Skip to content

adjust_model

This tool, part of the AllTools compilation, is designed to modify and enhance existing models by adding new components to their processing pipelines. It's especially useful for integrating new functionalities without rebuilding the entire model from scratch.

Common use cases include:

  • Adding Post-Calibration: You can add a post-calibration component to an existing model to fine-tune its outputs.
  • Adding Explainability: Easily integrate new modules for model explainability.

Command Line Interface:

The primary command for using adjust_model is:

1
adjust_model --rep <repository_path> --inModel <current_model_path> --out <out_model_path> --samples <path_to_samples_for_training_new_components> --preProcessors <pre_processor_addition_json> --postProcessors <post_processor_addition_json>

Arguments

  • --rep <repository_path>: The path to the data repository. Required unless you use --inCsv.
  • --inModel <current_model_path>: The path to the existing model file you want to modify.
  • --out <out_model_path>: The path where the new, adjusted model will be saved.
  • --samples <path_to_samples_for_training_new_components>: The path to the samples used for training the new components. Required if --learn_rep_proc is enabled or if --skip_model_apply is disabled.
  • --preProcessors <pre_processor_addition_json>: A JSON string or file containing a list of pre-model processors to add to the pipeline.
  • --postProcessors <post_processor_addition_json>: A JSON string or file containing a list of post-model processors to add to the pipeline.

Optional Flags

  • --inCsv: Provide your feature matrix directly in CSV format. This is an alternative to using a repository and samples.
  • --learn_rep_proc: When set to 1, this flag triggers the learn function for the new pre-processors. This is essential for components that require training. Note: If you use this flag, you must also provide --rep and --samples. (Default: 0)
  • --insert_rep_proc_first: When this flag is enabled, new pre-processors are added at the beginning of the existing pipeline. By default, they are appended to the end. (Default: 0)
  • --skip_model_apply: If 1, the model's pipeline won't be run to generate a feature matrix and samples. This can be useful for specific operations and can significantly speed up the process. (Default: 0)
  • --model_changes: This allows you to include instructions for modifying existing components within the model's pipeline. This can save you from having to use a separate change_model tool.

Note: You can use either --preProcessors or --postProcessors, or both.

JSON Templates and Examples

To add new components, you'll provide a JSON object defining the new processors. A "block" within this JSON corresponds to a specific processor. For example, a block might be a Calibrator processor.

Adding Pre-Processors

Here is a template for the --preProcessors argument. The "members" array is where you'll define each processor block.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
{ 
  "pre_processors": [
    {
      "action_type": "rp_set",
      "members": [
        // Your processor blocks go here
      ]
    }
  ] 
}
 

Adding Post-Processors

Here is a template for the --postProcessors argument.

1
2
3
4
5
{ 
  "post_processors": [
    // Your processor blocks go here
  ]
}

How to change existing components

If you want to change and not add new components, please refere to change_model