IO tools (text, CSV, HDF5, …)¶

Basic¶

filepath_or_buffervarious

Either a path to a file (a str, pathlib.Path, or py:py._path.local.LocalPath), URL (including http, ftp, and S3 locations), or any object with a read() method (such as an open file or StringIO).

sepstr, defaults to ',' for read_csv(), \t for read_table()

Delimiter to use. If sep is None, the C engine cannot automatically detect the separator, but the Python parsing engine can, meaning the latter will be used and automatically detect the separator by Python’s builtin sniffer tool, csv.Sniffer. In addition, separators longer than 1 character and different from '\s+' will be interpreted as regular expressions and will also force the use of the Python parsing engine. Note that regex delimiters are prone to ignoring quoted data. Regex example: '\\r\\t'.

delimiterstr, default None

Alternative argument name for sep.

delim_whitespaceboolean, default False

Specifies whether or not whitespace (e.g. ' ' or '\t') will be used as the delimiter. Equivalent to setting sep='\s+'. If this option is set to True, nothing should be passed in for the delimiter parameter.

Column and index locations and names¶

headerint or list of ints, default 'infer'

Row number(s) to use as the column names, and the start of the data. Default behavior is to infer the column names: if no names are passed the behavior is identical to header=0 and column names are inferred from the first line of the file, if column names are passed explicitly then the behavior is identical to header=None. Explicitly pass header=0 to be able to replace existing names.

The header can be a list of ints that specify row locations for a MultiIndex on the columns e.g. [0,1,3]. Intervening rows that are not specified will be skipped (e.g. 2 in this example is skipped). Note that this parameter ignores commented lines and empty lines if skip_blank_lines=True, so header=0 denotes the first line of data rather than the first line of the file.

namesarray-like, default None

List of column names to use. If file contains no header row, then you should explicitly pass header=None. Duplicates in this list are not allowed.

index_colint, str, sequence of int / str, or False, optional, default None

Column(s) to use as the row labels of the DataFrame, either given as string name or column index. If a sequence of int / str is given, a MultiIndex is used.

Note: index_col=False can be used to force pandas to not use the first column as the index, e.g. when you have a malformed file with delimiters at the end of each line.

The default value of None instructs pandas to guess. If the number of fields in the column header row is equal to the number of fields in the body of the data file, then a default index is used. If it is larger, then the first columns are used as index so that the remaining number of fields in the body are equal to the number of fields in the header.

The first row after the header is used to determine the number of columns, which will go into the index. If the subsequent rows contain less columns than the first row, they are filled with NaN.

This can be avoided through usecols. This ensures that the columns are taken as is and the trailing data are ignored.

usecolslist-like or callable, default None

Return a subset of the columns. If list-like, all elements must either be positional (i.e. integer indices into the document columns) or strings that correspond to column names provided either by the user in names or inferred from the document header row(s). If names are given, the document header row(s) are not taken into account. For example, a valid list-like usecols parameter would be [0, 1, 2] or ['foo', 'bar', 'baz'].

Element order is ignored, so usecols=[0, 1] is the same as [1, 0]. To instantiate a DataFrame from data with element order preserved use pd.read_csv(data, usecols=['foo', 'bar'])[['foo', 'bar']] for columns in ['foo', 'bar'] order or pd.read_csv(data, usecols=['foo', 'bar'])[['bar', 'foo']] for ['bar', 'foo'] order.

If callable, the callable function will be evaluated against the column names, returning names where the callable function evaluates to True:

In [1]: import pandas as pd

In [2]: from io import StringIO

In [3]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"

In [4]: pd.read_csv(StringIO(data))
Out[4]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3

In [5]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["COL1", "COL3"])
Out[5]: 
  col1  col3
0    a     1
1    a     2
2    c     3

Using this parameter results in much faster parsing time and lower memory usage when using the c engine. The Python engine loads the data first before deciding which columns to drop.

squeezeboolean, default False

If the parsed data only contains one column then return a Series.

Deprecated since version 1.4.0: Append .squeeze("columns") to the call to {func_name} to squeeze the data.

prefixstr, default None

Prefix to add to column numbers when no header, e.g. ‘X’ for X0, X1, …

Deprecated since version 1.4.0: Use a list comprehension on the DataFrame’s columns after calling read_csv.

In [6]: data = "col1,col2,col3\na,b,1"

In [7]: df = pd.read_csv(StringIO(data))

In [8]: df.columns = [f"pre_{col}" for col in df.columns]

In [9]: df
Out[9]: 
  pre_col1 pre_col2  pre_col3
0        a        b         1

mangle_dupe_colsboolean, default True

Duplicate columns will be specified as ‘X’, ‘X.1’…’X.N’, rather than ‘X’…’X’. Passing in False will cause data to be overwritten if there are duplicate names in the columns.

General parsing configuration¶

dtypeType name or dict of column -> type, default None

Data type for data or columns. E.g. {'a': np.float64, 'b': np.int32} (unsupported with engine='python'). Use str or object together with suitable na_values settings to preserve and not interpret dtype.

engine{'c', 'python', 'pyarrow'}

Parser engine to use. The C and pyarrow engines are faster, while the python engine is currently more feature-complete. Multithreading is currently only supported by the pyarrow engine.

New in version 1.4.0: The “pyarrow” engine was added as an experimental engine, and some features are unsupported, or may not work correctly, with this engine.

convertersdict, default None

Dict of functions for converting values in certain columns. Keys can either be integers or column labels.

true_valueslist, default None

Values to consider as True.

false_valueslist, default None

Values to consider as False.

skipinitialspaceboolean, default False

Skip spaces after delimiter.

skiprowslist-like or integer, default None

Line numbers to skip (0-indexed) or number of lines to skip (int) at the start of the file.

