IO tools (text, CSV, HDF5, …)

The pandas I/O API is a set of top level reader functions accessed like pandas.read_csv() that generally return a pandas object. The corresponding writer functions are object methods that are accessed like DataFrame.to_csv(). Below is a table containing available readers and writers.

Here is an informal performance comparison for some of these IO methods.

::: tip Note

For examples that use the StringIO class, make sure you import it according to your Python version, i.e. from StringIO import StringIO for Python 2 and from io import StringIO for Python 3.

:::

CSV & text files

The workhorse function for reading text files (a.k.a. flat files) is read_csv(). See the cookbook for some advanced strategies.

Parsing options

read_csv() accepts the following common arguments:

Basic

filepath_or_buffer : various

• Either a path to a file (a str, pathlib.Path, or 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).

sep : str, 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'.

delimiter : str, default None

• Alternative argument name for sep.

delim_whitespace : boolean, 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.

New in version 0.18.1: support for the Python parser.

Column and index locations and names

header : int 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.

names : array-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_col : int, str, sequence of int / str, or False, 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.

usecols : list-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). 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]: from io import StringIO, BytesIO
In [2]: data = ('col1,col2,col3\n'
   ...:         'a,b,1\n'
   ...:         'a,b,2\n'
   ...:         'c,d,3')
   ...: 
In [3]: pd.read_csv(StringIO(data))
Out[3]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3
In [4]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ['COL1', 'COL3'])
Out[4]: 
  col1  col3
0    a     1
1    a     2
2    c     3

Using this parameter results in much faster parsing time and lower memory usage.

squeeze : boolean, default False

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

prefix : str, default None

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

mangle_dupe_cols : boolean, 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

dtype : Type 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.

• New in version 0.20.0: support for the Python parser.

engine : {'c', 'python'}

• Parser engine to use. The C engine is faster while the Python engine is currently more feature-complete.

converters : dict, default None

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

true_values : list, default None

• Values to consider as True.

false_values : list, default None

• Values to consider as False.

skipinitialspace : boolean, default False

• Skip spaces after delimiter.

skiprows : list-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 [5]: data = ('col1,col2,col3\n'
   ...:         'a,b,1\n'
   ...:         'a,b,2\n'
   ...:         'c,d,3')
   ...: 
In [6]: pd.read_csv(StringIO(data))
Out[6]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3
In [7]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0)
Out[7]: 
  col1 col2  col3
0    a    b     2

skipfooter : int, default 0

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

nrows : int, default None

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

low_memory : boolean, 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_map : boolean, 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_values : scalar, 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_na : boolean, 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_filter : boolean, 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.

verbose : boolean, default False

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

skip_blank_lines : boolean, default True

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

Datetime handling

parse_dates : boolean 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_format : boolean, 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_col : boolean, default False

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

date_parser : function, 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.

dayfirst : boolean, default False

• DD/MM format dates, international and European format.

cache_dates : boolean, 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

iterator : boolean, default False

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

chunksize : int, default None

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

Quoting, compression, and file format

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

• For on-the-fly decompression of on-disk data. If ‘infer’, then use gzip, bz2, zip, or xz if filepath_or_buffer is a string ending in ‘.gz’, ‘.bz2’, ‘.zip’, or ‘.xz’, 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.

New in version 0.18.1: support for ‘zip’ and ‘xz’ compression.

Changed in version 0.24.0: ‘infer’ option added and set to default.

thousands : str, default None

• Thousands separator.

decimal : str, default '.'

• Character to recognize as decimal point. E.g. use ‘,’ for European data.

float_precision : string, 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.

lineterminator : str (length 1), default None

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

quotechar : str (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.

quoting : int or ``csv.QUOTE_instance, default0``*

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

doublequote : boolean, 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.

escapechar : str (length 1), default None

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

comment : str, 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 a,b,c 1,2,3’ with header=0 will result in ‘a,b,c’ being treated as the header.

encoding : str, default None

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

dialect : str 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_lines : boolean, default True

• 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.

warn_bad_lines : boolean, default True

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

Specifying column data types

You can indicate the data type for the whole DataFrame or individual columns:

