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.
reader
pandas.read_csv()
writer
DataFrame.to_csv()
readers
writers
Format Type
Data Description
Reader
Writer
text
CSV
read_csv
to_csv
Fixed-Width Text File
read_fwf
JSON
read_json
to_json
HTML
read_html
to_html
Local clipboard
read_clipboard
to_clipboard
binary
MS Excel
read_excel
to_excel
OpenDocument
HDF5 Format
read_hdf
to_hdf
Feather Format
read_feather
to_feather
Parquet Format
read_parquet
to_parquet
ORC Format
read_orc
Msgpack
read_msgpack
to_msgpack
Stata
read_stata
to_stata
SAS
read_sas
SPSS
read_spss
Python Pickle Format
read_pickle
to_pickle
SQL
read_sql
to_sql
Google BigQuery
read_gbq
to_gbq
Here is an informal performance comparison for some of these IO methods.
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.
StringIO
from StringIO import StringIO
from io import StringIO
The workhorse function for reading text files (a.k.a. flat files) is read_csv(). See the cookbook for some advanced strategies.
read_csv()
read_csv() accepts the following common arguments:
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).
str
pathlib.Path
py._path.local.LocalPath
read()
','
\t
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'.
None
csv.Sniffer
'\s+'
'\\r\\t'
Alternative argument name for sep.
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.
' '
'\t'
sep='\s+'
True
delimiter
'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.
header=0
header=None
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.
[0,1,3]
skip_blank_lines=True
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.
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.
DataFrame
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.
index_col=False
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'].
[0, 1, 2]
['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.
usecols=[0, 1]
[1, 0]
data
pd.read_csv(data, usecols=['foo', 'bar'])[['foo', 'bar']]
['foo', 'bar']
pd.read_csv(data, usecols=['foo', 'bar'])[['bar', 'foo']]
['bar', 'foo']
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\n' ...: 'a,b,1\n' ...: 'a,b,2\n' ...: 'c,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.
False
If the parsed data only contains one column then return a Series.
Series
Prefix to add to column numbers when no header, e.g. ‘X’ for X0, X1, …
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.
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.
{'a': np.float64, 'b': np.int32}
engine='python'
na_values
'c'
'python'
Parser engine to use. The C engine is faster while the Python engine is currently more feature-complete.
Dict of functions for converting values in certain columns. Keys can either be integers or column labels.
Values to consider as True.
Values to consider as False.
Skip spaces after delimiter.
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 [6]: data = ('col1,col2,col3\n' ...: 'a,b,1\n' ...: 'a,b,2\n' ...: 'c,d,3') ...: In [7]: pd.read_csv(StringIO(data)) Out[7]: col1 col2 col3 0 a b 1 1 a b 2 2 c d 3 In [8]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0) Out[8]: col1 col2 col3 0 a b 2
0
Number of lines at bottom of file to skip (unsupported with engine=’c’).
Number of rows of file to read. Useful for reading pieces of large files.
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)
dtype
chunksize
iterator
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.
filepath_or_buffer
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.
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.
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.
na_filter=False
Indicate number of NA values placed in non-numeric columns.
If True, skip over blank lines rather than interpreting as NaN values.
If True -> try parsing the index.
If [1, 2, 3] -> try parsing columns 1, 2, 3 each as a separate date column.
[1, 2, 3]
If [[1, 3]] -> combine columns 1 and 3 and parse as a single date column.
[[1, 3]]
If {'foo': [1, 3]} -> parse columns 1, 3 as date and call result ‘foo’. A fast-path exists for iso8601-formatted dates.
{'foo': [1, 3]}
If True and parse_dates is enabled for a column, attempt to infer the datetime format to speed up the processing.
If True and parse_dates specifies combining multiple columns then keep the original columns.
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.
dateutil.parser.parser
DD/MM format dates, international and European format.
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.
Return TextFileReader object for iteration or getting chunks with get_chunk().
get_chunk()
Return TextFileReader object for iteration. See iterating and chunking below.
'gzip'
'bz2'
'zip'
'xz'
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.
Changed in version 0.24.0: ‘infer’ option added and set to default.
Thousands separator.
'.'
Character to recognize as decimal point. E.g. use ',' for European data.
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.
high
round_trip
Character to break file into lines. Only valid with C parser.
The character used to denote the start and end of a quoted item. Quoted items can include the delimiter and it will be ignored.
csv.QUOTE_*
Control field quoting behavior per csv.QUOTE_* constants. Use one of QUOTE_MINIMAL (0), QUOTE_ALL (1), QUOTE_NONNUMERIC (2) or QUOTE_NONE (3).
QUOTE_MINIMAL
QUOTE_ALL
QUOTE_NONNUMERIC
QUOTE_NONE
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.
quotechar
quoting
One-character string used to escape delimiter when quoting is QUOTE_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.
comment='#'
Encoding to use for UTF when reading/writing (e.g. 'utf-8'). List of Python standard encodings.
'utf-8'
csv.Dialect
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.
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.
If error_bad_lines is False, and warn_bad_lines is True, a warning for each “bad line” will be output.
You can indicate the data type for the whole DataFrame or individual columns:
In [9]: import numpy as np In [10]: data = ('a,b,c,d\n' ....: '1,2,3,4\n' ....: '5,6,7,8\n' ....: '9,10,11') ....: In [11]: print(data) a,b,c,d 1,2,3,4 5,6,7,8 9,10,11 In [12]: df = pd.read_csv(StringIO(data), dtype=object) In [13]: df Out[13]: a b c d 0 1 2 3 4 1 5 6 7 8 2 9 10 11 NaN In [14]: df['a'][0] Out[14]: '1' In [15]: df = pd.read_csv(StringIO(data), ....: dtype={'b': object, 'c': np.float64, 'd': 'Int64'}) ....: In [16]: df.dtypes Out[16]: 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.
object
For instance, you can use the converters argument of read_csv():
converters
In [17]: data = ("col_1\n" ....: "1\n" ....: "2\n" ....: "'A'\n" ....: "4.22") ....: In [18]: df = pd.read_csv(StringIO(data), converters={'col_1': str}) In [19]: df Out[19]: col_1 0 1 1 2 2 'A' 3 4.22 In [20]: df['col_1'].apply(type).value_counts() Out[20]: <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,
to_numeric()
In [21]: df2 = pd.read_csv(StringIO(data)) In [22]: df2['col_1'] = pd.to_numeric(df2['col_1'], errors='coerce') In [23]: df2 Out[23]: col_1 0 1.00 1 2.00 2 NaN 3 4.22 In [24]: df2['col_1'].apply(type).value_counts() Out[24]: <class 'float'> 4 Name: col_1, dtype: int64
which will convert all valid parsing to floats, leaving the invalid parsing as NaN.
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.
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 [25]: col_1 = list(range(500000)) + ['a', 'b'] + list(range(500000)) In [26]: df = pd.DataFrame({'col_1': col_1}) In [27]: df.to_csv('foo.csv') In [28]: mixed_df = pd.read_csv('foo.csv') In [29]: mixed_df['col_1'].apply(type).value_counts() Out[29]: <class 'int'> 737858 <class 'str'> 262144 Name: col_1, dtype: int64 In [30]: mixed_df['col_1'].dtype Out[30]: 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.
int
Categorical columns can be parsed directly by specifying dtype='category' or dtype=CategoricalDtype(categories, ordered).
Categorical
dtype='category'
dtype=CategoricalDtype(categories, ordered)
In [31]: data = ('col1,col2,col3\n' ....: 'a,b,1\n' ....: 'a,b,2\n' ....: 'c,d,3') ....: In [32]: pd.read_csv(StringIO(data)) Out[32]: col1 col2 col3 0 a b 1 1 a b 2 2 c d 3 In [33]: pd.read_csv(StringIO(data)).dtypes Out[33]: col1 object col2 object col3 int64 dtype: object In [34]: pd.read_csv(StringIO(data), dtype='category').dtypes Out[34]: col1 category col2 category col3 category dtype: object
Individual columns can be parsed as a Categorical using a dict specification:
In [35]: pd.read_csv(StringIO(data), dtype={'col1': 'category'}).dtypes Out[35]: 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.
categories
CategoricalDtype
In [36]: from pandas.api.types import CategoricalDtype In [37]: dtype = CategoricalDtype(['d', 'c', 'b', 'a'], ordered=True) In [38]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).dtypes Out[38]: col1 category col2 object col3 int64 dtype: object
When using dtype=CategoricalDtype, “unexpected” values outside of dtype.categories are treated as missing values.
dtype=CategoricalDtype
dtype.categories
In [39]: dtype = CategoricalDtype(['a', 'b', 'd']) # No 'c' In [40]: pd.read_csv(StringIO(data), dtype={'col1': dtype}).col1 Out[40]: 0 a 1 a 2 NaN Name: col1, dtype: category Categories (3, object): [a, b, d]
This matches the behavior of Categorical.set_categories().
Categorical.set_categories()
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().
to_datetime()
When dtype is a CategoricalDtype with homogeneous categories ( all numeric, all datetimes, etc.), the conversion is done automatically.
In [41]: df = pd.read_csv(StringIO(data), dtype='category') In [42]: df.dtypes Out[42]: col1 category col2 category col3 category dtype: object In [43]: df['col3'] Out[43]: 0 1 1 2 2 3 Name: col3, dtype: category Categories (3, object): [1, 2, 3] In [44]: df['col3'].cat.categories = pd.to_numeric(df['col3'].cat.categories) In [45]: df['col3'] Out[45]: 0 1 1 2 2 3 Name: col3, dtype: category Categories (3, int64): [1, 2, 3]
A file may or may not have a header row. pandas assumes the first row should be used as the column names:
In [46]: data = ('a,b,c\n' ....: '1,2,3\n' ....: '4,5,6\n' ....: '7,8,9') ....: In [47]: print(data) a,b,c 1,2,3 4,5,6 7,8,9 In [48]: pd.read_csv(StringIO(data)) Out[48]: 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):
names
header
In [49]: print(data) a,b,c 1,2,3 4,5,6 7,8,9 In [50]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=0) Out[50]: foo bar baz 0 1 2 3 1 4 5 6 2 7 8 9 In [51]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=None) Out[51]: 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 [52]: data = ('skip this skip it\n' ....: 'a,b,c\n' ....: '1,2,3\n' ....: '4,5,6\n' ....: '7,8,9') ....: In [53]: pd.read_csv(StringIO(data), header=1) Out[53]: a b c 0 1 2 3 1 4 5 6 2 7 8 9
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.
If the file or header contains duplicate names, pandas will by default distinguish between them so as to prevent overwriting data:
In [54]: data = ('a,b,a\n' ....: '0,1,2\n' ....: '3,4,5') ....: In [55]: pd.read_csv(StringIO(data)) Out[55]: 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:
mangle_dupe_cols=True
mangle_dupe_cols=False
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:
ValueError
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
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 [56]: data = 'a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz' In [57]: pd.read_csv(StringIO(data)) Out[57]: a b c d 0 1 2 3 foo 1 4 5 6 bar 2 7 8 9 baz In [58]: pd.read_csv(StringIO(data), usecols=['b', 'd']) Out[58]: b d 0 2 foo 1 5 bar 2 8 baz In [59]: pd.read_csv(StringIO(data), usecols=[0, 2, 3]) Out[59]: a c d 0 1 3 foo 1 4 6 bar 2 7 9 baz In [60]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ['A', 'C']) Out[60]: 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 [61]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ['a', 'c']) Out[61]: 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.
If the comment parameter is specified, then completely commented lines will be ignored. By default, completely blank lines will be ignored as well.
comment
In [62]: data = ('\n' ....: 'a,b,c\n' ....: ' \n' ....: '# commented line\n' ....: '1,2,3\n' ....: '\n' ....: '4,5,6') ....: In [63]: print(data) a,b,c # commented line 1,2,3 4,5,6 In [64]: pd.read_csv(StringIO(data), comment='#') Out[64]: a b c 0 1 2 3 1 4 5 6
If skip_blank_lines=False, then read_csv will not ignore blank lines:
skip_blank_lines=False
In [65]: data = ('a,b,c\n' ....: '\n' ....: '1,2,3\n' ....: '\n' ....: '\n' ....: '4,5,6') ....: In [66]: pd.read_csv(StringIO(data), skip_blank_lines=False) Out[66]: 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
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):
skiprows
In [67]: data = ('#comment\n' ....: 'a,b,c\n' ....: 'A,B,C\n' ....: '1,2,3') ....: In [68]: pd.read_csv(StringIO(data), comment='#', header=1) Out[68]: A B C 0 1 2 3 In [69]: data = ('A,B,C\n' ....: '#comment\n' ....: 'a,b,c\n' ....: '1,2,3') ....: In [70]: pd.read_csv(StringIO(data), comment='#', skiprows=2) Out[70]: 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 [71]: 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 [72]: 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 [73]: pd.read_csv(StringIO(data), comment='#', skiprows=4, header=1) Out[73]: A B C 0 1.0 2.0 4.0 1 5.0 NaN 10.0
Sometimes comments or meta data may be included in a file:
In [74]: 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 [75]: df = pd.read_csv('tmp.csv') In [76]: df Out[76]: 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 [77]: df = pd.read_csv('tmp.csv', comment='#') In [78]: df Out[78]: ID level category 0 Patient1 123000 x 1 Patient2 23000 y 2 Patient3 1234018 z
The encoding argument should be used for encoded unicode data, which will result in byte strings being decoded to unicode in the result:
encoding
In [79]: from io import BytesIO In [80]: data = (b'word,length\n' ....: b'Tr\xc3\xa4umen,7\n' ....: b'Gr\xc3\xbc\xc3\x9fe,5') ....: In [81]: data = data.decode('utf8').encode('latin-1') In [82]: df = pd.read_csv(BytesIO(data), encoding='latin-1') In [83]: df Out[83]: word length 0 Träumen 7 1 Grüße 5 In [84]: df['word'][1] Out[84]: '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.
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 [85]: data = ('a,b,c\n' ....: '4,apple,bat,5.7\n' ....: '8,orange,cow,10') ....: In [86]: pd.read_csv(StringIO(data)) Out[86]: a b c 4 apple bat 5.7 8 orange cow 10.0
In [87]: data = ('index,a,b,c\n' ....: '4,apple,bat,5.7\n' ....: '8,orange,cow,10') ....: In [88]: pd.read_csv(StringIO(data), index_col=0) Out[88]: 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.
index_col
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 [89]: data = ('a,b,c\n' ....: '4,apple,bat,\n' ....: '8,orange,cow,') ....: In [90]: print(data) a,b,c 4,apple,bat, 8,orange,cow, In [91]: pd.read_csv(StringIO(data)) Out[91]: a b c 4 apple bat NaN 8 orange cow NaN In [92]: pd.read_csv(StringIO(data), index_col=False) Out[92]: 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 [93]: data = ('a,b,c\n' ....: '4,apple,bat,\n' ....: '8,orange,cow,') ....: In [94]: print(data) a,b,c 4,apple,bat, 8,orange,cow, In [95]: pd.read_csv(StringIO(data), usecols=['b', 'c']) Out[95]: b c 4 bat NaN 8 cow NaN In [96]: pd.read_csv(StringIO(data), usecols=['b', 'c'], index_col=0) Out[96]: b c 4 bat NaN 8 cow NaN
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.
parse_dates
date_parser
datetime
The simplest case is to just pass in parse_dates=True:
parse_dates=True
# Use a column as an index, and parse it as dates. In [97]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True) In [98]: df Out[98]: 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 [99]: df.index Out[99]: 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 [100]: 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 [101]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]]) In [102]: df Out[102]: 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:
keep_date_col
In [103]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]], .....: keep_date_col=True) .....: In [104]: df Out[104]: 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.
parse_dates=[1, 2]
parse_dates=[[1, 2]]
You can also use a dict to specify custom name 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) In [107]: df Out[107]: 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 [108]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]} In [109]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec, .....: index_col=0) # index is the nominal column .....: In [110]: df Out[110]: 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
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.
pd.read_csv
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.
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.
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 [111]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec, .....: date_parser=pd.io.date_converters.parse_date_time) .....: In [112]: df Out[112]: 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:
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'])).
date_parser(['2013', '2013'], ['1', '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'])).
date_parser(['2013 1', '2013 2'])
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.).
date_parser('2013', '1')
date_parser('2013', '2')
Note that performance-wise, you should try these methods of parsing dates in order:
Try to infer the format using infer_datetime_format=True (see section below).
infer_datetime_format=True
If you know the format, use pd.to_datetime(): date_parser=lambda x: pd.to_datetime(x, format=...).
pd.to_datetime()
date_parser=lambda x: pd.to_datetime(x, format=...)
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.
date_converters.py
generic_parser
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 [113]: content = """\ .....: a .....: 2000-01-01T00:00:00+05:00 .....: 2000-01-01T00:00:00+06:00""" .....: In [114]: df = pd.read_csv(StringIO(content), parse_dates=['a']) In [115]: df['a'] Out[115]: 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.
utc=True
In [116]: df = pd.read_csv(StringIO(content), parse_dates=['a'], .....: date_parser=lambda col: pd.to_datetime(col, utc=True)) .....: In [117]: df['a'] Out[117]: 0 1999-12-31 19:00:00+00:00 1 1999-12-31 18:00:00+00:00 Name: a, dtype: datetime64[ns, UTC]
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.
infer_datetime_format
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.
dayfirst
dayfirst=True
dayfirst=False
# Try to infer the format for the index column In [118]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True, .....: infer_datetime_format=True) .....: In [119]: df Out[119]: A B C date 2009-01-01 a 1 2 2009-01-02 b 3 4 2009-01-03 c 4 5
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 [120]: print(open('tmp.csv').read()) date,value,cat 1/6/2000,5,a 2/6/2000,10,b 3/6/2000,15,c In [121]: pd.read_csv('tmp.csv', parse_dates=[0]) Out[121]: date value cat 0 2000-01-06 5 a 1 2000-02-06 10 b 2 2000-03-06 15 c In [122]: pd.read_csv('tmp.csv', dayfirst=True, parse_dates=[0]) Out[122]: date value cat 0 2000-06-01 5 a 1 2000-06-02 10 b 2 2000-06-03 15 c
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:
float_precision
In [123]: val = '0.3066101993807095471566981359501369297504425048828125' In [124]: data = 'a,b,c\n1,2,{0}'.format(val) In [125]: abs(pd.read_csv(StringIO(data), engine='c', .....: float_precision=None)['c'][0] - float(val)) .....: Out[125]: 1.1102230246251565e-16 In [126]: abs(pd.read_csv(StringIO(data), engine='c', .....: float_precision='high')['c'][0] - float(val)) .....: Out[126]: 5.551115123125783e-17 In [127]: abs(pd.read_csv(StringIO(data), engine='c', .....: float_precision='round_trip')['c'][0] - float(val)) .....: Out[127]: 0.0
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:
thousands
By default, numbers with a thousands separator will be parsed as strings:
In [128]: print(open('tmp.csv').read()) ID|level|category Patient1|123,000|x Patient2|23,000|y Patient3|1,234,018|z In [129]: df = pd.read_csv('tmp.csv', sep='|') In [130]: df Out[130]: ID level category 0 Patient1 123,000 x 1 Patient2 23,000 y 2 Patient3 1,234,018 z In [131]: df.level.dtype Out[131]: dtype('O')
The thousands keyword allows integers to be parsed correctly:
In [132]: print(open('tmp.csv').read()) ID|level|category Patient1|123,000|x Patient2|23,000|y Patient3|1,234,018|z In [133]: df = pd.read_csv('tmp.csv', sep='|', thousands=',') In [134]: df Out[134]: ID level category 0 Patient1 123000 x 1 Patient2 23000 y 2 Patient3 1234018 z In [135]: df.level.dtype Out[135]: dtype('int64')
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).
float
5.0
integer
5
[5.0, 5]
To completely override the default values that are recognized as missing, specify keep_default_na=False.
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>', '#NA', 'NULL', 'null', 'NaN', '-NaN', 'nan', '-nan', ''].
['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A N/A', '#N/A', 'N/A', 'n/a', 'NA', '<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.
NA
pd.read_csv('path_to_file.csv', na_values=["Nope"])
The default values, in addition to the string "Nope" are recognized as NaN.
"Nope"
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.
inf
np.inf
-inf
-np.inf
Inf
Using the squeeze keyword, the parser will return output with a single column as a Series:
squeeze
In [136]: print(open('tmp.csv').read()) level Patient1,123000 Patient2,23000 Patient3,1234018 In [137]: output = pd.read_csv('tmp.csv', squeeze=True) In [138]: output Out[138]: Patient1 123000 Patient2 23000 Patient3 1234018 Name: level, dtype: int64 In [139]: type(output) Out[139]: pandas.core.series.Series
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:
TRUE
FALSE
true_values
false_values
In [140]: data = ('a,b,c\n' .....: '1,Yes,2\n' .....: '3,No,4') .....: In [141]: print(data) a,b,c 1,Yes,2 3,No,4 In [142]: pd.read_csv(StringIO(data)) Out[142]: a b c 0 1 Yes 2 1 3 No 4 In [143]: pd.read_csv(StringIO(data), true_values=['Yes'], false_values=['No']) Out[143]: a b c 0 1 True 2 1 3 False 4
Some files may have malformed lines with too few fields or too many. Lines with too few fields will have NA values filled in the trailing fields. Lines with too many fields will raise an error by default:
In [144]: data = ('a,b,c\n' .....: '1,2,3\n' .....: '4,5,6,7\n' .....: '8,9,10') .....: In [145]: pd.read_csv(StringIO(data)) --------------------------------------------------------------------------- ParserError Traceback (most recent call last) <ipython-input-145-6388c394e6b8> in <module> ----> 1 pd.read_csv(StringIO(data)) ~/scipy/pandas/pandas/io/parsers.py in parser_f(filepath_or_buffer, sep, delimiter, header, names, index_col, usecols, squeeze, prefix, mangle_dupe_cols, dtype, engine, converters, true_values, false_values, skipinitialspace, skiprows, skipfooter, nrows, na_values, keep_default_na, na_filter, verbose, skip_blank_lines, parse_dates, infer_datetime_format, keep_date_col, date_parser, dayfirst, cache_dates, iterator, chunksize, compression, thousands, decimal, lineterminator, quotechar, quoting, doublequote, escapechar, comment, encoding, dialect, error_bad_lines, warn_bad_lines, delim_whitespace, low_memory, memory_map, float_precision) 673 ) 674 --> 675 return _read(filepath_or_buffer, kwds) 676 677 parser_f.__name__ = name ~/scipy/pandas/pandas/io/parsers.py in _read(filepath_or_buffer, kwds) 452 453 try: --> 454 data = parser.read(nrows) 455 finally: 456 parser.close() ~/scipy/pandas/pandas/io/parsers.py in read(self, nrows) 1127 def read(self, nrows=None): 1128 nrows = _validate_integer("nrows", nrows) -> 1129 ret = self._engine.read(nrows) 1130 1131 # May alter columns / col_dict ~/scipy/pandas/pandas/io/parsers.py in read(self, nrows) 2027 def read(self, nrows=None): 2028 try: -> 2029 data = self._reader.read(nrows) 2030 except StopIteration: 2031 if self._first_chunk: ~/scipy/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.TextReader.read() ~/scipy/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.TextReader._read_low_memory() ~/scipy/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.TextReader._read_rows() ~/scipy/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.TextReader._tokenize_rows() ~/scipy/pandas/pandas/_libs/parsers.pyx in pandas._libs.parsers.raise_parser_error() ParserError: Error tokenizing data. C error: Expected 3 fields in line 3, saw 4
You can elect to skip bad lines:
In [29]: pd.read_csv(StringIO(data), error_bad_lines=False) Skipping line 3: expected 3 fields, saw 4 Out[29]: a b c 0 1 2 3 1 8 9 10
You can also use the usecols parameter to eliminate extraneous column data that appear in some lines but not others:
In [30]: pd.read_csv(StringIO(data), usecols=[0, 1, 2]) Out[30]: a b c 0 1 2 3 1 4 5 6 2 8 9 10
The dialect keyword gives greater flexibility in specifying the file format. By default it uses the Excel dialect but you can specify either the dialect name or a csv.Dialect instance.
dialect
Suppose you had data with unenclosed quotes:
In [146]: print(data) label1,label2,label3 index1,"a,c,e index2,b,d,f
By default, read_csv uses the Excel dialect and treats the double quote as the quote character, which causes it to fail when it finds a newline before it finds the closing double quote.
