries, including mxnet, numpy, onnx, and keras-mxnet. For more information on the runtime environment, including specific package versions, see SageMaker MXNet Containers.

If there are other packages you want to use with your script, you can include a requirements.txt file in the same directory as your training script to install other dependencies at runtime. Both requirements.txt and your training script should be put in the same folder. You must specify this folder in source_dir argument when creating an MXNet estimator.

The function of installing packages using requirements.txt is supported for MXNet versions 1.3.0 and higher during training.

When serving an MXNet model, support for this function varies with MXNet versions. For MXNet 1.6.0 or newer, requirements.txt must be under folder code. The SageMaker MXNet Estimator automatically saves code in model.tar.gz after training (assuming you set up your script and requirements.txt correctly as stipulated in the previous paragraph). In the case of bringing your own trained model for deployment, you must save requirements.txt under folder code in model.tar.gz yourself or specify it through dependencies. For MXNet 0.12.1-1.2.1, 1.4.0-1.4.1, requirements.txt is not supported for inference. For MXNet 1.3.0, requirements.txt must be in source_dir.

A requirements.txt file is a text file that contains a list of items that are installed by using pip install. You can also specify the version of an item to install. For information about the format of a requirements.txt file, see Requirements Files in the pip documentation.

Create an Estimator

You run MXNet training scripts on Amazon SageMaker by creating an MXNet estimator. When you call fit on an MXNet estimator, Amazon SageMaker starts a training job using your script as training code. The following code sample shows how you train a custom MXNet script “”.

mxnet_estimator = MXNet('',
                        hyperparameters={'batch-size': 100,
                                         'epochs': 10,
                                         'learning-rate': 0.1})'s3://my_bucket/my_training_data/')

For more information about the sagemaker.mxnet.MXNet estimator, see SageMaker MXNet Classes.

Distributed training

If you want to use parameter servers for distributed training, set the following parameter in your MXNet constructor:

distribution={'parameter_server': {'enabled': True}}

Then, when writing a distributed training script, use an MXNet kvstore to store and share model parameters. During training, Amazon SageMaker automatically starts an MXNet kvstore server and scheduler processes on hosts in your training job cluster. Your script runs as an MXNet worker task, with one server process on each host in your cluster. One host is selected arbitrarily to run the scheduler process.

To learn more about writing distributed MXNet programs, please see Distributed Training in the MXNet docs.

Call the fit Method

Start your training script by calling fit on an MXNet Estimator. fit takes both required and optional arguments. For what arguments can be passed into fit, see the API reference.

Deploy MXNet models

Once you have a trained MXNet model, you can host it in Amazon SageMaker by creating an Amazon SageMaker Endpoint. The endpoint runs a SageMaker-provided MXNet model server and hosts the model produced by your training script. This model can be one you trained in Amazon SageMaker or a pretrained one from somewhere else.

If you use the MXNet estimator to train the model, you can call deploy to create an Amazon SageMaker Endpoint:

# Train my estimator
mxnet_estimator = MXNet('',

# Deploy my estimator to an Amazon SageMaker Endpoint and get a Predictor
predictor = mxnet_estimator.deploy(instance_type='ml.m4.xlarge',

If using a pretrained model, create an MXNetModel object, and then call deploy to create the Amazon SageMaker Endpoint:

mxnet_model = MXNetModel(model_data='s3://my_bucket/pretrained_model/model.tar.gz',
predictor = mxnet_model.deploy(instance_type='ml.m4.xlarge',

In both cases, deploy returns a Predictor object, which you can use to do inference on the endpoint hosting your MXNet model.

Each Predictor provides a predict method, which can do inference with numpy arrays or Python lists. Inference arrays or lists are serialized and sent to the MXNet model server by an InvokeEndpoint SageMaker operation. predict returns the result of inference against your model. By default, the inference result is either a Python list or dictionary.

Elastic Inference

MXNet on Amazon SageMaker has support for Elastic Inference, which allows for inference acceleration to a hosted endpoint for a fraction of the cost of using a full GPU instance. In order to attach an Elastic Inference accelerator to your endpoint provide the accelerator type to accelerator_type to your deploy call.

predictor = mxnet_estimator.deploy(instance_type='ml.m4.xlarge',

Model Directory Structure

In general, if you use the same version of MXNet for both training and inference with the SageMaker Python SDK, the SDK should take care of ensuring that the contents of your model.tar.gz file are organized correctly.

