Parsing a data matrix containing HH:MM:SS.mmm times using numpy.loadtxt

Question

You could use the converter parameter to apply a function to each string in the first column. Calling a Python function once for each row may slow down np.loadtxt considerably, but this might still be a workable solution for moderate-sized files:

import numpy as np

def parse_date(datestr):
    return sum([multiplier*val for multiplier, val in
                zip((3600, 60, 1), map(float, datestr.split(':')))])


x = np.loadtxt('data', dtype={'names': ('time', 'magnitude'), 'formats': ('f8', 'f8')},
               converters={0:parse_date})
print(x)

Alternatively, you could parse the strings into floats after using loadtxt like this:

x = np.loadtxt('data', dtype={'names': ('time', 'magnitude'), 'formats': ('S12', 'f8')})
arr = np.char.split(x['time'], ':')
# http://stackoverflow.com/a/19459439/190597 (Jaime)
newarr = np.fromiter((tuple(row) for row in arr), dtype=[('', np.float)]*3,
                     count=len(arr)).view('float').reshape(-1, 3)
times = (newarr * [3600,60,1]).sum(axis=1)

y = np.empty_like(x, dtype={'names': ('time', 'magnitude'), 'formats': ('f8', 'f8')})
y['time'] = times
y['magnitude'] = x['magnitude']
print(y)

Edit: I created a test files of 10**6 lines to test which method is faster. The second method is a bit faster:

In [329]: %timeit using_fromiter()
1 loops, best of 3: 5.59 s per loop


In [328]: %timeit using_converter()
1 loops, best of 3: 6.88 s per loop

import os
import numpy as np

def create_data(N):
    data = np.random.random(size=N)*86400
    hours, remainder = data.__divmod__(3600)
    minutes, seconds = remainder.__divmod__(60)
    mag = np.arange(N)
    filename = os.path.expanduser('~/tmp/data')
    with open(filename, 'w') as f:
        for h,m,s,a in np.column_stack([hours, minutes, seconds, mag]):
            f.write('{h:d}:{m:d}:{s:.6f} {a}\n'.format(h=int(h), m=int(m), s=s, a=a))

def parse_date(datestr):
    return sum([multiplier*val for multiplier, val in
                zip((3600, 60, 1), map(float, datestr.split(':')))])

def using_converter():
    x = np.loadtxt('data', dtype={'names': ('time', 'magnitude'),
                                  'formats': ('f8', 'f8')},
                   converters={0:parse_date})
    return x

def using_fromiter():
    x = np.loadtxt('data', dtype={'names': ('time', 'magnitude'), 'formats': ('S12', 'f8')})
    arr = np.char.split(x['time'], ':')
    newarr = np.fromiter((tuple(row) for row in arr), dtype=[('', np.float)]*3,
                         count=len(arr)).view('float').reshape(-1, 3)
    times = (newarr * [3600,60,1]).sum(axis=1)

    y = np.empty_like(x, dtype={'names': ('time', 'magnitude'), 'formats': ('f8', 'f8')})
    y['time'] = times
    y['magnitude'] = x['magnitude']
    return y

create_data(10**6)