mllf.cli.workflow

High-level workflow utilities for preparing combos, training and running sims.

This module centralizes common steps used for contextual bandit training workflows:

  1. Generate combination directories from site/sub fragment files

  2. Split manifests into train/validation sets

  3. Build PyG graphs from RTF fragments or variables.py files

  4. Run quick training epochs for testing

  5. Execute simulations concurrently with Slurm support

  6. Archive completed runs

Functions

build_data_and_targets_from_combo(combo_dir)

Build PyG Data object and per-edge targets from a combo directory.

compress_runs(manifest, out_tar)

Create a gzipped tar archive of the directory containing combo runs.

create_and_manifest(input_dir, out_dir[, ...])

Generate combination directories and create a manifest file.

default_env_reward(actions, target_vals)

Compute reward as negative MSE between actions and target values.

graph_from_bias(bias)

Build an mllf Graph from a bias dict.

load_bias_from_variables(py_path)

Load the YAML bias mapping embedded inside a variables.py file.

load_manifest(manifest_path)

Load list of combo directories from manifest file.

run_from_config(config_path)

Execute a complete workflow based on a YAML configuration file.

run_quick_epoch_for_combo(combo_dir[, base_bias])

Run a single training epoch for demonstration/testing purposes.

save_graph_info_from_rtf(combo_dir, g)

Save graph_info.json from a Graph object built from RTF files.

split_manifest(manifest[, train_frac, seed])

Split a manifest file into training and validation sets.

write_variables_from_actions(combo_dir, ...)

Write a variables.py file from per-directed-edge policy actions.