For versions 1.4 and higher

For MXNet versions 1.4 and higher, the contents of model.tar.gz should be organized as follows:

  • Model files in the top-level directory

  • Inference script (and any other source files) in a directory named code/ (for more about the inference script, see The SageMaker MXNet Model Server)

  • Optional requirements file located at code/requirements.txt (for more about requirements files, see Use third-party libraries)

For example:

|- model-symbol.json
|- model-shapes.json
|- model-0000.params
|- code/
  |- requirements.txt  # only for versions 1.6.0 and higher

In this example, model-symbol.json, model-shapes.json, and model-0000.params are the model files saved from training, is the inference script, and requirements.txt is a requirements file.

The MXNet and MXNetModel classes repack model.tar.gz to include the inference script (and related files), as long as the framework_version is set to 1.4 or higher.

For versions 1.3 and lower

For MXNet versions 1.3 and lower, model.tar.gz should contain only the model files, while your inference script and optional requirements file are packed in a separate tarball, named sourcedir.tar.gz by default.

For example:

|- model-symbol.json
|- model-shapes.json
|- model-0000.params

|- requirements.txt  # only for versions 0.12.1-1.3.0

In this example, model-symbol.json, model-shapes.json, and model-0000.params are the model files saved from training, is the inference script, and requirements.txt is a requirements file.

The SageMaker MXNet Model Server

The MXNet endpoint you create with deploy runs a SageMaker MXNet model server. The model server loads the model provided and performs inference on the model in response to SageMaker InvokeEndpoint API calls.

You can configure two components of the model server: model loading and model serving. Model loading is the process of deserializing your saved model back into an MXNet model. Serving is the process of translating InvokeEndpoint requests to inference calls on the loaded model. These are configured by defining functions in the Python source file you pass to the MXNet or MXNetModel constructor.

Load a Model

Before a model can be served, it must be loaded. The model server loads your model by invoking the model_fn function in your inference script. If you don’t provide a model_fn function, the model server uses a default model_fn function. The default function works with MXNet Module model objects saved via the default save function.

If you wrote your own save logic, then you may need to write a custom model_fn function. The model_fn function must have the following signature:

def model_fn(model_dir)

Amazon SageMaker injects the directory where your model files and sub-directories have been mounted. Your model function should return a model object that can be used for model serving.

The following code snippet shows an example custom model_fn implementation. This returns an MXNet Gluon net model for resnet-34 inference. It loads the model parameters from a model.params file in the SageMaker model directory.

def model_fn(model_dir):
    """Load the Gluon model. Called when the hosting service starts.

        model_dir (str): The directory where model files are stored.

        mxnet.gluon.nn.Block: a Gluon network (for this example)
    net = models.get_model('resnet34_v2', ctx=mx.cpu(), pretrained=False, classes=10)
    net.load_params('%s/model.params' % model_dir, ctx=mx.cpu())
    return net

MXNet on Amazon SageMaker has support for Elastic Inference, which allows for inference acceleration to a hosted endpoint for a fraction of the cost of using a full GPU instance. In order to load and serve your MXNet model through Amazon Elastic Inference, import the eimx Python package and make one change in the code to partition your model and optimize it for the EIA back end, as shown here.

Based on the example above, the following code-snippet shows an example custom model_fn implementation, which enables loading and serving our MXNet model through Amazon Elastic Inference.

def model_fn(model_dir):
    """Load the Gluon model. Called when the hosting service starts.

        model_dir (str): The directory where model files are stored.

        mxnet.gluon.nn.Block: a Gluon network (for this example)
    net = models.get_model('resnet34_v2', ctx=mx.cpu(), pretrained=False, classes=10)
    net.load_params('%s/model.params' % model_dir, ctx=mx.cpu())
    net.hybridize(backend='EIA', static_alloc=True, static_shape=True)
    return net

If you are using MXNet 1.5.1 and earlier, the default_model_fn loads and serve your model through Elastic Inference, if applicable, within the Amazon SageMaker MXNet containers.

For more information on how to enable MXNet to interact with Amazon Elastic Inference, see Use Elastic Inference with MXNet.