We can get around this using dialect:
In [147]: import csv In [148]: dia = csv.excel() In [149]: dia.quoting = csv.QUOTE_NONE In [150]: pd.read_csv(StringIO(data), dialect=dia) Out[150]: label1 label2 label3 index1 "a c e index2 b d f
All of the dialect options can be specified separately by keyword arguments:
In [151]: data = 'a,b,c~1,2,3~4,5,6' In [152]: pd.read_csv(StringIO(data), lineterminator='~') Out[152]: a b c 0 1 2 3 1 4 5 6
Another common dialect option is skipinitialspace, to skip any whitespace after a delimiter:
skipinitialspace
In [153]: data = 'a, b, c\n1, 2, 3\n4, 5, 6' In [154]: print(data) a, b, c 1, 2, 3 4, 5, 6 In [155]: pd.read_csv(StringIO(data), skipinitialspace=True) Out[155]: a b c 0 1 2 3 1 4 5 6
The parsers make every attempt to “do the right thing” and not be fragile. Type inference is a pretty big deal. If a column can be coerced to integer dtype without altering the contents, the parser will do so. Any non-numeric columns will come through as object dtype as with the rest of pandas objects.
Quotes (and other escape characters) in embedded fields can be handled in any number of ways. One way is to use backslashes; to properly parse this data, you should pass the escapechar option:
escapechar
In [156]: data = 'a,b\n"hello, \\"Bob\\", nice to see you",5' In [157]: print(data) a,b "hello, \"Bob\", nice to see you",5 In [158]: pd.read_csv(StringIO(data), escapechar='\\') Out[158]: a b 0 hello, "Bob", nice to see you 5
While read_csv() reads delimited data, the read_fwf() function works with data files that have known and fixed column widths. The function parameters to read_fwf are largely the same as read_csv with two extra parameters, and a different usage of the delimiter parameter:
read_fwf()
colspecs: A list of pairs (tuples) giving the extents of the fixed-width fields of each line as half-open intervals (i.e., [from, to[ ). String value ‘infer’ can be used to instruct the parser to try detecting the column specifications from the first 100 rows of the data. Default behavior, if not specified, is to infer.
colspecs
widths: A list of field widths which can be used instead of ‘colspecs’ if the intervals are contiguous.
widths
delimiter: Characters to consider as filler characters in the fixed-width file. Can be used to specify the filler character of the fields if it is not spaces (e.g., ‘~’).
Consider a typical fixed-width data file:
In [159]: print(open('bar.csv').read()) id8141 360.242940 149.910199 11950.7 id1594 444.953632 166.985655 11788.4 id1849 364.136849 183.628767 11806.2 id1230 413.836124 184.375703 11916.8 id1948 502.953953 173.237159 12468.3
In order to parse this file into a DataFrame, we simply need to supply the column specifications to the read_fwf function along with the file name:
# Column specifications are a list of half-intervals In [160]: colspecs = [(0, 6), (8, 20), (21, 33), (34, 43)] In [161]: df = pd.read_fwf('bar.csv', colspecs=colspecs, header=None, index_col=0) In [162]: df Out[162]: 1 2 3 0 id8141 360.242940 149.910199 11950.7 id1594 444.953632 166.985655 11788.4 id1849 364.136849 183.628767 11806.2 id1230 413.836124 184.375703 11916.8 id1948 502.953953 173.237159 12468.3
Note how the parser automatically picks column names X.<column number> when header=None argument is specified. Alternatively, you can supply just the column widths for contiguous columns:
# Widths are a list of integers In [163]: widths = [6, 14, 13, 10] In [164]: df = pd.read_fwf('bar.csv', widths=widths, header=None) In [165]: df Out[165]: 0 1 2 3 0 id8141 360.242940 149.910199 11950.7 1 id1594 444.953632 166.985655 11788.4 2 id1849 364.136849 183.628767 11806.2 3 id1230 413.836124 184.375703 11916.8 4 id1948 502.953953 173.237159 12468.3
The parser will take care of extra white spaces around the columns so it’s ok to have extra separation between the columns in the file.
By default, read_fwf will try to infer the file’s colspecs by using the first 100 rows of the file. It can do it only in cases when the columns are aligned and correctly separated by the provided delimiter (default delimiter is whitespace).
In [166]: df = pd.read_fwf('bar.csv', header=None, index_col=0) In [167]: df Out[167]: 1 2 3 0 id8141 360.242940 149.910199 11950.7 id1594 444.953632 166.985655 11788.4 id1849 364.136849 183.628767 11806.2 id1230 413.836124 184.375703 11916.8 id1948 502.953953 173.237159 12468.3
read_fwf supports the dtype parameter for specifying the types of parsed columns to be different from the inferred type.
In [168]: pd.read_fwf('bar.csv', header=None, index_col=0).dtypes Out[168]: 1 float64 2 float64 3 float64 dtype: object In [169]: pd.read_fwf('bar.csv', header=None, dtype={2: 'object'}).dtypes Out[169]: 0 object 1 float64 2 object 3 float64 dtype: object
Consider a file with one less entry in the header than the number of data column:
In [170]: print(open('foo.csv').read()) A,B,C 20090101,a,1,2 20090102,b,3,4 20090103,c,4,5
In this special case, read_csv assumes that the first column is to be used as the index of the DataFrame:
In [171]: pd.read_csv('foo.csv') Out[171]: 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 [172]: df = pd.read_csv('foo.csv', parse_dates=True) In [173]: df.index Out[173]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', freq=None)
MultiIndex
Suppose you have data indexed by two columns:
In [174]: print(open('data/mindex_ex.csv').read()) year,indiv,zit,xit 1977,"A",1.2,.6 1977,"B",1.5,.5 1977,"C",1.7,.8 1978,"A",.2,.06 1978,"B",.7,.2 1978,"C",.8,.3 1978,"D",.9,.5 1978,"E",1.4,.9 1979,"C",.2,.15 1979,"D",.14,.05 1979,"E",.5,.15 1979,"F",1.2,.5 1979,"G",3.4,1.9 1979,"H",5.4,2.7 1979,"I",6.4,1.2
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 [175]: df = pd.read_csv("data/mindex_ex.csv", index_col=[0, 1]) In [176]: df Out[176]: zit xit year indiv 1977 A 1.20 0.60 B 1.50 0.50 C 1.70 0.80 1978 A 0.20 0.06 B 0.70 0.20 C 0.80 0.30 D 0.90 0.50 E 1.40 0.90 1979 C 0.20 0.15 D 0.14 0.05 E 0.50 0.15 F 1.20 0.50 G 3.40 1.90 H 5.40 2.70 I 6.40 1.20 In [177]: df.loc[1978] Out[177]: zit xit indiv A 0.2 0.06 B 0.7 0.20 C 0.8 0.30 D 0.9 0.50 E 1.4 0.90
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 [178]: from pandas._testing import makeCustomDataframe as mkdf In [179]: df = mkdf(5, 3, r_idx_nlevels=2, c_idx_nlevels=4) In [180]: df.to_csv('mi.csv') In [181]: 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 [182]: pd.read_csv('mi.csv', header=[0, 1, 2, 3], index_col=[0, 1]) Out[182]: 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 [183]: print(open('mi2.csv').read()) ,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 [184]: pd.read_csv('mi2.csv', header=[0, 1], index_col=0) Out[184]: 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.
df.to_csv(..., index=False)
read_csv is capable of inferring delimited (not necessarily comma-separated) files, as pandas uses the csv.Sniffer class of the csv module. For this, you have to specify sep=None.
sep=None
In [185]: print(open('tmp2.sv').read()) :0:1:2:3 0:0.4691122999071863:-0.2828633443286633:-1.5090585031735124:-1.1356323710171934 1:1.2121120250208506:-0.17321464905330858:0.11920871129693428:-1.0442359662799567 2:-0.8618489633477999:-2.1045692188948086:-0.4949292740687813:1.071803807037338 3:0.7215551622443669:-0.7067711336300845:-1.0395749851146963:0.27185988554282986 4:-0.42497232978883753:0.567020349793672:0.27623201927771873:-1.0874006912859915 5:-0.6736897080883706:0.1136484096888855:-1.4784265524372235:0.5249876671147047 6:0.4047052186802365:0.5770459859204836:-1.7150020161146375:-1.0392684835147725 7:-0.3706468582364464:-1.1578922506419993:-1.344311812731667:0.8448851414248841 8:1.0757697837155533:-0.10904997528022223:1.6435630703622064:-1.4693879595399115 9:0.35702056413309086:-0.6746001037299882:-1.776903716971867:-0.9689138124473498 In [186]: pd.read_csv('tmp2.sv', sep=None, engine='python') Out[186]: Unnamed: 0 0 1 2 3 0 0 0.469112 -0.282863 -1.509059 -1.135632 1 1 1.212112 -0.173215 0.119209 -1.044236 2 2 -0.861849 -2.104569 -0.494929 1.071804 3 3 0.721555 -0.706771 -1.039575 0.271860 4 4 -0.424972 0.567020 0.276232 -1.087401 5 5 -0.673690 0.113648 -1.478427 0.524988 6 6 0.404705 0.577046 -1.715002 -1.039268 7 7 -0.370647 -1.157892 -1.344312 0.844885 8 8 1.075770 -0.109050 1.643563 -1.469388 9 9 0.357021 -0.674600 -1.776904 -0.968914
It’s best to use concat() to combine multiple files. See the cookbook for an example.
concat()
Suppose you wish to iterate through a (potentially very large) file lazily rather than reading the entire file into memory, such as the following:
In [187]: print(open('tmp.sv').read()) |0|1|2|3 0|0.4691122999071863|-0.2828633443286633|-1.5090585031735124|-1.1356323710171934 1|1.2121120250208506|-0.17321464905330858|0.11920871129693428|-1.0442359662799567 2|-0.8618489633477999|-2.1045692188948086|-0.4949292740687813|1.071803807037338 3|0.7215551622443669|-0.7067711336300845|-1.0395749851146963|0.27185988554282986 4|-0.42497232978883753|0.567020349793672|0.27623201927771873|-1.0874006912859915 5|-0.6736897080883706|0.1136484096888855|-1.4784265524372235|0.5249876671147047 6|0.4047052186802365|0.5770459859204836|-1.7150020161146375|-1.0392684835147725 7|-0.3706468582364464|-1.1578922506419993|-1.344311812731667|0.8448851414248841 8|1.0757697837155533|-0.10904997528022223|1.6435630703622064|-1.4693879595399115 9|0.35702056413309086|-0.6746001037299882|-1.776903716971867|-0.9689138124473498 In [188]: table = pd.read_csv('tmp.sv', sep='|') In [189]: table Out[189]: Unnamed: 0 0 1 2 3 0 0 0.469112 -0.282863 -1.509059 -1.135632 1 1 1.212112 -0.173215 0.119209 -1.044236 2 2 -0.861849 -2.104569 -0.494929 1.071804 3 3 0.721555 -0.706771 -1.039575 0.271860 4 4 -0.424972 0.567020 0.276232 -1.087401 5 5 -0.673690 0.113648 -1.478427 0.524988 6 6 0.404705 0.577046 -1.715002 -1.039268 7 7 -0.370647 -1.157892 -1.344312 0.844885 8 8 1.075770 -0.109050 1.643563 -1.469388 9 9 0.357021 -0.674600 -1.776904 -0.968914
By specifying a chunksize to read_csv, the return value will be an iterable object of type TextFileReader:
TextFileReader
In [190]: reader = pd.read_csv('tmp.sv', sep='|', chunksize=4) In [191]: reader Out[191]: <pandas.io.parsers.TextFileReader at 0x7f6584e5fba8> In [192]: for chunk in reader: .....: print(chunk) .....: Unnamed: 0 0 1 2 3 0 0 0.469112 -0.282863 -1.509059 -1.135632 1 1 1.212112 -0.173215 0.119209 -1.044236 2 2 -0.861849 -2.104569 -0.494929 1.071804 3 3 0.721555 -0.706771 -1.039575 0.271860 Unnamed: 0 0 1 2 3 4 4 -0.424972 0.567020 0.276232 -1.087401 5 5 -0.673690 0.113648 -1.478427 0.524988 6 6 0.404705 0.577046 -1.715002 -1.039268 7 7 -0.370647 -1.157892 -1.344312 0.844885 Unnamed: 0 0 1 2 3 8 8 1.075770 -0.10905 1.643563 -1.469388 9 9 0.357021 -0.67460 -1.776904 -0.968914
Specifying iterator=True will also return the TextFileReader object:
iterator=True
In [193]: reader = pd.read_csv('tmp.sv', sep='|', iterator=True) In [194]: reader.get_chunk(5) Out[194]: Unnamed: 0 0 1 2 3 0 0 0.469112 -0.282863 -1.509059 -1.135632 1 1 1.212112 -0.173215 0.119209 -1.044236 2 2 -0.861849 -2.104569 -0.494929 1.071804 3 3 0.721555 -0.706771 -1.039575 0.271860 4 4 -0.424972 0.567020 0.276232 -1.087401
Under the hood pandas uses a fast and efficient parser implemented in C as well as a Python implementation which is currently more feature-complete. Where possible pandas uses the C parser (specified as engine='c'), but may fall back to Python if C-unsupported options are specified. Currently, C-unsupported options include:
engine='c'
sep other than a single character (e.g. regex separators)
sep
skipfooter
sep=None with delim_whitespace=False
delim_whitespace=False
Specifying any of the above options will produce a ParserWarning unless the python engine is selected explicitly using engine='python'.
ParserWarning
You can pass in a URL to a CSV file:
df = pd.read_csv('https://download.bls.gov/pub/time.series/cu/cu.item', sep='\t')
S3 URLs are handled as well but require installing the S3Fs library:
df = pd.read_csv('s3://pandas-test/tips.csv')
If your S3 bucket requires credentials you will need to set them as environment variables or in the ~/.aws/credentials config file, refer to the S3Fs documentation on credentials.
~/.aws/credentials
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=’‘
path_or_buf
sep : Field delimiter for the output file (default “,”)
na_rep: A string representation of a missing value (default ‘’)
na_rep
float_format: Format string for floating point numbers
float_format
columns: Columns to write (default None)
columns
header: Whether to write out the column names (default True)
index: whether to write row (index) names (default True)
index
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).
index_label
mode : Python write mode, default ‘w’
mode
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)
line_terminator
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)
doublequote
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
date_format
The DataFrame object has an instance method to_string which allows control over the string representation of the object. All arguments are optional:
to_string
buf default None, for example a StringIO object
buf
columns default None, which columns to write
col_space default None, minimum width of each column.
col_space
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
formatters
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.
sparsify
index_names default True, will print the names of the indices
index_names
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
justify
left
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.
length
Read and write JSON format files and strings.
A Series or DataFrame can be converted to a valid JSON string. Use to_json with optional parameters:
path_or_buf : the pathname or buffer to write the output This can be None in which case a JSON string is returned
orient :
orient
default is index
allowed values are {split, records, index}
split
records
default is columns
allowed values are {split, records, index, columns, values, table}
values
table
The format of the JSON string
dict like {index -> [index], columns -> [columns], data -> [values]}
list like [{column -> value}, … , {column -> value}]
dict like {index -> {column -> value}}
dict like {column -> {index -> value}}
just the values array
date_format : string, type of date conversion, ‘epoch’ for timestamp, ‘iso’ for ISO8601.
double_precision : The number of decimal places to use when encoding floating point values, default 10.
double_precision
force_ascii : force encoded string to be ASCII, default True.
force_ascii
date_unit : The time unit to encode to, governs timestamp and ISO8601 precision. One of ‘s’, ‘ms’, ‘us’ or ‘ns’ for seconds, milliseconds, microseconds and nanoseconds respectively. Default ‘ms’.
date_unit
default_handler : The handler to call if an object cannot otherwise be converted to a suitable format for JSON. Takes a single argument, which is the object to convert, and returns a serializable object.
default_handler
lines : If records orient, then will write each record per line as json.
lines
Note NaN’s, NaT’s and None will be converted to null and datetime objects will be converted based on the date_format and date_unit parameters.
NaT
null
In [195]: dfj = pd.DataFrame(np.random.randn(5, 2), columns=list('AB')) In [196]: json = dfj.to_json() In [197]: json Out[197]: '{"A":{"0":-1.2945235903,"1":0.2766617129,"2":-0.0139597524,"3":-0.0061535699,"4":0.8957173022},"B":{"0":0.4137381054,"1":-0.472034511,"2":-0.3625429925,"3":-0.923060654,"4":0.8052440254}}'
There are a number of different options for the format of the resulting JSON file / string. Consider the following DataFrame and Series:
In [198]: dfjo = pd.DataFrame(dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)), .....: columns=list('ABC'), index=list('xyz')) .....: In [199]: dfjo Out[199]: A B C x 1 4 7 y 2 5 8 z 3 6 9 In [200]: sjo = pd.Series(dict(x=15, y=16, z=17), name='D') In [201]: sjo Out[201]: 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 [202]: dfjo.to_json(orient="columns") Out[202]: '{"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 [203]: dfjo.to_json(orient="index") Out[203]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,"C":9}}' In [204]: sjo.to_json(orient="index") Out[204]: '{"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:
d3.js
In [205]: dfjo.to_json(orient="records") Out[205]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]' In [206]: sjo.to_json(orient="records") Out[206]: '[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 [207]: dfjo.to_json(orient="values") Out[207]: '[[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 [208]: dfjo.to_json(orient="split") Out[208]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,5,8],[3,6,9]]}' In [209]: sjo.to_json(orient="split") Out[209]: '{"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.
Any orient option that encodes to a JSON object will not preserve the ordering of index and column labels during round-trip serialization. If you wish to preserve label ordering use the split option as it uses ordered containers.
Writing in ISO date format:
In [210]: dfd = pd.DataFrame(np.random.randn(5, 2), columns=list('AB')) In [211]: dfd['date'] = pd.Timestamp('20130101') In [212]: dfd = dfd.sort_index(1, ascending=False) In [213]: json = dfd.to_json(date_format='iso') In [214]: json Out[214]: '{"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":2.5656459463,"1":1.3403088498,"2":-0.2261692849,"3":0.8138502857,"4":-0.8273169356},"A":{"0":-1.2064117817,"1":1.4312559863,"2":-1.1702987971,"3":0.4108345112,"4":0.1320031703}}'
Writing in ISO date format, with microseconds:
In [215]: json = dfd.to_json(date_format='iso', date_unit='us') In [216]: json Out[216]: '{"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":2.5656459463,"1":1.3403088498,"2":-0.2261692849,"3":0.8138502857,"4":-0.8273169356},"A":{"0":-1.2064117817,"1":1.4312559863,"2":-1.1702987971,"3":0.4108345112,"4":0.1320031703}}'
Epoch timestamps, in seconds:
In [217]: json = dfd.to_json(date_format='epoch', date_unit='s') In [218]: json Out[218]: '{"date":{"0":1356998400,"1":1356998400,"2":1356998400,"3":1356998400,"4":1356998400},"B":{"0":2.5656459463,"1":1.3403088498,"2":-0.2261692849,"3":0.8138502857,"4":-0.8273169356},"A":{"0":-1.2064117817,"1":1.4312559863,"2":-1.1702987971,"3":0.4108345112,"4":0.1320031703}}'
Writing to a file, with a date index and a date column:
In [219]: dfj2 = dfj.copy() In [220]: dfj2['date'] = pd.Timestamp('20130101') In [221]: dfj2['ints'] = list(range(5)) In [222]: dfj2['bools'] = True In [223]: dfj2.index = pd.date_range('20130101', periods=5) In [224]: dfj2.to_json('test.json') In [225]: with open('test.json') as fh: .....: print(fh.read()) .....: {"A":{"1356998400000":-1.2945235903,"1357084800000":0.2766617129,"1357171200000":-0.0139597524,"1357257600000":-0.0061535699,"1357344000000":0.8957173022},"B":{"1356998400000":0.4137381054,"1357084800000":-0.472034511,"1357171200000":-0.3625429925,"1357257600000":-0.923060654,"1357344000000":0.8052440254},"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}}
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.
np.complex
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.
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.
toDict
dict
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.
OverflowError
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 [226]: pd.DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json(default_handler=str) Out[226]: '{"0":{"0":"(1+0j)","1":"(2+0j)","2":"(1+2j)"}}'
Reading a JSON string to pandas object can take a number of parameters. The parser will try to parse a DataFrame if typ is not supplied or is None. To explicitly force Series parsing, pass typ=series
typ
typ=series
filepath_or_buffer : a VALID JSON string or file handle / StringIO. The string could be a URL. Valid URL schemes include http, ftp, S3, and file. For file URLs, a host is expected. For instance, a local file could be file ://localhost/path/to/table.json
typ : type of object to recover (series or frame), default ‘frame’
adhering to the JSON Table Schema
dtype : if True, infer dtypes, if a dict of column to dtype, then use those, if False, then don’t infer dtypes at all, default is True, apply only to the data.
convert_axes : boolean, try to convert the axes to the proper dtypes, default is True
convert_axes
convert_dates : a list of columns to parse for dates; If True, then try to parse date-like columns, default is True.
convert_dates
keep_default_dates : boolean, default True. If parsing dates, then parse the default date-like columns.
keep_default_dates
numpy : direct decoding to NumPy arrays. default is False; Supports numeric data only, although labels may be non-numeric. Also note that the JSON ordering MUST be the same for each term if numpy=True.
numpy
numpy=True
precise_float : boolean, default False. Set to enable usage of higher precision (strtod) function when decoding string to double values. Default (False) is to use fast but less precise builtin functionality.
precise_float
date_unit : string, the timestamp unit to detect if converting dates. Default None. By default the timestamp precision will be detected, if this is not desired then pass one of ‘s’, ‘ms’, ‘us’ or ‘ns’ to force timestamp precision to seconds, milliseconds, microseconds or nanoseconds respectively.
lines : reads file as one json object per line.
encoding : The encoding to use to decode py3 bytes.
chunksize : when used in combination with lines=True, return a JsonReader which reads in chunksize lines per iteration.
lines=True
The parser will raise one of ValueError/TypeError/AssertionError if the JSON is not parseable.
ValueError/TypeError/AssertionError
If a non-default orient was used when encoding to JSON be sure to pass the same option here so that decoding produces sensible results, see Orient Options for an overview.
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.
convert_axes=True
dtype=True
convert_dates=True
Large integer values may be converted to dates if convert_dates=True and the data and / or column labels appear ‘date-like’. The exact threshold depends on the date_unit specified. ‘date-like’ means that the column label meets one of the following criteria:
it ends with '_at' it ends with '_time' it begins with 'timestamp' it is 'modified' it is 'date'
it ends with '_at'
'_at'
it ends with '_time'
'_time'
it begins with 'timestamp'
'timestamp'
it is 'modified'
'modified'
it is 'date'
'date'
When reading JSON data, automatic coercing into dtypes has some quirks:
an index can be reconstructed in a different order from serialization, that is, the returned order is not guaranteed to be the same as before serialization a column that was float data will be converted to integer if it can be done safely, e.g. a column of 1. bool columns will be converted to integer on reconstruction
an index can be reconstructed in a different order from serialization, that is, the returned order is not guaranteed to be the same as before serialization
a column that was float data will be converted to integer if it can be done safely, e.g. a column of 1.
1.
bool columns will be converted to integer on reconstruction
Thus there are times where you may want to specify specific dtypes via the dtype keyword argument.
