A simple function to generate variable window sequences for LSTMs

lstm_utils.py

def _generate_start_and_end(num_elements, sequence_length):
    start_gen = chain(repeat(0, times=sequence_length), 
                      range(1, num_elements))
    end_gen = chain(range(1, sequence_length), 
                    range(sequence_length, num_elements))
    for start, stop in zip_longest(start_gen, end_gen, fillvalue=num_elements):
        yield start, stop

        
def generate_variable_window(timeseries_array, sequence_length):
    """ Generates a rolling window of the timeseries passed in
    
    The window will initially expand until it reaches the value specified by sequence_length then 
    it will continue to deliver sequences of length sequence_length from timeseries_array. 
    When it reaches the end of the timeseries_array it will shrink until it runs out of elements.
    
    
    Parameters
    ----------
    timeseries_array: 2D Numpy array where the columns are the features and the rows the samples
    sequence_length: integer representing the maximum size we want the sequence to grow to  
    
    
    """
    num_elements = timeseries_array.shape[0]
    for start, stop in _generate_start_and_end(num_elements, sequence_length):
        yield timeseries_array[start:stop, :]