Frontend Learning Book

Azure Machine Learning Service

Azure Machine Learning Service: Combination of Azure cloud services and Python SDK.

Create environment -> Prepare data -> Build models -> Train models -> Test models -> Manage models -> Track experiments -> Deploy models

Azure ML Studio

• Web portal in Azure Machine Learning
• Low-code and no-code options for project authoring
• Authoring tools: Notebooks, AutoML and Designer.
• Set of SDKs and cloud services to scale, monitor, and deploy ML models.

Three ways to machine learning projects in studio: * Azure Machine Learning designer * Automated Machine Learning UI * Notebooks

按代码量排序： 1. Automated ML 里配置自动化模型 2. Designer 里创建pipeline，拖拉拽step，用 Execute Python Script 嵌入代码 (见 aml2: AML Designer) 3. Notebooks 里用 PythonScriptStep 创建pipeline (见 aml1: AML Pipelines ) 4. Notebooks 里纯代码创建模型 (见 aml3: AML Notebook)

Compute Target

Attached to your Azure Machine Learning Workspace. Can be used to run pipelines and host deployed models as real-time endpoints or pipeline endpoints.

• Azure Machine Learning Compute: Training
• Azure Kubernetes Service: Deployment

优势

• 支持开源算法：Scikit-learn, TensorFlow, CNTK, PyTorch
• Automated 数据处理，Automated 算法选择，Automated 超参数选择
• Pipeline：串联step（数据准备/模型训练测试/模型部署），rerunnning指定step，不同step不同执行环境，创建可复用模板
• 集成 VS Code
• 部署模型 in ONNX format

ONNX (Open Neural Network Exchange) 的本质只是一套开放的 ML 模型标准，模型文件存储的只是网络的拓扑结构和权重，ONNX 的目的在于提供一个跨框架的模型中间表达框架，用于模型转换和部署

Build, Train and Deploy a Model步骤：

1. create Azure subscription -> resouce group -> workspace
2. import data -> prepare and clean data
3. train model -> test and tune model -> register model
4. compute target -> scoring file -> environment config file -> inference config
5. deploy model -> test model (web service)

Workspace

创建 Workspace： * Workspace name * Azure subscription ID * Resource group name * Location

每个Workspace配备： * Key Vault * Container Registry * Storage Account * Application Insights

Load Workspace

from azureml.core.workspace import Workspace
ws.write_config() # save workspace config file
ws = Workspace.from_config() # load an existing workspace
ws = Workspace.get(name="myworkspace", subscription_id='<subscription_id>', resource_group='myresourcegroup')

Build and Train a Model

Import Data

• AML Datasets (recommended): A reference to the data and its location together with a copy of its metadata
• Import Data Module: online data sources / data store / HTTP URL

Prepare and Clean Data

Transformation functions: * drop * dropna * rename * concat * transpose * replace * fillna * sum

Train Model

split data

from sklearn.model_selection import train_test_split

Automatic featurization 自动特征化:

• Generate more features – aka feature engineering 生成更多特征
• Impute missing values 插补缺少的值
• Transform – numeric to categorical 转换
• Encode – one-hot encoding and one-hot hash encoding 编码
• Drop features that have high cardinality or no variance 删除高基数或者无差异的特征
• Cluster distance 群集距离
• Word embeddings 单词嵌入，文档特征矢量

https://learn.microsoft.com/zh-cn/azure/machine-learning/how-to-configure-auto-features

Steps to use Automated ML: 1. Determine problem type：Classification,Regression,TimeSeriesforecasting 2. Decide between using Python SDK or the studio web experience 3. Determine the source and format of training data 4. Configure compute target for model training: Local computer, Azure ML Computes, Azure Databricks, remote VMs 5. Configure AutoML parameters 6. Submit the run 7. Examine the outcome

Save Model

Model 保存为 pickle 文件（.pkl） * Can be loaded later for predictions * Is the serialized version of the model. * Provides an efficient way to save model and transfer over the network. * Can be loaded and un-serialized using the load() method. * Can be used to load the model for further actions.

Register Model

Model registeration allows you to store and version the model. A registered model can be downloaded and deployed.