In [8]: data = ('a,b,c,d\n'
   ...:         '1,2,3,4\n'
   ...:         '5,6,7,8\n'
   ...:         '9,10,11')
   ...: 
In [9]: print(data)
a,b,c,d
1,2,3,4
5,6,7,8
9,10,11
In [10]: df = pd.read_csv(StringIO(data), dtype=object)
In [11]: df
Out[11]: 
   a   b   c    d
0  1   2   3    4
1  5   6   7    8
2  9  10  11  NaN
In [12]: df['a'][0]
Out[12]: '1'
In [13]: df = pd.read_csv(StringIO(data),
   ....:                  dtype={'b': object, 'c': np.float64, 'd': 'Int64'})
   ....: 
In [14]: df.dtypes
Out[14]: 
a      int64
b     object
c    float64
d      Int64
dtype: object

Fortunately, pandas offers more than one way to ensure that your column(s) contain only one dtype. If you’re unfamiliar with these concepts, you can see here to learn more about dtypes, and here to learn more about object conversion in pandas.

For instance, you can use the converters argument of read_csv():

In [15]: data = ("col_1\n"
   ....:         "1\n"
   ....:         "2\n"
   ....:         "'A'\n"
   ....:         "4.22")
   ....: 
In [16]: df = pd.read_csv(StringIO(data), converters={'col_1': str})
In [17]: df
Out[17]: 
  col_1
0     1
1     2
2   'A'
3  4.22
In [18]: df['col_1'].apply(type).value_counts()
Out[18]: 
<class 'str'>    4
Name: col_1, dtype: int64

Or you can use the to_numeric() function to coerce the dtypes after reading in the data,

In [19]: df2 = pd.read_csv(StringIO(data))
In [20]: df2['col_1'] = pd.to_numeric(df2['col_1'], errors='coerce')
In [21]: df2
Out[21]: 
   col_1
0   1.00
1   2.00
2    NaN
3   4.22
In [22]: df2['col_1'].apply(type).value_counts()
Out[22]: 
<class 'float'>    4
Name: col_1, dtype: int64

which will convert all valid parsing to floats, leaving the invalid parsing as NaN.

Ultimately, how you deal with reading in columns containing mixed dtypes depends on your specific needs. In the case above, if you wanted to NaN out the data anomalies, then to_numeric() is probably your best option. However, if you wanted for all the data to be coerced, no matter the type, then using the converters argument of read_csv() would certainly be worth trying.

New in version 0.20.0: support for the Python parser.

The dtype option is supported by the ‘python’ engine.

::: tip Note

In some cases, reading in abnormal data with columns containing mixed dtypes will result in an inconsistent dataset. If you rely on pandas to infer the dtypes of your columns, the parsing engine will go and infer the dtypes for different chunks of the data, rather than the whole dataset at once. Consequently, you can end up with column(s) with mixed dtypes. For example,

In [23]: col_1 = list(range(500000)) + ['a', 'b'] + list(range(500000))
In [24]: df = pd.DataFrame({'col_1': col_1})
In [25]: df.to_csv('foo.csv')
In [26]: mixed_df = pd.read_csv('foo.csv')
In [27]: mixed_df['col_1'].apply(type).value_counts()
Out[27]: 
<class 'int'>    737858
<class 'str'>    262144
Name: col_1, dtype: int64
In [28]: mixed_df['col_1'].dtype
Out[28]: dtype('O')

will result with mixed_df containing an int dtype for certain chunks of the column, and str for others due to the mixed dtypes from the data that was read in. It is important to note that the overall column will be marked with a dtype of object, which is used for columns with mixed dtypes.

:::

Specifying categorical dtype

New in version 0.19.0.

Categorical columns can be parsed directly by specifying dtype='category' or dtype=CategoricalDtype(categories, ordered).

In [29]: data = ('col1,col2,col3\n'
   ....:         'a,b,1\n'
   ....:         'a,b,2\n'
   ....:         'c,d,3')
   ....: 
In [30]: pd.read_csv(StringIO(data))
Out[30]: 
  col1 col2  col3
0    a    b     1
1    a    b     2
2    c    d     3
In [31]: pd.read_csv(StringIO(data)).dtypes
Out[31]: 
col1    object
col2    object
col3     int64
dtype: object
In [32]: pd.read_csv(StringIO(data), dtype='category').dtypes
Out[32]: 
col1    category
col2    category
col3    category
dtype: object

Individual columns can be parsed as a Categorical using a dict specification:

In [33]: pd.read_csv(StringIO(data), dtype={'col1': 'category'}).dtypes
Out[33]: 
col1    category
col2      object
col3       int64
dtype: object

New in version 0.21.0.

