Custom Objects

Azimuth uses Custom Objects to define how to integrate with models, datasets and metrics. Custom Objects are instantiated at runtime to load the object; the ML pipeline function will return a pipeline, and the Dataset function will return the dataset.

Our primary integration is through HuggingFace, but Azimuth supports any other type of dataset or model.

• Defining Datasets
• Defining a Model
• Defining Postprocessors
• Defining Metrics

What is a Custom Object?

A Custom Object is simply a path to a function and its arguments. When users supply their functions and classes, they should be added to azimuth_shr, as it is already mounted at startup on the Docker image.

from typing import Any, Dict, List, Optional, Union

from pydantic import BaseModel, Field


class CustomObject(BaseModel):
    class_name: str # (1)
    args: List[Union["CustomObject", Any]] = []
    kwargs: Dict[str, Union["CustomObject", Any]] = {}
    remote: Optional[str] = None # (2)

1. Name of the function or class that is located in remote. args and kwargs will be sent to the function/class.
2. Absolute path to class. class_name needs to be accessible from this path.

Example

Here is in example of two custom objects. In azimuth_shr/loading_resources.py, we will add two functions which will load a model and a dataset. The configuration file will then be defined as shown in config example.

azimuth_shr/loading_resources.pyConfig Example

import transformers
import datasets


def my_model(ckpt_path) -> transformers.Pipeline:
    pipeline = ...  # Load the pipeline from ckpt_path
    return pipeline


def my_dataset(ckpt_path) -> datasets.DatasetDict:
    dataset = ...  # Load the Dataset from ckpt_path
    return dataset

{
  "dataset": {
    "class_name": "loading_resources.my_dataset",
    "remote": "/azimuth_shr",
    "kwargs": {
      "ckpt_path": "/azimuth_shr/data/my_dataset" # (1)
    }
  },
  "pipelines": [
    {
      "model": {
        "class_name": "loading_resources.my_model",
        "remote": "/azimuth_shr",
        "kwargs": {
          "ckpt_path": "distilbert-base-uncased-finetuned-sst-2-english"
        }
      }
    }
  ]
}

1. Path to the dataset. Should be put under /azimuth_shr so it is mounted on the Docker image automatically.

Using the Config

If the function has an argument named azimuth_config, Azimuth will supply the config file to the function automatically. This can be useful if some attributes from the config are needed.

from azimuth.config import AzimuthConfig

# This is fine.
def my_model(num_classes, azimuth_config: AzimuthConfig):
    pass


# Also fine
def my_model(num_classes):
    pass

Install dependencies at runtime

To avoid needing to modify the Docker image to include your dependencies, Azimuth supports installing dependencies at runtime.

To do this, the field remote can be one of the following:

1. A package on Pypi
   1. Example: remote: "torchvision"
2. A folder with requirements.txt
3. A folder with setup.py

If none of the above matches, we append remote to PYTHONPATH, but no dependency will be installed.