If callable, the callable function will be evaluated against the row indices, returning True if the row should be skipped and False otherwise:

In [10]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"

In [11]: pd.read_csv(StringIO(data))
Out[11]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3

In [12]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0)
Out[12]: 
  col1 col2  col3
0    a    b     2

skipfooterint, default 0

Number of lines at bottom of file to skip (unsupported with engine=’c’).

nrowsint, default None

Number of rows of file to read. Useful for reading pieces of large files.

low_memoryboolean, default True

Internally process the file in chunks, resulting in lower memory use while parsing, but possibly mixed type inference. To ensure no mixed types either set False, or specify the type with the dtype parameter. Note that the entire file is read into a single DataFrame regardless, use the chunksize or iterator parameter to return the data in chunks. (Only valid with C parser)

memory_mapboolean, default False

If a filepath is provided for filepath_or_buffer, map the file object directly onto memory and access the data directly from there. Using this option can improve performance because there is no longer any I/O overhead.

NA and missing data handling¶

na_valuesscalar, str, list-like, or dict, default None

Additional strings to recognize as NA/NaN. If dict passed, specific per-column NA values. See na values const below for a list of the values interpreted as NaN by default.

keep_default_naboolean, default True

Whether or not to include the default NaN values when parsing the data. Depending on whether na_values is passed in, the behavior is as follows:

• If keep_default_na is True, and na_values are specified, na_values is appended to the default NaN values used for parsing.

• If keep_default_na is True, and na_values are not specified, only the default NaN values are used for parsing.

• If keep_default_na is False, and na_values are specified, only the NaN values specified na_values are used for parsing.

• If keep_default_na is False, and na_values are not specified, no strings will be parsed as NaN.

Note that if na_filter is passed in as False, the keep_default_na and na_values parameters will be ignored.

na_filterboolean, default True

Detect missing value markers (empty strings and the value of na_values). In data without any NAs, passing na_filter=False can improve the performance of reading a large file.

verboseboolean, default False

Indicate number of NA values placed in non-numeric columns.

skip_blank_linesboolean, default True

If True, skip over blank lines rather than interpreting as NaN values.

Datetime handling¶

parse_datesboolean or list of ints or names or list of lists or dict, default False.

• If True -> try parsing the index.

• If [1, 2, 3] -> try parsing columns 1, 2, 3 each as a separate date column.

• If [[1, 3]] -> combine columns 1 and 3 and parse as a single date column.

• If {'foo': [1, 3]} -> parse columns 1, 3 as date and call result ‘foo’. A fast-path exists for iso8601-formatted dates.

infer_datetime_formatboolean, default False

If True and parse_dates is enabled for a column, attempt to infer the datetime format to speed up the processing.

keep_date_colboolean, default False

If True and parse_dates specifies combining multiple columns then keep the original columns.

date_parserfunction, default None

Function to use for converting a sequence of string columns to an array of datetime instances. The default uses dateutil.parser.parser to do the conversion. pandas will try to call date_parser in three different ways, advancing to the next if an exception occurs: 1) Pass one or more arrays (as defined by parse_dates) as arguments; 2) concatenate (row-wise) the string values from the columns defined by parse_dates into a single array and pass that; and 3) call date_parser once for each row using one or more strings (corresponding to the columns defined by parse_dates) as arguments.

dayfirstboolean, default False

DD/MM format dates, international and European format.

cache_datesboolean, default True

If True, use a cache of unique, converted dates to apply the datetime conversion. May produce significant speed-up when parsing duplicate date strings, especially ones with timezone offsets.

New in version 0.25.0.

Iteration¶

iteratorboolean, default False

Return TextFileReader object for iteration or getting chunks with get_chunk().

chunksizeint, default None

Return TextFileReader object for iteration. See iterating and chunking below.

Quoting, compression, and file format¶

compression{'infer', 'gzip', 'bz2', 'zip', 'xz', 'zstd', None, dict}, default 'infer'

For on-the-fly decompression of on-disk data. If ‘infer’, then use gzip, bz2, zip, xz, or zstandard if filepath_or_buffer is path-like ending in ‘.gz’, ‘.bz2’, ‘.zip’, ‘.xz’, ‘.zst’, respectively, and no decompression otherwise. If using ‘zip’, the ZIP file must contain only one data file to be read in. Set to None for no decompression. Can also be a dict with key 'method' set to one of {'zip', 'gzip', 'bz2', 'zstd'} and other key-value pairs are forwarded to zipfile.ZipFile, gzip.GzipFile, bz2.BZ2File, or zstandard.ZstdDecompressor. As an example, the following could be passed for faster compression and to create a reproducible gzip archive: compression={'method': 'gzip', 'compresslevel': 1, 'mtime': 1}.

Changed in version 1.1.0: dict option extended to support gzip and bz2.

Changed in version 1.2.0: Previous versions forwarded dict entries for ‘gzip’ to gzip.open.

thousandsstr, default None

Thousands separator.

decimalstr, default '.'