Reading from a JSON string:
In [227]: pd.read_json(json) Out[227]: date B A 0 2013-01-01 2.565646 -1.206412 1 2013-01-01 1.340309 1.431256 2 2013-01-01 -0.226169 -1.170299 3 2013-01-01 0.813850 0.410835 4 2013-01-01 -0.827317 0.132003
Reading from a file:
In [228]: pd.read_json('test.json') Out[228]: A B date ints bools 2013-01-01 -1.294524 0.413738 2013-01-01 0 True 2013-01-02 0.276662 -0.472035 2013-01-01 1 True 2013-01-03 -0.013960 -0.362543 2013-01-01 2 True 2013-01-04 -0.006154 -0.923061 2013-01-01 3 True 2013-01-05 0.895717 0.805244 2013-01-01 4 True
Don’t convert any data (but still convert axes and dates):
In [229]: pd.read_json('test.json', dtype=object).dtypes Out[229]: A object B object date object ints object bools object dtype: object
Specify dtypes for conversion:
In [230]: pd.read_json('test.json', dtype={'A': 'float32', 'bools': 'int8'}).dtypes Out[230]: A float32 B float64 date datetime64[ns] ints int64 bools int8 dtype: object
Preserve string indices:
In [231]: si = pd.DataFrame(np.zeros((4, 4)), columns=list(range(4)), .....: index=[str(i) for i in range(4)]) .....: In [232]: si Out[232]: 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 [233]: si.index Out[233]: Index(['0', '1', '2', '3'], dtype='object') In [234]: si.columns Out[234]: Int64Index([0, 1, 2, 3], dtype='int64') In [235]: json = si.to_json() In [236]: sij = pd.read_json(json, convert_axes=False) In [237]: sij Out[237]: 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 [238]: sij.index Out[238]: Index(['0', '1', '2', '3'], dtype='object') In [239]: sij.columns Out[239]: Index(['0', '1', '2', '3'], dtype='object')
Dates written in nanoseconds need to be read back in nanoseconds:
In [240]: json = dfj2.to_json(date_unit='ns') # Try to parse timestamps as milliseconds -> Won't Work In [241]: dfju = pd.read_json(json, date_unit='ms') In [242]: dfju Out[242]: A B date ints bools 1356998400000000000 -1.294524 0.413738 1356998400000000000 0 True 1357084800000000000 0.276662 -0.472035 1356998400000000000 1 True 1357171200000000000 -0.013960 -0.362543 1356998400000000000 2 True 1357257600000000000 -0.006154 -0.923061 1356998400000000000 3 True 1357344000000000000 0.895717 0.805244 1356998400000000000 4 True # Let pandas detect the correct precision In [243]: dfju = pd.read_json(json) In [244]: dfju Out[244]: A B date ints bools 2013-01-01 -1.294524 0.413738 2013-01-01 0 True 2013-01-02 0.276662 -0.472035 2013-01-01 1 True 2013-01-03 -0.013960 -0.362543 2013-01-01 2 True 2013-01-04 -0.006154 -0.923061 2013-01-01 3 True 2013-01-05 0.895717 0.805244 2013-01-01 4 True # Or specify that all timestamps are in nanoseconds In [245]: dfju = pd.read_json(json, date_unit='ns') In [246]: dfju Out[246]: A B date ints bools 2013-01-01 -1.294524 0.413738 2013-01-01 0 True 2013-01-02 0.276662 -0.472035 2013-01-01 1 True 2013-01-03 -0.013960 -0.362543 2013-01-01 2 True 2013-01-04 -0.006154 -0.923061 2013-01-01 3 True 2013-01-05 0.895717 0.805244 2013-01-01 4 True
This param has been deprecated as of version 1.0.0 and will raise a FutureWarning.
FutureWarning
This supports numeric data only. Index and columns labels may be non-numeric, e.g. strings, dates etc.
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 [247]: randfloats = np.random.uniform(-100, 1000, 10000) In [248]: randfloats.shape = (1000, 10) In [249]: dffloats = pd.DataFrame(randfloats, columns=list('ABCDEFGHIJ')) In [250]: jsonfloats = dffloats.to_json()
In [251]: %timeit pd.read_json(jsonfloats) 11.9 ms +- 1.25 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
In [252]: %timeit pd.read_json(jsonfloats, numpy=True) 9.77 ms +- 1.5 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
The speedup is less noticeable for smaller datasets:
In [253]: jsonfloats = dffloats.head(100).to_json()
In [254]: %timeit pd.read_json(jsonfloats) 7.48 ms +- 1.34 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
In [255]: %timeit pd.read_json(jsonfloats, numpy=True) 6.1 ms +- 404 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
Direct NumPy decoding makes a number of assumptions and may fail or produce unexpected output if these assumptions are not satisfied:
data is numeric. data is uniform. The dtype is sniffed from the first value decoded. A ValueError may be raised, or incorrect output may be produced if this condition is not satisfied. labels are ordered. Labels are only read from the first container, it is assumed that each subsequent row / column has been encoded in the same order. This should be satisfied if the data was encoded using to_json but may not be the case if the JSON is from another source.
data is numeric.
data is uniform. The dtype is sniffed from the first value decoded. A ValueError may be raised, or incorrect output may be produced if this condition is not satisfied.
labels are ordered. Labels are only read from the first container, it is assumed that each subsequent row / column has been encoded in the same order. This should be satisfied if the data was encoded using to_json but may not be the case if the JSON is from another source.
pandas provides a utility function to take a dict or list of dicts and normalize this semi-structured data into a flat table.
In [256]: data = [{'id': 1, 'name': {'first': 'Coleen', 'last': 'Volk'}}, .....: {'name': {'given': 'Mose', 'family': 'Regner'}}, .....: {'id': 2, 'name': 'Faye Raker'}] .....: In [257]: pd.json_normalize(data) Out[257]: id name.first name.last name.given name.family name 0 1.0 Coleen Volk NaN NaN NaN 1 NaN NaN NaN Mose Regner NaN 2 2.0 NaN NaN NaN NaN Faye Raker
In [258]: data = [{'state': 'Florida', .....: 'shortname': 'FL', .....: 'info': {'governor': 'Rick Scott'}, .....: 'county': [{'name': 'Dade', 'population': 12345}, .....: {'name': 'Broward', 'population': 40000}, .....: {'name': 'Palm Beach', 'population': 60000}]}, .....: {'state': 'Ohio', .....: 'shortname': 'OH', .....: 'info': {'governor': 'John Kasich'}, .....: 'county': [{'name': 'Summit', 'population': 1234}, .....: {'name': 'Cuyahoga', 'population': 1337}]}] .....: In [259]: pd.json_normalize(data, 'county', ['state', 'shortname', ['info', 'governor']]) Out[259]: name population state shortname info.governor 0 Dade 12345 Florida FL Rick Scott 1 Broward 40000 Florida FL Rick Scott 2 Palm Beach 60000 Florida FL Rick Scott 3 Summit 1234 Ohio OH John Kasich 4 Cuyahoga 1337 Ohio OH John Kasich
The max_level parameter provides more control over which level to end normalization. With max_level=1 the following snippet normalizes until 1st nesting level of the provided dict.
In [260]: data = [{'CreatedBy': {'Name': 'User001'}, .....: 'Lookup': {'TextField': 'Some text', .....: 'UserField': {'Id': 'ID001', .....: 'Name': 'Name001'}}, .....: 'Image': {'a': 'b'} .....: }] .....: In [261]: pd.json_normalize(data, max_level=1) Out[261]: CreatedBy.Name Lookup.TextField Lookup.UserField Image.a 0 User001 Some text {'Id': 'ID001', 'Name': 'Name001'} b
pandas is able to read and write line-delimited json files that are common in data processing pipelines using Hadoop or Spark.
For line-delimited json files, pandas can also return an iterator which reads in chunksize lines at a time. This can be useful for large files or to read from a stream.
In [262]: jsonl = ''' .....: {"a": 1, "b": 2} .....: {"a": 3, "b": 4} .....: ''' .....: In [263]: df = pd.read_json(jsonl, lines=True) In [264]: df Out[264]: a b 0 1 2 1 3 4 In [265]: df.to_json(orient='records', lines=True) Out[265]: '{"a":1,"b":2}\n{"a":3,"b":4}' # reader is an iterator that returns `chunksize` lines each iteration In [266]: reader = pd.read_json(StringIO(jsonl), lines=True, chunksize=1) In [267]: reader Out[267]: <pandas.io.json._json.JsonReader at 0x7f65b39b75f8> In [268]: for chunk in reader: .....: print(chunk) .....: Empty DataFrame Columns: [] Index: [] a b 0 1 2 a b 1 3 4
Table Schema is a spec for describing tabular datasets as a JSON object. The JSON includes information on the field names, types, and other attributes. You can use the orient table to build a JSON string with two fields, schema and data.
schema
In [269]: df = pd.DataFrame({'A': [1, 2, 3], .....: 'B': ['a', 'b', 'c'], .....: 'C': pd.date_range('2016-01-01', freq='d', periods=3)}, .....: index=pd.Index(range(3), name='idx')) .....: In [270]: df Out[270]: A B C idx 0 1 a 2016-01-01 1 2 b 2016-01-02 2 3 c 2016-01-03 In [271]: df.to_json(orient='table', date_format="iso") Out[271]: '{"schema":{"fields":[{"name":"idx","type":"integer"},{"name":"A","type":"integer"},{"name":"B","type":"string"},{"name":"C","type":"datetime"}],"primaryKey":["idx"],"pandas_version":"0.20.0"},"data":[{"idx":0,"A":1,"B":"a","C":"2016-01-01T00:00:00.000Z"},{"idx":1,"A":2,"B":"b","C":"2016-01-02T00:00:00.000Z"},{"idx":2,"A":3,"B":"c","C":"2016-01-03T00:00:00.000Z"}]}'
The schema field contains the fields key, which itself contains a list of column name to type pairs, including the Index or MultiIndex (see below for a list of types). The schema field also contains a primaryKey field if the (Multi)index is unique.
fields
Index
primaryKey
The second field, data, contains the serialized data with the records orient. The index is included, and any datetimes are ISO 8601 formatted, as required by the Table Schema spec.
The full list of types supported are described in the Table Schema spec. This table shows the mapping from pandas types:
Pandas type
Table Schema type
int64
float64
number
bool
boolean
datetime64[ns]
timedelta64[ns]
duration
categorical
any
A few notes on the generated table schema:
The schema object contains a pandas_version field. This contains the version of pandas’ dialect of the schema, and will be incremented with each revision.
pandas_version
All dates are converted to UTC when serializing. Even timezone naive values, which are treated as UTC with an offset of 0.
In [272]: from pandas.io.json import build_table_schema In [273]: s = pd.Series(pd.date_range('2016', periods=4)) In [274]: build_table_schema(s) Out[274]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'datetime'}], 'primaryKey': ['index'], 'pandas_version': '0.20.0'}
datetimes with a timezone (before serializing), include an additional field tz with the time zone name (e.g. 'US/Central').
tz
'US/Central'
In [275]: s_tz = pd.Series(pd.date_range('2016', periods=12, .....: tz='US/Central')) .....: In [276]: build_table_schema(s_tz) Out[276]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'datetime', 'tz': 'US/Central'}], 'primaryKey': ['index'], 'pandas_version': '0.20.0'}
Periods are converted to timestamps before serialization, and so have the same behavior of being converted to UTC. In addition, periods will contain and additional field freq with the period’s frequency, e.g. 'A-DEC'.
freq
'A-DEC'
In [277]: s_per = pd.Series(1, index=pd.period_range('2016', freq='A-DEC', .....: periods=4)) .....: In [278]: build_table_schema(s_per) Out[278]: {'fields': [{'name': 'index', 'type': 'datetime', 'freq': 'A-DEC'}, {'name': 'values', 'type': 'integer'}], 'primaryKey': ['index'], 'pandas_version': '0.20.0'}
Categoricals use the any type and an enum constraint listing the set of possible values. Additionally, an ordered field is included:
enum
ordered
In [279]: s_cat = pd.Series(pd.Categorical(['a', 'b', 'a'])) In [280]: build_table_schema(s_cat) Out[280]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'any', 'constraints': {'enum': ['a', 'b']}, 'ordered': False}], 'primaryKey': ['index'], 'pandas_version': '0.20.0'}
A primaryKey field, containing an array of labels, is included if the index is unique:
In [281]: s_dupe = pd.Series([1, 2], index=[1, 1]) In [282]: build_table_schema(s_dupe) Out[282]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'integer'}], 'pandas_version': '0.20.0'}
The primaryKey behavior is the same with MultiIndexes, but in this case the primaryKey is an array:
In [283]: s_multi = pd.Series(1, index=pd.MultiIndex.from_product([('a', 'b'), .....: (0, 1)])) .....: In [284]: build_table_schema(s_multi) Out[284]: {'fields': [{'name': 'level_0', 'type': 'string'}, {'name': 'level_1', 'type': 'integer'}, {'name': 'values', 'type': 'integer'}], 'primaryKey': FrozenList(['level_0', 'level_1']), 'pandas_version': '0.20.0'}
The default naming roughly follows these rules:
For series, the object.name is used. If that’s none, then the name is values For DataFrames, the stringified version of the column name is used For Index (not MultiIndex), index.name is used, with a fallback to index if that is None. For MultiIndex, mi.names is used. If any level has no name, then level_<i> is used.
For series, the object.name is used. If that’s none, then the name is values
object.name
For DataFrames, the stringified version of the column name is used
DataFrames
For Index (not MultiIndex), index.name is used, with a fallback to index if that is None.
index.name
For MultiIndex, mi.names is used. If any level has no name, then level_<i> is used.
mi.names
level_<i>
New in version 0.23.0.
read_json also accepts orient='table' as an argument. This allows for the preservation of metadata such as dtypes and index names in a round-trippable manner.
orient='table'
In [285]: df = pd.DataFrame({'foo': [1, 2, 3, 4], .....: 'bar': ['a', 'b', 'c', 'd'], .....: 'baz': pd.date_range('2018-01-01', freq='d', periods=4), .....: 'qux': pd.Categorical(['a', 'b', 'c', 'c']) .....: }, index=pd.Index(range(4), name='idx')) .....: In [286]: df Out[286]: foo bar baz qux idx 0 1 a 2018-01-01 a 1 2 b 2018-01-02 b 2 3 c 2018-01-03 c 3 4 d 2018-01-04 c In [287]: df.dtypes Out[287]: foo int64 bar object baz datetime64[ns] qux category dtype: object In [288]: df.to_json('test.json', orient='table') In [289]: new_df = pd.read_json('test.json', orient='table') In [290]: new_df Out[290]: foo bar baz qux idx 0 1 a 2018-01-01 a 1 2 b 2018-01-02 b 2 3 c 2018-01-03 c 3 4 d 2018-01-04 c In [291]: new_df.dtypes Out[291]: foo int64 bar object baz datetime64[ns] qux category dtype: object
Please note that the literal string ‘index’ as the name of an Index is not round-trippable, nor are any names beginning with 'level_' within a MultiIndex. These are used by default in DataFrame.to_json() to indicate missing values and the subsequent read cannot distinguish the intent.
'level_'
DataFrame.to_json()
In [292]: df.index.name = 'index' In [293]: df.to_json('test.json', orient='table') In [294]: new_df = pd.read_json('test.json', orient='table') In [295]: print(new_df.index.name) None
We highly encourage you to read the HTML Table Parsing gotchas below regarding the issues surrounding the BeautifulSoup4/html5lib/lxml parsers.
The top-level read_html() function can accept an HTML string/file/URL and will parse HTML tables into list of pandas DataFrames. Let’s look at a few examples.
read_html()
read_html returns a list of DataFrame objects, even if there is only a single table contained in the HTML content.
list
Read a URL with no options:
In [296]: url = 'https://www.fdic.gov/bank/individual/failed/banklist.html' In [297]: dfs = pd.read_html(url) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-297-fafba0f26e7e> in <module> ----> 1 dfs = pd.read_html(url) ~/scipy/pandas/pandas/io/html.py in read_html(io, match, flavor, header, index_col, skiprows, attrs, parse_dates, thousands, encoding, decimal, converters, na_values, keep_default_na, displayed_only) 1097 na_values=na_values, 1098 keep_default_na=keep_default_na, -> 1099 displayed_only=displayed_only, 1100 ) ~/scipy/pandas/pandas/io/html.py in _parse(flavor, io, match, attrs, encoding, displayed_only, **kwargs) 889 retained = None 890 for flav in flavor: --> 891 parser = _parser_dispatch(flav) 892 p = parser(io, compiled_match, attrs, encoding, displayed_only) 893 ~/scipy/pandas/pandas/io/html.py in _parser_dispatch(flavor) 846 else: 847 if not _HAS_LXML: --> 848 raise ImportError("lxml not found, please install it") 849 return _valid_parsers[flavor] 850 ImportError: lxml not found, please install it In [298]: dfs Out[298]: ( Case Data 0 A 0.276232 1 A -1.087401 2 A -0.673690 3 B 0.113648, Case Data 4 A -1.478427 5 A 0.524988 6 B 0.404705, Case Data 7 A 0.577046 8 A -1.715002)
The data from the above URL changes every Monday so the resulting data above and the data below may be slightly different.
Read in the content of the file from the above URL and pass it to read_html as a string:
In [299]: with open(file_path, 'r') as f: .....: dfs = pd.read_html(f.read()) .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-299-088a5feb8d3b> in <module> 1 with open(file_path, 'r') as f: ----> 2 dfs = pd.read_html(f.read()) ~/scipy/pandas/pandas/io/html.py in read_html(io, match, flavor, header, index_col, skiprows, attrs, parse_dates, thousands, encoding, decimal, converters, na_values, keep_default_na, displayed_only) 1097 na_values=na_values, 1098 keep_default_na=keep_default_na, -> 1099 displayed_only=displayed_only, 1100 ) ~/scipy/pandas/pandas/io/html.py in _parse(flavor, io, match, attrs, encoding, displayed_only, **kwargs) 889 retained = None 890 for flav in flavor: --> 891 parser = _parser_dispatch(flav) 892 p = parser(io, compiled_match, attrs, encoding, displayed_only) 893 ~/scipy/pandas/pandas/io/html.py in _parser_dispatch(flavor) 846 else: 847 if not _HAS_LXML: --> 848 raise ImportError("lxml not found, please install it") 849 return _valid_parsers[flavor] 850 ImportError: lxml not found, please install it In [300]: dfs Out[300]: ( Case Data 0 A 0.276232 1 A -1.087401 2 A -0.673690 3 B 0.113648, Case Data 4 A -1.478427 5 A 0.524988 6 B 0.404705, Case Data 7 A 0.577046 8 A -1.715002)
You can even pass in an instance of StringIO if you so desire:
In [301]: with open(file_path, 'r') as f: .....: sio = StringIO(f.read()) .....: In [302]: dfs = pd.read_html(sio) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-302-b11088569efd> in <module> ----> 1 dfs = pd.read_html(sio) ~/scipy/pandas/pandas/io/html.py in read_html(io, match, flavor, header, index_col, skiprows, attrs, parse_dates, thousands, encoding, decimal, converters, na_values, keep_default_na, displayed_only) 1097 na_values=na_values, 1098 keep_default_na=keep_default_na, -> 1099 displayed_only=displayed_only, 1100 ) ~/scipy/pandas/pandas/io/html.py in _parse(flavor, io, match, attrs, encoding, displayed_only, **kwargs) 889 retained = None 890 for flav in flavor: --> 891 parser = _parser_dispatch(flav) 892 p = parser(io, compiled_match, attrs, encoding, displayed_only) 893 ~/scipy/pandas/pandas/io/html.py in _parser_dispatch(flavor) 846 else: 847 if not _HAS_LXML: --> 848 raise ImportError("lxml not found, please install it") 849 return _valid_parsers[flavor] 850 ImportError: lxml not found, please install it In [303]: dfs Out[303]: ( Case Data 0 A 0.276232 1 A -1.087401 2 A -0.673690 3 B 0.113648, Case Data 4 A -1.478427 5 A 0.524988 6 B 0.404705, Case Data 7 A 0.577046 8 A -1.715002)
The following examples are not run by the IPython evaluator due to the fact that having so many network-accessing functions slows down the documentation build. If you spot an error or an example that doesn’t run, please do not hesitate to report it over on pandas GitHub issues page.
Read a URL and match a table that contains specific text:
match = 'Metcalf Bank' df_list = pd.read_html(url, match=match)
Specify a header row (by default <th> or <td> elements located within a <thead> are used to form the column index, if multiple rows are contained within <thead> then a MultiIndex is created); if specified, the header row is taken from the data minus the parsed header elements (<th> elements).
<th>
<td>
<thead>
dfs = pd.read_html(url, header=0)
Specify an index column:
dfs = pd.read_html(url, index_col=0)
Specify a number of rows to skip:
dfs = pd.read_html(url, skiprows=0)
Specify a number of rows to skip using a list (xrange (Python 2 only) works as well):
xrange
dfs = pd.read_html(url, skiprows=range(2))
Specify an HTML attribute:
dfs1 = pd.read_html(url, attrs={'id': 'table'}) dfs2 = pd.read_html(url, attrs={'class': 'sortable'}) print(np.array_equal(dfs1[0], dfs2[0])) # Should be True
Specify values that should be converted to NaN:
dfs = pd.read_html(url, na_values=['No Acquirer'])
Specify whether to keep the default set of NaN values:
dfs = pd.read_html(url, keep_default_na=False)
Specify converters for columns. This is useful for numerical text data that has leading zeros. By default columns that are numerical are cast to numeric types and the leading zeros are lost. To avoid this, we can convert these columns to strings.
url_mcc = 'https://en.wikipedia.org/wiki/Mobile_country_code' dfs = pd.read_html(url_mcc, match='Telekom Albania', header=0, converters={'MNC': str})
Use some combination of the above:
dfs = pd.read_html(url, match='Metcalf Bank', index_col=0)
Read in pandas to_html output (with some loss of floating point precision):
df = pd.DataFrame(np.random.randn(2, 2)) s = df.to_html(float_format='{0:.40g}'.format) dfin = pd.read_html(s, index_col=0)
The lxml backend will raise an error on a failed parse if that is the only parser you provide. If you only have a single parser you can provide just a string, but it is considered good practice to pass a list with one string if, for example, the function expects a sequence of strings. You may use:
lxml
dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml'])
Or you could pass flavor='lxml' without a list:
flavor='lxml'
dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor='lxml')
However, if you have bs4 and html5lib installed and pass None or ['lxml', 'bs4'] then the parse will most likely succeed. Note that as soon as a parse succeeds, the function will return.
['lxml', 'bs4']
dfs = pd.read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml', 'bs4'])
DataFrame objects have an instance method to_html which renders the contents of the DataFrame as an HTML table. The function arguments are as in the method to_string described above.
Not all of the possible options for DataFrame.to_html are shown here for brevity’s sake. See to_html() for the full set of options.
DataFrame.to_html
to_html()
In [304]: df = pd.DataFrame(np.random.randn(2, 2)) In [305]: df Out[305]: 0 1 0 -0.184744 0.496971 1 -0.856240 1.857977 In [306]: print(df.to_html()) # raw html <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>0</th> <th>1</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>-0.184744</td> <td>0.496971</td> </tr> <tr> <th>1</th> <td>-0.856240</td> <td>1.857977</td> </tr> </tbody> </table>
HTML:
The columns argument will limit the columns shown:
In [307]: print(df.to_html(columns=[0])) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>0</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>-0.184744</td> </tr> <tr> <th>1</th> <td>-0.856240</td> </tr> </tbody> </table>
float_format takes a Python callable to control the precision of floating point values:
In [308]: print(df.to_html(float_format='{0:.10f}'.format)) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>0</th> <th>1</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>-0.1847438576</td> <td>0.4969711327</td> </tr> <tr> <th>1</th> <td>-0.8562396763</td> <td>1.8579766508</td> </tr> </tbody> </table>
bold_rows will make the row labels bold by default, but you can turn that off:
bold_rows
In [309]: print(df.to_html(bold_rows=False)) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>0</th> <th>1</th> </tr> </thead> <tbody> <tr> <td>0</td> <td>-0.184744</td> <td>0.496971</td> </tr> <tr> <td>1</td> <td>-0.856240</td> <td>1.857977</td> </tr> </tbody> </table>
The classes argument provides the ability to give the resulting HTML table CSS classes. Note that these classes are appended to the existing 'dataframe' class.
classes
'dataframe'
In [310]: print(df.to_html(classes=['awesome_table_class', 'even_more_awesome_class'])) <table border="1" class="dataframe awesome_table_class even_more_awesome_class"> <thead> <tr style="text-align: right;"> <th></th> <th>0</th> <th>1</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>-0.184744</td> <td>0.496971</td> </tr> <tr> <th>1</th> <td>-0.856240</td> <td>1.857977</td> </tr> </tbody> </table>
The render_links argument provides the ability to add hyperlinks to cells that contain URLs.
render_links
New in version 0.24.
