The challenge is to save not only the data buffer, but also the shape and dtype. np.fromstring reads the data buffer, but as a 1d array; you have to get the dtype and shape from else where.
In [184]: a=np.arange(12).reshape(3,4)
In [185]: np.fromstring(a.tostring(),int)
Out[185]: array([ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11])
In [186]: np.fromstring(a.tostring(),a.dtype).reshape(a.shape)
Out[186]:
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
A time honored mechanism to save Python objects is pickle, and numpy is pickle compliant:
In [169]: import pickle
In [170]: a=np.arange(12).reshape(3,4)
In [171]: s=pickle.dumps(a*2)
In [172]: s
Out[172]: "cnumpy.core.multiarray\n_reconstruct\np0\n(cnumpy\nndarray\np1\n(I0\ntp2\nS'b'\np3\ntp4\nRp5\n(I1\n(I3\nI4\ntp6\ncnumpy\ndtype\np7\n(S'i4'\np8\nI0\nI1\ntp9\nRp10\n(I3\nS'<'\np11\nNNNI-1\nI-1\nI0\ntp12\nbI00\nS'\\x00\\x00\\x00\\x00\\x02\\x00\\x00\\x00\\x04\\x00\\x00\\x00\\x06\\x00\\x00\\x00\\x08\\x00\\x00\\x00\\n\\x00\\x00\\x00\\x0c\\x00\\x00\\x00\\x0e\\x00\\x00\\x00\\x10\\x00\\x00\\x00\\x12\\x00\\x00\\x00\\x14\\x00\\x00\\x00\\x16\\x00\\x00\\x00'\np13\ntp14\nb."
In [173]: pickle.loads(s)
Out[173]:
array([[ 0, 2, 4, 6],
[ 8, 10, 12, 14],
[16, 18, 20, 22]])
There’s a numpy function that can read the pickle string:
In [181]: np.loads(s)
Out[181]:
array([[ 0, 2, 4, 6],
[ 8, 10, 12, 14],
[16, 18, 20, 22]])
You mentioned np.save to a string, but that you can’t use np.load. A way around that is to step further into the code, and use np.lib.npyio.format.
In [174]: import StringIO
In [175]: S=StringIO.StringIO() # a file like string buffer
In [176]: np.lib.npyio.format.write_array(S,a*3.3)
In [177]: S.seek(0) # rewind the string
In [178]: np.lib.npyio.format.read_array(S)
Out[178]:
array([[ 0. , 3.3, 6.6, 9.9],
[ 13.2, 16.5, 19.8, 23.1],
[ 26.4, 29.7, 33. , 36.3]])
The save string has a header with dtype and shape info:
In [179]: S.seek(0)
In [180]: S.readlines()
Out[180]:
["\x93NUMPY\x01\x00F\x00{'descr': '<f8', 'fortran_order': False, 'shape': (3, 4), } \n",
'\x00\x00\x00\x00\x00\x00\x00\x00ffffff\n',
'@ffffff\x1a@\xcc\xcc\xcc\xcc\xcc\xcc#@ffffff*@\x00\x00\x00\x00\x00\x800@\xcc\xcc\xcc\xcc\xcc\xcc3@\x99\x99\x99\x99\x99\x197@ffffff:@33333\xb3=@\x00\x00\x00\x00\x00\x80@@fffff&B@']
If you want a human readable string, you might try json.
In [196]: import json
In [197]: js=json.dumps(a.tolist())
In [198]: js
Out[198]: '[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]]'
In [199]: np.array(json.loads(js))
Out[199]:
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
Going to/from the list representation of the array is the most obvious use of json. Someone may have written a more elaborate json representation of arrays.
You could also go the csv format route – there have been lots of questions about reading/writing csv arrays.
'[[ 0.5544 0.4456], [ 0.8811 0.1189]]'
is a poor string representation for this purpose. It does look a lot like the str() of an array, but with , instead of \n. But there isn’t a clean way of parsing the nested [], and the missing delimiter is a pain. If it consistently uses , then json can convert it to list.
np.matrix accepts a MATLAB like string:
In [207]: np.matrix(' 0.5544, 0.4456;0.8811, 0.1189')
Out[207]:
matrix([[ 0.5544, 0.4456],
[ 0.8811, 0.1189]])
In [208]: str(np.matrix(' 0.5544, 0.4456;0.8811, 0.1189'))
Out[208]: '[[ 0.5544 0.4456]\n [ 0.8811 0.1189]]'