Character to recognize as decimal point. E.g. use ',' for European data.

float_precisionstring, default None

Specifies which converter the C engine should use for floating-point values. The options are None for the ordinary converter, high for the high-precision converter, and round_trip for the round-trip converter.

lineterminatorstr (length 1), default None

Character to break file into lines. Only valid with C parser.

quotecharstr (length 1)

The character used to denote the start and end of a quoted item. Quoted items can include the delimiter and it will be ignored.

quotingint or csv.QUOTE_* instance, default 0

Control field quoting behavior per csv.QUOTE_* constants. Use one of QUOTE_MINIMAL (0), QUOTE_ALL (1), QUOTE_NONNUMERIC (2) or QUOTE_NONE (3).

doublequoteboolean, default True

When quotechar is specified and quoting is not QUOTE_NONE, indicate whether or not to interpret two consecutive quotechar elements inside a field as a single quotechar element.

escapecharstr (length 1), default None

One-character string used to escape delimiter when quoting is QUOTE_NONE.

commentstr, default None

Indicates remainder of line should not be parsed. If found at the beginning of a line, the line will be ignored altogether. This parameter must be a single character. Like empty lines (as long as skip_blank_lines=True), fully commented lines are ignored by the parameter header but not by skiprows. For example, if comment='#', parsing ‘#empty\na,b,c\n1,2,3’ with header=0 will result in ‘a,b,c’ being treated as the header.

encodingstr, default None

Encoding to use for UTF when reading/writing (e.g. 'utf-8'). List of Python standard encodings.

dialectstr or csv.Dialect instance, default None

If provided, this parameter will override values (default or not) for the following parameters: delimiter, doublequote, escapechar, skipinitialspace, quotechar, and quoting. If it is necessary to override values, a ParserWarning will be issued. See csv.Dialect documentation for more details.

Error handling¶

error_bad_linesboolean, optional, default None

Lines with too many fields (e.g. a csv line with too many commas) will by default cause an exception to be raised, and no DataFrame will be returned. If False, then these “bad lines” will dropped from the DataFrame that is returned. See bad lines below.

Deprecated since version 1.3.0: The on_bad_lines parameter should be used instead to specify behavior upon encountering a bad line instead.

warn_bad_linesboolean, optional, default None

If error_bad_lines is False, and warn_bad_lines is True, a warning for each “bad line” will be output.

Deprecated since version 1.3.0: The on_bad_lines parameter should be used instead to specify behavior upon encountering a bad line instead.

on_bad_lines(‘error’, ‘warn’, ‘skip’), default ‘error’

Specifies what to do upon encountering a bad line (a line with too many fields). Allowed values are :

• ‘error’, raise an ParserError when a bad line is encountered.

• ‘warn’, print a warning when a bad line is encountered and skip that line.

• ‘skip’, skip bad lines without raising or warning when they are encountered.

New in version 1.3.0.

Handling column names¶

A file may or may not have a header row. pandas assumes the first row should be used as the column names:

In [50]: data = "a,b,c\n1,2,3\n4,5,6\n7,8,9"

In [51]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [52]: pd.read_csv(StringIO(data))
Out[52]: 
   a  b  c
0  1  2  3
1  4  5  6
2  7  8  9

By specifying the names argument in conjunction with header you can indicate other names to use and whether or not to throw away the header row (if any):

In [53]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9

In [54]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=0)
Out[54]: 
   foo  bar  baz
0    1    2    3
1    4    5    6
2    7    8    9

In [55]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=None)
Out[55]: 
  foo bar baz
0   a   b   c
1   1   2   3
2   4   5   6
3   7   8   9

If the header is in a row other than the first, pass the row number to header. This will skip the preceding rows:

In [56]: data = "skip this skip it\na,b,c\n1,2,3\n4,5,6\n7,8,9"

In [57]: pd.read_csv(StringIO(data), header=1)
Out[57]: 
   a  b  c
0  1  2  3
1  4  5  6
2  7  8  9

Filtering columns (usecols)¶

The usecols argument allows you to select any subset of the columns in a file, either using the column names, position numbers or a callable:

In [60]: data = "a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz"

In [61]: pd.read_csv(StringIO(data))
Out[61]: 
   a  b  c    d
0  1  2  3  foo
1  4  5  6  bar
2  7  8  9  baz

In [62]: pd.read_csv(StringIO(data), usecols=["b", "d"])
Out[62]: 
   b    d
0  2  foo
1  5  bar
2  8  baz

In [63]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])
Out[63]: 
   a  c    d
0  1  3  foo
1  4  6  bar
2  7  9  baz

In [64]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["A", "C"])
Out[64]: 
   a  c
0  1  3
1  4  6
2  7  9

The usecols argument can also be used to specify which columns not to use in the final result:

In [65]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ["a", "c"])
Out[65]: 
   b    d
0  2  foo
1  5  bar
2  8  baz

In this case, the callable is specifying that we exclude the “a” and “c” columns from the output.

Ignoring line comments and empty lines¶

If the comment parameter is specified, then completely commented lines will be ignored. By default, completely blank lines will be ignored as well.

In [66]: data = "\na,b,c\n  \n# commented line\n1,2,3\n\n4,5,6"

In [67]: print(data)

a,b,c
  
# commented line
1,2,3

4,5,6

In [68]: pd.read_csv(StringIO(data), comment="#")
Out[68]: 
   a  b  c
0  1  2  3
1  4  5  6

If skip_blank_lines=False, then read_csv will not ignore blank lines:

In [69]: data = "a,b,c\n\n1,2,3\n\n\n4,5,6"

In [70]: pd.read_csv(StringIO(data), skip_blank_lines=False)
Out[70]: 
     a    b    c
0  NaN  NaN  NaN
1  1.0  2.0  3.0
2  NaN  NaN  NaN
3  NaN  NaN  NaN
4  4.0  5.0  6.0

In [71]: data = "#comment\na,b,c\nA,B,C\n1,2,3"

In [72]: pd.read_csv(StringIO(data), comment="#", header=1)
Out[72]: 
   A  B  C
0  1  2  3

In [73]: data = "A,B,C\n#comment\na,b,c\n1,2,3"

In [74]: pd.read_csv(StringIO(data), comment="#", skiprows=2)
Out[74]: 
   a  b  c
0  1  2  3

