Using PyTorch with the SageMaker Python SDK

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

Supported versions of PyTorch: 0.4.0, 1.0.0.

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

You can visit the PyTorch repository at

Train a Model with PyTorch

To train a PyTorch model by using the SageMaker Python SDK:

  1. Prepare a training script
  2. Create a sagemaker.pytorch.PyTorch Estimator
  3. Call the estimator’s fit method

Prepare a PyTorch Training Script

Your PyTorch training script must be a Python 2.7 or 3.5 compatible source file.

Prepare your script in a separate source file than the notebook, terminal session, or source file you’re using to submit the script to SageMaker via a PyTorch Estimator. This will be discussed in further detail below.

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

  • SM_MODEL_DIR: A string representing the path to the directory to write model artifacts to. These artifacts are uploaded to S3 for model hosting.
  • SM_NUM_GPUS: An integer representing the number of GPUs available to the host.
  • SM_OUTPUT_DATA_DIR: A string representing the filesystem path to write output artifacts to. Output artifacts may include checkpoints, graphs, and other files to save, not including model artifacts. These artifacts are compressed and uploaded to S3 to the same S3 prefix as the model artifacts.

Suppose that two input channels, ‘train’ and ‘test’, were used in the call to the PyTorch estimator’s fit method, the following will be set, following the format “SM_CHANNEL_[channel_name]”:

  • SM_CHANNEL_TRAIN: A string representing the path to the directory containing data in the ‘train’ channel
  • SM_CHANNEL_TEST: Same as above, but for the ‘test’ channel.

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 hosted 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

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=50)
    parser.add_argument('--batch-size', type=int, default=64)
    parser.add_argument('--learning-rate', type=float, default=0.05)
    parser.add_argument('--use-cuda', type=bool, default=False)

    # Data, model, and output directories
    parser.add_argument('--output-data-dir', type=str, default=os.environ['SM_OUTPUT_DATA_DIR'])
    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 the 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 SageMaker does not inadvertently run your training code at the wrong point in execution.

Note that 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 PyTorch Estimator.

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

Save the Model

In order to save your trained PyTorch model for deployment on SageMaker, your training script should save your model to a certain filesystem path called model_dir. This value is accessible through the environment variable SM_MODEL_DIR. The following code demonstrates how to save a trained PyTorch model named model as model.pth at the :

import argparse
import os
import torch

if __name__=='__main__':
    # default to the value in environment variable `SM_MODEL_DIR`. Using args makes the script more portable.
    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])
    args, _ = parser.parse_known_args()

    # ... train `model`, then save it to `model_dir`
    with open(os.path.join(args.model_dir, 'model.pth'), 'wb') as f:, f)

After your training job is complete, SageMaker will compress and upload the serialized model to S3, and your model data will be available in the S3 output_path you specified when you created the PyTorch Estimator.

Using third-party libraries

When running your training script on SageMaker, it will have access to some pre-installed third-party libraries including torch, torchvisopm, and numpy. For more information on the runtime environment, including specific package versions, see SageMaker PyTorch Docker 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. 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 PyTorch training scripts on SageMaker by creating PyTorch Estimators. SageMaker training of your script is invoked when you call fit on a PyTorch Estimator. The following code sample shows how you train a custom PyTorch script “”, passing in three hyperparameters (‘epochs’, ‘batch-size’, and ‘learning-rate’), and using two input channel directories (‘train’ and ‘test’).

pytorch_estimator = PyTorch('',
                            hyperparameters = {'epochs': 20, 'batch-size': 64, 'learning-rate': 0.1}){'train': 's3://my-data-bucket/path/to/my/training/data',
                       'test': 's3://my-data-bucket/path/to/my/test/data'})

Call the fit Method

You start your training script by calling fit on a PyTorch Estimator. fit takes both required and optional arguments.

fit Required Arguments

  • inputs: This can take one of the following forms: A string S3 URI, for example s3://my-bucket/my-training-data. In this case, the S3 objects rooted at the my-training-data prefix will be available in the default train channel. A dict from string channel names to S3 URIs. In this case, the objects rooted at each S3 prefix will available as files in each channel directory.