Specifying dtype='category' will result in an unordered Categorical whose categories are the unique values observed in the data. For more control on the categories and order, create a CategoricalDtype ahead of time, and pass that for that column’s dtype.

In [34]: from pandas.api.types import CategoricalDtype
In [35]: dtype = CategoricalDtype(['d', 'c', 'b', 'a'], ordered=True)
In [36]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).dtypes
Out[36]: 
col1    category
col2      object
col3       int64
dtype: object

When using dtype=CategoricalDtype, “unexpected” values outside of dtype.categories are treated as missing values.

In [37]: dtype = CategoricalDtype(['a', 'b', 'd'])  # No 'c'
In [38]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).col1
Out[38]: 
0      a
1      a
2    NaN
Name: col1, dtype: category
Categories (3, object): [a, b, d]

This matches the behavior of Categorical.set_categories().

::: tip Note

With dtype='category', the resulting categories will always be parsed as strings (object dtype). If the categories are numeric they can be converted using the to_numeric() function, or as appropriate, another converter such as to_datetime().

When dtype is a CategoricalDtype with homogeneous categories ( all numeric, all datetimes, etc.), the conversion is done automatically.

In [39]: df = pd.read_csv(StringIO(data), dtype='category')
In [40]: df.dtypes
Out[40]: 
col1    category
col2    category
col3    category
dtype: object
In [41]: df['col3']
Out[41]: 
0    1
1    2
2    3
Name: col3, dtype: category
Categories (3, object): [1, 2, 3]
In [42]: df['col3'].cat.categories = pd.to_numeric(df['col3'].cat.categories)
In [43]: df['col3']
Out[43]: 
0    1
1    2
2    3
Name: col3, dtype: category
Categories (3, int64): [1, 2, 3]

:::

Naming and using columns

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 [44]: data = ('a,b,c\n'
   ....:         '1,2,3\n'
   ....:         '4,5,6\n'
   ....:         '7,8,9')
   ....: 
In [45]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [46]: pd.read_csv(StringIO(data))
Out[46]: 
   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 [47]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [48]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=0)
Out[48]: 
   foo  bar  baz
0    1    2    3
1    4    5    6
2    7    8    9
In [49]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=None)
Out[49]: 
  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 [50]: data = ('skip this skip it\n'
   ....:         'a,b,c\n'
   ....:         '1,2,3\n'
   ....:         '4,5,6\n'
   ....:         '7,8,9')
   ....: 
In [51]: pd.read_csv(StringIO(data), header=1)
Out[51]: 
   a  b  c
0  1  2  3
1  4  5  6
2  7  8  9

::: tip Note

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 non-blank line of the file, if column names are passed explicitly then the behavior is identical to header=None.

:::

Duplicate names parsing

If the file or header contains duplicate names, pandas will by default distinguish between them so as to prevent overwriting data:

In [52]: data = ('a,b,a\n'
   ....:         '0,1,2\n'
   ....:         '3,4,5')
   ....: 
In [53]: pd.read_csv(StringIO(data))
Out[53]: 
   a  b  a.1
0  0  1    2
1  3  4    5

There is no more duplicate data because mangle_dupe_cols=True by default, which modifies a series of duplicate columns ‘X’, …, ‘X’ to become ‘X’, ‘X.1’, …, ‘X.N’. If mangle_dupe_cols=False, duplicate data can arise:

In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
Out[3]:
   a  b  a
0  2  1  2
1  5  4  5

To prevent users from encountering this problem with duplicate data, a ValueError exception is raised if mangle_dupe_cols != True:

In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
...
ValueError: Setting mangle_dupe_cols=False is not supported yet

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:

New in version 0.20.0: support for callable usecols arguments

In [54]: data = 'a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz'
In [55]: pd.read_csv(StringIO(data))
Out[55]: 
   a  b  c    d
0  1  2  3  foo
1  4  5  6  bar
2  7  8  9  baz
In [56]: pd.read_csv(StringIO(data), usecols=['b', 'd'])
Out[56]: 
   b    d
0  2  foo
1  5  bar
2  8  baz
In [57]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])
Out[57]: 
   a  c    d
0  1  3  foo
1  4  6  bar
2  7  9  baz
In [58]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ['A', 'C'])
Out[58]: 
   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 [59]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ['a', 'c'])
Out[59]: 
   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.

