Model Hyperparameter¶

All hyperparameter values are stored separately from the model code in models/hyperparameters.yaml.

This design avoids hard-coded values and provides a central place to manage model settings.

The workspace config also includes reproducibility.seed in specs/pynnlf_config.yaml. When this value is set, PyNNLF seeds common random number generators and overrides seed-like hyperparameter keys such as seed, xgb_seed, random_seed, and random_state at runtime. The experiment summary records the central seed in a1_experiment_result.csv, and a2_hyperparameter.csv records the effective hyperparameters used for the run.

Format¶

PyNNLF expects a YAML mapping:

The top-level key must match the full model file stem, such as m7_ann.
Each hyperparameter set is stored under an hp_no key such as hp1, hp2, or hp3.

Example:

m7_ann:
  hp1:
    seed: 99
    hidden_size: 10
    activation_function: relu
    learning_rate: 0.001
    solver: adam
    epochs: 500
  hp2:
    seed: 99
    hidden_size: 10
    activation_function: relu
    learning_rate: 0.01
    solver: adam
    epochs: 500

How to Modify Model Hyperparameter¶

If you want to modify the learning rate of the ANN model, add a new hyperparameter set with a new hp_no:

m7_ann:
  hp1:
    seed: 99
    hidden_size: 10
    activation_function: relu
    learning_rate: 0.001
    solver: adam
    epochs: 500
  hp2:
    seed: 99
    hidden_size: 10
    activation_function: relu
    learning_rate: 0.01
    solver: adam
    epochs: 500
  hp3:
    seed: 99
    hidden_size: 10
    activation_function: relu
    learning_rate: 0.1
    solver: adam
    epochs: 500

Then select the new hyperparameter set in example_project/specs/experiment.yaml:

model: m7
hyperparameter: hp3

The experiment spec uses the short model ID such as m7, while models/hyperparameters.yaml uses the full model file stem such as m7_ann.