For example:


fit Optional Arguments

  • wait: Defaults to True, whether to block and wait for the training script to complete before returning.
  • logs: Defaults to True, whether to show logs produced by training job in the Python session. Only meaningful when wait is True.

Distributed PyTorch Training

You can run a multi-machine, distributed PyTorch training using the PyTorch Estimator. By default, PyTorch objects will submit single-machine training jobs to SageMaker. If you set train_instance_count to be greater than one, multi-machine training jobs will be launched when fit is called. When you run multi-machine training, SageMaker will import your training script and run it on each host in the cluster.

To initialize distributed training in your script you would call dist.init_process_group providing desired backend and rank and setting ‘WORLD_SIZE’ environment variable similar to how you would do it outside of SageMaker using environment variable initialization:

if args.distributed:
    # Initialize the distributed environment.
    world_size = len(args.hosts)
    os.environ['WORLD_SIZE'] = str(world_size)
    host_rank = args.hosts.index(args.current_host)
    dist.init_process_group(backend=args.backend, rank=host_rank)

SageMaker sets ‘MASTER_ADDR’ and ‘MASTER_PORT’ environment variables for you, but you can overwrite them.

Supported backends: - gloo and tcp for cpu instances - gloo and nccl for gpu instances

Deploy PyTorch Models

After a PyTorch Estimator has been fit, you can host the newly created model in SageMaker.

After calling fit, you can call deploy on a PyTorch Estimator to create a SageMaker Endpoint. The Endpoint runs a SageMaker-provided PyTorch model server and hosts the model produced by your training script, which was run when you called fit. This was the model you saved to model_dir.

deploy returns a Predictor object, which you can use to do inference on the Endpoint hosting your PyTorch 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 PyTorch model server by an InvokeEndpoint SageMaker operation.

predict returns the result of inference against your model. By default, the inference result a NumPy array.

# Train my estimator
pytorch_estimator = PyTorch(entry_point='',

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

# `data` is a NumPy array or a Python list.
# `response` is a NumPy array.
response = predictor.predict(data)

You use the SageMaker PyTorch model server to host your PyTorch model when you call deploy on an PyTorch Estimator. The model server runs inside a SageMaker Endpoint, which your call to deploy creates. You can access the name of the Endpoint by the name property on the returned Predictor.

The SageMaker PyTorch Model Server

The PyTorch Endpoint you create with deploy runs a SageMaker PyTorch model server. The model server loads the model that was saved by your training script and performs inference on the model in response to SageMaker InvokeEndpoint API calls.

You can configure two components of the SageMaker PyTorch model server: Model loading and model serving. Model loading is the process of deserializing your saved model back into a PyTorch model. Serving is the process of translating InvokeEndpoint requests to inference calls on the loaded model.

You configure the PyTorch model server by defining functions in the Python source file you passed to the PyTorch constructor.

Load a Model

Before a model can be served, it must be loaded. The SageMaker PyTorch model server loads your model by invoking a model_fn function that you must provide in your script. The model_fn should have the following signature:

def model_fn(model_dir)

SageMaker will inject the directory where your model files and sub-directories, saved by save, 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 model_fn implementation. It loads the model parameters from a model.pth file in the SageMaker model directory model_dir.

import torch
import os

def model_fn(model_dir):
    model = Your_Model()
    with open(os.path.join(model_dir, 'model.pth'), 'rb') as f:
    return model

Serve a PyTorch Model

After the SageMaker model server has loaded your model by calling model_fn, SageMaker will serve your model. Model serving is the process of responding to inference requests, received by SageMaker InvokeEndpoint API calls. The SageMaker PyTorch model server breaks request handling into three steps:

  • input processing,
  • prediction, and
  • output processing.

In a similar way to model loading, you configure these steps by defining functions in your Python source file.

Each step involves invoking a python function, with information about the request and the return value from the previous function in the chain. Inside the SageMaker PyTorch 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
output = output_fn(prediction, response_content_type)

The above code sample shows the three function definitions:

  • 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 SageMaker PyTorch model server provides default implementations of these functions. You can provide your own implementations for these functions in your hosting script. If you omit any definition then the SageMaker PyTorch model server will use its default implementation for that function.