Comments and empty lines

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 [60]: data = ('\n'
   ....:         'a,b,c\n'
   ....:         '  \n'
   ....:         '# commented line\n'
   ....:         '1,2,3\n'
   ....:         '\n'
   ....:         '4,5,6')
   ....: 
In [61]: print(data)
a,b,c
# commented line
1,2,3
4,5,6
In [62]: pd.read_csv(StringIO(data), comment='#')
Out[62]: 
   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 [63]: data = ('a,b,c\n'
   ....:         '\n'
   ....:         '1,2,3\n'
   ....:         '\n'
   ....:         '\n'
   ....:         '4,5,6')
   ....: 
In [64]: pd.read_csv(StringIO(data), skip_blank_lines=False)
Out[64]: 
     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

::: danger Warning

The presence of ignored lines might create ambiguities involving line numbers; the parameter header uses row numbers (ignoring commented/empty lines), while skiprows uses line numbers (including commented/empty lines):

In [65]: data = ('#comment\n'
   ....:         'a,b,c\n'
   ....:         'A,B,C\n'
   ....:         '1,2,3')
   ....: 
In [66]: pd.read_csv(StringIO(data), comment='#', header=1)
Out[66]: 
   A  B  C
0  1  2  3
In [67]: data = ('A,B,C\n'
   ....:         '#comment\n'
   ....:         'a,b,c\n'
   ....:         '1,2,3')
   ....: 
In [68]: pd.read_csv(StringIO(data), comment='#', skiprows=2)
Out[68]: 
   a  b  c
0  1  2  3

If both header and skiprows are specified, header will be relative to the end of skiprows. For example:

:::

In [69]: 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 [70]: 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 [71]: pd.read_csv(StringIO(data), comment='#', skiprows=4, header=1)
Out[71]: 
     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 [72]: 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 [73]: df = pd.read_csv('tmp.csv')
In [74]: df
Out[74]: 
         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 [75]: df = pd.read_csv('tmp.csv', comment='#')
In [76]: df
Out[76]: 
         ID    level category
0  Patient1   123000       x 
1  Patient2    23000       y 
2  Patient3  1234018       z

Dealing with Unicode data

The encoding argument should be used for encoded unicode data, which will result in byte strings being decoded to unicode in the result:

In [77]: data = (b'word,length\n'
   ....:         b'Tr\xc3\xa4umen,7\n'
   ....:         b'Gr\xc3\xbc\xc3\x9fe,5')
   ....: 
In [78]: data = data.decode('utf8').encode('latin-1')
In [79]: df = pd.read_csv(BytesIO(data), encoding='latin-1')
In [80]: df
Out[80]: 
      word  length
0  Träumen       7
1    Grüße       5
In [81]: df['word'][1]
Out[81]: 'Grüße'

Some formats which encode all characters as multiple bytes, like UTF-16, won’t parse correctly at all without specifying the encoding. Full list of Python standard encodings.

Index columns and trailing delimiters

If a file has one more column of data than the number of column names, the first column will be used as the DataFrame’s row names:

In [82]: data = ('a,b,c\n'
   ....:         '4,apple,bat,5.7\n'
   ....:         '8,orange,cow,10')
   ....: 
In [83]: pd.read_csv(StringIO(data))
Out[83]: 
        a    b     c
4   apple  bat   5.7
8  orange  cow  10.0

In [84]: data = ('index,a,b,c\n'
   ....:         '4,apple,bat,5.7\n'
   ....:         '8,orange,cow,10')
   ....: 
In [85]: pd.read_csv(StringIO(data), index_col=0)
Out[85]: 
            a    b     c
index                   
4       apple  bat   5.7
8      orange  cow  10.0

Ordinarily, you can achieve this behavior using the index_col option.

There are some exception cases when a file has been prepared with delimiters at the end of each data line, confusing the parser. To explicitly disable the index column inference and discard the last column, pass index_col=False:

In [86]: data = ('a,b,c\n'
   ....:         '4,apple,bat,\n'
   ....:         '8,orange,cow,')
   ....: 
In [87]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,
In [88]: pd.read_csv(StringIO(data))
Out[88]: 
        a    b   c
4   apple  bat NaN
8  orange  cow NaN
In [89]: pd.read_csv(StringIO(data), index_col=False)
Out[89]: 
   a       b    c
0  4   apple  bat
1  8  orange  cow

If a subset of data is being parsed using the usecols option, the index_col specification is based on that subset, not the original data.

