Utils Module

visualize_training.utils.break_list_by_lengths(lst, lengths)

Break the list in chunks of specified length

Parameters:

  • lst (list) – input list

  • lengths (int) – lengh of one chunk

Returns:

list of chunks of specified length

Return type:

result

visualize_training.utils.characterize_all_transitions(model, data, best_predictions, cols, lengths, phases, top_n)

Characterize phase transitions for all the phase permutations.

Parameters:

  • model – HMM model object

  • data (list) – List of dataframes used to train HMM model.

  • best_predictions (list) – List of list of predictions broken into respective lengths of dataframes.

  • lengths (list) – List of lengths of dataframes.

  • cols (list) – List of columns to be returned in sorted order

  • phases (_type_) – _description_

  • top_n (_type_) – _description_

Returns:

_description_

Return type:

_type_

visualize_training.utils.characterize_transition_between_phases(model, data, best_predictions, cols, lengths, i, j)

Compute the average derivative for each feature, sort features by highest absolute value

Parameters:

  • model – HMM model object

  • data (list) – List of dataframes used to train HMM model.

  • best_predictions (list) – List of list of predictions broken into respective lengths of dataframes.

  • lengths (list) – List of lengths of dataframes.

  • cols (list) – List of columns to be returned in sorted order

  • i (str) – Initial phase from where the transition is originating.

  • j (str) – Final phase where the transition is reaching.

visualize_training.utils.find_i_followed_by_j(lst, i, j)

Find transition in the estimated hidden state

Returns:

indexes of predictions where i is followed by j

Return type:

indexes

visualize_training.utils.get_derivatives(X, model)

Compute the derivative d/dz_t p(s_t = k | z_{1:t}) for the entire forward lattice.

Parameters:

  • X (dataframe) – Data used to train HMM model

  • model – HMM model object

Returns:

derivates (list)

visualize_training.utils.get_difference_bt_means(model, phase_1, phase_2)

Returns the difference between mean values of two different phases

Parameters:

  • model – Trained HMM Model

  • phase_1 (int) – Source Phase of a transition

  • phase_2 (int) – Destination phase of a transition

visualize_training.utils.get_features_for_transition(model, data, best_predictions, lengths, phase_1, phase_2)

Return Top N features responsible for a state transition.

For each time a transition (phase_1 -> phase_2) happens, compute the derivatives for each feature.

This computation is slightly inefficient, in that it computes the entire forward lattice of derivatives.

In practice, this inefficiency doesn’t seem to be an issue in terms of runtime.

Parameters:

  • model – HMM Model object

  • data (list) – List of dataframes used to train HMM model.

  • best_predictions (list) – List of list of predictions broken into respective lengths of dataframes.

  • lengths (list) – List of lengths of dataframes.

  • phase_1 (str) – Initial phase from where the transition is originating.

  • phase_2 (str) – Final phase where the transition is reaching.

Returns:

_description_

Return type:

_type_

visualize_training.utils.load_model(model_path)

Load trained HMM model from directory

Parameters:

model_path (str) – Path to the directory of the model

Returns:

HMM Model object

visualize_training.utils.make_hmm_data(data_dir, cols, sort=True, first_n=1000, sort_col='epoch')

Data Preprocessing method similar to the one present in HMM class. Separate method intended to be used independently on need basis.

Parameters:

  • data_dir (str) – Path to data files.

  • cols (list) – List of columns to be returned.

  • sort (bool) – Whether to sort the rows based on sort_col or not. Defaults to True.

  • first_n (int) – No of rows to be returned for each data file. Defaults to 1000.

  • sort_col (str) – Column name based on which sorting needs to be done. Defaults to “epoch”

Returns:

List of dataframes for training the HMM test_dfs (list): List of dataframes for testing the HMM train_data (array): All the dataframes in train_dfs stacked vertically test_data (array): All the dataframes in test_dfs stacked vertically

Return type:

train_dfs (list)

visualize_training.utils.save_model(model_path, model)

Saves trained HMM model to directory

Parameters:

  • model_path (str) – Directory path where model needs to be saved.

  • model – Trained Model object.

visualize_training.utils.softmax_with_overflow(logits)

log-sum-exp. convert logits into probabilities using softmax