In [75]: data = (
   ....:     "# empty\n"
   ....:     "# second empty line\n"
   ....:     "# third emptyline\n"
   ....:     "X,Y,Z\n"
   ....:     "1,2,3\n"
   ....:     "A,B,C\n"
   ....:     "1,2.,4.\n"
   ....:     "5.,NaN,10.0\n"
   ....: )
   ....: 

In [76]: print(data)
# empty
# second empty line
# third emptyline
X,Y,Z
1,2,3
A,B,C
1,2.,4.
5.,NaN,10.0


In [77]: pd.read_csv(StringIO(data), comment="#", skiprows=4, header=1)
Out[77]: 
     A    B     C
0  1.0  2.0   4.0
1  5.0  NaN  10.0

Comments¶

Sometimes comments or meta data may be included in a file:

In [78]: print(open("tmp.csv").read())
ID,level,category
Patient1,123000,x # really unpleasant
Patient2,23000,y # wouldn't take his medicine
Patient3,1234018,z # awesome

By default, the parser includes the comments in the output:

In [79]: df = pd.read_csv("tmp.csv")

In [80]: df
Out[80]: 
         ID    level                        category
0  Patient1   123000           x # really unpleasant
1  Patient2    23000  y # wouldn't take his medicine
2  Patient3  1234018                     z # awesome

We can suppress the comments using the comment keyword:

In [81]: df = pd.read_csv("tmp.csv", comment="#")

In [82]: df
Out[82]: 
         ID    level category
0  Patient1   123000       x 
1  Patient2    23000       y 
2  Patient3  1234018       z 

Specifying date columns¶

To better facilitate working with datetime data, read_csv() uses the keyword arguments parse_dates and date_parser to allow users to specify a variety of columns and date/time formats to turn the input text data into datetime objects.

The simplest case is to just pass in parse_dates=True:

In [101]: f = open("foo.csv", "w")

In [102]: f.write("date,A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5")
Out[102]: 55

In [103]: f.close()

# Use a column as an index, and parse it as dates.
In [104]: df = pd.read_csv("foo.csv", index_col=0, parse_dates=True)

In [105]: df
Out[105]: 
            A  B  C
date               
2009-01-01  a  1  2
2009-01-02  b  3  4
2009-01-03  c  4  5

# These are Python datetime objects
In [106]: df.index
Out[106]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', name='date', freq=None)

It is often the case that we may want to store date and time data separately, or store various date fields separately. the parse_dates keyword can be used to specify a combination of columns to parse the dates and/or times from.

You can specify a list of column lists to parse_dates, the resulting date columns will be prepended to the output (so as to not affect the existing column order) and the new column names will be the concatenation of the component column names:

In [107]: data = (
   .....:     "KORD,19990127, 19:00:00, 18:56:00, 0.8100\n"
   .....:     "KORD,19990127, 20:00:00, 19:56:00, 0.0100\n"
   .....:     "KORD,19990127, 21:00:00, 20:56:00, -0.5900\n"
   .....:     "KORD,19990127, 21:00:00, 21:18:00, -0.9900\n"
   .....:     "KORD,19990127, 22:00:00, 21:56:00, -0.5900\n"
   .....:     "KORD,19990127, 23:00:00, 22:56:00, -0.5900"
   .....: )
   .....: 

In [108]: with open("tmp.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [109]: df = pd.read_csv("tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]])

In [110]: df
Out[110]: 
                  1_2                 1_3     0     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

By default the parser removes the component date columns, but you can choose to retain them via the keep_date_col keyword:

In [111]: df = pd.read_csv(
   .....:     "tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]], keep_date_col=True
   .....: )
   .....: 

In [112]: df
Out[112]: 
                  1_2                 1_3     0         1          2          3     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  19990127   19:00:00   18:56:00  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  19990127   20:00:00   19:56:00  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD  19990127   21:00:00   20:56:00 -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD  19990127   21:00:00   21:18:00 -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD  19990127   22:00:00   21:56:00 -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD  19990127   23:00:00   22:56:00 -0.59

Note that if you wish to combine multiple columns into a single date column, a nested list must be used. In other words, parse_dates=[1, 2] indicates that the second and third columns should each be parsed as separate date columns while parse_dates=[[1, 2]] means the two columns should be parsed into a single column.

You can also use a dict to specify custom name columns:

In [113]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}

In [114]: df = pd.read_csv("tmp.csv", header=None, parse_dates=date_spec)

In [115]: df
Out[115]: 
              nominal              actual     0     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

It is important to remember that if multiple text columns are to be parsed into a single date column, then a new column is prepended to the data. The index_col specification is based off of this new set of columns rather than the original data columns:

In [116]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}

In [117]: df = pd.read_csv(
   .....:     "tmp.csv", header=None, parse_dates=date_spec, index_col=0
   .....: )  # index is the nominal column
   .....: 

In [118]: df
Out[118]: 
                                 actual     0     4
nominal                                            
1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

Date parsing functions¶

Finally, the parser allows you to specify a custom date_parser function to take full advantage of the flexibility of the date parsing API:

In [119]: df = pd.read_csv(
   .....:     "tmp.csv", header=None, parse_dates=date_spec, date_parser=pd.to_datetime
   .....: )
   .....: 

In [120]: df
Out[120]: 
              nominal              actual     0     4
0 1999-01-27 19:00:00 1999-01-27 18:56:00  KORD  0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00  KORD  0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00  KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00  KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00  KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00  KORD -0.59

pandas will try to call the date_parser function in three different ways. If an exception is raised, the next one is tried:

1. date_parser is first called with one or more arrays as arguments, as defined using parse_dates (e.g., date_parser(['2013', '2013'], ['1', '2'])).

