aml-constructor

Azure Machine Learning Pipeline Constructor

aml-constructor - or shortly - azuremlconstructor allows you to create Azure Machine Learning(shortly - AML) Pipeline. azuremlconstructor based on the Azure Machine Learning SDK, and implements main operations of the Pipeline creation. You can create pipelines with AML Steps, which can take DataInputs. In azuremlconstructor pipeline creation consists of 3 steps:

0. Preporation

It's highly recommended to create separated folder your pipeline projects. And also, virtual environment(venv) - article on RealPython. You can create separated venv for future AML projects. It's specially useful if you are working with different kinds of libraries: data science oriented, web and so on.

1. Pipeline initialisation

Something like project initialisation. You choose pipeline name, directory and credential .env file. For storing azuremlconstructor has denv storage - or EnvBank. Initialise pipeline as:

python -m azuremlconstructor init [path] -n myfirstpipe -e denv_name

Here -n shows pipeline name, path - directory in which pipeline will be created - by default = ., -e - dotenv name. I will talk about denv's a little bit later. After this, in the passed directory will be created named as pipeline passed name.

myfirstpipe
---|settings/
------|settings.py
------|.amlignore
------|.env
------|conda_dependencies.yml

Inside the directory settings directory which contains: settings.py, .amlignore, .env and conda_dependencies.yml files. conda_dependencies.yml will be used for environment creation on AML side. .amlignore something like .gitignore but for AML. .env is file form of our EnvBank instance. -e is optional, if it's skipped, will be created .env template with necessary fields, which you have to fill before running pipeline.

settings.py:

This module contains all necessary configuractions:

from azuremlconstructor.input import FileInputSchema, PathInputSchema
from azuremlconstructor.core import StepSchema

# --------------------------| Module Names |----------------------------
AML_MODULE_NAME: str =       'aml'
SCRIPT_MODULE_NAME: str =    'script'
DATALOADER_MODULE_NAME: str = 'data_loader'



# ---------------------------| General |---------------------------------

NAME = "{{pipe_name}}"
DESCRIPTION = "Your pipeline description"


# ---------------------------| DataInputs |-------------------------------

file = FileInputSchema(
                       name='name', 
                       datastore_name='datastore', 
                       path_on_datastore='', 
                       files = ['file.ext'], 
                       data_reference_name = ''
   )

path = PathInputSchema(
                       name='name', 
                       datastore_name='datastore', 
                       path_on_datastore='',
                       data_reference_name=''
   )
# ---------------------------| Steps |---------------------------------
step1 = StepSchema(
                       name='step_name', 
                       compute_target='compute_name', 
                       input_data=[file, path], 
                       allow_reuse=False
           )
STEPS = [step1, ]

# ---------------------------| extra |---------------------------------

# 'submit' option will apply if set `is_active = True`

EXTRA = {
           'continue_on_step_failure': False,
           'submit': {'is_active': False, 'experiment_name': 'DebugPipeline', 'job_name': NAME, 'tags': None, 'kwargs': None}
}

Lets look at the variables we have here.

AML_MODULE_NAME - initially, pipeline project has 3 main scripts: dataloader.py - loads all the DataInputs into the pipeline, aml.py - main script of the pipeline, loaded data inputs imported here automaticaly, script.py - just empty script for implement your deep logic. You are free for remove this module or add so many as you need, however - the entry point of project is aml.py. AML_MODULE_NAME is the name of aml.py module. And the same thing for DATALOADER_MODULE_NAME and SCRIPT_MODULE_NAME.

NAME - name of your pipeline.

DESCRIPTION - description of the pipeline.

PathInputSchema and FileInputSchema DataInput of your pipeline. You create instances of the classes and pass into StepSchema class. Each StepSchema class is abstraction of PythonScriptStep. All steps must be inside STEPS list.

EXTRA options

There are extra - additional options that can be helpfull.

• continue_on_step_failure - Indicates whether to continue execution of other steps in the PipelineRun if a step fails; the default is false. If True, only steps that have no dependency on the output of the failed step will continue execution.

• submit - submit options. Pipeline will be submitted, if is_active is True.

After filling settings, you have to apply your settings.

2. Apply Settings

python -m azuremlconstructor apply <path_to_pipeline>

Applying pipeline means - create structure based on the settings.py module. For each step will be created directory inside pipeline directory and each directory will contain: aml.py, dataloader.py and script.py.

