AWS DeepLens
Developer Guide

Use Amazon SageMaker to Provision a Pre-trained Model for a Sample Project

In this example, you start with a SqueezeNet object detection model and use Amazon SageMaker to train it to perform binary classification to determine whether an object is a hot dog. The example shows you how to edit a model to perform binary classification, and explains learning rate and epochs. We have provided a Jupyter notebook instance, which is open source software for interactive computing. It includes the editing code to execute and explanations for the entire process.

After training the model, you import its artifacts into AWS DeepLens, and create a project. You then watch as your AWS DeepLens detects and identifies hot dogs.

Step 1: Create an Amazon S3 Bucket

Before you begin, be sure that you have created an AWS account, and the required IAM users and roles.

  1. Sign in to the AWS Management Console and open the Amazon S3 console at

  2. Choose Create bucket.

  3. On the Name and region tab:

    1. Name the bucket deeplens-sagemaker-<some_string>, where <some_string> can be any string to make the bucket name unique across AWS. An example would be your account name or ID.


      The bucket name must begin with deeplens-sagemaker- to work with the default access policies. Otherwise, the services will not be able to access it.

    2. Verify that you are in the US East (N. Virginia) region.

    3. Choose Next.

  4. On the Configure options tab, choose Next.

  5. On the Set permissions tab, choose Grant Amazon S3 Log Delivery group write access to this bucket from the Manage system permissions drop-down menu, then choose Next.

  6. On the Review tab, review your settings then choose Create bucket.

  7. On the created bucket's page, choose the Overview tab and then choose Create folder.

  8. Name the folder test then choose Save.

    To use the folder with the example notebook instance to be created next, you must name the folder as test, unless you change the default folder name in the notebook.

Step 2: Create an Amazon SageMaker Notebook Instance

Create an Amazon SageMaker notebook instance.

  1. Open the Amazon SageMaker console at

    Make sure that you have chosen the us-east-1 — US East (N. Virginia) Region.

  2. Choose Notebook instances from the navigation pane and then choose Create notebook instance.

  3. On the Create notebook instance page, then do the following:

    1. Under Notebook instance settings, type a name for the notebook in the Notebook instance name input field; for example, <your-name>-hotdog.

    2. For Notebook instance type, choose ml.t2.medium.

    3. For Permissions and encryption, under IAM role, choose Create a new role, if this is the first time you run a notebook, type deeplens-sagemaker in the Specific S3 buckets input field, and then choose Create role.

      After you've created the first notebook instance for AWS DeepLens, you can choose an available IAM role from the Use existing role list.

      If you've already created the AWSDeepLensSageMaker role as part of the setup, choose Enter a custom IAM role ARN, paste the Amazon Resource Name (ARN) of your Amazon SageMaker role in the Custom IAM role ARN box. You can find the ARN of your Amazon SageMaker role as follows:

      1. Open the IAM console at

      2. In the navigation pane, choose Roles.

      3. Find the AWSDeepLensSagemakerRole and choose its name. This takes you to the role's Summary page.

      4. On the Summary page, locate and copy the Role ARN. The ARN will look something like this:

        arn:aws:iam::account id:role/AWSDeepLensSagemakerRole
    4. All other settings are optional. Skip them for this exercise.


      If you want to access resources in your VPC from the notebook instance, choose a VPC and a SubnetId. For Security Group, choose the default security group of the VPC. The inbound and outbound rules of the default security group are sufficient for the exercises in this guide.

    5. Choose Create notebookinstance.

      Your new notebook instance is now available on the Notebooks page.

Step 3: Edit the Model in Amazon SageMaker

In this step, you open the <your-name>-hotdog notebook and edit the object detection model so it recognizes a hot dog. The notebook contains explanations to help you through each step.

  1. Open the Amazon SageMaker console at

  2. Make sure the US East (N. Virginia) Region is chosen.

  3. In the navigation pane, choose Notebook instances.

  4. On the Notebooks page, choose the radio button to the left of the notebook instance that you just created (<your-name>-hotdog). When the notebook's status is InService, choose Open Jupiter.