2. If #1 fails, date_parser is called with all the columns concatenated row-wise into a single array (e.g., date_parser(['2013 1', '2013 2'])).

Note that performance-wise, you should try these methods of parsing dates in order:

1. Try to infer the format using infer_datetime_format=True (see section below).

2. If you know the format, use pd.to_datetime(): date_parser=lambda x: pd.to_datetime(x, format=...).

3. If you have a really non-standard format, use a custom date_parser function. For optimal performance, this should be vectorized, i.e., it should accept arrays as arguments.

Parsing a CSV with mixed timezones¶

pandas cannot natively represent a column or index with mixed timezones. If your CSV file contains columns with a mixture of timezones, the default result will be an object-dtype column with strings, even with parse_dates.

In [121]: content = """\
   .....: a
   .....: 2000-01-01T00:00:00+05:00
   .....: 2000-01-01T00:00:00+06:00"""
   .....: 

In [122]: df = pd.read_csv(StringIO(content), parse_dates=["a"])

In [123]: df["a"]
Out[123]: 
0    2000-01-01 00:00:00+05:00
1    2000-01-01 00:00:00+06:00
Name: a, dtype: object

To parse the mixed-timezone values as a datetime column, pass a partially-applied to_datetime() with utc=True as the date_parser.

In [124]: df = pd.read_csv(
   .....:     StringIO(content),
   .....:     parse_dates=["a"],
   .....:     date_parser=lambda col: pd.to_datetime(col, utc=True),
   .....: )
   .....: 

In [125]: df["a"]
Out[125]: 
0   1999-12-31 19:00:00+00:00
1   1999-12-31 18:00:00+00:00
Name: a, dtype: datetime64[ns, UTC]

Inferring datetime format¶

If you have parse_dates enabled for some or all of your columns, and your datetime strings are all formatted the same way, you may get a large speed up by setting infer_datetime_format=True. If set, pandas will attempt to guess the format of your datetime strings, and then use a faster means of parsing the strings. 5-10x parsing speeds have been observed. pandas will fallback to the usual parsing if either the format cannot be guessed or the format that was guessed cannot properly parse the entire column of strings. So in general, infer_datetime_format should not have any negative consequences if enabled.

Here are some examples of datetime strings that can be guessed (All representing December 30th, 2011 at 00:00:00):

• “20111230”

• “2011/12/30”

• “20111230 00:00:00”

• “12/30/2011 00:00:00”

• “30/Dec/2011 00:00:00”

• “30/December/2011 00:00:00”

Note that infer_datetime_format is sensitive to dayfirst. With dayfirst=True, it will guess “01/12/2011” to be December 1st. With dayfirst=False (default) it will guess “01/12/2011” to be January 12th.

# Try to infer the format for the index column
In [126]: df = pd.read_csv(
   .....:     "foo.csv",
   .....:     index_col=0,
   .....:     parse_dates=True,
   .....:     infer_datetime_format=True,
   .....: )
   .....: 

In [127]: df
Out[127]: 
            A  B  C
date               
2009-01-01  a  1  2
2009-01-02  b  3  4
2009-01-03  c  4  5

International date formats¶

While US date formats tend to be MM/DD/YYYY, many international formats use DD/MM/YYYY instead. For convenience, a dayfirst keyword is provided:

In [128]: data = "date,value,cat\n1/6/2000,5,a\n2/6/2000,10,b\n3/6/2000,15,c"

In [129]: print(data)
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c

In [130]: with open("tmp.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [131]: pd.read_csv("tmp.csv", parse_dates=[0])
Out[131]: 
        date  value cat
0 2000-01-06      5   a
1 2000-02-06     10   b
2 2000-03-06     15   c

In [132]: pd.read_csv("tmp.csv", dayfirst=True, parse_dates=[0])
Out[132]: 
        date  value cat
0 2000-06-01      5   a
1 2000-06-02     10   b
2 2000-06-03     15   c

Writing CSVs to binary file objects¶

df.to_csv(..., mode="wb") allows writing a CSV to a file object opened binary mode. In most cases, it is not necessary to specify mode as Pandas will auto-detect whether the file object is opened in text or binary mode.

In [133]: import io

In [134]: data = pd.DataFrame([0, 1, 2])

In [135]: buffer = io.BytesIO()

In [136]: data.to_csv(buffer, encoding="utf-8", compression="gzip")

Files with an “implicit” index column¶

Consider a file with one less entry in the header than the number of data column:

In [190]: data = "A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5"

In [191]: print(data)
A,B,C
20090101,a,1,2
20090102,b,3,4
20090103,c,4,5

In [192]: with open("foo.csv", "w") as f:
   .....:     f.write(data)
   .....: 

In this special case, read_csv assumes that the first column is to be used as the index of the DataFrame:

In [193]: pd.read_csv("foo.csv")
Out[193]: 
          A  B  C
20090101  a  1  2
20090102  b  3  4
20090103  c  4  5

Note that the dates weren’t automatically parsed. In that case you would need to do as before:

In [194]: df = pd.read_csv("foo.csv", parse_dates=True)

In [195]: df.index
Out[195]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', freq=None)

Reading an index with a MultiIndex¶

Suppose you have data indexed by two columns:

In [196]: data = 'year,indiv,zit,xit\n1977,"A",1.2,.6\n1977,"B",1.5,.5'

In [197]: print(data)
year,indiv,zit,xit
1977,"A",1.2,.6
1977,"B",1.5,.5

In [198]: with open("mindex_ex.csv", mode="w") as f:
   .....:     f.write(data)
   .....: 

The index_col argument to read_csv can take a list of column numbers to turn multiple columns into a MultiIndex for the index of the returned object:

In [199]: df = pd.read_csv("mindex_ex.csv", index_col=[0, 1])

In [200]: df
Out[200]: 
            zit  xit
year indiv          
1977 A      1.2  0.6
     B      1.5  0.5

In [201]: df.loc[1977]
Out[201]: 
       zit  xit
indiv          
A      1.2  0.6
B      1.5  0.5

Reading columns with a MultiIndex¶

By specifying list of row locations for the header argument, you can read in a MultiIndex for the columns. Specifying non-consecutive rows will skip the intervening rows.

