SageMaker Example for Keras

Requirements

• Python 3.6
• Tensorflow 1.12

Setting Environment Variable

Required

SAGEMAKER_ROLE='arn:aws:iam::1234567890'

Optional

S3_BUCKET_NAME='sagemaker-example'

Dataset Preparation

Download MNIST dataset and upload to S3.

python prepare_mnist.py

Training

Create a SageMaker training job. Then, the job name is mnist-yyyy-mm-dd-HH-MM.

python create_training_job.py

Source code used for training is stored in S3. And, after training is complete, the output and model are stored in S3.

• source code: s3://bucket-name/training/job-name/source/sourcedir.tar.gz
• output: s3://bucket-name/training/job-name/output/output.tar.gz
• model: s3://bucket-name/training/job-name/output/model.tar.gz

Deployment

Deploy model to TensorFlow Serving-based server in SageMaker.

python deploy_model.py <model_data>

model_data is s3 path to model.tar.gz. For example, s3://bucket-name/training/job-name/output/model.tar.gz.

Models to be deployed are available not only for models trained in sagemaker, but also for BYO models.

Endpoint Delete

python delete_endpoint.py <endpoint_name>

Inference

python infer.py <endpoint_name>

TODO

• Inference
  • REST (API Gateway)
  • Batch transform
• Custom Image