(Ref: https://docs.microsoft.com/en-us/azure/machine-learning/reference-yaml-environment#yaml-syntax)
The environment item in an AzureML cli v2 job specification yml file can leverage an AzureML curated environment directly, referenced as follows:
environment: azureml:AzureML-pytorch-1.10-ubuntu18.04-py38-cuda11-gpu@latest
A list of available curated environments can be found here, or from wtihin the environments tab in AzureML Studio.
Alternatively, the environment could also be specified as a combination of a docker image address, and optionally, additional conda dependencies.
⚠️ Be aware of the packages already present in curated environments, and keep the conda specification strictly to few additional dependencies to avoid package conflicts.
environment:
image: azureml/curated/pytorch-1.10-ubuntu18.04-py38-cuda11-gpu:latest
conda_file: local/path/to/few-additional-dependencies.yml
The list of docker images available from the microsoft repository is here: mcr.microsoft.com/v2/_catalog
It is also possible to start from an image from public docker repositories. The image should at minimum support python and conda:
environment:
image: continuumio/miniconda3:latest
conda_file: local/path/to/custom-conda.yml
Lastly, the environment can be built from a local Dockerfile:
environment:
build:
path: ../local/docker-context-dir/
dockerfile_path: Dockerfile