Compute Target: * AKS: Real-time High-Scale Production * AML Compute: Batch prediction from AML Pipelines * ACI: Quick testing * AZure IoT Edge: IoT * FPGA: Web

Deploy a Trained Model

Scoring File (.py)

two functions in score.py: * init() 加载模型 * run(input_data) 预测数据

# score.py
from sklearn.externals import joblib
from sklearn.linear_model import <algorithm_name> from azureml.core.model import Model
def init():
  global model
  model_path = Model.get_model_path ('<pkl_file_name>')
  model = joblib.load(model_path)
def run(raw_data):
  data = json.loads(raw_data)['data']
  data = numpy.array(data)
  result = model.predict(data)
  return result.tolist()

Environment Configuration File (.yml)

• Defines the conda environment for the model. (conda-dependency file)
• Used to make sure all model dependencies are included in the container image.

from azureml.core.conda_dependencies import CondaDependencies
env1 = CondaDependencies.create(conda_packages=['scikit-learn'])
with open("<yml_file_name>","w") as f: f.write(env1.serialize_to_string())

inference configuration

describes how to set up the web-service containing your model.

from azureml.core.model import InferenceConfig
inference_config = InferenceConfig(runtime= "python", entry_script="score.py",
conda_file="<yml_file_name>")

Deploy to Compute Target

service = model.deploy( workspace=ws, name='<service_name>', deployment_config=aciconfig, models=[model],
inference_config = inference_config )

Consume Deployed Model

A web service is created if image is deployed to ACI/AKS/FPGA

Deploying as web service creates a REST API. Can use HTTP to send requests and receive responses

import json
test_input = json.dumps({'data':[[i1, i2, ...]]}) 
prediction = service.run(input_data = test_input) Print(prediction)

Monitor

Azure monitor is a full stack monitoring service in Azure.

• Metrics：可以添加 Model Deploy Failed 到 dashboard
• Alert：Alert rule: Model Deploy Failed 发送邮件到选择的 action group

AML Pipelines

1.Install AML SDK -> create workspace

2.Set up data store

workspaceblobstore (default) / workspacefilestore

datastore = ws.get_default_datastore # SDK v1
default_store = ml_client.datastores.get_default() # SDK v2

3.Upload data to storage

4.Create Data Reference object

from azureml.data.data_reference import DataReference
stored_data = DataReference(
  datastore = MyDefault_Store,
  data_reference_name = "FlightDelay_dataset", path_on_datastore = ("FlightDelay/part-00000")
)

5.run configuration (pip package dependencies / SDK version)

6.Create PipelineData object

from azureml.pipeline.core import PipelineData
normalized_data = PipelineData(
  name = "normalized_flight_delay_data",
  data_store = default_store
)

7.Construct pipeline steps

need to define: * input / output * compute target / run configuration * other required parameters

DataCleaningStep = PythonScriptStep(
  name="Clean Data", 
  script_name="cleanse_datasets.py",
  arguments=[<script_inputs>], 
  inputs=[<DataReference_or_PipelineData>], 
  outputs=[cleansed_data],
  compute_target=MyAml_Compute_Target, runconfig=MyAml_Run_Config, source_directory=<path_to_Python_script>, 
  allow_reuse=True
)

8.Build pipeline

from azureml.pipeline.core import Pipeline
PPLines_Steps = [CleaningStep, NormalizingStep] # 定义step顺序
MyPipeline = Pipeline(workspace = ws, steps = PPLines_Steps)

9.Submit pipeline

from azureml.core import Experiment
PPLine_Run = Experiment(ws, "exp_name").submit(MyPipeline,regenerate_outputs = True)

10.View progress

from azureml.widgets import RunDetails
RunDetails(PPLine_Run).show()
PPLine_Run.wait_for_completion()

# get step output
def fetch_df(step, output_name):
    output_data = step.get_output_data(output_name)    
    download_path = './outputs/' + output_name
    # download to local
    output_data.download(download_path, overwrite=True)
    df_path = get_download_path(download_path, output_name) + '/processed.parquet'
    return pd.read_parquet(df_path)
y_test = fetch_df(split_step, output_split_test.name)[['cost']]

MLOps

Train Model -> Package Model -> Validate Model -> Deploy Model -> Monitor Model -> Retrain Model

模型衰退(Model Decay) 历史的模型效果并不能保证未来的模型效果

7 Principles of MLOps * Version control code, data, and experimentation outputs * Use multiple environment * Manage infrastructure and configurations-as-code * Track and manage machine learning experiments * Test code, validate data integrity, model quality * Machine learning continuous integration and delivery * Monitor services, models, and data

MLOPs solution Architecture

Azure Event Grid

AML events:

• Run completion
• Model registration
• Model deployment
• Data drift detection
• Run status change

Use cases:

• Send emails on run completion
• Use an azure function after a model is registered
• Trigger an ML pipeline when drift is detected