Bring your own container on Amazon SageMaker Lab

1. create a cloud9 environment in east-us-1 with name: sagemaker-container-workshop and type t2.micro.

2. In cloud9, bash shell exec: git clone https://github.com/awslabs/amazon-sagemaker-examples.git

3. cp -r amazon-sagemaker-examples/advanced_functionality/scikit_bring_your_own/ /home/ec2-user/environment/.

4. cd scikit_bring_your_own/container/

5. Build the docker image: ./build_and_push.sh <image-name>. Image name suggest to use scikit-<your-name> format.

6. docker images: you will see scikit-<your-name> with latest TAG in your cloud9 and ecr respository.

local training

1. cd local_test
2. chmod +x *.sh
3. ./train_local.sh scikit-<your-name>
4. ls test_dir/model/ to check the model output.

error debug

1. mv test_dir/input/config/hyperparameters.json test_dir/input/config/hyperparameters.json.bak
2. ./train_local.sh scikit-<your-name> You will see error in the console
3. cat test_dir/output/failure to see failure
4. mv test_dir/input/config/hyperparameters.json.bak test_dir/input/config/hyperparameters.json
5. ./train_local.sh scikit-<your-name>

Local Server Inference

1. ./serve_local.sh scikit-<your-name> > output.log
2. Open new shell terminal in cloud9
3. cd scikit_bring_your_own/container/local_test/
4. ./predict.sh payload.csv text/csv

Train and deploy using the Amazon SageMaker console

Upload full data set to s3

• remember to change the bucket name sagemaker-iris-dataset-<your-id>-yyyymmdd

cd /home/ec2-user/environment/scikit_bring_your_own/container/local_test/test_dir
aws s3api create-bucket --bucket sagemaker-iris-dataset-beyoung-20200806 --create-bucket-configuration LocationConstraint=us-west-2
aws s3 cp . s3://sagemaker-iris-dataset-beyoung-20200806 --recursive

s3 bucket name as sagemaker-iris-dataset-<your-id>-yyyymmdd/data/training/

Create Training Job

• Job name: scikit-<your-name>-yyyymmdd
• Algorithm: Custom
• Input mode: File
• Training image: <Amazon ECR path>:<tag>. Can get it from docker images
• Hyperparameters: max_leaf_nodes value: 8
• Training
  • channel name: training
  • s3 location: sagemaker-iris-dataset-<your-id>/data/training/
• Output data configuration: s3 output path: s3://sagemaker-iris-dataset-<your-id>/models/
• create training job.
• After training job complete, you will see model in s3://<bucket name>/models/<job name>/output/model.tar.gz You can get it from trainig jobs detail page.

Deploy an endpoint

Create model

• model name: scikit-<your-name>-yyyymmdd
• click create the model

Create an endpoint

• name: scikit-<your-name>-yyyymmdd

• new endpoint configuration: scikit-<your-name>-yyyymmdd

• create endpoint configration

• click create endpoint

Test the endpoint

In cloud9, install pip3:

• sudo easy_install-3.6 pip
• sudo /usr/local/bin/pip3 install boto3 pandas

Download the test-endpoint-sample.py in the same github folder.

wget "https://raw.githubusercontent.com/HKT-SSA/bring-your-own-container-on-sm/master/test-endpoint-sample.py" \
-O "/home/ec2-user/environment/test-endpoint-sample.py"

• modify the bucket name and endpoint
• run python3 test-endpoint-sample.py

Clean Up

• SageMaker Endpoint
• s3
• cloud9

Reference

• AWS Blog