In [311]: url_df = pd.DataFrame({ .....: 'name': ['Python', 'Pandas'], .....: 'url': ['https://www.python.org/', 'https://pandas.pydata.org']}) .....: In [312]: print(url_df.to_html(render_links=True)) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>name</th> <th>url</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>Python</td> <td><a href="https://www.python.org/" target="_blank">https://www.python.org/</a></td> </tr> <tr> <th>1</th> <td>Pandas</td> <td><a href="https://pandas.pydata.org" target="_blank">https://pandas.pydata.org</a></td> </tr> </tbody> </table>
Finally, the escape argument allows you to control whether the “<”, “>” and “&” characters escaped in the resulting HTML (by default it is True). So to get the HTML without escaped characters pass escape=False
escape
escape=False
In [313]: df = pd.DataFrame({'a': list('&<>'), 'b': np.random.randn(3)})
Escaped:
In [314]: print(df.to_html()) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>a</th> <th>b</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>&</td> <td>-0.474063</td> </tr> <tr> <th>1</th> <td><</td> <td>-0.230305</td> </tr> <tr> <th>2</th> <td>></td> <td>-0.400654</td> </tr> </tbody> </table>
Not escaped:
In [315]: print(df.to_html(escape=False)) <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>a</th> <th>b</th> </tr> </thead> <tbody> <tr> <th>0</th> <td>&</td> <td>-0.474063</td> </tr> <tr> <th>1</th> <td><</td> <td>-0.230305</td> </tr> <tr> <th>2</th> <td>></td> <td>-0.400654</td> </tr> </tbody> </table>
Some browsers may not show a difference in the rendering of the previous two HTML tables.
There are some versioning issues surrounding the libraries that are used to parse HTML tables in the top-level pandas io function read_html.
Issues with lxml
Benefits
lxml is very fast. lxml requires Cython to install correctly.
lxml is very fast.
lxml requires Cython to install correctly.
Drawbacks
lxml does not make any guarantees about the results of its parse unless it is given strictly valid markup. In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend will use html5lib if lxml fails to parse It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will still get a valid result (provided everything else is valid) even if lxml fails.
lxml does not make any guarantees about the results of its parse unless it is given strictly valid markup.
In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend will use html5lib if lxml fails to parse
It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will still get a valid result (provided everything else is valid) even if lxml fails.
Issues with BeautifulSoup4 using lxml as a backend
The above issues hold here as well since BeautifulSoup4 is essentially just a wrapper around a parser backend.
Issues with BeautifulSoup4 using html5lib as a backend
html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way rather than just, e.g., dropping an element without notifying you. html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is “correct”, since the process of fixing markup does not have a single definition. html5lib is pure Python and requires no additional build steps beyond its own installation.
html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way rather than just, e.g., dropping an element without notifying you.
html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is “correct”, since the process of fixing markup does not have a single definition.
html5lib is pure Python and requires no additional build steps beyond its own installation.
The biggest drawback to using html5lib is that it is slow as molasses. However consider the fact that many tables on the web are not big enough for the parsing algorithm runtime to matter. It is more likely that the bottleneck will be in the process of reading the raw text from the URL over the web, i.e., IO (input-output). For very large tables, this might not be true.
The read_excel() method can read Excel 2003 (.xls) files using the xlrd Python module. Excel 2007+ (.xlsx) files can be read using either xlrd or openpyxl. The to_excel() instance method is used for saving a DataFrame to Excel. Generally the semantics are similar to working with csv data. See the cookbook for some advanced strategies.
read_excel()
.xls
xlrd
.xlsx
openpyxl
to_excel()
In the most basic use-case, read_excel takes a path to an Excel file, and the sheet_name indicating which sheet to parse.
sheet_name
# Returns a DataFrame pd.read_excel('path_to_file.xls', sheet_name='Sheet1')
ExcelFile
To facilitate working with multiple sheets from the same file, the ExcelFile class can be used to wrap the file and can be passed into read_excel There will be a performance benefit for reading multiple sheets as the file is read into memory only once.
xlsx = pd.ExcelFile('path_to_file.xls') df = pd.read_excel(xlsx, 'Sheet1')
The ExcelFile class can also be used as a context manager.
with pd.ExcelFile('path_to_file.xls') as xls: df1 = pd.read_excel(xls, 'Sheet1') df2 = pd.read_excel(xls, 'Sheet2')
The sheet_names property will generate a list of the sheet names in the file.
sheet_names
The primary use-case for an ExcelFile is parsing multiple sheets with different parameters:
data = {} # For when Sheet1's format differs from Sheet2 with pd.ExcelFile('path_to_file.xls') as xls: data['Sheet1'] = pd.read_excel(xls, 'Sheet1', index_col=None, na_values=['NA']) data['Sheet2'] = pd.read_excel(xls, 'Sheet2', index_col=1)
Note that if the same parsing parameters are used for all sheets, a list of sheet names can simply be passed to read_excel with no loss in performance.
# using the ExcelFile class data = {} with pd.ExcelFile('path_to_file.xls') as xls: data['Sheet1'] = pd.read_excel(xls, 'Sheet1', index_col=None, na_values=['NA']) data['Sheet2'] = pd.read_excel(xls, 'Sheet2', index_col=None, na_values=['NA']) # equivalent using the read_excel function data = pd.read_excel('path_to_file.xls', ['Sheet1', 'Sheet2'], index_col=None, na_values=['NA'])
ExcelFile can also be called with a xlrd.book.Book object as a parameter. This allows the user to control how the excel file is read. For example, sheets can be loaded on demand by calling xlrd.open_workbook() with on_demand=True.
xlrd.book.Book
xlrd.open_workbook()
on_demand=True
import xlrd xlrd_book = xlrd.open_workbook('path_to_file.xls', on_demand=True) with pd.ExcelFile(xlrd_book) as xls: df1 = pd.read_excel(xls, 'Sheet1') df2 = pd.read_excel(xls, 'Sheet2')
The second argument is sheet_name, not to be confused with ExcelFile.sheet_names.
ExcelFile.sheet_names
An ExcelFile’s attribute sheet_names provides access to a list of sheets.
The arguments sheet_name allows specifying the sheet or sheets to read.
The default value for sheet_name is 0, indicating to read the first sheet
Pass a string to refer to the name of a particular sheet in the workbook.
Pass an integer to refer to the index of a sheet. Indices follow Python convention, beginning at 0.
Pass a list of either strings or integers, to return a dictionary of specified sheets.
Pass a None to return a dictionary of all available sheets.
# Returns a DataFrame pd.read_excel('path_to_file.xls', 'Sheet1', index_col=None, na_values=['NA'])
Using the sheet index:
# Returns a DataFrame pd.read_excel('path_to_file.xls', 0, index_col=None, na_values=['NA'])
Using all default values:
# Returns a DataFrame pd.read_excel('path_to_file.xls')
Using None to get all sheets:
# Returns a dictionary of DataFrames pd.read_excel('path_to_file.xls', sheet_name=None)
Using a list to get multiple sheets:
# Returns the 1st and 4th sheet, as a dictionary of DataFrames. pd.read_excel('path_to_file.xls', sheet_name=['Sheet1', 3])
read_excel can read more than one sheet, by setting sheet_name to either a list of sheet names, a list of sheet positions, or None to read all sheets. Sheets can be specified by sheet index or sheet name, using an integer or string, respectively.
read_excel can read a MultiIndex index, by passing a list of columns to index_col and a MultiIndex column by passing a list of rows to header. If either the index or columns have serialized level names those will be read in as well by specifying the rows/columns that make up the levels.
For example, to read in a MultiIndex index without names:
In [316]: df = pd.DataFrame({'a': [1, 2, 3, 4], 'b': [5, 6, 7, 8]}, .....: index=pd.MultiIndex.from_product([['a', 'b'], ['c', 'd']])) .....: In [317]: df.to_excel('path_to_file.xlsx') --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-317-02ef2d16bef7> in <module> ----> 1 df.to_excel('path_to_file.xlsx') ~/scipy/pandas/pandas/core/generic.py in to_excel(self, excel_writer, sheet_name, na_rep, float_format, columns, header, index, index_label, startrow, startcol, engine, merge_cells, encoding, inf_rep, verbose, freeze_panes) 2144 startcol=startcol, 2145 freeze_panes=freeze_panes, -> 2146 engine=engine, 2147 ) 2148 ~/scipy/pandas/pandas/io/formats/excel.py in write(self, writer, sheet_name, startrow, startcol, freeze_panes, engine) 724 need_save = False 725 else: --> 726 writer = ExcelWriter(stringify_path(writer), engine=engine) 727 need_save = True 728 ~/scipy/pandas/pandas/io/excel/_openpyxl.py in __init__(self, path, engine, mode, **engine_kwargs) 16 def __init__(self, path, engine=None, mode="w", **engine_kwargs): 17 # Use the openpyxl module as the Excel writer. ---> 18 from openpyxl.workbook import Workbook 19 20 super().__init__(path, mode=mode, **engine_kwargs) ModuleNotFoundError: No module named 'openpyxl' In [318]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1]) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-318-cb6d294acdd1> in <module> ----> 1 df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1]) ~/scipy/pandas/pandas/io/excel/_base.py in read_excel(io, sheet_name, header, names, index_col, usecols, squeeze, dtype, engine, converters, true_values, false_values, skiprows, nrows, na_values, keep_default_na, verbose, parse_dates, date_parser, thousands, comment, skipfooter, convert_float, mangle_dupe_cols, **kwds) 301 302 if not isinstance(io, ExcelFile): --> 303 io = ExcelFile(io, engine=engine) 304 elif engine and engine != io.engine: 305 raise ValueError( ~/scipy/pandas/pandas/io/excel/_base.py in __init__(self, io, engine) 812 self._io = stringify_path(io) 813 --> 814 self._reader = self._engines[engine](self._io) 815 816 def __fspath__(self): ~/scipy/pandas/pandas/io/excel/_xlrd.py in __init__(self, filepath_or_buffer) 18 """ 19 err_msg = "Install xlrd >= 1.0.0 for Excel support" ---> 20 import_optional_dependency("xlrd", extra=err_msg) 21 super().__init__(filepath_or_buffer) 22 ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'xlrd'. Install xlrd >= 1.0.0 for Excel support Use pip or conda to install xlrd. In [319]: df Out[319]: a b a c 1 5 d 2 6 b c 3 7 d 4 8
If the index has level names, they will parsed as well, using the same parameters.
In [320]: df.index = df.index.set_names(['lvl1', 'lvl2']) In [321]: df.to_excel('path_to_file.xlsx') --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-321-02ef2d16bef7> in <module> ----> 1 df.to_excel('path_to_file.xlsx') ~/scipy/pandas/pandas/core/generic.py in to_excel(self, excel_writer, sheet_name, na_rep, float_format, columns, header, index, index_label, startrow, startcol, engine, merge_cells, encoding, inf_rep, verbose, freeze_panes) 2144 startcol=startcol, 2145 freeze_panes=freeze_panes, -> 2146 engine=engine, 2147 ) 2148 ~/scipy/pandas/pandas/io/formats/excel.py in write(self, writer, sheet_name, startrow, startcol, freeze_panes, engine) 724 need_save = False 725 else: --> 726 writer = ExcelWriter(stringify_path(writer), engine=engine) 727 need_save = True 728 ~/scipy/pandas/pandas/io/excel/_openpyxl.py in __init__(self, path, engine, mode, **engine_kwargs) 16 def __init__(self, path, engine=None, mode="w", **engine_kwargs): 17 # Use the openpyxl module as the Excel writer. ---> 18 from openpyxl.workbook import Workbook 19 20 super().__init__(path, mode=mode, **engine_kwargs) ModuleNotFoundError: No module named 'openpyxl' In [322]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1]) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-322-cb6d294acdd1> in <module> ----> 1 df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1]) ~/scipy/pandas/pandas/io/excel/_base.py in read_excel(io, sheet_name, header, names, index_col, usecols, squeeze, dtype, engine, converters, true_values, false_values, skiprows, nrows, na_values, keep_default_na, verbose, parse_dates, date_parser, thousands, comment, skipfooter, convert_float, mangle_dupe_cols, **kwds) 301 302 if not isinstance(io, ExcelFile): --> 303 io = ExcelFile(io, engine=engine) 304 elif engine and engine != io.engine: 305 raise ValueError( ~/scipy/pandas/pandas/io/excel/_base.py in __init__(self, io, engine) 812 self._io = stringify_path(io) 813 --> 814 self._reader = self._engines[engine](self._io) 815 816 def __fspath__(self): ~/scipy/pandas/pandas/io/excel/_xlrd.py in __init__(self, filepath_or_buffer) 18 """ 19 err_msg = "Install xlrd >= 1.0.0 for Excel support" ---> 20 import_optional_dependency("xlrd", extra=err_msg) 21 super().__init__(filepath_or_buffer) 22 ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'xlrd'. Install xlrd >= 1.0.0 for Excel support Use pip or conda to install xlrd. In [323]: df Out[323]: a b lvl1 lvl2 a c 1 5 d 2 6 b c 3 7 d 4 8
If the source file has both MultiIndex index and columns, lists specifying each should be passed to index_col and header:
In [324]: df.columns = pd.MultiIndex.from_product([['a'], ['b', 'd']], .....: names=['c1', 'c2']) .....: In [325]: df.to_excel('path_to_file.xlsx') --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-325-02ef2d16bef7> in <module> ----> 1 df.to_excel('path_to_file.xlsx') ~/scipy/pandas/pandas/core/generic.py in to_excel(self, excel_writer, sheet_name, na_rep, float_format, columns, header, index, index_label, startrow, startcol, engine, merge_cells, encoding, inf_rep, verbose, freeze_panes) 2144 startcol=startcol, 2145 freeze_panes=freeze_panes, -> 2146 engine=engine, 2147 ) 2148 ~/scipy/pandas/pandas/io/formats/excel.py in write(self, writer, sheet_name, startrow, startcol, freeze_panes, engine) 724 need_save = False 725 else: --> 726 writer = ExcelWriter(stringify_path(writer), engine=engine) 727 need_save = True 728 ~/scipy/pandas/pandas/io/excel/_openpyxl.py in __init__(self, path, engine, mode, **engine_kwargs) 16 def __init__(self, path, engine=None, mode="w", **engine_kwargs): 17 # Use the openpyxl module as the Excel writer. ---> 18 from openpyxl.workbook import Workbook 19 20 super().__init__(path, mode=mode, **engine_kwargs) ModuleNotFoundError: No module named 'openpyxl' In [326]: df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1], header=[0, 1]) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-326-3a6768bb42fd> in <module> ----> 1 df = pd.read_excel('path_to_file.xlsx', index_col=[0, 1], header=[0, 1]) ~/scipy/pandas/pandas/io/excel/_base.py in read_excel(io, sheet_name, header, names, index_col, usecols, squeeze, dtype, engine, converters, true_values, false_values, skiprows, nrows, na_values, keep_default_na, verbose, parse_dates, date_parser, thousands, comment, skipfooter, convert_float, mangle_dupe_cols, **kwds) 301 302 if not isinstance(io, ExcelFile): --> 303 io = ExcelFile(io, engine=engine) 304 elif engine and engine != io.engine: 305 raise ValueError( ~/scipy/pandas/pandas/io/excel/_base.py in __init__(self, io, engine) 812 self._io = stringify_path(io) 813 --> 814 self._reader = self._engines[engine](self._io) 815 816 def __fspath__(self): ~/scipy/pandas/pandas/io/excel/_xlrd.py in __init__(self, filepath_or_buffer) 18 """ 19 err_msg = "Install xlrd >= 1.0.0 for Excel support" ---> 20 import_optional_dependency("xlrd", extra=err_msg) 21 super().__init__(filepath_or_buffer) 22 ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'xlrd'. Install xlrd >= 1.0.0 for Excel support Use pip or conda to install xlrd. In [327]: df Out[327]: c1 a c2 b d lvl1 lvl2 a c 1 5 d 2 6 b c 3 7 d 4 8
It is often the case that users will insert columns to do temporary computations in Excel and you may not want to read in those columns. read_excel takes a usecols keyword to allow you to specify a subset of columns to parse.
Deprecated since version 0.24.0.
Passing in an integer for usecols has been deprecated. Please pass in a list of ints from 0 to usecols inclusive instead.
If usecols is an integer, then it is assumed to indicate the last column to be parsed.
pd.read_excel('path_to_file.xls', 'Sheet1', usecols=2)
You can also specify a comma-delimited set of Excel columns and ranges as a string:
pd.read_excel('path_to_file.xls', 'Sheet1', usecols='A,C:E')
If usecols is a list of integers, then it is assumed to be the file column indices to be parsed.
pd.read_excel('path_to_file.xls', 'Sheet1', usecols=[0, 2, 3])
Element order is ignored, so usecols=[0, 1] is the same as [1, 0].
If usecols is a list of strings, it is assumed that each string corresponds to a column name provided either by the user in names or inferred from the document header row(s). Those strings define which columns will be parsed:
pd.read_excel('path_to_file.xls', 'Sheet1', usecols=['foo', 'bar'])
Element order is ignored, so usecols=['baz', 'joe'] is the same as ['joe', 'baz'].
usecols=['baz', 'joe']
['joe', 'baz']
If usecols is callable, the callable function will be evaluated against the column names, returning names where the callable function evaluates to True.
pd.read_excel('path_to_file.xls', 'Sheet1', usecols=lambda x: x.isalpha())
Datetime-like values are normally automatically converted to the appropriate dtype when reading the excel file. But if you have a column of strings that look like dates (but are not actually formatted as dates in excel), you can use the parse_dates keyword to parse those strings to datetimes:
pd.read_excel('path_to_file.xls', 'Sheet1', parse_dates=['date_strings'])
It is possible to transform the contents of Excel cells via the converters option. For instance, to convert a column to boolean:
pd.read_excel('path_to_file.xls', 'Sheet1', converters={'MyBools': bool})
This options handles missing values and treats exceptions in the converters as missing data. Transformations are applied cell by cell rather than to the column as a whole, so the array dtype is not guaranteed. For instance, a column of integers with missing values cannot be transformed to an array with integer dtype, because NaN is strictly a float. You can manually mask missing data to recover integer dtype:
def cfun(x): return int(x) if x else -1 pd.read_excel('path_to_file.xls', 'Sheet1', converters={'MyInts': cfun})
As an alternative to converters, the type for an entire column can be specified using the dtype keyword, which takes a dictionary mapping column names to types. To interpret data with no type inference, use the type str or object.
pd.read_excel('path_to_file.xls', dtype={'MyInts': 'int64', 'MyText': str})
To write a DataFrame object to a sheet of an Excel file, you can use the to_excel instance method. The arguments are largely the same as to_csv described above, the first argument being the name of the excel file, and the optional second argument the name of the sheet to which the DataFrame should be written. For example:
df.to_excel('path_to_file.xlsx', sheet_name='Sheet1')
Files with a .xls extension will be written using xlwt and those with a .xlsx extension will be written using xlsxwriter (if available) or openpyxl.
xlwt
xlsxwriter
The DataFrame will be written in a way that tries to mimic the REPL output. The index_label will be placed in the second row instead of the first. You can place it in the first row by setting the merge_cells option in to_excel() to False:
merge_cells
df.to_excel('path_to_file.xlsx', index_label='label', merge_cells=False)
In order to write separate DataFrames to separate sheets in a single Excel file, one can pass an ExcelWriter.
ExcelWriter
with pd.ExcelWriter('path_to_file.xlsx') as writer: df1.to_excel(writer, sheet_name='Sheet1') df2.to_excel(writer, sheet_name='Sheet2')
Wringing a little more performance out of read_excel Internally, Excel stores all numeric data as floats. Because this can produce unexpected behavior when reading in data, pandas defaults to trying to convert integers to floats if it doesn’t lose information (1.0 --> 1). You can pass convert_float=False to disable this behavior, which may give a slight performance improvement.
1.0 --> 1
convert_float=False
Pandas supports writing Excel files to buffer-like objects such as StringIO or BytesIO using ExcelWriter.
BytesIO
# Safe import for either Python 2.x or 3.x try: from io import BytesIO except ImportError: from cStringIO import StringIO as BytesIO bio = BytesIO() # By setting the 'engine' in the ExcelWriter constructor. writer = pd.ExcelWriter(bio, engine='xlsxwriter') df.to_excel(writer, sheet_name='Sheet1') # Save the workbook writer.save() # Seek to the beginning and read to copy the workbook to a variable in memory bio.seek(0) workbook = bio.read()
engine is optional but recommended. Setting the engine determines the version of workbook produced. Setting engine='xlrd' will produce an Excel 2003-format workbook (xls). Using either 'openpyxl' or 'xlsxwriter' will produce an Excel 2007-format workbook (xlsx). If omitted, an Excel 2007-formatted workbook is produced.
engine
engine='xlrd'
'openpyxl'
'xlsxwriter'
Pandas chooses an Excel writer via two methods:
the engine keyword argument
the filename extension (via the default specified in config options)
By default, pandas uses the XlsxWriter for .xlsx, openpyxl for .xlsm, and xlwt for .xls files. If you have multiple engines installed, you can set the default engine through setting the config options io.excel.xlsx.writer and io.excel.xls.writer. pandas will fall back on openpyxl for .xlsx files if Xlsxwriter is not available.
.xlsm
io.excel.xlsx.writer
io.excel.xls.writer
To specify which writer you want to use, you can pass an engine keyword argument to to_excel and to ExcelWriter. The built-in engines are:
openpyxl: version 2.4 or higher is required
# By setting the 'engine' in the DataFrame 'to_excel()' methods. df.to_excel('path_to_file.xlsx', sheet_name='Sheet1', engine='xlsxwriter') # By setting the 'engine' in the ExcelWriter constructor. writer = pd.ExcelWriter('path_to_file.xlsx', engine='xlsxwriter') # Or via pandas configuration. from pandas import options # noqa: E402 options.io.excel.xlsx.writer = 'xlsxwriter' df.to_excel('path_to_file.xlsx', sheet_name='Sheet1')
The look and feel of Excel worksheets created from pandas can be modified using the following parameters on the DataFrame’s to_excel method.
float_format : Format string for floating point numbers (default None).
freeze_panes : A tuple of two integers representing the bottommost row and rightmost column to freeze. Each of these parameters is one-based, so (1, 1) will freeze the first row and first column (default None).
freeze_panes
Using the Xlsxwriter engine provides many options for controlling the format of an Excel worksheet created with the to_excel method. Excellent examples can be found in the Xlsxwriter documentation here: https://xlsxwriter.readthedocs.io/working_with_pandas.html
New in version 0.25.
The read_excel() method can also read OpenDocument spreadsheets using the odfpy module. The semantics and features for reading OpenDocument spreadsheets match what can be done for Excel files using engine='odf'.
odfpy
engine='odf'
# Returns a DataFrame pd.read_excel('path_to_file.ods', engine='odf')
Currently pandas only supports reading OpenDocument spreadsheets. Writing is not implemented.
A handy way to grab data is to use the read_clipboard() method, which takes the contents of the clipboard buffer and passes them to the read_csv method. For instance, you can copy the following text to the clipboard (CTRL-C on many operating systems):
read_clipboard()
A B C x 1 4 p y 2 5 q z 3 6 r
And then import the data directly to a DataFrame by calling:
>>> clipdf = pd.read_clipboard() >>> clipdf A B C x 1 4 p y 2 5 q z 3 6 r
The to_clipboard method can be used to write the contents of a DataFrame to the clipboard. Following which you can paste the clipboard contents into other applications (CTRL-V on many operating systems). Here we illustrate writing a DataFrame into clipboard and reading it back.
