osl_dynamics.config_api.wrappers.train_mdynemo#

osl_dynamics.config_api.wrappers.train_mdynemo(data, output_dir, config_kwargs, init_kwargs=None, fit_kwargs=None, corrs_init_kwargs=None, save_inf_params=True)#

Train M-DyNeMo. This function will:

  1. Build an mdynemo.Model object.

  2. Initialize the mode correlations using sliding window and KMeans.

  3. Initialize the parameters of the model using the method set by init_method in the config (Model.initialization).

  4. Perform full training.

  5. Save the inferred parameters (mode time courses, means, stds and corrs) if save_inf_params=True.

This function will create two directories:

  • <output_dir>/model, which contains the trained model.

  • <output_dir>/inf_params, which contains the inferred parameters.

Parameters:
dataosl_dynamics.data.Data

Data object for training the model.

output_dirstr

Path to output directory.

config_kwargsdict

Keyword arguments to pass to osl_dynamics.models.mdynemo.Config. Defaults to:

{
    'n_channels': data.n_channels,
    'sequence_length': 200,
    'inference_n_units': 64,
    'inference_normalization': 'layer',
    'model_n_units': 64,
    'model_normalization': 'layer',
    'do_kl_annealing': True,
    'kl_annealing_curve': 'tanh',
    'kl_annealing_sharpness': 10,
    'n_kl_annealing_epochs': 20,
    'batch_size': 128,
    'learning_rate': 0.01,
    'lr_decay': 0.1,
    'n_epochs': 40,
}.
init_kwargsdict, optional

Keyword arguments to pass to the initialization method (Model.initialization). Defaults to:

{'n_init': 5, 'n_epochs': 5, 'take': 1}.
fit_kwargsdict, optional

Keyword arguments to pass to the Model.fit.

corrs_init_kwargsdict, optional

Keyword arguments to pass to the mode correlations initialization. Defaults to:

{
    'window_length': data.sampling_frequency * 2,
    'step_size': data.sampling_frequency // 25,
    'random_state': None,
    'n_init': 'auto',
    'init': 'k-means++',
}.
save_inf_paramsbool, optional

Should we save the inferred parameters?