Ksama Arora

When which compute used?:

• For experimentation: Compute instance or Serverless Spark compute (make use of Spark’s distributed compute power)
• For production: Compute cluster (automatically scales) or Azure ML Serverless compute
• For deployment:
  • Batch predictions - Compute clusters or Azure ML serverless compute
  • Real-time predictions - Azure ML containers or Kubernetes cluster

Create Compute Targets*

1) Using Python SDK:

Create Compute Instance:

A compute instance can be assigned to one user, as it can’t handle parallel workloads. Can schedule to start/stop compute instance or configure to automatically shut down when been idle for set amount of time

IMP:

• Configure compute instance using YAML file given scale down to one instance after one hour of activity

  # Set YAML file parameters
  min_instances: 1 
  idle_time_before_scale_down: 3600 # 1 hour = (60 min x 60 sec) = 3600 seconds
  

from azure.ai.ml.entities import ComputeInstance

# Compute Instances need to have a unique name across the region.
ci_basic_name = "basic-ci-12345"
ci_basic = ComputeInstance(
    name=ci_basic_name, 
    size="STANDARD_DS3_v2"
)

ml_client.begin_create_or_update(ci_basic).result()

Create a Compute Cluster:

3 main parameters to be defined:

• size: specify size of VM and also specify u want to use CPU or GPU
• max_instances: specify maximum number of nodes i.e. the number of parallel workloads compute cluster can handle
• tier: specify tier of VM - low priority or dedicated

NOTE: Increase the idle time before scale down to keep the compute cluster active between pipeline runs, minimizing start-up time for rapid experimentation.

3 scenarios to use compute cluster

• Running a pipeline job you built in the Designer.
• Running an Automated Machine Learning job.
• Running a script as a job.

from azure.ai.ml.entities import AmlCompute

cluster_basic = AmlCompute(
    name="cpu-cluster",
    type="amlcompute",
    size="STANDARD_DS3_v2",
    location="westus",
    min_instances=0,
    max_instances=2,
    idle_time_before_scale_down=120,
    tier="low_priority",
)
ml_client.begin_create_or_update(cluster_basic).result()

• Sample code - set compute target as compute cluster

from azure.ai.ml import command

# configure job
job = command(
    code="./src",
    command="python diabetes-training.py",
    environment="AzureML-sklearn-0.24-ubuntu18.04-py37-cpu@latest",
    compute="cpu-cluster",
    display_name="train-with-cluster",
    experiment_name="diabetes-training"
    )

# submit job
returned_job = ml_client.create_or_update(job)
aml_url = returned_job.studio_url
print("Monitor your job at", aml_url)

2) Using Azure CLI:*

Create Compute Instance

• Settings:
  • Compute name: Name of compute instance. Has to be unique and fewer than 24 characters.
  • Virtual machine size: STANDARD_DS11_V2
  • Compute type (instance or cluster): ComputeInstance
  • Azure Machine Learning workspace name: mlw-dp100-labs
  • Resource group: rg-dp100-labs

az ml compute create --name "ci1572" --size STANDARD_DS11_V2 --type ComputeInstance -w mlw-dp100-labs -g rg-dp100-labs

Create Compute Cluster

• Settings
  • Compute name: aml-cluster
  • Virtual machine size: STANDARD_DS11_V2
  • Compute type: AmlCompute (Creates a compute cluster)
  • Maximum instances: Maximum number of nodes
  • Azure Machine Learning workspace name: mlw-dp100-labs
  • Resource group: rg-dp100-labs

az ml compute create --name "aml-cluster" --size STANDARD_DS11_V2 --max-instances 2 --type AmlCompute -w mlw-dp100-labs -g rg-dp100-labs

Create compute target (NOTE IMP)

• METHOD 1: Creating compute directly without using YAML file

  az ml compute create --name aml-cluster --size STANDARD_DS3_v2 --min-instances 0 --max-instances 5 --type AmlCompute --resource-group my-resource-group --workspace-name my-workspace
  

• METHOD 2: Creating compute using YAML file
  • Can use YAML files to define configuration - becomes easier to organize and automate tasks
  • NOTE: Designed for automating tasks. By using YAML files to define how the model must be trained, the ML tasks will be repeatable, consistent and reliable.
  • yml file configuration

    $schema: https://azuremlschemas.azureedge.net/latest/amlCompute.schema.json 
    name: aml-cluster
    type: amlcompute
    size: STANDARD_DS3_v2
    min_instances: 0
    max_instances: 5
    

  • When u saved the YAML file as compute.yml, you can create compute target by following command

    az ml compute create --file compute.yml --resource-group my-resource-group --workspace-name my-workspace