The RealTimePredictor used by PyTorch in the SageMaker Python SDK serializes NumPy arrays to the NPY format by default, with Content-Type application/x-npy. The SageMaker PyTorch model server can deserialize NPY-formatted data (along with JSON and CSV data).

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

  • Prediction on models that implement the __call__ method
  • Serialization and deserialization of torch.Tensor.

The default input_fn and output_fn are meant to make it easy to predict on torch.Tensors. If your model expects a torch.Tensor and returns a torch.Tensor, then these functions do not have to be overridden when sending NPY-formatted data.

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 a SageMaker PyTorch model server, the model server receives two pieces of information:

  • The request Content-Type, for example “application/x-npy”
  • The request data body, a byte array

The SageMaker PyTorch model server will invoke an input_fn function in your hosting script, 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 a Python string

The SageMaker PyTorch model server provides a default implementation of input_fn. This function deserializes JSON, CSV, or NPY encoded data into a torch.Tensor.

Default NPY deserialization requires request_body to follow the NPY format. For PyTorch, the Python SDK defaults to sending prediction requests with this format.

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 model loaded in model_fn. The list’s shape must be identical to the model’s input shape, for all dimensions after the first (which first dimension is the batch size).

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.

The example below shows a custom input_fn for preparing pickled torch.Tensor.

import numpy as np
import torch
from six import BytesIO

def input_fn(request_body, request_content_type):
    """An input_fn that loads a pickled tensor"""
    if request_content_type == 'application/python-pickle':
        return torch.load(BytesIO(request_body))
        # Handle other content-types here or raise an Exception
        # if the content type is not supported.

Get Predictions from a PyTorch Model

After the inference request has been deserialized by input_fn, the SageMaker PyTorch model server invokes predict_fn on the return value of input_fn.

As with input_fn, you can define your own predict_fn or use the SageMaker PyTorch model server 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 invokes the loaded model’s __call__ function on input_object, and returns the resulting value. The return-type should be a torch.Tensor to be compatible with the default output_fn.

The example below shows an overridden predict_fn:

import torch
import numpy as np

def predict_fn(input_data, model):
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    with torch.no_grad():
        return model(

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 will be a torch.Tensor.
  • The second argument is the loaded model.
  • 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 a torch.Tensor.
Process Model Output

After invoking predict_fn, the model server invokes output_fn, passing in the return value from predict_fn and the content type for the response, as specified by the InvokeEndpoint request.

The output_fn has the following signature:

def output_fn(prediction, content_type)

Where prediction is the result of invoking predict_fn and the content type for the response, as specified by the InvokeEndpoint request. The function should return a byte array of data serialized to content_type.

The default implementation expects prediction to be a torch.Tensor and can serialize the result to JSON, CSV, or NPY. It accepts response content types of “application/json”, “text/csv”, and “application/x-npy”.

Working with Existing Model Data and Training Jobs

Attach to existing training jobs

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

my_training_job_name = 'MyAwesomePyTorchTrainingJob'
pytorch_estimator = PyTorch.attach(my_training_job_name)

After attaching, if the training job has finished with job status “Completed”, it can be deployed to create a SageMaker Endpoint and return a Predictor. If the training job is in progress, attach will block and display log messages from the training job, until the training job completes.

The attach method accepts the following arguments:

  • training_job_name: The name of the training job to attach to.
  • sagemaker_session: The Session used to interact with SageMaker

Deploy Endpoints from model data

In addition to attaching to existing training jobs, you can deploy models directly from model data in S3. The following code sample shows how to do this, using the PyTorchModel class.

pytorch_model = PyTorchModel(model_data='s3://bucket/model.tar.gz', role='SageMakerRole',

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

The PyTorchModel constructor takes the following arguments:

