Combining hydra configurations with f3dasm

hydra is an open-source configuration management framework that is widely used in machine learning and other software development domains. It is designed to help developers manage and organize complex configuration settings for their projects, making it easier to experiment with different configurations, manage multiple environments, and maintain reproducibility in their work.

hydra can be seamlessly integrated with the worfklows in f3dasm to manage the configuration settings for the project.

Domain from a hydra configuration file

If you are using hydra to manage your configuration files, you can create a Domain from a configuration file. Your config needs to have a key (e.g. 'domain') that has two keys: 'input_space' and 'output_space'. Each design space dictionary can have parameter names (e.g. 'param_1') as keys and a dictionary with an optional parameter type ('type') and the corresponding arguments as values:

domain:
  input:
    param_1:
      type: float
      low: -1.0
      high: 1.0
    param_2:
      type: int
      low: 1
      high: 10
    param_3:
      type: category
      categories: ['red', 'blue', 'green', 'yellow', 'purple']
    param_4:
      type: constant
      value: some_value
  output:
    y:
      to_disk: False

In order to run the following code snippet, you need to have a configuration file named 'config.yaml' in the current working directory.

with initialize(version_base=None, config_path="."):
    config = compose(config_name="config")

domain = Domain.from_yaml(config.domain)

ExperimentData from a hydra configuration file

If you are using hydra for configuring your experiments, you can use it to construct an ExperimentData object from the information in the 'config.yaml' file with the from_yaml() method:

ExperimentData from file

You can create an ExperimentData object in the same way as the from_file() method, but with the 'from_file' key in the 'config.yaml' file:

experimentdata:
    from_file: ./example_project_dir

with initialize(version_base=None, config_path="."):
    config = compose(config_name="config")


experiment_data = ExperimentData.from_yaml(config.experimentdata)