Specifying dtype float32 with pandas.read_csv on pandas 0.10.1
Question:
I’m attempting to read a simple space-separated file with pandas read_csv method. However, pandas doesn’t seem to be obeying my dtype argument. Maybe I’m incorrectly specifying it?
I’ve distilled down my somewhat complicated call to read_csv to this simple test case. I’m actually using the converters argument in my ‘real’ scenario but I removed this for simplicity.
Below is my ipython session:
>>> cat test.out
a b
0.76398 0.81394
0.32136 0.91063
>>> import pandas
>>> import numpy
>>> x = pandas.read_csv('test.out', dtype={'a': numpy.float32}, delim_whitespace=True)
>>> x
a b
0 0.76398 0.81394
1 0.32136 0.91063
>>> x.a.dtype
dtype('float64')
I’ve also tried this using this with a dtype of numpy.int32 or numpy.int64. These choices result in an exception:
AttributeError: 'NoneType' object has no attribute 'dtype'
I’m assuming the AttributeError is because pandas will not automatically try to convert/truncate the float values into an integer?
I’m running on a 32-bit machine with a 32-bit version of Python.
>>> !uname -a
Linux ubuntu 3.0.0-13-generic #22-Ubuntu SMP Wed Nov 2 13:25:36 UTC 2011 i686 i686 i386 GNU/Linux
>>> import platform
>>> platform.architecture()
('32bit', 'ELF')
>>> pandas.__version__
'0.10.1'
Answers:
0.10.1 doesn’t really support float32 very much
see this http://pandas.pydata.org/pandas-docs/dev/whatsnew.html#dtype-specification
you can do this in 0.11 like this:
# dont' use dtype converters explicity for the columns you care about
# they will be converted to float64 if possible, or object if they cannot
df = pd.read_csv('test.csv'.....)
#### this is optional and related to the issue you posted ####
# force anything that is not a numeric to nan
# columns are the list of columns that you are interesetd in
df[columns] = df[columns].convert_objects(convert_numeric=True)
# astype
df[columns] = df[columns].astype('float32')
see http://pandas.pydata.org/pandas-docs/dev/basics.html#object-conversion
Its not as efficient as doing it directly in read_csv (but that requires
some low-level changes)
I have confirmed that with 0.11-dev, this DOES work (on 32-bit and 64-bit, results are the same)
In [5]: x = pd.read_csv(StringIO.StringIO(data), dtype={'a': np.float32}, delim_whitespace=True)
In [6]: x
Out[6]:
a b
0 0.76398 0.81394
1 0.32136 0.91063
In [7]: x.dtypes
Out[7]:
a float32
b float64
dtype: object
In [8]: pd.__version__
Out[8]: '0.11.0.dev-385ff82'
In [9]: quit()
vagrant@precise32:~/pandas$ uname -a
Linux precise32 3.2.0-23-generic-pae #36-Ubuntu SMP Tue Apr 10 22:19:09 UTC 2012 i686 i686 i386 GNU/Linux
In [22]: df.a.dtype = pd.np.float32
In [23]: df.a.dtype
Out[23]: dtype('float32')
the above works fine for me under pandas 0.10.1
I’m attempting to read a simple space-separated file with pandas read_csv method. However, pandas doesn’t seem to be obeying my dtype argument. Maybe I’m incorrectly specifying it?
I’ve distilled down my somewhat complicated call to read_csv to this simple test case. I’m actually using the converters argument in my ‘real’ scenario but I removed this for simplicity.
Below is my ipython session:
>>> cat test.out
a b
0.76398 0.81394
0.32136 0.91063
>>> import pandas
>>> import numpy
>>> x = pandas.read_csv('test.out', dtype={'a': numpy.float32}, delim_whitespace=True)
>>> x
a b
0 0.76398 0.81394
1 0.32136 0.91063
>>> x.a.dtype
dtype('float64')
I’ve also tried this using this with a dtype of numpy.int32 or numpy.int64. These choices result in an exception:
AttributeError: 'NoneType' object has no attribute 'dtype'
I’m assuming the AttributeError is because pandas will not automatically try to convert/truncate the float values into an integer?
I’m running on a 32-bit machine with a 32-bit version of Python.
>>> !uname -a
Linux ubuntu 3.0.0-13-generic #22-Ubuntu SMP Wed Nov 2 13:25:36 UTC 2011 i686 i686 i386 GNU/Linux
>>> import platform
>>> platform.architecture()
('32bit', 'ELF')
>>> pandas.__version__
'0.10.1'
0.10.1 doesn’t really support float32 very much
see this http://pandas.pydata.org/pandas-docs/dev/whatsnew.html#dtype-specification
you can do this in 0.11 like this:
# dont' use dtype converters explicity for the columns you care about
# they will be converted to float64 if possible, or object if they cannot
df = pd.read_csv('test.csv'.....)
#### this is optional and related to the issue you posted ####
# force anything that is not a numeric to nan
# columns are the list of columns that you are interesetd in
df[columns] = df[columns].convert_objects(convert_numeric=True)
# astype
df[columns] = df[columns].astype('float32')
see http://pandas.pydata.org/pandas-docs/dev/basics.html#object-conversion
Its not as efficient as doing it directly in read_csv (but that requires
some low-level changes)
I have confirmed that with 0.11-dev, this DOES work (on 32-bit and 64-bit, results are the same)
In [5]: x = pd.read_csv(StringIO.StringIO(data), dtype={'a': np.float32}, delim_whitespace=True)
In [6]: x
Out[6]:
a b
0 0.76398 0.81394
1 0.32136 0.91063
In [7]: x.dtypes
Out[7]:
a float32
b float64
dtype: object
In [8]: pd.__version__
Out[8]: '0.11.0.dev-385ff82'
In [9]: quit()
vagrant@precise32:~/pandas$ uname -a
Linux precise32 3.2.0-23-generic-pae #36-Ubuntu SMP Tue Apr 10 22:19:09 UTC 2012 i686 i686 i386 GNU/Linux
In [22]: df.a.dtype = pd.np.float32
In [23]: df.a.dtype
Out[23]: dtype('float32')
the above works fine for me under pandas 0.10.1