In [202]: from pandas._testing import makeCustomDataframe as mkdf

In [203]: df = mkdf(5, 3, r_idx_nlevels=2, c_idx_nlevels=4)

In [204]: df.to_csv("mi.csv")

In [205]: print(open("mi.csv").read())
C0,,C_l0_g0,C_l0_g1,C_l0_g2
C1,,C_l1_g0,C_l1_g1,C_l1_g2
C2,,C_l2_g0,C_l2_g1,C_l2_g2
C3,,C_l3_g0,C_l3_g1,C_l3_g2
R0,R1,,,
R_l0_g0,R_l1_g0,R0C0,R0C1,R0C2
R_l0_g1,R_l1_g1,R1C0,R1C1,R1C2
R_l0_g2,R_l1_g2,R2C0,R2C1,R2C2
R_l0_g3,R_l1_g3,R3C0,R3C1,R3C2
R_l0_g4,R_l1_g4,R4C0,R4C1,R4C2


In [206]: pd.read_csv("mi.csv", header=[0, 1, 2, 3], index_col=[0, 1])
Out[206]: 
C0              C_l0_g0 C_l0_g1 C_l0_g2
C1              C_l1_g0 C_l1_g1 C_l1_g2
C2              C_l2_g0 C_l2_g1 C_l2_g2
C3              C_l3_g0 C_l3_g1 C_l3_g2
R0      R1                             
R_l0_g0 R_l1_g0    R0C0    R0C1    R0C2
R_l0_g1 R_l1_g1    R1C0    R1C1    R1C2
R_l0_g2 R_l1_g2    R2C0    R2C1    R2C2
R_l0_g3 R_l1_g3    R3C0    R3C1    R3C2
R_l0_g4 R_l1_g4    R4C0    R4C1    R4C2

read_csv is also able to interpret a more common format of multi-columns indices.

In [207]: data = ",a,a,a,b,c,c\n,q,r,s,t,u,v\none,1,2,3,4,5,6\ntwo,7,8,9,10,11,12"

In [208]: print(data)
,a,a,a,b,c,c
,q,r,s,t,u,v
one,1,2,3,4,5,6
two,7,8,9,10,11,12

In [209]: with open("mi2.csv", "w") as fh:
   .....:     fh.write(data)
   .....: 

In [210]: pd.read_csv("mi2.csv", header=[0, 1], index_col=0)
Out[210]: 
     a         b   c    
     q  r  s   t   u   v
one  1  2  3   4   5   6
two  7  8  9  10  11  12

Note: If an index_col is not specified (e.g. you don’t have an index, or wrote it with df.to_csv(..., index=False), then any names on the columns index will be lost.

Writing to CSV format¶

The Series and DataFrame objects have an instance method to_csv which allows storing the contents of the object as a comma-separated-values file. The function takes a number of arguments. Only the first is required.

• path_or_buf: A string path to the file to write or a file object. If a file object it must be opened with newline=''

• sep : Field delimiter for the output file (default “,”)

• na_rep: A string representation of a missing value (default ‘’)

• float_format: Format string for floating point numbers

• columns: Columns to write (default None)

• header: Whether to write out the column names (default True)

• index: whether to write row (index) names (default True)

• index_label: Column label(s) for index column(s) if desired. If None (default), and header and index are True, then the index names are used. (A sequence should be given if the DataFrame uses MultiIndex).

• mode : Python write mode, default ‘w’

• encoding: a string representing the encoding to use if the contents are non-ASCII, for Python versions prior to 3

• line_terminator: Character sequence denoting line end (default os.linesep)

• quoting: Set quoting rules as in csv module (default csv.QUOTE_MINIMAL). Note that if you have set a float_format then floats are converted to strings and csv.QUOTE_NONNUMERIC will treat them as non-numeric

• quotechar: Character used to quote fields (default ‘”’)

• doublequote: Control quoting of quotechar in fields (default True)

• escapechar: Character used to escape sep and quotechar when appropriate (default None)

• chunksize: Number of rows to write at a time

• date_format: Format string for datetime objects

Writing a formatted string¶

The DataFrame object has an instance method to_string which allows control over the string representation of the object. All arguments are optional:

• buf default None, for example a StringIO object

• columns default None, which columns to write

• col_space default None, minimum width of each column.

• na_rep default NaN, representation of NA value

• formatters default None, a dictionary (by column) of functions each of which takes a single argument and returns a formatted string

• float_format default None, a function which takes a single (float) argument and returns a formatted string; to be applied to floats in the DataFrame.

• sparsify default True, set to False for a DataFrame with a hierarchical index to print every MultiIndex key at each row.

• index_names default True, will print the names of the indices

• index default True, will print the index (ie, row labels)

• header default True, will print the column labels

• justify default left, will print column headers left- or right-justified

The Series object also has a to_string method, but with only the buf, na_rep, float_format arguments. There is also a length argument which, if set to True, will additionally output the length of the Series.