  • model_dat: An S3 location of a SageMaker model data .tar.gz file
  • image: A Docker image URI
  • role: An IAM role name or Arn for SageMaker to access AWS resources on your behalf.
  • predictor_cls: A function to call to create a predictor. If not None, deploy will return the result of invoking this function on the created endpoint name
  • env: Environment variables to run with image when hosted in SageMaker.
  • name: The model name. If None, a default model name will be selected on each deploy.
  • entry_point: Path (absolute or relative) to the Python file which should be executed as the entry point to model hosting.
  • source_dir: Optional. Path (absolute or relative) to a directory with any other training source code dependencies including tne entry point file. Structure within this directory will be preserved when training on SageMaker.
  • enable_cloudwatch_metrics: Optional. If true, training and hosting containers will generate Cloudwatch metrics under the AWS/SageMakerContainer namespace.
  • container_log_level: Log level to use within the container. Valid values are defined in the Python logging module.
  • code_location: Optional. Name of the S3 bucket where your custom code will be uploaded to. If not specified, will use the SageMaker default bucket created by sagemaker.Session.
  • sagemaker_session: The SageMaker Session object, used for SageMaker interaction

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

Assuming you have a local directory containg your model data named “my_model” you can tar and gzip compress the file and upload to 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 S3 in the bucket “my-bucket”, with the key “my-path/model.tar.gz”.

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

PyTorch Training Examples

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

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

PyTorch Estimators

The PyTorch constructor takes both required and optional arguments.

Required arguments

The following are required arguments to the PyTorch constructor. When you create a PyTorch object, you must include these in the constructor, either positionally or as keyword arguments.

  • entry_point Path (absolute or relative) to the Python file which should be executed as the entry point to training.
  • role An AWS IAM role (either name or full ARN). The Amazon SageMaker training jobs and APIs that create Amazon SageMaker endpoints use this role to access training data and model artifacts. After the endpoint is created, the inference code might use the IAM role, if accessing AWS resource.
  • train_instance_count Number of Amazon EC2 instances to use for training.
  • train_instance_type Type of EC2 instance to use for training, for example, ‘ml.m4.xlarge’.

Optional arguments

The following are optional arguments. When you create a PyTorch object, you can specify these as keyword arguments.

  • source_dir Path (absolute or relative) to a directory with any other training source code dependencies including the entry point file. Structure within this directory will be preserved when training on SageMaker.

  • dependencies (list[str]) A list of paths to directories (absolute or relative) with

    any additional libraries that will be exported to the container (default: []). The library folders will be copied to SageMaker in the same folder where the entrypoint is copied. If the `source_dir` points to S3, code will be uploaded and the S3 location will be used instead. Example:

    The following call >>> PyTorch(entry_point=’’, dependencies=[‘my/libs/common’, ‘virtual-env’]) results in the following inside the container:

    >>> $ ls
    >>> opt/ml/code
    >>>     ├──
    >>>     ├── common
    >>>     └── virtual-env
  • hyperparameters Hyperparameters that will be used for training. Will be made accessible as a dict[str, str] to the training code on SageMaker. For convenience, accepts other types besides strings, but str will be called on keys and values to convert them before training.

  • py_version Python version you want to use for executing your model training code.

  • framework_version PyTorch version you want to use for executing your model training code. You can find the list of supported versions in SageMaker PyTorch Docker Containers.

  • train_volume_size Size in GB of the EBS volume to use for storing input data during training. Must be large enough to store training data if input_mode=’File’ is used (which is the default).

  • train_max_run Timeout in seconds for training, after which Amazon SageMaker terminates the job regardless of its current status.

  • input_mode The input mode that the algorithm supports. Valid modes: ‘File’ - Amazon SageMaker copies the training dataset from the S3 location to a directory in the Docker container. ‘Pipe’ - Amazon SageMaker streams data directly from S3 to the container via a Unix named pipe.

  • output_path S3 location where you want the training result (model artifacts and optional output files) saved. If not specified, results are stored to a default bucket. If the bucket with the specific name does not exist, the estimator creates the bucket during the fit method execution.

  • output_kms_key Optional KMS key ID to optionally encrypt training output with.

  • job_name Name to assign for the training job that the fit` method launches. If not specified, the estimator generates a default job name, based on the training image name and current timestamp

  • image_name An alternative docker image to use for training and serving. If specified, the estimator will use this image for training and hosting, instead of selecting the appropriate SageMaker official image based on framework_version and py_version. Refer to: SageMaker PyTorch Docker Containers for details on what the Official images support and where to find the source code to build your custom image.

SageMaker PyTorch Docker Containers

For information about SageMaker PyTorch containers, see the SageMaker PyTorch containers repository.

For information about SageMaker PyTorch container dependencies, see SageMaker PyTorch Containers.