  5. Open a new tab in your browser and navigate to

  6. Download the .zip file or clone the Git repository with the following command. If you downloaded the .zip file, locate it and extract all.

    git clone

    If you downloaded the .zip file, locate it and extract all.

You now upload the training file and use it to edit the model.

  1. On the Jupyter tab, choose Upload.

  2. Navigate to the extracted file deeplens-hotdog-or-not-hotdog.ipynb then choose Open, then choose Upload.

  3. Locate and choose the deeplens-hotdog-or-not-hotdog notebook.

  4. In the upper right corner of the Jupyter screen, verify that the kernal is conda_mxnet_p36. If it isn't, change the kernal.

  5. In the deeplens-hotdog-or-not-hotdog.ipynb file, search for bucket= 'your S3 bucket name here'. Replace 'your s3 bucket name here' with the name of your S3 bucket, for example deeplens-sagemaker-<your-name>.

    Return to the top of the file.

    For each step (In [#]:) in the file:

    1. Read the step's description.

    2. If the block has code in it, place your cursor in the code block and run the code block. To run a code block in Jupyter, use Ctrl+<Enter> (macOS Cnd_<Enter>) or choose the run icon ().


      Each step is numbered in a fashion such as In [1]:. While the block is executing, that changes to In [*]:. When the block finishes executing it returns to In [1]:. Do not move on to the next code block while the current block is still running.

  6. After you finish editing the model, return to the Amazon S3 console, choose your bucket name, choose the test folder, and then verify that the following artifacts of the edited model are stored in your S3 bucket's test folder.

    • Hotdog_or_not_model-0000.params

    • Hotdog_or_not_model-symbol.json

Step 4: Optimize the Model

Now that you have a trained mxNet model there is one final step that is required before you run the model on the AWS DeepLens’s GPU. The trained mxNet model does not come in a computationally optimized format. If we deploy the model in the original format it will run on the CPU via mxNet at sub optimal speed. In order to run the model at optimal speed on the GPU we need to perform model optimization. For instructions on how to optimize your MXNet model, see Optimize a Custom Model.

Step 5: Import the Model

Import the edited model into AWS DeepLens.

  1. Open the AWS DeepLens console by naviating to from a web browser.

  2. Choose Models, then choose Import model.

  3. For Import model to AWS DeepLens, choose Externally trained model.

  4. For Model settings, do the following:

    1. For Model artifact, type the path to the artifacts that you uploaded to the Amazon S3 bucket in the previous step. The path begins with s3://deeplens-. For example, if you followed the naming in Step 1, the path will be s3://deeplens-sagemaker-<your-name>/<dir>.

    2. For Model name, type a name for the model.

    3. For Model description, type a description.

  5. Choose Import model.

Step 6: Create an Inference Lambda Function

Use the AWS Lambda console to create a Lambda function that uses your model. For specific instructions with sample code, see Create and Publish an AWS DeepLens Inference Lambda Function.

Step 7: Create a New AWS DeepLens Project

Now create a new AWS DeepLens project and add the edited model to it.

  1. Open the AWS DeepLens console by naviating to from a web browser.

  2. Choose Projects.

  3. Choose Create new project.

  4. For Choose project type, choose Create a new blank project, then choose Next.

  5. For Project information, type a name and description for this project.

  6. For Project content, choose Add model.

  7. Search for and choose the model that you just created, then choose Add model.

  8. Choose Add function, then choose, for example, deeplens-hotdog-no-hotdog, then choose Add function.

  9. Choose Create.

Step 8: Review and Deploy the Project

  1. Open the AWS DeepLens console by naviating to from a web browser.

  2. From the list of projects, choose the project that you just created, then choose Deploy to device.

  3. Choose your AWS DeepLens as your target device, then choose Review.

  4. Review the settings, then choose Deploy.


    Deploying a project incurs costs for the various AWS services that are used.

Step 9: View Your Model's Output

To view your model's output, follow the instructions at Viewing AWS DeepLens Output Streams.