support for string types in array2tree

[maxnoe]

Support for numpy string dtypes would be great.

[ndawe]

Do you mean for the conversion of numpy arrays into TTrees?

[maxnoe]

Yes, most important for me would be the fixed length string types like numpy.dtype('S10'),
to save something like datestamps into the rootfiles.

This would ownly need conversion from numpy arrays with dtype(Sx) to char* arrays with char[x+1] entries.

[ndawe]

Yes, this is certainly possible. I'll start implementing it. Another solution in the meantime would be to convert the date string into an integer timestamp like http://stackoverflow.com/a/9637908 and array2root or array2tree will convert that.

[maxnoe]

Thats what I'm doing right now, but there are some things better described as String (source names etc).

Great!

[ndawe]

Preliminary code is implemented in c26bd7c. Can you give this a try? I'm planning on extending that to handle fixed-length arrays of all basic types.

[ndawe]

Install from my master branch:

pip install --upgrade --user https://github.com/ndawe/root_numpy/zipball/master

[maxnoe]

Writing works, thanks for that.

But reading is not working properly, the strings are converted into a list of integers,
the ascii codes of the letters to be precise.

import numpy as np
import root_numpy as rnp


data = np.array(['2015-01-01', '2015-01-02'], dtype=[('date', 'S10')])

rnp.array2root(data, 'test.root', treename='data', mode='recreate')

data = rnp.root2array('test.root', treename='data')

print(data)

The Result is:

[[50 48 49 53 45 48 49 45 48 49]
 [50 48 49 53 45 48 49 45 48 50]]

[ndawe]

Yes, I forgot to mention this. root_numpy treats arrays of each type "equally" so it really is an array of char bytes (np.int8). If you want to view the char array as strings then use data.view('S10'). That interprets the underlying bytes as strings without making a copy. root_numpy needs to pick one default view since the original array could be of the equivalent type np.int8, but there is no way of determining the original view used on the input data before the array2tree+tree2array cycle. So if you originally viewed the data as S, then this extra view() step is required.

[ndawe]

Extending your example:

>>> import numpy as np
>>> import root_numpy as rnp
>>> data = np.array(['2015-01-01', '2015-01-02'], dtype=[('date', 'S10')])
>>> rnp.array2root(data, 'test.root', treename='data', mode='recreate')
>>> data = rnp.root2array('test.root', treename='data')
>>> data
array([([50, 48, 49, 53, 45, 48, 49, 45, 48, 49],),
       ([50, 48, 49, 53, 45, 48, 49, 45, 48, 50],)], 
      dtype=[('date', 'i1', (10,))])
>>> data.view('S10')
array(['2015-01-01', '2015-01-02'], 
      dtype='|S10')

[maxnoe]

Why is there no way? The root datatype should provide the information. One could use the same mapping of TYPES_NUMPY2ROOT the other way around

[ndawe]

In the ROOT file np.dtype(('i1', (10,))) and np.dtype('S10') both end up as char[10] and then from char[10] alone it is not possible to determine the original view.

[maxnoe]

Oh, I see you used the 'B' root datatype, shouldn't be the 'C' datatype be more suited and solve the problem?

[ndawe]

Yes, you are right. We can use C to distinguish the two purposes. The bytes are the same either way, but C assumes null termination.

[ndawe]

Got it working. You're welcome to test for yourself:

pip install --upgrade --user https://github.com/ndawe/root_numpy/zipball/master

>>> import numpy as np
>>> import root_numpy as rnp
>>> data = np.array(['2015-01-01', '2015-01-02'], dtype=[('date', 'S10')])
>>> rnp.array2root(data, 'test.root', treename='data', mode='recreate')
>>> rnp.root2array('test.root', treename='data')
array([('2015-01-01',), ('2015-01-02',)], 
      dtype=[('date', 'S10')])