>>> df = pd.DataFrame({'A': [1, 2, 3], ... 'B': [4, 5, 6], ... 'C': ['p', 'q', 'r']}, ... index=['x', 'y', 'z']) >>> df A B C x 1 4 p y 2 5 q z 3 6 r >>> df.to_clipboard() >>> pd.read_clipboard() A B C x 1 4 p y 2 5 q z 3 6 r
We can see that we got the same content back, which we had earlier written to the clipboard.
You may need to install xclip or xsel (with PyQt5, PyQt4 or qtpy) on Linux to use these methods.
All pandas objects are equipped with to_pickle methods which use Python’s cPickle module to save data structures to disk using the pickle format.
cPickle
In [328]: df Out[328]: c1 a c2 b d lvl1 lvl2 a c 1 5 d 2 6 b c 3 7 d 4 8 In [329]: df.to_pickle('foo.pkl')
The read_pickle function in the pandas namespace can be used to load any pickled pandas object (or any other pickled object) from file:
pandas
In [330]: pd.read_pickle('foo.pkl') Out[330]: c1 a c2 b d lvl1 lvl2 a c 1 5 d 2 6 b c 3 7 d 4 8
Loading pickled data received from untrusted sources can be unsafe.
See: https://docs.python.org/3/library/pickle.html
read_pickle() is only guaranteed backwards compatible back to pandas version 0.20.3
read_pickle()
read_pickle(), DataFrame.to_pickle() and Series.to_pickle() can read and write compressed pickle files. The compression types of gzip, bz2, xz are supported for reading and writing. The zip file format only supports reading and must contain only one data file to be read.
DataFrame.to_pickle()
Series.to_pickle()
gzip
bz2
xz
zip
The compression type can be an explicit parameter or be inferred from the file extension. If ‘infer’, then use gzip, bz2, zip, or xz if filename ends in '.gz', '.bz2', '.zip', or '.xz', respectively.
'.gz'
'.bz2'
'.zip'
'.xz'
In [331]: df = pd.DataFrame({ .....: 'A': np.random.randn(1000), .....: 'B': 'foo', .....: 'C': pd.date_range('20130101', periods=1000, freq='s')}) .....: In [332]: df Out[332]: A B C 0 -0.288267 foo 2013-01-01 00:00:00 1 -0.084905 foo 2013-01-01 00:00:01 2 0.004772 foo 2013-01-01 00:00:02 3 1.382989 foo 2013-01-01 00:00:03 4 0.343635 foo 2013-01-01 00:00:04 .. ... ... ... 995 -0.220893 foo 2013-01-01 00:16:35 996 0.492996 foo 2013-01-01 00:16:36 997 -0.461625 foo 2013-01-01 00:16:37 998 1.361779 foo 2013-01-01 00:16:38 999 -1.197988 foo 2013-01-01 00:16:39 [1000 rows x 3 columns]
Using an explicit compression type:
In [333]: df.to_pickle("data.pkl.compress", compression="gzip") In [334]: rt = pd.read_pickle("data.pkl.compress", compression="gzip") In [335]: rt Out[335]: A B C 0 -0.288267 foo 2013-01-01 00:00:00 1 -0.084905 foo 2013-01-01 00:00:01 2 0.004772 foo 2013-01-01 00:00:02 3 1.382989 foo 2013-01-01 00:00:03 4 0.343635 foo 2013-01-01 00:00:04 .. ... ... ... 995 -0.220893 foo 2013-01-01 00:16:35 996 0.492996 foo 2013-01-01 00:16:36 997 -0.461625 foo 2013-01-01 00:16:37 998 1.361779 foo 2013-01-01 00:16:38 999 -1.197988 foo 2013-01-01 00:16:39 [1000 rows x 3 columns]
Inferring compression type from the extension:
In [336]: df.to_pickle("data.pkl.xz", compression="infer") In [337]: rt = pd.read_pickle("data.pkl.xz", compression="infer") In [338]: rt Out[338]: A B C 0 -0.288267 foo 2013-01-01 00:00:00 1 -0.084905 foo 2013-01-01 00:00:01 2 0.004772 foo 2013-01-01 00:00:02 3 1.382989 foo 2013-01-01 00:00:03 4 0.343635 foo 2013-01-01 00:00:04 .. ... ... ... 995 -0.220893 foo 2013-01-01 00:16:35 996 0.492996 foo 2013-01-01 00:16:36 997 -0.461625 foo 2013-01-01 00:16:37 998 1.361779 foo 2013-01-01 00:16:38 999 -1.197988 foo 2013-01-01 00:16:39 [1000 rows x 3 columns]
The default is to ‘infer’:
In [339]: df.to_pickle("data.pkl.gz") In [340]: rt = pd.read_pickle("data.pkl.gz") In [341]: rt Out[341]: A B C 0 -0.288267 foo 2013-01-01 00:00:00 1 -0.084905 foo 2013-01-01 00:00:01 2 0.004772 foo 2013-01-01 00:00:02 3 1.382989 foo 2013-01-01 00:00:03 4 0.343635 foo 2013-01-01 00:00:04 .. ... ... ... 995 -0.220893 foo 2013-01-01 00:16:35 996 0.492996 foo 2013-01-01 00:16:36 997 -0.461625 foo 2013-01-01 00:16:37 998 1.361779 foo 2013-01-01 00:16:38 999 -1.197988 foo 2013-01-01 00:16:39 [1000 rows x 3 columns] In [342]: df["A"].to_pickle("s1.pkl.bz2") In [343]: rt = pd.read_pickle("s1.pkl.bz2") In [344]: rt Out[344]: 0 -0.288267 1 -0.084905 2 0.004772 3 1.382989 4 0.343635 ... 995 -0.220893 996 0.492996 997 -0.461625 998 1.361779 999 -1.197988 Name: A, Length: 1000, dtype: float64
pandas support for msgpack has been removed in version 1.0.0. It is recommended to use pyarrow for on-the-wire transmission of pandas objects.
msgpack
Example pyarrow usage:
>>> import pandas as pd >>> import pyarrow as pa >>> df = pd.DataFrame({'A': [1, 2, 3]}) >>> context = pa.default_serialization_context() >>> df_bytestring = context.serialize(df).to_buffer().to_pybytes()
For documentation on pyarrow, see here.
HDFStore is a dict-like object which reads and writes pandas using the high performance HDF5 format using the excellent PyTables library. See the cookbook for some advanced strategies
HDFStore
pandas requires PyTables >= 3.0.0. There is a indexing bug in PyTables < 3.2 which may appear when querying stores using an index. If you see a subset of results being returned, upgrade to PyTables >= 3.2. Stores created previously will need to be rewritten using the updated version.
PyTables
In [345]: store = pd.HDFStore('store.h5') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-345-54ac839dc497> in <module> ----> 1 store = pd.HDFStore('store.h5') ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [346]: print(store) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-346-34a6349b9096> in <module> ----> 1 print(store) NameError: name 'store' is not defined
Objects can be written to the file just like adding key-value pairs to a dict:
In [347]: index = pd.date_range('1/1/2000', periods=8) In [348]: s = pd.Series(np.random.randn(5), index=['a', 'b', 'c', 'd', 'e']) In [349]: df = pd.DataFrame(np.random.randn(8, 3), index=index, .....: columns=['A', 'B', 'C']) .....: # store.put('s', s) is an equivalent method In [350]: store['s'] = s --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-350-2bdbf26fda71> in <module> ----> 1 store['s'] = s NameError: name 'store' is not defined In [351]: store['df'] = df --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-351-02ded5cb794d> in <module> ----> 1 store['df'] = df NameError: name 'store' is not defined In [352]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-352-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined
In a current or later Python session, you can retrieve stored objects:
# store.get('df') is an equivalent method In [353]: store['df'] --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-353-1cc723079d4d> in <module> ----> 1 store['df'] NameError: name 'store' is not defined # dotted (attribute) access provides get as well In [354]: store.df --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-354-53088afd4d0c> in <module> ----> 1 store.df NameError: name 'store' is not defined
Deletion of the object specified by the key:
# store.remove('df') is an equivalent method In [355]: del store['df'] --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-355-ebd3cc5c3ec8> in <module> ----> 1 del store['df'] NameError: name 'store' is not defined In [356]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-356-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined
Closing a Store and using a context manager:
In [357]: store.close() --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-357-7d09d9e55ce9> in <module> ----> 1 store.close() NameError: name 'store' is not defined In [358]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-358-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined In [359]: store.is_open --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-359-11e2ce0fc911> in <module> ----> 1 store.is_open NameError: name 'store' is not defined # Working with, and automatically closing the store using a context manager In [360]: with pd.HDFStore('store.h5') as store: .....: store.keys() .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-360-e12c83740f64> in <module> ----> 1 with pd.HDFStore('store.h5') as store: 2 store.keys() ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables.
HDFStore supports a top-level API using read_hdf for reading and to_hdf for writing, similar to how read_csv and to_csv work.
In [361]: df_tl = pd.DataFrame({'A': list(range(5)), 'B': list(range(5))}) In [362]: df_tl.to_hdf('store_tl.h5', 'table', append=True) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-362-6a90542df09b> in <module> ----> 1 df_tl.to_hdf('store_tl.h5', 'table', append=True) ~/scipy/pandas/pandas/core/generic.py in to_hdf(self, path_or_buf, key, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 2468 data_columns=data_columns, 2469 errors=errors, -> 2470 encoding=encoding, 2471 ) 2472 ~/scipy/pandas/pandas/io/pytables.py in to_hdf(path_or_buf, key, value, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 278 if isinstance(path_or_buf, str): 279 with HDFStore( --> 280 path_or_buf, mode=mode, complevel=complevel, complib=complib 281 ) as store: 282 f(store) ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [363]: pd.read_hdf('store_tl.h5', 'table', where=['index>2']) --------------------------------------------------------------------------- FileNotFoundError Traceback (most recent call last) <ipython-input-363-48dfc76f8b8f> in <module> ----> 1 pd.read_hdf('store_tl.h5', 'table', where=['index>2']) ~/scipy/pandas/pandas/io/pytables.py in read_hdf(path_or_buf, key, mode, errors, where, start, stop, columns, iterator, chunksize, **kwargs) 393 394 if not exists: --> 395 raise FileNotFoundError(f"File {path_or_buf} does not exist") 396 397 store = HDFStore(path_or_buf, mode=mode, errors=errors, **kwargs) FileNotFoundError: File store_tl.h5 does not exist
HDFStore will by default not drop rows that are all missing. This behavior can be changed by setting dropna=True.
dropna=True
In [364]: df_with_missing = pd.DataFrame({'col1': [0, np.nan, 2], .....: 'col2': [1, np.nan, np.nan]}) .....: In [365]: df_with_missing Out[365]: col1 col2 0 0.0 1.0 1 NaN NaN 2 2.0 NaN In [366]: df_with_missing.to_hdf('file.h5', 'df_with_missing', .....: format='table', mode='w') .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-366-d760b349c4a7> in <module> 1 df_with_missing.to_hdf('file.h5', 'df_with_missing', ----> 2 format='table', mode='w') ~/scipy/pandas/pandas/core/generic.py in to_hdf(self, path_or_buf, key, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 2468 data_columns=data_columns, 2469 errors=errors, -> 2470 encoding=encoding, 2471 ) 2472 ~/scipy/pandas/pandas/io/pytables.py in to_hdf(path_or_buf, key, value, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 278 if isinstance(path_or_buf, str): 279 with HDFStore( --> 280 path_or_buf, mode=mode, complevel=complevel, complib=complib 281 ) as store: 282 f(store) ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [367]: pd.read_hdf('file.h5', 'df_with_missing') --------------------------------------------------------------------------- FileNotFoundError Traceback (most recent call last) <ipython-input-367-59c725e9cc6c> in <module> ----> 1 pd.read_hdf('file.h5', 'df_with_missing') ~/scipy/pandas/pandas/io/pytables.py in read_hdf(path_or_buf, key, mode, errors, where, start, stop, columns, iterator, chunksize, **kwargs) 393 394 if not exists: --> 395 raise FileNotFoundError(f"File {path_or_buf} does not exist") 396 397 store = HDFStore(path_or_buf, mode=mode, errors=errors, **kwargs) FileNotFoundError: File file.h5 does not exist In [368]: df_with_missing.to_hdf('file.h5', 'df_with_missing', .....: format='table', mode='w', dropna=True) .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-368-263c86d5a881> in <module> 1 df_with_missing.to_hdf('file.h5', 'df_with_missing', ----> 2 format='table', mode='w', dropna=True) ~/scipy/pandas/pandas/core/generic.py in to_hdf(self, path_or_buf, key, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 2468 data_columns=data_columns, 2469 errors=errors, -> 2470 encoding=encoding, 2471 ) 2472 ~/scipy/pandas/pandas/io/pytables.py in to_hdf(path_or_buf, key, value, mode, complevel, complib, append, format, index, min_itemsize, nan_rep, dropna, data_columns, errors, encoding) 278 if isinstance(path_or_buf, str): 279 with HDFStore( --> 280 path_or_buf, mode=mode, complevel=complevel, complib=complib 281 ) as store: 282 f(store) ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [369]: pd.read_hdf('file.h5', 'df_with_missing') --------------------------------------------------------------------------- FileNotFoundError Traceback (most recent call last) <ipython-input-369-59c725e9cc6c> in <module> ----> 1 pd.read_hdf('file.h5', 'df_with_missing') ~/scipy/pandas/pandas/io/pytables.py in read_hdf(path_or_buf, key, mode, errors, where, start, stop, columns, iterator, chunksize, **kwargs) 393 394 if not exists: --> 395 raise FileNotFoundError(f"File {path_or_buf} does not exist") 396 397 store = HDFStore(path_or_buf, mode=mode, errors=errors, **kwargs) FileNotFoundError: File file.h5 does not exist
The examples above show storing using put, which write the HDF5 to PyTables in a fixed array format, called the fixed format. These types of stores are not appendable once written (though you can simply remove them and rewrite). Nor are they queryable; they must be retrieved in their entirety. They also do not support dataframes with non-unique column names. The fixed format stores offer very fast writing and slightly faster reading than table stores. This format is specified by default when using put or to_hdf or by format='fixed' or format='f'.
put
fixed
format='fixed'
format='f'
A fixed format will raise a TypeError if you try to retrieve using a where:
TypeError
where
>>> pd.DataFrame(np.random.randn(10, 2)).to_hdf('test_fixed.h5', 'df') >>> pd.read_hdf('test_fixed.h5', 'df', where='index>5') TypeError: cannot pass a where specification when reading a fixed format. this store must be selected in its entirety
HDFStore supports another PyTables format on disk, the table format. Conceptually a table is shaped very much like a DataFrame, with rows and columns. A table may be appended to in the same or other sessions. In addition, delete and query type operations are supported. This format is specified by format='table' or format='t' to append or put or to_hdf.
format='table'
format='t'
append
This format can be set as an option as well pd.set_option('io.hdf.default_format','table') to enable put/append/to_hdf to by default store in the table format.
pd.set_option('io.hdf.default_format','table')
put/append/to_hdf
In [370]: store = pd.HDFStore('store.h5') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-370-54ac839dc497> in <module> ----> 1 store = pd.HDFStore('store.h5') ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [371]: df1 = df[0:4] In [372]: df2 = df[4:] # append data (creates a table automatically) In [373]: store.append('df', df1) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-373-5d2670dce5e2> in <module> ----> 1 store.append('df', df1) NameError: name 'store' is not defined In [374]: store.append('df', df2) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-374-d0e0be4a91a9> in <module> ----> 1 store.append('df', df2) NameError: name 'store' is not defined In [375]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-375-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined # select the entire object In [376]: store.select('df') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-376-217756ae0052> in <module> ----> 1 store.select('df') NameError: name 'store' is not defined # the type of stored data In [377]: store.root.df._v_attrs.pandas_type --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-377-ef62b5b0aab6> in <module> ----> 1 store.root.df._v_attrs.pandas_type NameError: name 'store' is not defined
You can also create a table by passing format='table' or format='t' to a put operation.
Keys to a store can be specified as a string. These can be in a hierarchical path-name like format (e.g. foo/bar/bah), which will generate a hierarchy of sub-stores (or Groups in PyTables parlance). Keys can be specified without the leading ‘/’ and are always absolute (e.g. ‘foo’ refers to ‘/foo’). Removal operations can remove everything in the sub-store and below, so be careful.
foo/bar/bah
Groups
In [378]: store.put('foo/bar/bah', df) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-378-e34dc8f9055e> in <module> ----> 1 store.put('foo/bar/bah', df) NameError: name 'store' is not defined In [379]: store.append('food/orange', df) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-379-dcad8aaa721a> in <module> ----> 1 store.append('food/orange', df) NameError: name 'store' is not defined In [380]: store.append('food/apple', df) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-380-b3de712d0d57> in <module> ----> 1 store.append('food/apple', df) NameError: name 'store' is not defined In [381]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-381-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined # a list of keys are returned In [382]: store.keys() --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-382-0aa828311813> in <module> ----> 1 store.keys() NameError: name 'store' is not defined # remove all nodes under this level In [383]: store.remove('food') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-383-f156e07a9855> in <module> ----> 1 store.remove('food') NameError: name 'store' is not defined In [384]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-384-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined
You can walk through the group hierarchy using the walk method which will yield a tuple for each group key along with the relative keys of its contents.
walk
New in version 0.24.0.
In [385]: for (path, subgroups, subkeys) in store.walk(): .....: for subgroup in subgroups: .....: print('GROUP: {}/{}'.format(path, subgroup)) .....: for subkey in subkeys: .....: key = '/'.join([path, subkey]) .....: print('KEY: {}'.format(key)) .....: print(store.get(key)) .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-385-2d918dbe0f51> in <module> ----> 1 for (path, subgroups, subkeys) in store.walk(): 2 for subgroup in subgroups: 3 print('GROUP: {}/{}'.format(path, subgroup)) 4 for subkey in subkeys: 5 key = '/'.join([path, subkey]) NameError: name 'store' is not defined
Hierarchical keys cannot be retrieved as dotted (attribute) access as described above for items stored under the root node.
In [8]: store.foo.bar.bah AttributeError: 'HDFStore' object has no attribute 'foo' # you can directly access the actual PyTables node but using the root node In [9]: store.root.foo.bar.bah Out[9]: /foo/bar/bah (Group) '' children := ['block0_items' (Array), 'block0_values' (Array), 'axis0' (Array), 'axis1' (Array)]
Instead, use explicit string based keys:
In [386]: store['foo/bar/bah'] --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-386-d4ee51f6bb0c> in <module> ----> 1 store['foo/bar/bah'] NameError: name 'store' is not defined
Storing mixed-dtype data is supported. Strings are stored as a fixed-width using the maximum size of the appended column. Subsequent attempts at appending longer strings will raise a ValueError.
Passing min_itemsize={`values`: size} as a parameter to append will set a larger minimum for the string columns. Storing floats, strings, ints, bools, datetime64 are currently supported. For string columns, passing nan_rep = 'nan' to append will change the default nan representation on disk (which converts to/from np.nan), this defaults to nan.
min_itemsize={`values`: size}
floats, strings, ints, bools, datetime64
nan_rep = 'nan'
In [387]: df_mixed = pd.DataFrame({'A': np.random.randn(8), .....: 'B': np.random.randn(8), .....: 'C': np.array(np.random.randn(8), dtype='float32'), .....: 'string': 'string', .....: 'int': 1, .....: 'bool': True, .....: 'datetime64': pd.Timestamp('20010102')}, .....: index=list(range(8))) .....: In [388]: df_mixed.loc[df_mixed.index[3:5], .....: ['A', 'B', 'string', 'datetime64']] = np.nan .....: In [389]: store.append('df_mixed', df_mixed, min_itemsize={'values': 50}) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-389-84f39588825d> in <module> ----> 1 store.append('df_mixed', df_mixed, min_itemsize={'values': 50}) NameError: name 'store' is not defined In [390]: df_mixed1 = store.select('df_mixed') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-390-8540a95eefa5> in <module> ----> 1 df_mixed1 = store.select('df_mixed') NameError: name 'store' is not defined In [391]: df_mixed1 --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-391-7c2d528034e6> in <module> ----> 1 df_mixed1 NameError: name 'df_mixed1' is not defined In [392]: df_mixed1.dtypes.value_counts() --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-392-7de1c772ff05> in <module> ----> 1 df_mixed1.dtypes.value_counts() NameError: name 'df_mixed1' is not defined # we have provided a minimum string column size In [393]: store.root.df_mixed.table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-393-eb0bb4c187f7> in <module> ----> 1 store.root.df_mixed.table NameError: name 'store' is not defined
Storing MultiIndex DataFrames as tables is very similar to storing/selecting from homogeneous index DataFrames.
In [394]: index = pd.MultiIndex(levels=[['foo', 'bar', 'baz', 'qux'], .....: ['one', 'two', 'three']], .....: codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], .....: [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]], .....: names=['foo', 'bar']) .....: In [395]: df_mi = pd.DataFrame(np.random.randn(10, 3), index=index, .....: columns=['A', 'B', 'C']) .....: In [396]: df_mi Out[396]: A B C foo bar foo one 0.667450 0.169405 -1.358046 two -0.105563 0.492195 0.076693 three 0.213685 -0.285283 -1.210529 bar one -1.408386 0.941577 -0.342447 two 0.222031 0.052607 2.093214 baz two 1.064908 1.778161 -0.913867 three -0.030004 -0.399846 -1.234765 qux one 0.081323 -0.268494 0.168016 two -0.898283 -0.218499 1.408028 three -1.267828 -0.689263 0.520995 In [397]: store.append('df_mi', df_mi) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-397-3b47a988298a> in <module> ----> 1 store.append('df_mi', df_mi) NameError: name 'store' is not defined In [398]: store.select('df_mi') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-398-07196865d252> in <module> ----> 1 store.select('df_mi') NameError: name 'store' is not defined # the levels are automatically included as data columns In [399]: store.select('df_mi', 'foo=bar') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-399-515b984b10eb> in <module> ----> 1 store.select('df_mi', 'foo=bar') NameError: name 'store' is not defined
The index keyword is reserved and cannot be use as a level name.
select and delete operations have an optional criterion that can be specified to select/delete only a subset of the data. This allows one to have a very large on-disk table and retrieve only a portion of the data.
select
delete
A query is specified using the Term class under the hood, as a boolean expression.
Term
index and columns are supported indexers of DataFrames.
if data_columns are specified, these can be used as additional indexers.
data_columns
level name in a MultiIndex, with default name level_0, level_1, … if not provided.
level_0
level_1
Valid comparison operators are:
=, ==, !=, >, >=, <, <=
Valid boolean expressions are combined with:
| : or
|
& : and
&
( and ) : for grouping
(
)
These rules are similar to how boolean expressions are used in pandas for indexing.
= will be automatically expanded to the comparison operator ==
=
==
~ is the not operator, but can only be used in very limited circumstances
~
If a list/tuple of expressions is passed they will be combined via &
The following are valid expressions:
'index >= date'
"columns = ['A', 'D']"
"columns in ['A', 'D']"
'columns = A'
'columns == A'
"~(columns = ['A', 'B'])"
'index > df.index[3] & string = "bar"'
'(index > df.index[3] & index <= df.index[6]) | string = "bar"'
"ts >= Timestamp('2012-02-01')"
"major_axis>=20130101"
The indexers are on the left-hand side of the sub-expression:
indexers
columns, major_axis, ts
major_axis
ts
The right-hand side of the sub-expression (after a comparison operator) can be:
functions that will be evaluated, e.g. Timestamp('2012-02-01')
Timestamp('2012-02-01')
strings, e.g. "bar"
"bar"
date-like, e.g. 20130101, or "20130101"
20130101
"20130101"
lists, e.g. "['A', 'B']"
"['A', 'B']"
variables that are defined in the local names space, e.g. date
date
Passing a string to a query by interpolating it into the query expression is not recommended. Simply assign the string of interest to a variable and use that variable in an expression. For example, do this
string = "HolyMoly'" store.select('df', 'index == string')
instead of this
string = "HolyMoly'" store.select('df', 'index == %s' % string)
The latter will not work and will raise a SyntaxError.Note that there’s a single quote followed by a double quote in the string variable.