Orient options¶

There are a number of different options for the format of the resulting JSON file / string. Consider the following DataFrame and Series:

In [224]: dfjo = pd.DataFrame(
   .....:     dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)),
   .....:     columns=list("ABC"),
   .....:     index=list("xyz"),
   .....: )
   .....: 

In [225]: dfjo
Out[225]: 
   A  B  C
x  1  4  7
y  2  5  8
z  3  6  9

In [226]: sjo = pd.Series(dict(x=15, y=16, z=17), name="D")

In [227]: sjo
Out[227]: 
x    15
y    16
z    17
Name: D, dtype: int64

Column oriented (the default for DataFrame) serializes the data as nested JSON objects with column labels acting as the primary index:

In [228]: dfjo.to_json(orient="columns")
Out[228]: '{"A":{"x":1,"y":2,"z":3},"B":{"x":4,"y":5,"z":6},"C":{"x":7,"y":8,"z":9}}'

# Not available for Series

Index oriented (the default for Series) similar to column oriented but the index labels are now primary:

In [229]: dfjo.to_json(orient="index")
Out[229]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,"C":9}}'

In [230]: sjo.to_json(orient="index")
Out[230]: '{"x":15,"y":16,"z":17}'

Record oriented serializes the data to a JSON array of column -> value records, index labels are not included. This is useful for passing DataFrame data to plotting libraries, for example the JavaScript library d3.js:

In [231]: dfjo.to_json(orient="records")
Out[231]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]'

In [232]: sjo.to_json(orient="records")
Out[232]: '[15,16,17]'

Value oriented is a bare-bones option which serializes to nested JSON arrays of values only, column and index labels are not included:

In [233]: dfjo.to_json(orient="values")
Out[233]: '[[1,4,7],[2,5,8],[3,6,9]]'

# Not available for Series

Split oriented serializes to a JSON object containing separate entries for values, index and columns. Name is also included for Series:

In [234]: dfjo.to_json(orient="split")
Out[234]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,5,8],[3,6,9]]}'

In [235]: sjo.to_json(orient="split")
Out[235]: '{"name":"D","index":["x","y","z"],"data":[15,16,17]}'

Table oriented serializes to the JSON Table Schema, allowing for the preservation of metadata including but not limited to dtypes and index names.

Date handling¶

Writing in ISO date format:

In [236]: dfd = pd.DataFrame(np.random.randn(5, 2), columns=list("AB"))

In [237]: dfd["date"] = pd.Timestamp("20130101")

In [238]: dfd = dfd.sort_index(axis=1, ascending=False)

In [239]: json = dfd.to_json(date_format="iso")

In [240]: json
Out[240]: '{"date":{"0":"2013-01-01T00:00:00.000Z","1":"2013-01-01T00:00:00.000Z","2":"2013-01-01T00:00:00.000Z","3":"2013-01-01T00:00:00.000Z","4":"2013-01-01T00:00:00.000Z"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Writing in ISO date format, with microseconds:

In [241]: json = dfd.to_json(date_format="iso", date_unit="us")

In [242]: json
Out[242]: '{"date":{"0":"2013-01-01T00:00:00.000000Z","1":"2013-01-01T00:00:00.000000Z","2":"2013-01-01T00:00:00.000000Z","3":"2013-01-01T00:00:00.000000Z","4":"2013-01-01T00:00:00.000000Z"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Epoch timestamps, in seconds:

In [243]: json = dfd.to_json(date_format="epoch", date_unit="s")

In [244]: json
Out[244]: '{"date":{"0":1356998400,"1":1356998400,"2":1356998400,"3":1356998400,"4":1356998400},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'

Writing to a file, with a date index and a date column:

In [245]: dfj2 = dfj.copy()

In [246]: dfj2["date"] = pd.Timestamp("20130101")

In [247]: dfj2["ints"] = list(range(5))

In [248]: dfj2["bools"] = True

In [249]: dfj2.index = pd.date_range("20130101", periods=5)

In [250]: dfj2.to_json("test.json")

In [251]: with open("test.json") as fh:
   .....:     print(fh.read())
   .....: 
{"A":{"1356998400000":-0.1213062281,"1357084800000":0.6957746499,"1357171200000":0.9597255933,"1357257600000":-0.6199759194,"1357344000000":-0.7323393705},"B":{"1356998400000":-0.0978826728,"1357084800000":0.3417343559,"1357171200000":-1.1103361029,"1357257600000":0.1497483186,"1357344000000":0.6877383895},"date":{"1356998400000":1356998400000,"1357084800000":1356998400000,"1357171200000":1356998400000,"1357257600000":1356998400000,"1357344000000":1356998400000},"ints":{"1356998400000":0,"1357084800000":1,"1357171200000":2,"1357257600000":3,"1357344000000":4},"bools":{"1356998400000":true,"1357084800000":true,"1357171200000":true,"1357257600000":true,"1357344000000":true}}

Fallback behavior¶

If the JSON serializer cannot handle the container contents directly it will fall back in the following manner:

• if the dtype is unsupported (e.g. np.complex_) then the default_handler, if provided, will be called for each value, otherwise an exception is raised.

• if an object is unsupported it will attempt the following:

  • check if the object has defined a toDict method and call it. A toDict method should return a dict which will then be JSON serialized.

  • invoke the default_handler if one was provided.

  • convert the object to a dict by traversing its contents. However this will often fail with an OverflowError or give unexpected results.

In general the best approach for unsupported objects or dtypes is to provide a default_handler. For example:

>>> DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json()  # raises
RuntimeError: Unhandled numpy dtype 15

can be dealt with by specifying a simple default_handler:

In [252]: pd.DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json(default_handler=str)
Out[252]: '{"0":{"0":"(1+0j)","1":"(2+0j)","2":"(1+2j)"}}'

Data conversion¶

The default of convert_axes=True, dtype=True, and convert_dates=True will try to parse the axes, and all of the data into appropriate types, including dates. If you need to override specific dtypes, pass a dict to dtype. convert_axes should only be set to False if you need to preserve string-like numbers (e.g. ‘1’, ‘2’) in an axes.

