This is a self-paced workshop designed for anyone who is interested in building self-driving cars using Amazon SageMaker.
AWS has published several blog posts that walkthrough the process of building one. In the first blog post of the autonomous vehicle series, you built your Donkey vehicle and deployed your pilot server onto an Amazon EC2 instance. In the second blog post, you learned to drive the Donkey car, and the Donkey car learned to self-drive. In the third blog post, you learned about the process of streaming telemetry from the Donkey vehicle into AWS using AWS IoT. In the forth blog post, you learned the concept of behavioral cloning with Convolutional Neural Networks (CNNs).
In this workshop, we will go through how to setup Amazon SageMaker with custom algorithm to train the model for your autonomous vehicle.
You have built a car by following the instructions in Assemble hardware and Install software section of the documentation from donkeycar, with one exception that you use the repo from https://github.com/chankh/donkey instead.
$ git clone https://github.com/chankh/donkey donkeycar
$ pip install -e donkeycar
You will also need to have an AWS account to go through this workshop.
Previously we would copy the data from the Pi to our Amazon EC2 instance. However with Amazon SageMaker, you don't need an Amazon EC2 instance. Instead the data will be pulled from Amazon S3. So let's first create our stack containing all required AWS resources by choosing one of the regions below. This diagram shows the architecture that will be created by our CloudFormation template.
| Region | Launch Template |
|---|---|
| N. Virginia (us-east-1) |  |
| Oregon (us-west-2) | |
| Tokyo (ap-northeast-1) | |
Once the stack is created, status is CREATE_COMPLETE, you can find the AWS IoT Endpoint URL and S3 Bucket Name under the Outputs tab. Save these values as we will need them later when configuring our donkeycar and training our model.
You will need to calibrate your car by following the instructions described in Calibrate your car.
In your AWS Management Console, go to Services then IoT Core. In the navigation panel on the left, click Manage and Things. You should see a Thing has been created for you.
Click onto the donkeycar item. Next, you will need to create a certificate for our donkeycar to connect to AWS IoT. To do that, click on Security on the navigation panel, and then click Create certificate.
After AWS IoT has created the certificates, you need to download all these certificates into your donkeycar. For the root CA for AWS IoT, you will have to download Amazon Root CA 1 since we are using Amazon Trusted Services Endpoint for connection. Click Activate to enable this certificate. Click Attach a policy to attach a policy so that our donkeycar has the right permission to access the platform.
Next, you will be asked to attach a policy to the certificate. The policy is already created for you. Click the checkbox and then click Done.
You have now completed the setup on AWS, next you need to copy the certificates that you have downloaded to your donkeycar and update the config.py file with the appropriate values.
#AWS IOT
IOT_ENABLED = True
VEHICLE_ID = 'donkey'
AWS_ENDPOINT = <replace with endpoint from AWS IoT console>
CA_PATH = os.path.join(CAR_PATH, '<path to the root CA PEM file>')
PRIVATE_KEY_PATH = os.path.join(CAR_PATH, '<path to the private key file>')
CERTIFICATE_PATH = os.path.join(CAR_PATH, '<path to the certificate file>')
Save the changes and you are good to go.
The first step in teaching your car to drive is laying out a track. After your track is ready, you’ll need to drive your car through its first round to start training its neural network. Feel free to use your creativity in the design, but make sure that you incorporate both left and right hand turns of various degrees to give your neural network exposure to many different situations. Keep in mind that your car uses its front facing camera to see the track you lay out, so you’ll need to make sure the tape is easily distinguishable from the floor.
Make sure you collect good data.
- Practice driving around the track a couple times without recording data.
- When you're confident you can drive 10 laps without mistake press Start Recording
- If you crash or run off the track press Stop Car immediately to stop recording. A little bad data won't affect your autopilot.
- After you've collected 10-20 laps of good data (5-20k images) you can stop your car with Ctrl-c in the ssh session for your car.
- The data you've collected is sent to AWS and stored in your Amazon S3 bucket.
Training the model with Amazon SageMaker
Amazon SageMaker makes it easy to train and deploy ML models. In your AWS Management Console, go to Services and Amazon SageMaker. Navigate to Notebook Instances on the left, we have created a notebook for you. Click Open to launch your notebook.
Download the notebook from here and upload to your notebook environment in Amazon SageMaker. After that click on the name donkeycar.ipynb to open this notebook.
In the notebook, you need to update the value of data_location to the name of your S3 bucket,
which you can retrieve from the Outputs in CloudFormation.
After that, you can run the notebook by clicking Cell and Run all.
Amazon SageMaker saves the model to a S3 location. You can now download the model from S3 into your donkeycar and extract the contents from the tarball. This should give you a file named donkeycar.
$ curl -O <S3 location>
$ tar -zxf model.tar.gz
Run your autonomous vehicle using the downloaded model.
$ python manage.py drive --model donkeycar
Conclusion
Amazon SageMaker provides an easy to use platform for building ML models. In this post, you've seen how it's possible to bring your own algorithms to SageMaker and start training immediately. It is so easy that even someone without knowledge on ML is able to do it without any special training.