better quantities support

[twmr]

I want to use the quantities package and pandas to process scientific data. However, pandas strips the unit(s) of the data stored in numpy arrays if I create a dataframe out of them:

a = np.random.rand(10)*pq.s
b = np.random.rand(10)*pq.A
df = pd.DataFrame({'current':b, 't':a}, columns=['t','current'])

In [1]: df
Out[1]: 
         t   current
0  0.663397  0.435423
1  0.038498  0.101763
2  0.960983  0.091785
3  0.262863  0.364734
4  0.154440  0.274169
5  0.953129  0.052678
6  0.389961  0.272535
7  0.961604  0.559451
8  0.747192  0.438268
9  0.789207  0.568685

What do you think, should pandas have support for writing the unit of each column in the corresponding columnname if the column is a quantities array, or should it be part of the quantities package.

In [1]: df
Out[1]: 
       t [s]  current [A]
0  0.663397  0.435423
1  0.038498  0.101763
2  0.960983  0.091785
....

[changhiskhan]

related #1071

[changhiskhan]

idea solution would be #2485

note that pandas does allow arbitrary objects as data, so this works:

In [1]: import quantities as pq

In [2]: a = [x for x in np.random.rand(10)*pq.s]
   ...: b = [x for x in np.random.rand(10)*pq.A]
   ...: df = pd.DataFrame({'current':b, 't':a}, columns=['t','current'])

In [3]: df
Out[3]: 
                   t           current
0   0.947453713439 s  0.862354858891 A
1  0.0715388489745 s  0.289714978433 A
2   0.355878981889 s  0.104722774751 A
3  0.0348524553738 s  0.583777806824 A
4  0.0138462074388 s  0.696935444594 A
5  0.0618976659102 s  0.517191317629 A
6   0.280860457598 s  0.811183797669 A
7     0.7259184522 s  0.270850455923 A
8   0.148336783722 s  0.203341988353 A
9   0.619437945726 s  0.895585882586 A

Which means this works, and is the most wonderful thing I've seen
all week:

In [4]: df**2
Out[4]: 
                        t               current
0      0.89766853911 s**2   0.743655902652 A**2
1   0.00511780691259 s**2  0.0839347687284 A**2
2      0.12664984975 s**2  0.0109668595515 A**2
3   0.00121469364558 s**2    0.34079652774 A**2
4  0.000191717460437 s**2   0.485719013931 A**2
5   0.00383132104514 s**2   0.267486859031 A**2
6    0.0788825966422 s**2     0.6580191536 A**2
7     0.526957599245 s**2  0.0733599694737 A**2
8    0.0220038014049 s**2  0.0413479642275 A**2
9     0.383703368605 s**2   0.802074073088 A**2

cgs ye shall torment me no more.

not familiar with quantities internals, so this may have degraded performance
since the new type is probably calculated per datum rather then per vector operation.
still. very cool.

I'll try and create a PR with a new keyword for the dataframe constructor which would
have the constructor do the conversion from X-array to array of X's.
Might enable other packages that subclass from numpy to work as well.

edit: on second thought, not generally useful.

[mikofski]

+1, and/or Pint support as well. I will contribute as I can

[jreback]

closing in favor of #10349

[5igno]

Just as a comment, the approach suggested here leaves room to the possibility of having quantities that have inconsistent values from one row to the next. The measurement unit should be something that describes the content of each column, as e.g. some column metadata.