Reading from a JSON string:

In [253]: pd.read_json(json)
Out[253]: 
        date         B         A
0 2013-01-01  0.403310  0.176444
1 2013-01-01  0.301624 -0.154951
2 2013-01-01 -1.369849 -2.179861
3 2013-01-01  1.462696 -0.954208
4 2013-01-01 -0.826591 -1.743161

Reading from a file:

In [254]: pd.read_json("test.json")
Out[254]: 
                   A         B       date  ints  bools
2013-01-01 -0.121306 -0.097883 2013-01-01     0   True
2013-01-02  0.695775  0.341734 2013-01-01     1   True
2013-01-03  0.959726 -1.110336 2013-01-01     2   True
2013-01-04 -0.619976  0.149748 2013-01-01     3   True
2013-01-05 -0.732339  0.687738 2013-01-01     4   True

Don’t convert any data (but still convert axes and dates):

In [255]: pd.read_json("test.json", dtype=object).dtypes
Out[255]: 
A        object
B        object
date     object
ints     object
bools    object
dtype: object

Specify dtypes for conversion:

In [256]: pd.read_json("test.json", dtype={"A": "float32", "bools": "int8"}).dtypes
Out[256]: 
A               float32
B               float64
date     datetime64[ns]
ints              int64
bools              int8
dtype: object

Preserve string indices:

In [257]: si = pd.DataFrame(
   .....:     np.zeros((4, 4)), columns=list(range(4)), index=[str(i) for i in range(4)]
   .....: )
   .....: 

In [258]: si
Out[258]: 
     0    1    2    3
0  0.0  0.0  0.0  0.0
1  0.0  0.0  0.0  0.0
2  0.0  0.0  0.0  0.0
3  0.0  0.0  0.0  0.0

In [259]: si.index
Out[259]: Index(['0', '1', '2', '3'], dtype='object')

In [260]: si.columns
Out[260]: Int64Index([0, 1, 2, 3], dtype='int64')

In [261]: json = si.to_json()

In [262]: sij = pd.read_json(json, convert_axes=False)

In [263]: sij
Out[263]: 
   0  1  2  3
0  0  0  0  0
1  0  0  0  0
2  0  0  0  0
3  0  0  0  0

In [264]: sij.index
Out[264]: Index(['0', '1', '2', '3'], dtype='object')

In [265]: sij.columns
Out[265]: Index(['0', '1', '2', '3'], dtype='object')

Dates written in nanoseconds need to be read back in nanoseconds:

In [266]: json = dfj2.to_json(date_unit="ns")

# Try to parse timestamps as milliseconds -> Won't Work
In [267]: dfju = pd.read_json(json, date_unit="ms")

In [268]: dfju
Out[268]: 
                            A         B                 date  ints  bools
1356998400000000000 -0.121306 -0.097883  1356998400000000000     0   True
1357084800000000000  0.695775  0.341734  1356998400000000000     1   True
1357171200000000000  0.959726 -1.110336  1356998400000000000     2   True
1357257600000000000 -0.619976  0.149748  1356998400000000000     3   True
1357344000000000000 -0.732339  0.687738  1356998400000000000     4   True

# Let pandas detect the correct precision
In [269]: dfju = pd.read_json(json)

In [270]: dfju
Out[270]: 
                   A         B       date  ints  bools
2013-01-01 -0.121306 -0.097883 2013-01-01     0   True
2013-01-02  0.695775  0.341734 2013-01-01     1   True
2013-01-03  0.959726 -1.110336 2013-01-01     2   True
2013-01-04 -0.619976  0.149748 2013-01-01     3   True
2013-01-05 -0.732339  0.687738 2013-01-01     4   True

# Or specify that all timestamps are in nanoseconds
In [271]: dfju = pd.read_json(json, date_unit="ns")

In [272]: dfju
Out[272]: 
                   A         B       date  ints  bools
2013-01-01 -0.121306 -0.097883 2013-01-01     0   True
2013-01-02  0.695775  0.341734 2013-01-01     1   True
2013-01-03  0.959726 -1.110336 2013-01-01     2   True
2013-01-04 -0.619976  0.149748 2013-01-01     3   True
2013-01-05 -0.732339  0.687738 2013-01-01     4   True

The Numpy parameter¶

If numpy=True is passed to read_json an attempt will be made to sniff an appropriate dtype during deserialization and to subsequently decode directly to NumPy arrays, bypassing the need for intermediate Python objects.

This can provide speedups if you are deserialising a large amount of numeric data:

In [273]: randfloats = np.random.uniform(-100, 1000, 10000)

In [274]: randfloats.shape = (1000, 10)

In [275]: dffloats = pd.DataFrame(randfloats, columns=list("ABCDEFGHIJ"))

In [276]: jsonfloats = dffloats.to_json()

In [277]: %timeit pd.read_json(jsonfloats)
27.1 ms +- 7.28 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)

In [278]: %timeit pd.read_json(jsonfloats, numpy=True)
22.9 ms +- 3.41 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)

The speedup is less noticeable for smaller datasets:

In [279]: jsonfloats = dffloats.head(100).to_json()

In [280]: %timeit pd.read_json(jsonfloats)
12.1 ms +- 3.66 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)

In [281]: %timeit pd.read_json(jsonfloats, numpy=True)
7.65 ms +- 432 us per loop (mean +- std. dev. of 7 runs, 100 loops each)