After applying, your project structure will be like this:

myfirstpipe
---|settings/
------| settings.py
------| .amlignore
------| .env
------| conda_dependencies.yml
---| step_name/
------| dataloader.py
------| aml.py
------| script.py
---| step2_name/
------| dataloader.py
------| aml.py
------| script.py

Note: names of the modules setted in the settings.py module.

3. Run Pipeline

bash python -m azuremlconstructor run <path_to_pipeline>

This command will publish your pipeline into your AML. Additionally, can submit according to the EXTRA.submit option.

EnvBank

For work on AML pipeline you have to use your credentials: workspace_name, resource_group, subscription_id, build_id, environment_name and tenant_id. In amltor these variables store as instances of EnvBank, which is encrypted jsonlike file. You can create, retrieve or remove EnvBank instances(I'll name it as denv). In this purpose you've to use denv command.

Create denv

You can create denv in 2 ways: pass path of existing .env file or in interactive mode - via terminal. In the first case:

python -m azuremlconstructor denv create -p <path_to_.env file> -n <new_name>

Then you'll type new password twise for encryption. After that, denv will save into local storage and you will be able to use it for future pipeline creation.

For create denv in interactive mode, you have to pass -i or --interactive arg:

python -m azuremlconstructor denv create -i

After that you have to type each asked field and set password.

Get denv

For retrieve denv use:

python -m azuremlconstructor denv get -n <name_of_denv>

For list all existing denv names add --all argument:

python -m azuremlconstructor denv get --all

Note: for view the denv, you have to type password.

Remove denv

For removing denv:

python -m azuremlconstructor denv rm -n <name_of_denv>

DataInputs

DataInputs can be files or paths from AML Datastore. Whole process is creating DataReference object behind the scenes... All inputs will be loaded in the dataloader.py and imported into aml.py module. Lets look at azuremlconstructor DataInputs.

PathInputSchema

Allows you to create data reference link to any directory inside the datastore. class looks like this:

class PathInputSchema:
    name: str
    datastore_name: str
    path_on_datastore: str
    data_reference_name: str

Where: name name of your PathInput, this name will be used as variable name for importing. datastore_name - Datastore name, path_on_datastore - target path related to the Datastore. data_reference_name - data reference name for DataReference class, optional - if empty, will be used name.

FileInputSchema

Allows you to mount files from Datastore. Behind the scines, very similar to PathInput, but with file oriented additions.

class FileInputSchema:
    name: str
    datastore_name: str
    path_on_datastore: str
    data_reference_name: str
    files: List[str]

First 4 fields as previous. files - you can list file or files as list, which will be mounted from Datastore. If you want to get one file, pass as string, for more files - list of strings. File inputs will be assigned to variable names - generated on the base of file name itself. You can use FileInputSchema.files dict notation, which allows you pass {'file_name.extention': 'variable_name', 'file_name2.extention': 'variable_name2', ...} for map files with variable names to use. Remember that, variable names must be unique in the scope of step. When you pass multiple filename, they must be on the same path.

Supported file types: azuremlconstructor uses pandas pandas.read_... methods for read the mounted files. At the moment, suported file types:

csv, parquet, excell sheet, json

Slugged file names will be used as variable names for importing files.

Other commands

Update

You can update project according to the settings.py. Updates will affect whole project if passed --overwrite. Otherwise, user have to choose what to do with already existing modules - overwrite, skeep or cancel updating. It can be useful when you maked some changes into settings.py and don't want to overwrite whole pipeline structure by scratch, in this case you can use update:

python -m azuremlconstructor update <path_to_pipe> --overwrite [Optional]

Rename

python -m azuremlconstructor rename <path_to_pipe> -n <new_name>

Renames pipeline into new_name. Renaming pipeline means: rename pipeline project directory, change NAME variable in settings.py and edit ENVIRONMENT_FILE in the .env file.

Some usefull utils

azuremlconstructor.utils module has a banch of usefull tools, that can be usefull. - utils.upload_data(datastore_name: str, files: List[str], target_path: str=".") - uploads file(s) to the blob; - recursive read_concat functions: utils.read_concat_csvfiles: List[str], return_types: bool=False, sep: str = ',', utils.read_concat_parquet(files: List[str], return_types: bool=False, engine: Literal['fastparquet', 'pyarrow'] = 'fastparquet'), utils.recursive_glob_list(folders: List[str], file_ext: str='parquet'). Each function has doc