SyntaxError
string
If you must interpolate, use the '%r' format specifier
'%r'
store.select('df', 'index == %r' % string)
which will quote string.
Here are some examples:
In [400]: dfq = pd.DataFrame(np.random.randn(10, 4), columns=list('ABCD'), .....: index=pd.date_range('20130101', periods=10)) .....: In [401]: store.append('dfq', dfq, format='table', data_columns=True) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-401-82c5e5a63f3a> in <module> ----> 1 store.append('dfq', dfq, format='table', data_columns=True) NameError: name 'store' is not defined
Use boolean expressions, with in-line function evaluation.
In [402]: store.select('dfq', "index>pd.Timestamp('20130104') & columns=['A', 'B']") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-402-31fefd59cc8b> in <module> ----> 1 store.select('dfq', "index>pd.Timestamp('20130104') & columns=['A', 'B']") NameError: name 'store' is not defined
Use inline column reference.
In [403]: store.select('dfq', where="A>0 or C>0") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-403-9335f8ce7ae3> in <module> ----> 1 store.select('dfq', where="A>0 or C>0") NameError: name 'store' is not defined
The columns keyword can be supplied to select a list of columns to be returned, this is equivalent to passing a 'columns=list_of_columns_to_filter':
'columns=list_of_columns_to_filter'
In [404]: store.select('df', "columns=['A', 'B']") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-404-8f20fee3a6a5> in <module> ----> 1 store.select('df', "columns=['A', 'B']") NameError: name 'store' is not defined
start and stop parameters can be specified to limit the total search space. These are in terms of the total number of rows in a table.
start
stop
select will raise a ValueError if the query expression has an unknown variable reference. Usually this means that you are trying to select on a column that is not a data_column.
select will raise a SyntaxError if the query expression is not valid.
You can store and query using the timedelta64[ns] type. Terms can be specified in the format: <float>(<unit>), where float may be signed (and fractional), and unit can be D,s,ms,us,ns for the timedelta. Here’s an example:
<float>(<unit>)
D,s,ms,us,ns
In [405]: from datetime import timedelta In [406]: dftd = pd.DataFrame({'A': pd.Timestamp('20130101'), .....: 'B': [pd.Timestamp('20130101') + timedelta(days=i, .....: seconds=10) .....: for i in range(10)]}) .....: In [407]: dftd['C'] = dftd['A'] - dftd['B'] In [408]: dftd Out[408]: A B C 0 2013-01-01 2013-01-01 00:00:10 -1 days +23:59:50 1 2013-01-01 2013-01-02 00:00:10 -2 days +23:59:50 2 2013-01-01 2013-01-03 00:00:10 -3 days +23:59:50 3 2013-01-01 2013-01-04 00:00:10 -4 days +23:59:50 4 2013-01-01 2013-01-05 00:00:10 -5 days +23:59:50 5 2013-01-01 2013-01-06 00:00:10 -6 days +23:59:50 6 2013-01-01 2013-01-07 00:00:10 -7 days +23:59:50 7 2013-01-01 2013-01-08 00:00:10 -8 days +23:59:50 8 2013-01-01 2013-01-09 00:00:10 -9 days +23:59:50 9 2013-01-01 2013-01-10 00:00:10 -10 days +23:59:50 In [409]: store.append('dftd', dftd, data_columns=True) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-409-2f9a092b0b0c> in <module> ----> 1 store.append('dftd', dftd, data_columns=True) NameError: name 'store' is not defined In [410]: store.select('dftd', "C<'-3.5D'") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-410-2a9409d9d4cc> in <module> ----> 1 store.select('dftd', "C<'-3.5D'") NameError: name 'store' is not defined
Selecting from a MultiIndex can be achieved by using the name of the level.
In [411]: df_mi.index.names Out[411]: FrozenList(['foo', 'bar']) In [412]: store.select('df_mi', "foo=baz and bar=two") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-412-ce8999831bda> in <module> ----> 1 store.select('df_mi', "foo=baz and bar=two") NameError: name 'store' is not defined
If the MultiIndex levels names are None, the levels are automatically made available via the level_n keyword with n the level of the MultiIndex you want to select from.
level_n
n
In [413]: index = pd.MultiIndex( .....: levels=[["foo", "bar", "baz", "qux"], ["one", "two", "three"]], .....: codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]], .....: ) .....: In [414]: df_mi_2 = pd.DataFrame(np.random.randn(10, 3), .....: index=index, columns=["A", "B", "C"]) .....: In [415]: df_mi_2 Out[415]: A B C foo one 0.856838 1.491776 0.001283 two 0.701816 -1.097917 0.102588 three 0.661740 0.443531 0.559313 bar one -0.459055 -1.222598 -0.455304 two -0.781163 0.826204 -0.530057 baz two 0.296135 1.366810 1.073372 three -0.994957 0.755314 2.119746 qux one -2.628174 -0.089460 -0.133636 two 0.337920 -0.634027 0.421107 three 0.604303 1.053434 1.109090 In [416]: store.append("df_mi_2", df_mi_2) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-416-a1e1938eea12> in <module> ----> 1 store.append("df_mi_2", df_mi_2) NameError: name 'store' is not defined # the levels are automatically included as data columns with keyword level_n In [417]: store.select("df_mi_2", "level_0=foo and level_1=two") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-417-2dd3927ca4dd> in <module> ----> 1 store.select("df_mi_2", "level_0=foo and level_1=two") NameError: name 'store' is not defined
You can create/modify an index for a table with create_table_index after data is already in the table (after and append/put operation). Creating a table index is highly encouraged. This will speed your queries a great deal when you use a select with the indexed dimension as the where.
create_table_index
append/put
Indexes are automagically created on the indexables and any data columns you specify. This behavior can be turned off by passing index=False to append.
index=False
# we have automagically already created an index (in the first section) In [418]: i = store.root.df.table.cols.index.index --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-418-df460de23e31> in <module> ----> 1 i = store.root.df.table.cols.index.index NameError: name 'store' is not defined In [419]: i.optlevel, i.kind --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-419-18f1f93c3266> in <module> ----> 1 i.optlevel, i.kind AttributeError: 'DatetimeIndex' object has no attribute 'optlevel' # change an index by passing new parameters In [420]: store.create_table_index('df', optlevel=9, kind='full') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-420-9d3aefb100d8> in <module> ----> 1 store.create_table_index('df', optlevel=9, kind='full') NameError: name 'store' is not defined In [421]: i = store.root.df.table.cols.index.index --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-421-df460de23e31> in <module> ----> 1 i = store.root.df.table.cols.index.index NameError: name 'store' is not defined In [422]: i.optlevel, i.kind --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-422-18f1f93c3266> in <module> ----> 1 i.optlevel, i.kind AttributeError: 'DatetimeIndex' object has no attribute 'optlevel'
Oftentimes when appending large amounts of data to a store, it is useful to turn off index creation for each append, then recreate at the end.
In [423]: df_1 = pd.DataFrame(np.random.randn(10, 2), columns=list('AB')) In [424]: df_2 = pd.DataFrame(np.random.randn(10, 2), columns=list('AB')) In [425]: st = pd.HDFStore('appends.h5', mode='w') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-425-be0dd03dac58> in <module> ----> 1 st = pd.HDFStore('appends.h5', mode='w') ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [426]: st.append('df', df_1, data_columns=['B'], index=False) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-426-57eea5b3943f> in <module> ----> 1 st.append('df', df_1, data_columns=['B'], index=False) NameError: name 'st' is not defined In [427]: st.append('df', df_2, data_columns=['B'], index=False) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-427-b819d1a883ed> in <module> ----> 1 st.append('df', df_2, data_columns=['B'], index=False) NameError: name 'st' is not defined In [428]: st.get_storer('df').table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-428-a9163e299a37> in <module> ----> 1 st.get_storer('df').table NameError: name 'st' is not defined
Then create the index when finished appending.
In [429]: st.create_table_index('df', columns=['B'], optlevel=9, kind='full') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-429-99f7165ea423> in <module> ----> 1 st.create_table_index('df', columns=['B'], optlevel=9, kind='full') NameError: name 'st' is not defined In [430]: st.get_storer('df').table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-430-a9163e299a37> in <module> ----> 1 st.get_storer('df').table NameError: name 'st' is not defined In [431]: st.close() --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-431-5daf6d1e2155> in <module> ----> 1 st.close() NameError: name 'st' is not defined
See here for how to create a completely-sorted-index (CSI) on an existing store.
You can designate (and index) certain columns that you want to be able to perform queries (other than the indexable columns, which you can always query). For instance say you want to perform this common operation, on-disk, and return just the frame that matches this query. You can specify data_columns = True to force all columns to be data_columns.
data_columns = True
In [432]: df_dc = df.copy() In [433]: df_dc['string'] = 'foo' In [434]: df_dc.loc[df_dc.index[4:6], 'string'] = np.nan In [435]: df_dc.loc[df_dc.index[7:9], 'string'] = 'bar' In [436]: df_dc['string2'] = 'cool' In [437]: df_dc.loc[df_dc.index[1:3], ['B', 'C']] = 1.0 In [438]: df_dc Out[438]: A B C string string2 2000-01-01 1.334065 0.521036 0.930384 foo cool 2000-01-02 -1.613932 1.000000 1.000000 foo cool 2000-01-03 -0.585314 1.000000 1.000000 foo cool 2000-01-04 0.632369 -1.249657 0.975593 foo cool 2000-01-05 1.060617 -0.143682 0.218423 NaN cool 2000-01-06 3.050329 1.317933 -0.963725 NaN cool 2000-01-07 -0.539452 -0.771133 0.023751 foo cool 2000-01-08 0.649464 -1.736427 0.197288 bar cool # on-disk operations In [439]: store.append('df_dc', df_dc, data_columns=['B', 'C', 'string', 'string2']) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-439-f4f595672983> in <module> ----> 1 store.append('df_dc', df_dc, data_columns=['B', 'C', 'string', 'string2']) NameError: name 'store' is not defined In [440]: store.select('df_dc', where='B > 0') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-440-ec03c35decfc> in <module> ----> 1 store.select('df_dc', where='B > 0') NameError: name 'store' is not defined # getting creative In [441]: store.select('df_dc', 'B > 0 & C > 0 & string == foo') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-441-ff1ed859b7a9> in <module> ----> 1 store.select('df_dc', 'B > 0 & C > 0 & string == foo') NameError: name 'store' is not defined # this is in-memory version of this type of selection In [442]: df_dc[(df_dc.B > 0) & (df_dc.C > 0) & (df_dc.string == 'foo')] Out[442]: A B C string string2 2000-01-01 1.334065 0.521036 0.930384 foo cool 2000-01-02 -1.613932 1.000000 1.000000 foo cool 2000-01-03 -0.585314 1.000000 1.000000 foo cool # we have automagically created this index and the B/C/string/string2 # columns are stored separately as ``PyTables`` columns In [443]: store.root.df_dc.table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-443-77ef75f2a289> in <module> ----> 1 store.root.df_dc.table NameError: name 'store' is not defined
There is some performance degradation by making lots of columns into data columns, so it is up to the user to designate these. In addition, you cannot change data columns (nor indexables) after the first append/put operation (Of course you can simply read in the data and create a new table!).
You can pass iterator=True or chunksize=number_in_a_chunk to select and select_as_multiple to return an iterator on the results. The default is 50,000 rows returned in a chunk.
chunksize=number_in_a_chunk
select_as_multiple
In [444]: for df in store.select('df', chunksize=3): .....: print(df) .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-444-4e0e7b0a8a32> in <module> ----> 1 for df in store.select('df', chunksize=3): 2 print(df) NameError: name 'store' is not defined
You can also use the iterator with read_hdf which will open, then automatically close the store when finished iterating.
for df in pd.read_hdf('store.h5', 'df', chunksize=3): print(df)
Note, that the chunksize keyword applies to the source rows. So if you are doing a query, then the chunksize will subdivide the total rows in the table and the query applied, returning an iterator on potentially unequal sized chunks.
Here is a recipe for generating a query and using it to create equal sized return chunks.
In [445]: dfeq = pd.DataFrame({'number': np.arange(1, 11)}) In [446]: dfeq Out[446]: number 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 10 In [447]: store.append('dfeq', dfeq, data_columns=['number']) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-447-7e47358a30c7> in <module> ----> 1 store.append('dfeq', dfeq, data_columns=['number']) NameError: name 'store' is not defined In [448]: def chunks(l, n): .....: return [l[i:i + n] for i in range(0, len(l), n)] .....: In [449]: evens = [2, 4, 6, 8, 10] In [450]: coordinates = store.select_as_coordinates('dfeq', 'number=evens') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-450-18616f8f97fd> in <module> ----> 1 coordinates = store.select_as_coordinates('dfeq', 'number=evens') NameError: name 'store' is not defined In [451]: for c in chunks(coordinates, 2): .....: print(store.select('dfeq', where=c)) .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-451-caf9949119a3> in <module> ----> 1 for c in chunks(coordinates, 2): 2 print(store.select('dfeq', where=c)) NameError: name 'coordinates' is not defined
To retrieve a single indexable or data column, use the method select_column. This will, for example, enable you to get the index very quickly. These return a Series of the result, indexed by the row number. These do not currently accept the where selector.
select_column
In [452]: store.select_column('df_dc', 'index') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-452-471e895df9a6> in <module> ----> 1 store.select_column('df_dc', 'index') NameError: name 'store' is not defined In [453]: store.select_column('df_dc', 'string') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-453-bca5961a1371> in <module> ----> 1 store.select_column('df_dc', 'string') NameError: name 'store' is not defined
Sometimes you want to get the coordinates (a.k.a the index locations) of your query. This returns an Int64Index of the resulting locations. These coordinates can also be passed to subsequent where operations.
Int64Index
In [454]: df_coord = pd.DataFrame(np.random.randn(1000, 2), .....: index=pd.date_range('20000101', periods=1000)) .....: In [455]: store.append('df_coord', df_coord) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-455-92517ada9e7b> in <module> ----> 1 store.append('df_coord', df_coord) NameError: name 'store' is not defined In [456]: c = store.select_as_coordinates('df_coord', 'index > 20020101') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-456-5a37e7982f8c> in <module> ----> 1 c = store.select_as_coordinates('df_coord', 'index > 20020101') NameError: name 'store' is not defined In [457]: c Out[457]: [b, c, a, b] Categories (3, object): [b, c, a] In [458]: store.select('df_coord', where=c) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-458-12a0dfd76804> in <module> ----> 1 store.select('df_coord', where=c) NameError: name 'store' is not defined
Sometime your query can involve creating a list of rows to select. Usually this mask would be a resulting index from an indexing operation. This example selects the months of a datetimeindex which are 5.
mask
In [459]: df_mask = pd.DataFrame(np.random.randn(1000, 2), .....: index=pd.date_range('20000101', periods=1000)) .....: In [460]: store.append('df_mask', df_mask) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-460-d845638bd163> in <module> ----> 1 store.append('df_mask', df_mask) NameError: name 'store' is not defined In [461]: c = store.select_column('df_mask', 'index') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-461-02fa60af8861> in <module> ----> 1 c = store.select_column('df_mask', 'index') NameError: name 'store' is not defined In [462]: where = c[pd.DatetimeIndex(c).month == 5].index --------------------------------------------------------------------------- TypeError Traceback (most recent call last) ~/scipy/pandas/pandas/core/arrays/datetimes.py in objects_to_datetime64ns(data, dayfirst, yearfirst, utc, errors, require_iso8601, allow_object) 1852 try: -> 1853 values, tz_parsed = conversion.datetime_to_datetime64(data) 1854 # If tzaware, these values represent unix timestamps, so we ~/scipy/pandas/pandas/_libs/tslibs/conversion.pyx in pandas._libs.tslibs.conversion.datetime_to_datetime64() TypeError: Unrecognized value type: <class 'str'> During handling of the above exception, another exception occurred: ValueError Traceback (most recent call last) <ipython-input-462-74cb6f2edb69> in <module> ----> 1 where = c[pd.DatetimeIndex(c).month == 5].index ~/scipy/pandas/pandas/core/indexes/datetimes.py in __new__(cls, data, freq, tz, normalize, closed, ambiguous, dayfirst, yearfirst, dtype, copy, name) 258 dayfirst=dayfirst, 259 yearfirst=yearfirst, --> 260 ambiguous=ambiguous, 261 ) 262 ~/scipy/pandas/pandas/core/arrays/datetimes.py in _from_sequence(cls, data, dtype, copy, tz, freq, dayfirst, yearfirst, ambiguous) 315 dayfirst=dayfirst, 316 yearfirst=yearfirst, --> 317 ambiguous=ambiguous, 318 ) 319 ~/scipy/pandas/pandas/core/arrays/datetimes.py in sequence_to_dt64ns(data, dtype, copy, tz, dayfirst, yearfirst, ambiguous) 1748 # or M8[ns] to denote wall times 1749 data, inferred_tz = objects_to_datetime64ns( -> 1750 data, dayfirst=dayfirst, yearfirst=yearfirst 1751 ) 1752 tz = maybe_infer_tz(tz, inferred_tz) ~/scipy/pandas/pandas/core/arrays/datetimes.py in objects_to_datetime64ns(data, dayfirst, yearfirst, utc, errors, require_iso8601, allow_object) 1856 return values.view("i8"), tz_parsed 1857 except (ValueError, TypeError): -> 1858 raise e 1859 1860 if tz_parsed is not None: ~/scipy/pandas/pandas/core/arrays/datetimes.py in objects_to_datetime64ns(data, dayfirst, yearfirst, utc, errors, require_iso8601, allow_object) 1847 dayfirst=dayfirst, 1848 yearfirst=yearfirst, -> 1849 require_iso8601=require_iso8601, 1850 ) 1851 except ValueError as e: ~/scipy/pandas/pandas/_libs/tslib.pyx in pandas._libs.tslib.array_to_datetime() ~/scipy/pandas/pandas/_libs/tslib.pyx in pandas._libs.tslib.array_to_datetime() ~/scipy/pandas/pandas/_libs/tslib.pyx in pandas._libs.tslib.array_to_datetime_object() ~/scipy/pandas/pandas/_libs/tslib.pyx in pandas._libs.tslib.array_to_datetime_object() ~/scipy/pandas/pandas/_libs/tslibs/parsing.pyx in pandas._libs.tslibs.parsing.parse_datetime_string() ~/miniconda3/envs/docs/lib/python3.7/site-packages/dateutil/parser/_parser.py in parse(timestr, parserinfo, **kwargs) 1356 return parser(parserinfo).parse(timestr, **kwargs) 1357 else: -> 1358 return DEFAULTPARSER.parse(timestr, **kwargs) 1359 1360 ~/miniconda3/envs/docs/lib/python3.7/site-packages/dateutil/parser/_parser.py in parse(self, timestr, default, ignoretz, tzinfos, **kwargs) 647 648 if res is None: --> 649 raise ValueError("Unknown string format:", timestr) 650 651 if len(res) == 0: ValueError: ('Unknown string format:', 'b') In [463]: store.select('df_mask', where=where) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-463-13e54c09efc9> in <module> ----> 1 store.select('df_mask', where=where) NameError: name 'store' is not defined
If you want to inspect the stored object, retrieve via get_storer. You could use this programmatically to say get the number of rows in an object.
get_storer
In [464]: store.get_storer('df_dc').nrows --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-464-f54320d10554> in <module> ----> 1 store.get_storer('df_dc').nrows NameError: name 'store' is not defined
The methods append_to_multiple and select_as_multiple can perform appending/selecting from multiple tables at once. The idea is to have one table (call it the selector table) that you index most/all of the columns, and perform your queries. The other table(s) are data tables with an index matching the selector table’s index. You can then perform a very fast query on the selector table, yet get lots of data back. This method is similar to having a very wide table, but enables more efficient queries.
append_to_multiple
The append_to_multiple method splits a given single DataFrame into multiple tables according to d, a dictionary that maps the table names to a list of ‘columns’ you want in that table. If None is used in place of a list, that table will have the remaining unspecified columns of the given DataFrame. The argument selector defines which table is the selector table (which you can make queries from). The argument dropna will drop rows from the input DataFrame to ensure tables are synchronized. This means that if a row for one of the tables being written to is entirely np.NaN, that row will be dropped from all tables.
d
selector
dropna
np.NaN
If dropna is False, THE USER IS RESPONSIBLE FOR SYNCHRONIZING THE TABLES. Remember that entirely np.Nan rows are not written to the HDFStore, so if you choose to call dropna=False, some tables may have more rows than others, and therefore select_as_multiple may not work or it may return unexpected results.
np.Nan
dropna=False
In [465]: df_mt = pd.DataFrame(np.random.randn(8, 6), .....: index=pd.date_range('1/1/2000', periods=8), .....: columns=['A', 'B', 'C', 'D', 'E', 'F']) .....: In [466]: df_mt['foo'] = 'bar' In [467]: df_mt.loc[df_mt.index[1], ('A', 'B')] = np.nan # you can also create the tables individually In [468]: store.append_to_multiple({'df1_mt': ['A', 'B'], 'df2_mt': None}, .....: df_mt, selector='df1_mt') .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-468-0f2ea287af29> in <module> ----> 1 store.append_to_multiple({'df1_mt': ['A', 'B'], 'df2_mt': None}, 2 df_mt, selector='df1_mt') NameError: name 'store' is not defined In [469]: store --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-469-1cc0c364c02d> in <module> ----> 1 store NameError: name 'store' is not defined # individual tables were created In [470]: store.select('df1_mt') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-470-cd604a6b0b71> in <module> ----> 1 store.select('df1_mt') NameError: name 'store' is not defined In [471]: store.select('df2_mt') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-471-a3ed0f55544c> in <module> ----> 1 store.select('df2_mt') NameError: name 'store' is not defined # as a multiple In [472]: store.select_as_multiple(['df1_mt', 'df2_mt'], where=['A>0', 'B>0'], .....: selector='df1_mt') .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-472-e97f27782b41> in <module> ----> 1 store.select_as_multiple(['df1_mt', 'df2_mt'], where=['A>0', 'B>0'], 2 selector='df1_mt') NameError: name 'store' is not defined
You can delete from a table selectively by specifying a where. In deleting rows, it is important to understand the PyTables deletes rows by erasing the rows, then moving the following data. Thus deleting can potentially be a very expensive operation depending on the orientation of your data. To get optimal performance, it’s worthwhile to have the dimension you are deleting be the first of the indexables.
indexables
Data is ordered (on the disk) in terms of the indexables. Here’s a simple use case. You store panel-type data, with dates in the major_axis and ids in the minor_axis. The data is then interleaved like this:
minor_axis
id_1
id_2
.
id_n
It should be clear that a delete operation on the major_axis will be fairly quick, as one chunk is removed, then the following data moved. On the other hand a delete operation on the minor_axis will be very expensive. In this case it would almost certainly be faster to rewrite the table using a where that selects all but the missing data.
Please note that HDF5 DOES NOT RECLAIM SPACE in the h5 files automatically. Thus, repeatedly deleting (or removing nodes) and adding again, WILL TEND TO INCREASE THE FILE SIZE.
To repack and clean the file, use ptrepack.