In [90]: data = ('a,b,c\n'
   ....:         '4,apple,bat,\n'
   ....:         '8,orange,cow,')
   ....: 
In [91]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,
In [92]: pd.read_csv(StringIO(data), usecols=['b', 'c'])
Out[92]: 
     b   c
4  bat NaN
8  cow NaN
In [93]: pd.read_csv(StringIO(data), usecols=['b', 'c'], index_col=0)
Out[93]: 
     b   c
4  bat NaN
8  cow NaN

Date Handling

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:

# Use a column as an index, and parse it as dates.
In [94]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True)
In [95]: df
Out[95]: 
            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 [96]: df.index
Out[96]: 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 [97]: print(open('tmp.csv').read())
KORD,19990127, 19:00:00, 18:56:00, 0.8100
KORD,19990127, 20:00:00, 19:56:00, 0.0100
KORD,19990127, 21:00:00, 20:56:00, -0.5900
KORD,19990127, 21:00:00, 21:18:00, -0.9900
KORD,19990127, 22:00:00, 21:56:00, -0.5900
KORD,19990127, 23:00:00, 22:56:00, -0.5900
In [98]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]])
In [99]: df
Out[99]: 
                  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 [100]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]],
   .....:                  keep_date_col=True)
   .....: 
In [101]: df
Out[101]: 
                  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 [102]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}
In [103]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec)
In [104]: df
Out[104]: 
              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 [105]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}
In [106]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,
   .....:                  index_col=0)  # index is the nominal column
   .....: 
In [107]: df
Out[107]: 
                                 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

::: tip Note

If a column or index contains an unparsable date, the entire column or index will be returned unaltered as an object data type. For non-standard datetime parsing, use to_datetime() after pd.read_csv.

:::

::: tip Note

read_csv has a fast_path for parsing datetime strings in iso8601 format, e.g “2000-01-01T00:01:02+00:00” and similar variations. If you can arrange for your data to store datetimes in this format, load times will be significantly faster, ~20x has been observed.

:::

::: tip Note

When passing a dict as the parse_dates argument, the order of the columns prepended is not guaranteed, because dict objects do not impose an ordering on their keys. On Python 2.7+ you may use collections.OrderedDict instead of a regular dict if this matters to you. Because of this, when using a dict for ‘parse_dates’ in conjunction with the index_col argument, it’s best to specify index_col as a column label rather then as an index on the resulting frame.

:::

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 [108]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,
   .....:                  date_parser=pd.io.date_converters.parse_date_time)
   .....: 
In [109]: df
Out[109]: 
              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'])).
3. If #2 fails, date_parser is called once for every row with one or more string arguments from the columns indicated with parse_dates (e.g., date_parser('2013', '1') for the first row, date_parser('2013', '2') for the second, etc.).

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.

You can explore the date parsing functionality in date_converters.py and add your own. We would love to turn this module into a community supported set of date/time parsers. To get you started, date_converters.py contains functions to parse dual date and time columns, year/month/day columns, and year/month/day/hour/minute/second columns. It also contains a generic_parser function so you can curry it with a function that deals with a single date rather than the entire array.

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 [110]: content = """\
   .....: a
   .....: 2000-01-01T00:00:00+05:00
   .....: 2000-01-01T00:00:00+06:00"""
   .....: 
In [111]: df = pd.read_csv(StringIO(content), parse_dates=['a'])
In [112]: df['a']
Out[112]: 
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 [113]: df = pd.read_csv(StringIO(content), parse_dates=['a'],
   .....:                  date_parser=lambda col: pd.to_datetime(col, utc=True))
   .....: 
In [114]: df['a']
Out[114]: 
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 [115]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True,
   .....:                  infer_datetime_format=True)
   .....: 
In [116]: df
Out[116]: 
            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 [117]: print(open('tmp.csv').read())
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c
In [118]: pd.read_csv('tmp.csv', parse_dates=[0])
Out[118]: 
        date  value cat
0 2000-01-06      5   a
1 2000-02-06     10   b
2 2000-03-06     15   c
In [119]: pd.read_csv('tmp.csv', dayfirst=True, parse_dates=[0])
Out[119]: 
        date  value cat
0 2000-06-01      5   a
1 2000-06-02     10   b
2 2000-06-03     15   c

Specifying method for floating-point conversion

The parameter float_precision can be specified in order to use a specific floating-point converter during parsing with the C engine. The options are the ordinary converter, the high-precision converter, and the round-trip converter (which is guaranteed to round-trip values after writing to a file). For example:

