time_series_window.py

import pandas as pd
import numpy as np
import time
import random

'''
This class will return a random time series based window of data. It intakes a dataframe
that is already pre sorted into the correct time pattern where index 0 is timestep 0
and index N is timestep N. The function window_maker will incrementally step through
the entire dataframe returning the next window of data. The function random_window_maker
will return a random window of data that is time continuous. All outputs are numpy arrays.
'''

class time_series_windows(object):
    def __init__(self, dataframe, window_size):
        self.df = dataframe
        self.last_rec_start = 0
        self.window_size = window_size
        self.row_stop = self.last_rec_start + self.window_size
        self.batch = np.zeros((0, window_size*dataframe.shape[1]))
        self.batch_counter = 0

    def window_maker(self):
        if self.row_stop <= self.df.shape[0]:
            record = self.df.loc[self.last_rec_start:self.row_stop-1].as_matrix().flatten()
            if len(record) == self.window_size*self.df.shape[1]:
                self.batch_counter += 1
                self.last_rec_start += 1
                self.row_stop += 1
                return record
        else:
            self.batch_counter = 0
            self.last_rec_start = 0
            self.row_stop = self.last_rec_start + self.window_size
            return self.window_maker()

    def make_batch(self, batch_size):
        batch_start_point = random.randrange(0, self.df.shape[0]-batch_size)
        self.batch_df = self.df.iloc[batch_start_point:,:].reset_index(drop=True)
        self.window_batch = np.empty((0,self.df.shape[1]*self.window_size), float)
        for i in range(batch_size):
            self.window_batch = np.append(self.window_batch,
                                          self.window_maker().reshape((1, self.df.shape[1]*self.window_size)), axis=0)
        return self.window_batch

#
    def random_window_maker(self):

        _ = False
        while _ is False:
            random_start = random.randrange(0, self.df.shape[0])
            self.last_rec_start = random_start
            self.row_stop = self.last_rec_start + self.window_size
            if self.row_stop <= self.df.shape[0]:
                record = self.df.loc[self.last_rec_start:self.row_stop - 1].as_matrix().flatten()
                _ = True
                return record


if __name__ == "__main__":
    df = pd.DataFrame(np.arange(0.0, 160.0).reshape((80, 2)))
    df['target'] = ['target' + str(i) for i in range(80)]
    # print(df.shape)
    window_size = 3
    win = time_series_windows(df, window_size)
    for i in range(10):
        data = win.window_maker()
        X = data[:-1]
        y = data[-1]
        print(i,'X out:', X)
        print(i, 'X out:', y)
    print('out:', win.random_window_maker())

    print(win.make_batch(8))