condition_polytomous_response

def condition_polytomous_response(dataset, trim_ends=True, _reference=1.0):
    """ Recodes polytomous responses into linear indices.

    Takes an input array of ordinal values and converts it into
    an array of linear indices to access difficulty parameters through 
    fancy indexing. 

    Args:
        dataset:  [n_items x n_takers] 2d array of ordinal responses
        trim_ends:  (boolean) trims responses that are either all no or all yes

    Returns:
        dataset: (2d array) ordinal values converted to linear indices
        beta_length: (1d array) number of unique values per item
    """
    # Remove all no / yes endorsements
    min_value, max_value = dataset.min(), dataset.max()
    n_items = dataset.shape[0]

    if trim_ends:
        raw_score_sums = dataset.sum(0)
        mask = ((raw_score_sums == (n_items * min_value)) | 
                (raw_score_sums == (n_items * max_value)))
        dataset = dataset[:, ~mask]
    
    betas_length = np.zeros((n_items,), dtype='int')
    the_output = dataset.copy()
    the_output -= min_value
    
    # Loop over rows, determine the number of unique
    # responses, and replace with linear indexing
    cnt = 0
    for ndx, item in enumerate(the_output):
        values, indices = np.unique(item, return_inverse=True)
        betas_length[ndx] = values.size

        # Recode from zero to N-1
        values = np.arange(0, betas_length[ndx]) + cnt * _reference
        the_output[ndx] = values[indices]

        # Update linear index
        cnt += betas_length[ndx]

    return the_output, betas_length