Serve an MXNet Model

After the MXNet model server loads your model by calling either the default model_fn or the implementation in your script, it serves your model. Model serving is the process of responding to inference requests received by SageMaker InvokeEndpoint API calls. Defining how to handle these requests can be done in one of two ways:

  • using input_fn, predict_fn, and output_fn, some of which may be your own implementations

  • writing your own transform_fn for handling input processing, prediction, and output processing

Use input_fn, predict_fn, and output_fn

The SageMaker MXNet model server breaks request handling into three steps:

  • input processing

  • prediction

  • output processing

Just like with model_fn, you configure these steps by defining functions in your Python source file.

Each step has its own Python function, which takes in information about the request and the return value from the previous function in the chain. Inside the MXNet model server, the process looks like:

# Deserialize the Invoke request body into an object we can perform prediction on
input_object = input_fn(request_body, request_content_type)

# Perform prediction on the deserialized object, with the loaded model
prediction = predict_fn(input_object, model)

# Serialize the prediction result into the desired response content type
ouput = output_fn(prediction, response_content_type)

The above code sample shows the three function definitions that correlate to the three steps mentioned above:

  • input_fn: Takes request data and deserializes the data into an object for prediction.

  • predict_fn: Takes the deserialized request object and performs inference against the loaded model.

  • output_fn: Takes the result of prediction and serializes this according to the response content type.

The MXNet model server provides default implementations of these functions. These work with both Gluon API and Module API model objects. The following content types are supported:

  • Gluon API: ‘application/json’, ‘application/x-npy’

  • Module API: ‘application/json’, ‘application/x-npy’, ‘text-csv’

You can also provide your own implementations for these functions in your training script. If you omit any definition, the MXNet model server uses its default implementation for that function.

If you rely solely on the SageMaker MXNet model server defaults, you get the following functionality:

  • Prediction on MXNet Gluon API net and Module API Module objects.

  • Deserialization from CSV and JSON to NDArrayIters.

  • Serialization of NDArrayIters to CSV or JSON.

In the following sections, we describe the default implementations of input_fn, predict_fn, and output_fn. We describe the input arguments and expected return types of each, so you can define your own implementations.

Process Model Input

When an InvokeEndpoint operation is made against an endpoint running an MXNet model server, the model server receives two pieces of information:

  • The request’s content type, e.g. ‘application/json’

  • The request data body as a byte array

The MXNet model server invokes input_fn, passing in this information. If you define an input_fn function definition, it should return an object that can be passed to predict_fn and have the following signature:

def input_fn(request_body, request_content_type)

Where request_body is a byte buffer and request_content_type is the content type of the request.

The MXNet model server provides a default implementation of input_fn. This function deserializes JSON or CSV encoded data into an MXNet NDArrayIter (external API docs) multi-dimensional array iterator. This works with the default predict_fn implementation, which expects an NDArrayIter as input.

Default JSON deserialization requires request_body contain a single JSON list. Sending multiple JSON objects within the same request_body is not supported. The list must have a dimensionality compatible with the MXNet net or Module object. Specifically, after the list is loaded, it’s either padded or split to fit the first dimension of the model input shape. The list’s shape must be identical to the model’s input shape, for all dimensions after the first.

Default CSV deserialization requires request_body contain one or more lines of CSV numerical data. The data is loaded into a two-dimensional array, where each line break defines the boundaries of the first dimension. This two-dimensional array is then re-shaped to be compatible with the shape expected by the model object. Specifically, the first dimension is kept unchanged, but the second dimension is reshaped to be consistent with the shape of all dimensions in the model, following the first dimension.

If you provide your own implementation of input_fn, you should abide by the input_fn signature. If you want to use this with the default predict_fn, then you should return an NDArrayIter. The NDArrayIter should have a shape identical to the shape of the model being predicted on. The example below shows a custom input_fn for preparing pickled numpy arrays.

import numpy as np
import mxnet as mx

def input_fn(request_body, request_content_type):
    """An input_fn that loads a pickled numpy array"""
    if request_content_type == 'application/python-pickle':
        array = np.load(StringIO(request_body))
        # Handle other content-types here or raise an Exception
        # if the content type is not supported.
Predict from a Deployed Model

After the inference request has been deserialized by input_fn, the MXNet model server invokes predict_fn. As with the other functions, you can define your own predict_fn or use the model server’s default.

The predict_fn function has the following signature:

def predict_fn(input_object, model)

Where input_object is the object returned from input_fn and model is the model loaded by model_fn.