PyTables allows the stored data to be compressed. This applies to all kinds of stores, not just tables. Two parameters are used to control compression: complevel and complib.
complevel
complib
complevel=0 and complevel=None disables compression and 0<complevel<10 enables compression.
complevel=0
complevel=None
0<complevel<10
specified the default library zlib is used. A compression library usually optimizes for either good compression rates or speed and the results will depend on the type of data. Which type of compression to choose depends on your specific needs and data. The list of supported compression libraries:
zlib
zlib: The default compression library. A classic in terms of compression, achieves good compression rates but is somewhat slow. lzo: Fast compression and decompression. bzip2: Good compression rates. blosc: Fast compression and decompression. Support for alternative blosc compressors: blosc:blosclz This is the default compressor for blosc blosc:lz4: A compact, very popular and fast compressor. blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of speed. blosc:snappy: A popular compressor used in many places. blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression ratios. blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among the others above, and at reasonably fast speed. If complib is defined as something other than the listed libraries a ValueError exception is issued.
zlib: The default compression library. A classic in terms of compression, achieves good compression rates but is somewhat slow.
lzo: Fast compression and decompression.
bzip2: Good compression rates.
blosc: Fast compression and decompression.
Support for alternative blosc compressors: blosc:blosclz This is the default compressor for blosc blosc:lz4: A compact, very popular and fast compressor. blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of speed. blosc:snappy: A popular compressor used in many places. blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression ratios. blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among the others above, and at reasonably fast speed.
Support for alternative blosc compressors:
blosc:blosclz This is the default compressor for blosc
blosc
blosc:lz4: A compact, very popular and fast compressor.
blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of speed.
blosc:snappy: A popular compressor used in many places.
blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression ratios.
blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among the others above, and at reasonably fast speed.
If complib is defined as something other than the listed libraries a ValueError exception is issued.
If the library specified with the complib option is missing on your platform, compression defaults to zlib without further ado.
Enable compression for all objects within the file:
store_compressed = pd.HDFStore('store_compressed.h5', complevel=9, complib='blosc:blosclz')
Or on-the-fly compression (this only applies to tables) in stores where compression is not enabled:
store.append('df', df, complib='zlib', complevel=5)
PyTables offers better write performance when tables are compressed after they are written, as opposed to turning on compression at the very beginning. You can use the supplied PyTables utility ptrepack. In addition, ptrepack can change compression levels after the fact.
ptrepack
ptrepack --chunkshape=auto --propindexes --complevel=9 --complib=blosc in.h5 out.h5
Furthermore ptrepack in.h5 out.h5 will repack the file to allow you to reuse previously deleted space. Alternatively, one can simply remove the file and write again, or use the copy method.
ptrepack in.h5 out.h5
copy
HDFStore is not-threadsafe for writing. The underlying PyTables only supports concurrent reads (via threading or processes). If you need reading and writing at the same time, you need to serialize these operations in a single thread in a single process. You will corrupt your data otherwise. See the (GH2397) for more information.
If you use locks to manage write access between multiple processes, you may want to use fsync() before releasing write locks. For convenience you can use store.flush(fsync=True) to do this for you.
fsync()
store.flush(fsync=True)
Once a table is created columns (DataFrame) are fixed; only exactly the same columns can be appended
Be aware that timezones (e.g., pytz.timezone('US/Eastern')) are not necessarily equal across timezone versions. So if data is localized to a specific timezone in the HDFStore using one version of a timezone library and that data is updated with another version, the data will be converted to UTC since these timezones are not considered equal. Either use the same version of timezone library or use tz_convert with the updated timezone definition.
pytz.timezone('US/Eastern')
tz_convert
PyTables will show a NaturalNameWarning if a column name cannot be used as an attribute selector. Natural identifiers contain only letters, numbers, and underscores, and may not begin with a number. Other identifiers cannot be used in a where clause and are generally a bad idea.
NaturalNameWarning
HDFStore will map an object dtype to the PyTables underlying dtype. This means the following types are known to work:
Type
Represents missing values
floating : float64, float32, float16
float64, float32, float16
np.nan
integer : int64, int32, int8, uint64,uint32, uint8
int64, int32, int8, uint64,uint32, uint8
categorical : see the section below
object : strings
strings
unicode columns are not supported, and WILL FAIL.
unicode
You can write data that contains category dtypes to a HDFStore. Queries work the same as if it was an object array. However, the category dtyped data is stored in a more efficient manner.
category
In [473]: dfcat = pd.DataFrame({'A': pd.Series(list('aabbcdba')).astype('category'), .....: 'B': np.random.randn(8)}) .....: In [474]: dfcat Out[474]: A B 0 a 0.477849 1 a 0.283128 2 b -2.045700 3 b -0.338206 4 c -0.423113 5 d 2.314361 6 b -0.033100 7 a -0.965461 In [475]: dfcat.dtypes Out[475]: A category B float64 dtype: object In [476]: cstore = pd.HDFStore('cats.h5', mode='w') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-476-f0dcb969ab57> in <module> ----> 1 cstore = pd.HDFStore('cats.h5', mode='w') ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [477]: cstore.append('dfcat', dfcat, format='table', data_columns=['A']) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-477-1811d2336854> in <module> ----> 1 cstore.append('dfcat', dfcat, format='table', data_columns=['A']) NameError: name 'cstore' is not defined In [478]: result = cstore.select('dfcat', where="A in ['b', 'c']") --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-478-efe8bd5bc969> in <module> ----> 1 result = cstore.select('dfcat', where="A in ['b', 'c']") NameError: name 'cstore' is not defined In [479]: result Out[479]: metrics b_sum c_mean a 0 2.0 0.5 1 2.0 0.5 2 2.0 0.5 In [480]: result.dtypes Out[480]: metrics b_sum float64 c_mean float64 dtype: object
min_itemsize
The underlying implementation of HDFStore uses a fixed column width (itemsize) for string columns. A string column itemsize is calculated as the maximum of the length of data (for that column) that is passed to the HDFStore, in the first append. Subsequent appends, may introduce a string for a column larger than the column can hold, an Exception will be raised (otherwise you could have a silent truncation of these columns, leading to loss of information). In the future we may relax this and allow a user-specified truncation to occur.
Pass min_itemsize on the first table creation to a-priori specify the minimum length of a particular string column. min_itemsize can be an integer, or a dict mapping a column name to an integer. You can pass values as a key to allow all indexables or data_columns to have this min_itemsize.
Passing a min_itemsize dict will cause all passed columns to be created as data_columns automatically.
If you are not passing any data_columns, then the min_itemsize will be the maximum of the length of any string passed
In [481]: dfs = pd.DataFrame({'A': 'foo', 'B': 'bar'}, index=list(range(5))) In [482]: dfs Out[482]: A B 0 foo bar 1 foo bar 2 foo bar 3 foo bar 4 foo bar # A and B have a size of 30 In [483]: store.append('dfs', dfs, min_itemsize=30) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-483-52a3ef06393c> in <module> ----> 1 store.append('dfs', dfs, min_itemsize=30) NameError: name 'store' is not defined In [484]: store.get_storer('dfs').table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-484-934029d84104> in <module> ----> 1 store.get_storer('dfs').table NameError: name 'store' is not defined # A is created as a data_column with a size of 30 # B is size is calculated In [485]: store.append('dfs2', dfs, min_itemsize={'A': 30}) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-485-c510144990e0> in <module> ----> 1 store.append('dfs2', dfs, min_itemsize={'A': 30}) NameError: name 'store' is not defined In [486]: store.get_storer('dfs2').table --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-486-301bf7828112> in <module> ----> 1 store.get_storer('dfs2').table NameError: name 'store' is not defined
nan_rep
String columns will serialize a np.nan (a missing value) with the nan_rep string representation. This defaults to the string value nan. You could inadvertently turn an actual nan value into a missing value.
nan
In [487]: dfss = pd.DataFrame({'A': ['foo', 'bar', 'nan']}) In [488]: dfss Out[488]: A 0 foo 1 bar 2 nan In [489]: store.append('dfss', dfss) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-489-7dd606fe45ee> in <module> ----> 1 store.append('dfss', dfss) NameError: name 'store' is not defined In [490]: store.select('dfss') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-490-1dce3f313c79> in <module> ----> 1 store.select('dfss') NameError: name 'store' is not defined # here you need to specify a different nan rep In [491]: store.append('dfss2', dfss, nan_rep='_nan_') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-491-87550d41b8ff> in <module> ----> 1 store.append('dfss2', dfss, nan_rep='_nan_') NameError: name 'store' is not defined In [492]: store.select('dfss2') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-492-2ccfb6815466> in <module> ----> 1 store.select('dfss2') NameError: name 'store' is not defined
HDFStore writes table format objects in specific formats suitable for producing loss-less round trips to pandas objects. For external compatibility, HDFStore can read native PyTables format tables.
It is possible to write an HDFStore object that can easily be imported into R using the rhdf5 library (Package website). Create a table format store like this:
R
rhdf5
In [493]: df_for_r = pd.DataFrame({"first": np.random.rand(100), .....: "second": np.random.rand(100), .....: "class": np.random.randint(0, 2, (100, ))}, .....: index=range(100)) .....: In [494]: df_for_r.head() Out[494]: first second class 0 0.864919 0.852910 0 1 0.030579 0.412962 1 2 0.015226 0.978410 0 3 0.498512 0.686761 0 4 0.232163 0.328185 1 In [495]: store_export = pd.HDFStore('export.h5') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-495-d8f1d2b3a753> in <module> ----> 1 store_export = pd.HDFStore('export.h5') ~/scipy/pandas/pandas/io/pytables.py in __init__(self, path, mode, complevel, complib, fletcher32, **kwargs) 516 raise ValueError("format is not a defined argument for HDFStore") 517 --> 518 tables = import_optional_dependency("tables") 519 520 if complib is not None and complib not in tables.filters.all_complibs: ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'tables'. Use pip or conda to install tables. In [496]: store_export.append('df_for_r', df_for_r, data_columns=df_dc.columns) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-496-d2f250518f37> in <module> ----> 1 store_export.append('df_for_r', df_for_r, data_columns=df_dc.columns) NameError: name 'store_export' is not defined In [497]: store_export --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-497-9a52a1914703> in <module> ----> 1 store_export NameError: name 'store_export' is not defined
In R this file can be read into a data.frame object using the rhdf5 library. The following example function reads the corresponding column names and data values from the values and assembles them into a data.frame:
data.frame
# Load values and column names for all datasets from corresponding nodes and # insert them into one data.frame object. library(rhdf5) loadhdf5data <- function(h5File) { listing <- h5ls(h5File) # Find all data nodes, values are stored in *_values and corresponding column # titles in *_items data_nodes <- grep("_values", listing$name) name_nodes <- grep("_items", listing$name) data_paths = paste(listing$group[data_nodes], listing$name[data_nodes], sep = "/") name_paths = paste(listing$group[name_nodes], listing$name[name_nodes], sep = "/") columns = list() for (idx in seq(data_paths)) { # NOTE: matrices returned by h5read have to be transposed to obtain # required Fortran order! data <- data.frame(t(h5read(h5File, data_paths[idx]))) names <- t(h5read(h5File, name_paths[idx])) entry <- data.frame(data) colnames(entry) <- names columns <- append(columns, entry) } data <- data.frame(columns) return(data) }
Now you can import the DataFrame into R:
> data = loadhdf5data("transfer.hdf5") > head(data) first second class 1 0.4170220047 0.3266449 0 2 0.7203244934 0.5270581 0 3 0.0001143748 0.8859421 1 4 0.3023325726 0.3572698 1 5 0.1467558908 0.9085352 1 6 0.0923385948 0.6233601 1
The R function lists the entire HDF5 file’s contents and assembles the data.frame object from all matching nodes, so use this only as a starting point if you have stored multiple DataFrame objects to a single HDF5 file.
tables format come with a writing performance penalty as compared to fixed stores. The benefit is the ability to append/delete and query (potentially very large amounts of data). Write times are generally longer as compared with regular stores. Query times can be quite fast, especially on an indexed axis.
tables
You can pass chunksize=<int> to append, specifying the write chunksize (default is 50000). This will significantly lower your memory usage on writing.
chunksize=<int>
You can pass expectedrows=<int> to the first append, to set the TOTAL number of rows that PyTables will expect. This will optimize read/write performance.
expectedrows=<int>
Duplicate rows can be written to tables, but are filtered out in selection (with the last items being selected; thus a table is unique on major, minor pairs)
A PerformanceWarning will be raised if you are attempting to store types that will be pickled by PyTables (rather than stored as endemic types). See Here for more information and some solutions.
PerformanceWarning
Feather provides binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy.
Feather is designed to faithfully serialize and de-serialize DataFrames, supporting all of the pandas dtypes, including extension dtypes such as categorical and datetime with tz.
Several caveats.
This is a newer library, and the format, though stable, is not guaranteed to be backward compatible to the earlier versions.
The format will NOT write an Index, or MultiIndex for the DataFrame and will raise an error if a non-default one is provided. You can .reset_index() to store the index or .reset_index(drop=True) to ignore it.
.reset_index()
.reset_index(drop=True)
Duplicate column names and non-string columns names are not supported
Non supported types include Period and actual Python object types. These will raise a helpful error message on an attempt at serialization.
Period
See the Full Documentation.
In [498]: df = pd.DataFrame({'a': list('abc'), .....: 'b': list(range(1, 4)), .....: 'c': np.arange(3, 6).astype('u1'), .....: 'd': np.arange(4.0, 7.0, dtype='float64'), .....: 'e': [True, False, True], .....: 'f': pd.Categorical(list('abc')), .....: 'g': pd.date_range('20130101', periods=3), .....: 'h': pd.date_range('20130101', periods=3, tz='US/Eastern'), .....: 'i': pd.date_range('20130101', periods=3, freq='ns')}) .....: In [499]: df Out[499]: a b c d e f g h i 0 a 1 3 4.0 True a 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01 00:00:00.000000000 1 b 2 4 5.0 False b 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01 00:00:00.000000001 2 c 3 5 6.0 True c 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01 00:00:00.000000002 In [500]: df.dtypes Out[500]: a object b int64 c uint8 d float64 e bool f category g datetime64[ns] h datetime64[ns, US/Eastern] i datetime64[ns] dtype: object
Write to a feather file.
In [501]: df.to_feather('example.feather') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-501-b832ec9cf6be> in <module> ----> 1 df.to_feather('example.feather') ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_feather(self, path) 1992 from pandas.io.feather_format import to_feather 1993 -> 1994 to_feather(self, path) 1995 1996 @Appender( ~/scipy/pandas/pandas/io/feather_format.py in to_feather(df, path) 18 19 """ ---> 20 import_optional_dependency("pyarrow") 21 from pyarrow import feather 22 ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. Use pip or conda to install pyarrow.
Read from a feather file.
In [502]: result = pd.read_feather('example.feather') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-502-9efa72931ac9> in <module> ----> 1 result = pd.read_feather('example.feather') ~/scipy/pandas/pandas/io/feather_format.py in read_feather(path, columns, use_threads) 93 type of object stored in file 94 """ ---> 95 import_optional_dependency("pyarrow") 96 from pyarrow import feather 97 ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. Use pip or conda to install pyarrow. In [503]: result Out[503]: metrics b_sum c_mean a 0 2.0 0.5 1 2.0 0.5 2 2.0 0.5 # we preserve dtypes In [504]: result.dtypes Out[504]: metrics b_sum float64 c_mean float64 dtype: object
Apache Parquet provides a partitioned binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy. Parquet can use a variety of compression techniques to shrink the file size as much as possible while still maintaining good read performance.
Parquet is designed to faithfully serialize and de-serialize DataFrame s, supporting all of the pandas dtypes, including extension dtypes such as datetime with tz.
Duplicate column names and non-string columns names are not supported.
The pyarrow engine always writes the index to the output, but fastparquet only writes non-default indexes. This extra column can cause problems for non-Pandas consumers that are not expecting it. You can force including or omitting indexes with the index argument, regardless of the underlying engine.
pyarrow
fastparquet
Index level names, if specified, must be strings.
In the pyarrow engine, categorical dtypes for non-string types can be serialized to parquet, but will de-serialize as their primitive dtype.
The pyarrow engine preserves the ordered flag of categorical dtypes with string types. fastparquet does not preserve the ordered flag.
Non supported types include Interval and actual Python object types. These will raise a helpful error message on an attempt at serialization. Period type is supported with pyarrow >= 0.16.0.
Interval
The pyarrow engine preserves extension data types such as the nullable integer and string data type (requiring pyarrow >= 0.16.0, and requiring the extension type to implement the needed protocols, see the extension types documentation).
You can specify an engine to direct the serialization. This can be one of pyarrow, or fastparquet, or auto. If the engine is NOT specified, then the pd.options.io.parquet.engine option is checked; if this is also auto, then pyarrow is tried, and falling back to fastparquet.
auto
pd.options.io.parquet.engine
See the documentation for pyarrow and fastparquet.
These engines are very similar and should read/write nearly identical parquet format files. Currently pyarrow does not support timedelta data, fastparquet>=0.1.4 supports timezone aware datetimes. These libraries differ by having different underlying dependencies (fastparquet by using numba, while pyarrow uses a c-library).
fastparquet>=0.1.4
numba
In [505]: df = pd.DataFrame({'a': list('abc'), .....: 'b': list(range(1, 4)), .....: 'c': np.arange(3, 6).astype('u1'), .....: 'd': np.arange(4.0, 7.0, dtype='float64'), .....: 'e': [True, False, True], .....: 'f': pd.date_range('20130101', periods=3), .....: 'g': pd.date_range('20130101', periods=3, tz='US/Eastern'), .....: 'h': pd.Categorical(list('abc')), .....: 'i': pd.Categorical(list('abc'), ordered=True)}) .....: In [506]: df Out[506]: a b c d e f g h i 0 a 1 3 4.0 True 2013-01-01 2013-01-01 00:00:00-05:00 a a 1 b 2 4 5.0 False 2013-01-02 2013-01-02 00:00:00-05:00 b b 2 c 3 5 6.0 True 2013-01-03 2013-01-03 00:00:00-05:00 c c In [507]: df.dtypes Out[507]: a object b int64 c uint8 d float64 e bool f datetime64[ns] g datetime64[ns, US/Eastern] h category i category dtype: object
Write to a parquet file.
In [508]: df.to_parquet('example_pa.parquet', engine='pyarrow') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-508-96e1a42ae5ed> in <module> ----> 1 df.to_parquet('example_pa.parquet', engine='pyarrow') ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_parquet(self, path, engine, compression, index, partition_cols, **kwargs) 2114 index=index, 2115 partition_cols=partition_cols, -> 2116 **kwargs, 2117 ) 2118 ~/scipy/pandas/pandas/io/parquet.py in to_parquet(df, path, engine, compression, index, partition_cols, **kwargs) 253 if isinstance(partition_cols, str): 254 partition_cols = [partition_cols] --> 255 impl = get_engine(engine) 256 return impl.write( 257 df, ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 37 38 if engine == "pyarrow": ---> 39 return PyArrowImpl() 40 elif engine == "fastparquet": 41 return FastParquetImpl() ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 72 def __init__(self): 73 import_optional_dependency( ---> 74 "pyarrow", extra="pyarrow is required for parquet support." 75 ) 76 import pyarrow.parquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet support. Use pip or conda to install pyarrow. In [509]: df.to_parquet('example_fp.parquet', engine='fastparquet') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-509-8ba55b61d48c> in <module> ----> 1 df.to_parquet('example_fp.parquet', engine='fastparquet') ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_parquet(self, path, engine, compression, index, partition_cols, **kwargs) 2114 index=index, 2115 partition_cols=partition_cols, -> 2116 **kwargs, 2117 ) 2118 ~/scipy/pandas/pandas/io/parquet.py in to_parquet(df, path, engine, compression, index, partition_cols, **kwargs) 253 if isinstance(partition_cols, str): 254 partition_cols = [partition_cols] --> 255 impl = get_engine(engine) 256 return impl.write( 257 df, ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 39 return PyArrowImpl() 40 elif engine == "fastparquet": ---> 41 return FastParquetImpl() 42 43 raise ValueError("engine must be one of 'pyarrow', 'fastparquet'") ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 135 # we need to import on first use 136 fastparquet = import_optional_dependency( --> 137 "fastparquet", extra="fastparquet is required for parquet support." 138 ) 139 self.api = fastparquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'fastparquet'. fastparquet is required for parquet support. Use pip or conda to install fastparquet.
Read from a parquet file.
In [510]: result = pd.read_parquet('example_fp.parquet', engine='fastparquet') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-510-d4bce7d5df53> in <module> ----> 1 result = pd.read_parquet('example_fp.parquet', engine='fastparquet') ~/scipy/pandas/pandas/io/parquet.py in read_parquet(path, engine, columns, **kwargs) 305 """ 306 --> 307 impl = get_engine(engine) 308 return impl.read(path, columns=columns, **kwargs) ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 39 return PyArrowImpl() 40 elif engine == "fastparquet": ---> 41 return FastParquetImpl() 42 43 raise ValueError("engine must be one of 'pyarrow', 'fastparquet'") ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 135 # we need to import on first use 136 fastparquet = import_optional_dependency( --> 137 "fastparquet", extra="fastparquet is required for parquet support." 138 ) 139 self.api = fastparquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'fastparquet'. fastparquet is required for parquet support. Use pip or conda to install fastparquet. In [511]: result = pd.read_parquet('example_pa.parquet', engine='pyarrow') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-511-f9990387c867> in <module> ----> 1 result = pd.read_parquet('example_pa.parquet', engine='pyarrow') ~/scipy/pandas/pandas/io/parquet.py in read_parquet(path, engine, columns, **kwargs) 305 """ 306 --> 307 impl = get_engine(engine) 308 return impl.read(path, columns=columns, **kwargs) ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 37 38 if engine == "pyarrow": ---> 39 return PyArrowImpl() 40 elif engine == "fastparquet": 41 return FastParquetImpl() ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 72 def __init__(self): 73 import_optional_dependency( ---> 74 "pyarrow", extra="pyarrow is required for parquet support." 75 ) 76 import pyarrow.parquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet support. Use pip or conda to install pyarrow. In [512]: result.dtypes Out[512]: metrics b_sum float64 c_mean float64 dtype: object
Read only certain columns of a parquet file.
In [513]: result = pd.read_parquet('example_fp.parquet', .....: engine='fastparquet', columns=['a', 'b']) .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-513-e4f7add80b89> in <module> 1 result = pd.read_parquet('example_fp.parquet', ----> 2 engine='fastparquet', columns=['a', 'b']) ~/scipy/pandas/pandas/io/parquet.py in read_parquet(path, engine, columns, **kwargs) 305 """ 306 --> 307 impl = get_engine(engine) 308 return impl.read(path, columns=columns, **kwargs) ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 39 return PyArrowImpl() 40 elif engine == "fastparquet": ---> 41 return FastParquetImpl() 42 43 raise ValueError("engine must be one of 'pyarrow', 'fastparquet'") ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 135 # we need to import on first use 136 fastparquet = import_optional_dependency( --> 137 "fastparquet", extra="fastparquet is required for parquet support." 138 ) 139 self.api = fastparquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'fastparquet'. fastparquet is required for parquet support. Use pip or conda to install fastparquet. In [514]: result = pd.read_parquet('example_pa.parquet', .....: engine='pyarrow', columns=['a', 'b']) .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-514-be22c54a7770> in <module> 1 result = pd.read_parquet('example_pa.parquet', ----> 2 engine='pyarrow', columns=['a', 'b']) ~/scipy/pandas/pandas/io/parquet.py in read_parquet(path, engine, columns, **kwargs) 305 """ 306 --> 307 impl = get_engine(engine) 308 return impl.read(path, columns=columns, **kwargs) ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 37 38 if engine == "pyarrow": ---> 39 return PyArrowImpl() 40 elif engine == "fastparquet": 41 return FastParquetImpl() ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 72 def __init__(self): 73 import_optional_dependency( ---> 74 "pyarrow", extra="pyarrow is required for parquet support." 75 ) 76 import pyarrow.parquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet support. Use pip or conda to install pyarrow. In [515]: result.dtypes Out[515]: metrics b_sum float64 c_mean float64 dtype: object
Serializing a DataFrame to parquet may include the implicit index as one or more columns in the output file. Thus, this code:
In [516]: df = pd.DataFrame({'a': [1, 2], 'b': [3, 4]}) In [517]: df.to_parquet('test.parquet', engine='pyarrow') --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-517-bebaa85a69fa> in <module> ----> 1 df.to_parquet('test.parquet', engine='pyarrow') ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_parquet(self, path, engine, compression, index, partition_cols, **kwargs) 2114 index=index, 2115 partition_cols=partition_cols, -> 2116 **kwargs, 2117 ) 2118 ~/scipy/pandas/pandas/io/parquet.py in to_parquet(df, path, engine, compression, index, partition_cols, **kwargs) 253 if isinstance(partition_cols, str): 254 partition_cols = [partition_cols] --> 255 impl = get_engine(engine) 256 return impl.write( 257 df, ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 37 38 if engine == "pyarrow": ---> 39 return PyArrowImpl() 40 elif engine == "fastparquet": 41 return FastParquetImpl() ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 72 def __init__(self): 73 import_optional_dependency( ---> 74 "pyarrow", extra="pyarrow is required for parquet support." 75 ) 76 import pyarrow.parquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet support. Use pip or conda to install pyarrow.
creates a parquet file with three columns if you use pyarrow for serialization: a, b, and __index_level_0__. If you’re using fastparquet, the index may or may not be written to the file.
a
b
__index_level_0__
This unexpected extra column causes some databases like Amazon Redshift to reject the file, because that column doesn’t exist in the target table.
