What is the proper way of converting integer dates to datetime64 in numpy? I tried:
import numpy
a = numpy.array([20090913, 20101020, 20110125])
numpy.datetime64(a.astype("S8"))
but get an incorrect conversion. How about reading them in correctly as numpy.datetime64 objects using numpy.loadtxt (they are coming from a csv file)?
You problem is that datetime64 expects a string in the format yyyy-mm-dd, while the type conversion produces strings in the format yyyymmdd. I would suggest something like this:
conversion = lambda x: "%s-%s-%s" % (x[:4], x[4:6], x[6:])
np_conversion = numpy.frompyfunc(conversion,1,1)
b = np_conversion(a.astype('S10'))
numpy.datetime64(b)
However it's not working for me (I have numpy 1.6.1), it fails with the message "NotImplementedError: Not implemented for this type". Unless that is implemented in 1.7, I can only suggest a pure Python solution:
numpy.datetime64(numpy.array([conversion(str(x)) for x in a], dtype="S10"))
...or pre-processing your input, to deliver the dates in the expected format.
Edit: I can also offer an alternative solution, using vectorize, but I don't know very well how it works, so I don't know what's going wrong:
>>> conversion = vectorize(lambda x: "%s-%s-%s" % (x[:4], x[4:6], x[6:]), otypes=['S10'])
>>> conversion(a.astype('S10'))
array(['2009', '2010', '2011'],
dtype='|S4')
For some reason it's ignoring the otypes and outputting |S4 instead of |S10. Sorry I can't help more, but this should provide a starting point for searching other solutions.
Update: Thanks to OP feedback, I thought of a new possibility. This should work as expected:
>>> conversion = lambda x: numpy.datetime64(str(x))
>>> np_conversion = numpy.frompyfunc(conversion, 1, 1)
>>> np_conversion(a)
array([2009-09-13 00:00:00, 2010-10-20 00:00:00, 2011-01-25 00:00:00], dtype=object)
# Works too:
>>> conversion = lambda x: numpy.datetime64("%s-%s-%s" % (x/10000, x/100%100, x%100))
Weird how, in this case, datetime64 works fine with or without the dashes...
Oddly, this works: numpy.datetime64(a.astype("S8").tolist()), while this does not: numpy.datetime64(a.astype("S8")). The first method is still a bit less convoluted than: numpy.array([numpy.datetime64(str(i)) for i in a]). I asked why in this question.