In [120]: val = '0.3066101993807095471566981359501369297504425048828125'
In [121]: data = 'a,b,c\n1,2,{0}'.format(val)
In [122]: abs(pd.read_csv(StringIO(data), engine='c',
   .....:                 float_precision=None)['c'][0] - float(val))
   .....: 
Out[122]: 1.1102230246251565e-16
In [123]: abs(pd.read_csv(StringIO(data), engine='c',
   .....:                 float_precision='high')['c'][0] - float(val))
   .....: 
Out[123]: 5.551115123125783e-17
In [124]: abs(pd.read_csv(StringIO(data), engine='c',
   .....:                 float_precision='round_trip')['c'][0] - float(val))
   .....: 
Out[124]: 0.0

Thousand separators

For large numbers that have been written with a thousands separator, you can set the thousands keyword to a string of length 1 so that integers will be parsed correctly:

By default, numbers with a thousands separator will be parsed as strings:

In [125]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y
Patient3|1,234,018|z
In [126]: df = pd.read_csv('tmp.csv', sep='|')
In [127]: df
Out[127]: 
         ID      level category
0  Patient1    123,000        x
1  Patient2     23,000        y
2  Patient3  1,234,018        z
In [128]: df.level.dtype
Out[128]: dtype('O')

The thousands keyword allows integers to be parsed correctly:

In [129]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y
Patient3|1,234,018|z
In [130]: df = pd.read_csv('tmp.csv', sep='|', thousands=',')
In [131]: df
Out[131]: 
         ID    level category
0  Patient1   123000        x
1  Patient2    23000        y
2  Patient3  1234018        z
In [132]: df.level.dtype
Out[132]: dtype('int64')

NA values

To control which values are parsed as missing values (which are signified by NaN), specify a string in na_values. If you specify a list of strings, then all values in it are considered to be missing values. If you specify a number (a float, like 5.0 or an integer like 5), the corresponding equivalent values will also imply a missing value (in this case effectively [5.0, 5] are recognized as NaN).

To completely override the default values that are recognized as missing, specify keep_default_na=False.

The default NaN recognized values are ['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A N/A', '#N/A', 'N/A', 'n/a', 'NA', '#NA', 'NULL', 'null', 'NaN', '-NaN', 'nan', '-nan', ''].

Let us consider some examples:

pd.read_csv('path_to_file.csv', na_values=[5])

In the example above 5 and 5.0 will be recognized as NaN, in addition to the defaults. A string will first be interpreted as a numerical 5, then as a NaN.

pd.read_csv('path_to_file.csv', keep_default_na=False, na_values=[""])

Above, only an empty field will be recognized as NaN.

pd.read_csv('path_to_file.csv', keep_default_na=False, na_values=["NA", "0"])

Above, both NA and 0 as strings are NaN.

pd.read_csv('path_to_file.csv', na_values=["Nope"])

The default values, in addition to the string "Nope" are recognized as NaN.

Infinity

inf like values will be parsed as np.inf (positive infinity), and -inf as -np.inf (negative infinity). These will ignore the case of the value, meaning Inf, will also be parsed as np.inf.

Returning Series

Using the squeeze keyword, the parser will return output with a single column as a Series:

In [133]: print(open('tmp.csv').read())
level
Patient1,123000
Patient2,23000
Patient3,1234018
In [134]: output = pd.read_csv('tmp.csv', squeeze=True)
In [135]: output
Out[135]: 
Patient1     123000
Patient2      23000
Patient3    1234018
Name: level, dtype: int64
In [136]: type(output)
Out[136]: pandas.core.series.Series

Boolean values

The common values True, False, TRUE, and FALSE are all recognized as boolean. Occasionally you might want to recognize other values as being boolean. To do this, use the true_values and false_values options as follows:

In [137]: data = ('a,b,c\n'
   .....:         '1,Yes,2\n'
   .....:         '3,No,4')
   .....: 
In [138]: print(data)
a,b,c
1,Yes,2
3,No,4
In [139]: pd.read_csv(StringIO(data))
Out[139]: 
   a    b  c
0  1  Yes  2
1  3   No  4
In [140]: pd.read_csv(StringIO(data), true_values=['Yes'], false_values=['No'])
Out[140]: 
   a      b  c
0  1   True  2
1  3  False  4