BUG: should astype only TRY to convert string columns?

[jreback]

should astype TRY to convert string columns (e.g. if the value of the string column is like '5'), or skip if its an exception (and just return it), rather than raise an exception on the whole operation?

I can fix for 0.10.2

In [4]: df = pd.DataFrame({ 'a' : 'foo', 'b' : 1. },index=np.arange(10))
In [6]: df .dtypes
Out[6]: 
a     object
b    float64

In [8]: df.astype('float64')
/mnt/home/jreback/pandas/pandas/core/internals.pyc in astype(self, dtype)
    613         new_blocks = []
    614         for block in self.blocks:
--> 615             newb = make_block(com._astype_nansafe(block.values, dtype),
    616                               block.items, block.ref_items)
    617             new_blocks.append(newb)

/mnt/home/jreback/pandas/pandas/core/common.pyc in _astype_nansafe(arr, dtype)
   1058         return lib.astype_intsafe(arr.ravel(), dtype).reshape(arr.shape)
   1059 
-> 1060     return arr.astype(dtype)
   1061 
   1062 

ValueError: could not convert string to float: foo

[changhiskhan]

I think the default behavior should stay as is, but I see nothing wrong with adding a raise keyword that is True by default and can be set to False so errors are passed silently and whatever can be converted is converted

[jreback]

ok....next question....what if you astype stuff that loses meaning...not sure what is the 'correct' behavior here
(this is using dtypes branch)

(Pdb) mn['big_float'] = 1234566789.
(Pdb) mn
   a  b  float32  int32   big_float
0  1  2        1      1  1234566789
1  1  2        1      1  1234566789
2  1  2        1      1  1234566789
3  1  2        1      1  1234566789
4  1  2        1      1  1234566789
5  1  2        1      1  1234566789
6  1  2        1      1  1234566789
7  1  2        1      1  1234566789
8  1  2        1      1  1234566789
9  1  2        1      1  1234566789
(Pdb) mn.dtypes
a            float64
b              int64
float32      float32
int32          int32
big_float    float64
(Pdb) mn.astype('uint8')
   a  b  float32  int32  big_float
0  1  2        1      1        133
1  1  2        1      1        133
2  1  2        1      1        133
3  1  2        1      1        133
4  1  2        1      1        133
5  1  2        1      1        133
6  1  2        1      1        133
7  1  2        1      1        133
8  1  2        1      1        133
9  1  2        1      1        133
(Pdb) mn.astype('uint8').dtypes
a            uint8
b            uint8
float32      uint8
int32        uint8
big_float    uint8

[jreback]

raise_on_error in astype added in #2708

[jreback]

I think I fixed this. Will raise on dtypes that change the shape of the block or itemsize (numpy basically tells us).