The default implementation of predict_fn requires input_object be an NDArrayIter, which is the return-type of the default input_fn. It also requires that model be either an MXNet Gluon API net object or a Module API Module object.

The default implementation performs inference with the input NDArrayIter on the Gluon or Module object. If the model is a Gluon net it performs: net.forward(input_object). If the model is a Module object it performs module.predict(input_object). In both cases, it returns the result of that call.

If you implement your own prediction function, you should take care to ensure that:

  • The first argument is expected to be the return value from input_fn. If you use the default input_fn, this is an NDArrayIter.

  • The second argument is the loaded model. If you use the default model_fn implementation, this is an MXNet Module object. Otherwise, it is the return value of your model_fn implementation.

  • The return value should be of the correct type to be passed as the first argument to output_fn. If you use the default output_fn, this should be an NDArrayIter.

Process Model Output

After invoking predict_fn, the model server invokes output_fn, passing in the return value from predict_fn and the InvokeEndpoint requested response content type.

The output_fn has the following signature:

def output_fn(prediction, content_type)

Where prediction is the result of invoking predict_fn and content_type is the requested response content type for InvokeEndpoint. The function should return an array of bytes serialized to the expected content type.

The default implementation expects prediction to be an NDArray and can serialize the result to either JSON or CSV. It accepts response content types of “application/json” and “text/csv”.

Use transform_fn

If you would rather not structure your code around the three methods described above, you can instead define your own transform_fn to handle inference requests. An error is thrown if a transform_fn is present in conjunction with any input_fn, predict_fn, and/or output_fn. transform_fn has the following signature:

def transform_fn(model, request_body, content_type, accept_type)

Where model is the model objected loaded by model_fn, request_body is the data from the inference request, content_type is the content type of the request, and accept_type is the request content type for the response.

This one function should handle processing the input, performing a prediction, and processing the output. The return object should be one of the following:

For versions 1.4 and higher:

  • a tuple with two items: the response data and accept_type (the content type of the response data), or

  • the response data: (the content type of the response is set to either the accept header in the initial request or default to “application/json”)

For versions 1.3 and lower:

For an example inference script using this structure, see the mxnet_gluon_sentiment notebook.

Work with Existing Model Data and Training Jobs

Attach to Existing Training Jobs

You can attach an MXNet Estimator to an existing training job using the attach method.

my_training_job_name = 'MyAwesomeMXNetTrainingJob'
mxnet_estimator = MXNet.attach(my_training_job_name)

After attaching, if the training job’s status is “Complete”, it can be deployed to create an Amazon SageMaker Endpoint and return a Predictor. If the training job is in progress, attach blocks and displays log messages from the training job until the training job completes.

For information about arguments that attach accepts, see the function’s API reference.

Deploy Endpoints from Model Data

As well as attaching to existing training jobs, you can deploy models directly from model data in Amazon S3. The following code sample shows how to do this, using the MXNetModel class.

mxnet_model = MXNetModel(model_data='s3://bucket/model.tar.gz', role='SageMakerRole', entry_point='')

predictor = mxnet_model.deploy(instance_type='ml.c4.xlarge', initial_instance_count=1)

For information about arguments that the MXNetModel constructor accepts, see the class’s API reference.

Your model data must be a .tar.gz file in Amazon S3. Amazon SageMaker Training Job model data is saved to .tar.gz files in Amazon S3, however if you have local data you want to deploy, you can prepare the data yourself.

Assuming you have a local directory containing your model data named “my_model” you can tar and gzip compress the file and upload to Amazon S3 using the following commands:

tar -czf model.tar.gz my_model
aws s3 cp model.tar.gz s3://my-bucket/my-path/model.tar.gz

This uploads the contents of my_model to a gzip-compressed tar file to Amazon S3 in the bucket “my-bucket”, with the key “my-path/model.tar.gz”.

To run this command, you need the AWS CLI tool installed. Please refer to our FAQ for more information on installing this.


Amazon provides several example Jupyter notebooks that demonstrate end-to-end training on Amazon SageMaker using MXNet. Please refer to:

These are also available in Amazon SageMaker Notebook Instance hosted Jupyter notebooks under the “sample notebooks” folder.

SageMaker MXNet Classes

For information about the different MXNet-related classes in the SageMaker Python SDK, see