Import data and python scripts in azure ml entry script when deploying models

I found the solution:

The solution was right there I just had to read more in the official documentation. Below I have explained the solution to my problem.

When deploying a model on Azure Machine Learning Studio we have to prepare 3 things:

1. Entry Script is the actual python script that makes predictions.

2. Deployment Configuration can be thought of as the computer where your model will run.

3. Inference Configuration defines the software dependencies of your model. It is here, where we can also provide text/static files.

The inferece_config takes an argument called source_directory. We can provide a path to a folder to the source_directory argument like below

inference_config = InferenceConfig(
    entry_script="<entry-script.py>",
    environment=myenv, 
    source_directory="<path-to-your-folder>",
)

We can place all our requirements in this folder, including text, static, CSV, python scripts, etc files. The entry script must also be placed in this folder. This folder will be directly copied by azure in the deployment image that gets created. This way we can access these files in the entry script using regular python syntax.

For my solution:

If I have my inference_config like this

inference_config = InferenceConfig(
    entry_script="score.py", 
    environment=myenv, 
    source_directory='./assets',
)

And have this in the ./assets folder

assets
├── my_model_class.py
├── score.py
└── data
    └── vocaublary.txt

1 directory, 3 files

Then my entry script (score.py) will look something like this:

from assets.my_model_class import MyModelClass
    
def init():
  global my_model
  model_path = os.path.join(os.getenv('AZUREML_MODEL_DIR'), 
                            'my_model.pkl',)

  my_model = MyModelClass(vocab_path='assets/data/vocaublary.txt', 
                          model_path=model_path,)

def run(json_data):
  try:
    data = json.loads(json_data)
    prediction = my_model.predict(data)
    return {'data': prediction}
  except Exception as e:
    // do error handling
    return error

Note: We have to import from assets.my_model_class even though score.py is present in assets folder because, after deployment the CWD(Current Working Directory) changes to be the parent folder of assets folder

The additional files related to a model can be uploaded directly to Azure ML Workspace(with GUI) in source-directory, which is HOME > Notebooks. (Refer to the image).

Add files into this(source-directory) using, Add Files button

The uploaded files can be accessed os.getEnv() method.

from azureml.core import Workspace
from azureml.core import Environment
import os

# Connect to WS
ws = Workspace(subscription_id="sid",
               resource_group="rg",
               workspace_name="wsname")

# Creating inference Config for demo
env = Environment(name='environment_name')
inf_config = InferenceConfig(environment=env, source_directory=os.getenv('ENV_NAME'), entry_script='./score.py')

For txt/static files - you could upload to Azure Blobs and consume them on fly. And for classes, you can encapsulate your classes inside the modules.

There are 2 options you could do here

Option 1 :

You could publish your module to pip and then consume it.

Option 2 :

The "Private wheel files" solution as described in this article. The below is the sample environment and pls look closely under the pip section.

name: project_environment
dependencies:
  # The python interpreter version.

  # Currently Azure ML only supports 3.5.2 and later.

- python=3.6.2

- pip:
  - scikit-learn==0.22.1
  - azureml-defaults
  - https://myworkspaceid.blob.core/azureml/Environment/azureml-private-packages/my-wheel-file-name.whl
channels:
- conda-forge

Saw a reference code and similar discussion in this thread.