If you want to omit a dataframe’s indexes when writing, pass index=False to to_parquet():
to_parquet()
In [518]: df.to_parquet('test.parquet', index=False) --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-518-92139e5cd2c3> in <module> ----> 1 df.to_parquet('test.parquet', index=False) ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_parquet(self, path, engine, compression, index, partition_cols, **kwargs) 2114 index=index, 2115 partition_cols=partition_cols, -> 2116 **kwargs, 2117 ) 2118 ~/scipy/pandas/pandas/io/parquet.py in to_parquet(df, path, engine, compression, index, partition_cols, **kwargs) 253 if isinstance(partition_cols, str): 254 partition_cols = [partition_cols] --> 255 impl = get_engine(engine) 256 return impl.write( 257 df, ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 31 32 raise ImportError( ---> 33 "Unable to find a usable engine; " 34 "tried using: 'pyarrow', 'fastparquet'.\n" 35 "pyarrow or fastparquet is required for parquet support" ImportError: Unable to find a usable engine; tried using: 'pyarrow', 'fastparquet'. pyarrow or fastparquet is required for parquet support
This creates a parquet file with just the two expected columns, a and b. If your DataFrame has a custom index, you won’t get it back when you load this file into a DataFrame.
Passing index=True will always write the index, even if that’s not the underlying engine’s default behavior.
index=True
Parquet supports partitioning of data based on the values of one or more columns.
In [519]: df = pd.DataFrame({'a': [0, 0, 1, 1], 'b': [0, 1, 0, 1]}) In [520]: df.to_parquet(path='test', engine='pyarrow', .....: partition_cols=['a'], compression=None) .....: --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-520-dea796fa072e> in <module> 1 df.to_parquet(path='test', engine='pyarrow', ----> 2 partition_cols=['a'], compression=None) ~/scipy/pandas/pandas/util/_decorators.py in wrapper(*args, **kwargs) 212 else: 213 kwargs[new_arg_name] = new_arg_value --> 214 return func(*args, **kwargs) 215 216 return cast(F, wrapper) ~/scipy/pandas/pandas/core/frame.py in to_parquet(self, path, engine, compression, index, partition_cols, **kwargs) 2114 index=index, 2115 partition_cols=partition_cols, -> 2116 **kwargs, 2117 ) 2118 ~/scipy/pandas/pandas/io/parquet.py in to_parquet(df, path, engine, compression, index, partition_cols, **kwargs) 253 if isinstance(partition_cols, str): 254 partition_cols = [partition_cols] --> 255 impl = get_engine(engine) 256 return impl.write( 257 df, ~/scipy/pandas/pandas/io/parquet.py in get_engine(engine) 37 38 if engine == "pyarrow": ---> 39 return PyArrowImpl() 40 elif engine == "fastparquet": 41 return FastParquetImpl() ~/scipy/pandas/pandas/io/parquet.py in __init__(self) 72 def __init__(self): 73 import_optional_dependency( ---> 74 "pyarrow", extra="pyarrow is required for parquet support." 75 ) 76 import pyarrow.parquet ~/scipy/pandas/pandas/compat/_optional.py in import_optional_dependency(name, extra, raise_on_missing, on_version) 89 except ImportError: 90 if raise_on_missing: ---> 91 raise ImportError(msg) from None 92 else: 93 return None ImportError: Missing optional dependency 'pyarrow'. pyarrow is required for parquet support. Use pip or conda to install pyarrow.
The path specifies the parent directory to which data will be saved. The partition_cols are the column names by which the dataset will be partitioned. Columns are partitioned in the order they are given. The partition splits are determined by the unique values in the partition columns. The above example creates a partitioned dataset that may look like:
test ├── a=0 │ ├── 0bac803e32dc42ae83fddfd029cbdebc.parquet │ └── ... └── a=1 ├── e6ab24a4f45147b49b54a662f0c412a3.parquet └── ...
New in version 1.0.0.
Similar to the parquet format, the ORC Format is a binary columnar serialization for data frames. It is designed to make reading data frames efficient. Pandas provides only a reader for the ORC format, read_orc(). This requires the pyarrow library.
read_orc()
The pandas.io.sql module provides a collection of query wrappers to both facilitate data retrieval and to reduce dependency on DB-specific API. Database abstraction is provided by SQLAlchemy if installed. In addition you will need a driver library for your database. Examples of such drivers are psycopg2 for PostgreSQL or pymysql for MySQL. For SQLite this is included in Python’s standard library by default. You can find an overview of supported drivers for each SQL dialect in the SQLAlchemy docs.
pandas.io.sql
If SQLAlchemy is not installed, a fallback is only provided for sqlite (and for mysql for backwards compatibility, but this is deprecated and will be removed in a future version). This mode requires a Python database adapter which respect the Python DB-API.
See also some cookbook examples for some advanced strategies.
The key functions are:
read_sql_table(table_name, con[, schema, …])
read_sql_table
Read SQL database table into a DataFrame.
read_sql_query(sql, con[, index_col, …])
read_sql_query
Read SQL query into a DataFrame.
read_sql(sql, con[, index_col, …])
Read SQL query or database table into a DataFrame.
DataFrame.to_sql(self, name, con[, schema, …])
DataFrame.to_sql
Write records stored in a DataFrame to a SQL database.
The function read_sql() is a convenience wrapper around read_sql_table() and read_sql_query() (and for backward compatibility) and will delegate to specific function depending on the provided input (database table name or sql query). Table names do not need to be quoted if they have special characters.
read_sql()
read_sql_table()
read_sql_query()
In the following example, we use the SQlite SQL database engine. You can use a temporary SQLite database where data are stored in “memory”.
To connect with SQLAlchemy you use the create_engine() function to create an engine object from database URI. You only need to create the engine once per database you are connecting to. For more information on create_engine() and the URI formatting, see the examples below and the SQLAlchemy documentation
create_engine()
In [521]: from sqlalchemy import create_engine --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-521-172ae2805bcd> in <module> ----> 1 from sqlalchemy import create_engine ModuleNotFoundError: No module named 'sqlalchemy' # Create your engine. In [522]: engine = create_engine('sqlite:///:memory:') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-522-b507e441a3a6> in <module> ----> 1 engine = create_engine('sqlite:///:memory:') NameError: name 'create_engine' is not defined
If you want to manage your own connections you can pass one of those instead:
with engine.connect() as conn, conn.begin(): data = pd.read_sql_table('data', conn)
Assuming the following data is in a DataFrame data, we can insert it into the database using to_sql().
to_sql()
id
Date
Col_1
Col_2
Col_3
26
2012-10-18
X
25.7
42
2012-10-19
Y
-12.4
63
2012-10-20
Z
5.73
In [523]: data Out[523]: id Date Col_1 Col_2 Col_3 0 26 2010-10-18 X 27.50 True 1 42 2010-10-19 Y -12.50 False 2 63 2010-10-20 Z 5.73 True In [524]: data.to_sql('data', engine) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-524-f5cd9ab3b7dc> in <module> ----> 1 data.to_sql('data', engine) NameError: name 'engine' is not defined
With some databases, writing large DataFrames can result in errors due to packet size limitations being exceeded. This can be avoided by setting the chunksize parameter when calling to_sql. For example, the following writes data to the database in batches of 1000 rows at a time:
In [525]: data.to_sql('data_chunked', engine, chunksize=1000) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-525-4aa1ee94beb2> in <module> ----> 1 data.to_sql('data_chunked', engine, chunksize=1000) NameError: name 'engine' is not defined
to_sql() will try to map your data to an appropriate SQL data type based on the dtype of the data. When you have columns of dtype object, pandas will try to infer the data type.
You can always override the default type by specifying the desired SQL type of any of the columns by using the dtype argument. This argument needs a dictionary mapping column names to SQLAlchemy types (or strings for the sqlite3 fallback mode). For example, specifying to use the sqlalchemy String type instead of the default Text type for string columns:
String
Text
In [526]: from sqlalchemy.types import String --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-526-df98ab998107> in <module> ----> 1 from sqlalchemy.types import String ModuleNotFoundError: No module named 'sqlalchemy' In [527]: data.to_sql('data_dtype', engine, dtype={'Col_1': String}) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-527-5b0f75b1bc50> in <module> ----> 1 data.to_sql('data_dtype', engine, dtype={'Col_1': String}) NameError: name 'engine' is not defined
Due to the limited support for timedelta’s in the different database flavors, columns with type timedelta64 will be written as integer values as nanoseconds to the database and a warning will be raised.
timedelta64
Columns of category dtype will be converted to the dense representation as you would get with np.asarray(categorical) (e.g. for string categories this gives an array of strings). Because of this, reading the database table back in does not generate a categorical.
np.asarray(categorical)
Using SQLAlchemy, to_sql() is capable of writing datetime data that is timezone naive or timezone aware. However, the resulting data stored in the database ultimately depends on the supported data type for datetime data of the database system being used.
The following table lists supported data types for datetime data for some common databases. Other database dialects may have different data types for datetime data.
Database
SQL Datetime Types
Timezone Support
SQLite
TEXT
No
MySQL
TIMESTAMP or DATETIME
TIMESTAMP
DATETIME
PostgreSQL
TIMESTAMP or TIMESTAMP WITH TIME ZONE
TIMESTAMP WITH TIME ZONE
Yes
When writing timezone aware data to databases that do not support timezones, the data will be written as timezone naive timestamps that are in local time with respect to the timezone.
read_sql_table() is also capable of reading datetime data that is timezone aware or naive. When reading TIMESTAMP WITH TIME ZONE types, pandas will convert the data to UTC.
The parameter method controls the SQL insertion clause used. Possible values are:
method
None: Uses standard SQL INSERT clause (one per row).
INSERT
'multi': Pass multiple values in a single INSERT clause. It uses a special SQL syntax not supported by all backends. This usually provides better performance for analytic databases like Presto and Redshift, but has worse performance for traditional SQL backend if the table contains many columns. For more information check the SQLAlchemy documention.
'multi'
callable with signature (pd_table, conn, keys, data_iter): This can be used to implement a more performant insertion method based on specific backend dialect features.
(pd_table, conn, keys, data_iter)
Example of a callable using PostgreSQL COPY clause:
# Alternative to_sql() *method* for DBs that support COPY FROM import csv from io import StringIO def psql_insert_copy(table, conn, keys, data_iter): """ Execute SQL statement inserting data Parameters ---------- table : pandas.io.sql.SQLTable conn : sqlalchemy.engine.Engine or sqlalchemy.engine.Connection keys : list of str Column names data_iter : Iterable that iterates the values to be inserted """ # gets a DBAPI connection that can provide a cursor dbapi_conn = conn.connection with dbapi_conn.cursor() as cur: s_buf = StringIO() writer = csv.writer(s_buf) writer.writerows(data_iter) s_buf.seek(0) columns = ', '.join('"{}"'.format(k) for k in keys) if table.schema: table_name = '{}.{}'.format(table.schema, table.name) else: table_name = table.name sql = 'COPY {} ({}) FROM STDIN WITH CSV'.format( table_name, columns) cur.copy_expert(sql=sql, file=s_buf)
read_sql_table() will read a database table given the table name and optionally a subset of columns to read.
In order to use read_sql_table(), you must have the SQLAlchemy optional dependency installed.
In [528]: pd.read_sql_table('data', engine) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-528-5d7b378a0e2a> in <module> ----> 1 pd.read_sql_table('data', engine) NameError: name 'engine' is not defined
Note that pandas infers column dtypes from query outputs, and not by looking up data types in the physical database schema. For example, assume userid is an integer column in a table. Then, intuitively, select userid ... will return integer-valued series, while select cast(userid as text) ... will return object-valued (str) series. Accordingly, if the query output is empty, then all resulting columns will be returned as object-valued (since they are most general). If you foresee that your query will sometimes generate an empty result, you may want to explicitly typecast afterwards to ensure dtype integrity.
userid
select userid ...
select cast(userid as text) ...
You can also specify the name of the column as the DataFrame index, and specify a subset of columns to be read.
In [529]: pd.read_sql_table('data', engine, index_col='id') --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-529-5ce6d94466f1> in <module> ----> 1 pd.read_sql_table('data', engine, index_col='id') NameError: name 'engine' is not defined In [530]: pd.read_sql_table('data', engine, columns=['Col_1', 'Col_2']) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-530-390a3360f767> in <module> ----> 1 pd.read_sql_table('data', engine, columns=['Col_1', 'Col_2']) NameError: name 'engine' is not defined
And you can explicitly force columns to be parsed as dates:
In [531]: pd.read_sql_table('data', engine, parse_dates=['Date']) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-531-c8d1ce2459c5> in <module> ----> 1 pd.read_sql_table('data', engine, parse_dates=['Date']) NameError: name 'engine' is not defined
If needed you can explicitly specify a format string, or a dict of arguments to pass to pandas.to_datetime():
pandas.to_datetime()
pd.read_sql_table('data', engine, parse_dates={'Date': '%Y-%m-%d'}) pd.read_sql_table('data', engine, parse_dates={'Date': {'format': '%Y-%m-%d %H:%M:%S'}})
You can check if a table exists using has_table()
has_table()
Reading from and writing to different schema’s is supported through the schema keyword in the read_sql_table() and to_sql() functions. Note however that this depends on the database flavor (sqlite does not have schema’s). For example:
df.to_sql('table', engine, schema='other_schema') pd.read_sql_table('table', engine, schema='other_schema')
You can query using raw SQL in the read_sql_query() function. In this case you must use the SQL variant appropriate for your database. When using SQLAlchemy, you can also pass SQLAlchemy Expression language constructs, which are database-agnostic.
In [532]: pd.read_sql_query('SELECT * FROM data', engine) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-532-e4e9aacb096d> in <module> ----> 1 pd.read_sql_query('SELECT * FROM data', engine) NameError: name 'engine' is not defined
Of course, you can specify a more “complex” query.
In [533]: pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;", engine) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-533-4d9921b5c402> in <module> ----> 1 pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;", engine) NameError: name 'engine' is not defined
The read_sql_query() function supports a chunksize argument. Specifying this will return an iterator through chunks of the query result:
In [534]: df = pd.DataFrame(np.random.randn(20, 3), columns=list('abc')) In [535]: df.to_sql('data_chunks', engine, index=False) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-535-c1a842c1a9cc> in <module> ----> 1 df.to_sql('data_chunks', engine, index=False) NameError: name 'engine' is not defined
In [536]: for chunk in pd.read_sql_query("SELECT * FROM data_chunks", .....: engine, chunksize=5): .....: print(chunk) .....: --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-536-2dee3444bb39> in <module> 1 for chunk in pd.read_sql_query("SELECT * FROM data_chunks", ----> 2 engine, chunksize=5): 3 print(chunk) NameError: name 'engine' is not defined
You can also run a plain query without creating a DataFrame with execute(). This is useful for queries that don’t return values, such as INSERT. This is functionally equivalent to calling execute on the SQLAlchemy engine or db connection object. Again, you must use the SQL syntax variant appropriate for your database.
execute()
execute
from pandas.io import sql sql.execute('SELECT * FROM table_name', engine) sql.execute('INSERT INTO table_name VALUES(?, ?, ?)', engine, params=[('id', 1, 12.2, True)])
To connect with SQLAlchemy you use the create_engine() function to create an engine object from database URI. You only need to create the engine once per database you are connecting to.
from sqlalchemy import create_engine engine = create_engine('postgresql://scott:tiger@localhost:5432/mydatabase') engine = create_engine('mysql+mysqldb://scott:tiger@localhost/foo') engine = create_engine('oracle://scott:tiger@127.0.0.1:1521/sidname') engine = create_engine('mssql+pyodbc://mydsn') # sqlite://<nohostname>/<path> # where <path> is relative: engine = create_engine('sqlite:///foo.db') # or absolute, starting with a slash: engine = create_engine('sqlite:////absolute/path/to/foo.db')
For more information see the examples the SQLAlchemy documentation
You can use SQLAlchemy constructs to describe your query.
Use sqlalchemy.text() to specify query parameters in a backend-neutral way
sqlalchemy.text()
In [537]: import sqlalchemy as sa --------------------------------------------------------------------------- ModuleNotFoundError Traceback (most recent call last) <ipython-input-537-014e07789c18> in <module> ----> 1 import sqlalchemy as sa ModuleNotFoundError: No module named 'sqlalchemy' In [538]: pd.read_sql(sa.text('SELECT * FROM data where Col_1=:col1'), .....: engine, params={'col1': 'X'}) .....: --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-538-9ac795479a2a> in <module> ----> 1 pd.read_sql(sa.text('SELECT * FROM data where Col_1=:col1'), 2 engine, params={'col1': 'X'}) ~/scipy/pandas/pandas/core/generic.py in __getattr__(self, name) 5148 if self._info_axis._can_hold_identifiers_and_holds_name(name): 5149 return self[name] -> 5150 return object.__getattribute__(self, name) 5151 5152 def __setattr__(self, name: str, value) -> None: AttributeError: 'Series' object has no attribute 'text'
If you have an SQLAlchemy description of your database you can express where conditions using SQLAlchemy expressions
In [539]: metadata = sa.MetaData() --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-539-4f1045301553> in <module> ----> 1 metadata = sa.MetaData() ~/scipy/pandas/pandas/core/generic.py in __getattr__(self, name) 5148 if self._info_axis._can_hold_identifiers_and_holds_name(name): 5149 return self[name] -> 5150 return object.__getattribute__(self, name) 5151 5152 def __setattr__(self, name: str, value) -> None: AttributeError: 'Series' object has no attribute 'MetaData' In [540]: data_table = sa.Table('data', metadata, .....: sa.Column('index', sa.Integer), .....: sa.Column('Date', sa.DateTime), .....: sa.Column('Col_1', sa.String), .....: sa.Column('Col_2', sa.Float), .....: sa.Column('Col_3', sa.Boolean), .....: ) .....: --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-540-ff63ad105bfc> in <module> ----> 1 data_table = sa.Table('data', metadata, 2 sa.Column('index', sa.Integer), 3 sa.Column('Date', sa.DateTime), 4 sa.Column('Col_1', sa.String), 5 sa.Column('Col_2', sa.Float), ~/scipy/pandas/pandas/core/generic.py in __getattr__(self, name) 5148 if self._info_axis._can_hold_identifiers_and_holds_name(name): 5149 return self[name] -> 5150 return object.__getattribute__(self, name) 5151 5152 def __setattr__(self, name: str, value) -> None: AttributeError: 'Series' object has no attribute 'Table' In [541]: pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 is True), engine) --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-541-19a3f87b04e3> in <module> ----> 1 pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 is True), engine) ~/scipy/pandas/pandas/core/generic.py in __getattr__(self, name) 5148 if self._info_axis._can_hold_identifiers_and_holds_name(name): 5149 return self[name] -> 5150 return object.__getattribute__(self, name) 5151 5152 def __setattr__(self, name: str, value) -> None: AttributeError: 'Series' object has no attribute 'select'
You can combine SQLAlchemy expressions with parameters passed to read_sql() using sqlalchemy.bindparam()
sqlalchemy.bindparam()
In [542]: import datetime as dt In [543]: expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam('date')) --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) <ipython-input-543-f8eaa7637fd8> in <module> ----> 1 expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam('date')) ~/scipy/pandas/pandas/core/generic.py in __getattr__(self, name) 5148 if self._info_axis._can_hold_identifiers_and_holds_name(name): 5149 return self[name] -> 5150 return object.__getattribute__(self, name) 5151 5152 def __setattr__(self, name: str, value) -> None: AttributeError: 'Series' object has no attribute 'select' In [544]: pd.read_sql(expr, engine, params={'date': dt.datetime(2010, 10, 18)}) --------------------------------------------------------------------------- NameError Traceback (most recent call last) <ipython-input-544-c6c4def788a1> in <module> ----> 1 pd.read_sql(expr, engine, params={'date': dt.datetime(2010, 10, 18)}) NameError: name 'engine' is not defined
The use of sqlite is supported without using SQLAlchemy. This mode requires a Python database adapter which respect the Python DB-API.
You can create connections like so:
import sqlite3 con = sqlite3.connect(':memory:')
And then issue the following queries:
data.to_sql('data', con) pd.read_sql_query("SELECT * FROM data", con)
Starting in 0.20.0, pandas has split off Google BigQuery support into the separate package pandas-gbq. You can pip install pandas-gbq to get it.
pandas-gbq
pip install pandas-gbq
The pandas-gbq package provides functionality to read/write from Google BigQuery.
pandas integrates with this external package. if pandas-gbq is installed, you can use the pandas methods pd.read_gbq and DataFrame.to_gbq, which will call the respective functions from pandas-gbq.
pd.read_gbq
DataFrame.to_gbq
Full documentation can be found here.
The method to_stata() will write a DataFrame into a .dta file. The format version of this file is always 115 (Stata 12).
to_stata()
In [545]: df = pd.DataFrame(np.random.randn(10, 2), columns=list('AB')) In [546]: df.to_stata('stata.dta')
Stata data files have limited data type support; only strings with 244 or fewer characters, int8, int16, int32, float32 and float64 can be stored in .dta files. Additionally, Stata reserves certain values to represent missing data. Exporting a non-missing value that is outside of the permitted range in Stata for a particular data type will retype the variable to the next larger size. For example, int8 values are restricted to lie between -127 and 100 in Stata, and so variables with values above 100 will trigger a conversion to int16. nan values in floating points data types are stored as the basic missing data type (. in Stata).
int8
int16
int32
float32
.dta
It is not possible to export missing data values for integer data types.
The Stata writer gracefully handles other data types including int64, bool, uint8, uint16, uint32 by casting to the smallest supported type that can represent the data. For example, data with a type of uint8 will be cast to int8 if all values are less than 100 (the upper bound for non-missing int8 data in Stata), or, if values are outside of this range, the variable is cast to int16.
uint8
uint16
uint32
Conversion from int64 to float64 may result in a loss of precision if int64 values are larger than 2**53.
StataWriter and to_stata() only support fixed width strings containing up to 244 characters, a limitation imposed by the version 115 dta file format. Attempting to write Stata dta files with strings longer than 244 characters raises a ValueError.
StataWriter
The top-level function read_stata will read a dta file and return either a DataFrame or a StataReader that can be used to read the file incrementally.
StataReader
In [547]: pd.read_stata('stata.dta') Out[547]: index A B 0 0 0.608228 1.064810 1 1 -0.780506 -2.736887 2 2 0.143539 1.170191 3 3 -1.573076 0.075792 4 4 -1.722223 -0.774650 5 5 0.803627 0.221665 6 6 0.584637 0.147264 7 7 1.057825 -0.284136 8 8 0.912395 1.552808 9 9 0.189376 -0.109830
Specifying a chunksize yields a StataReader instance that can be used to read chunksize lines from the file at a time. The StataReader object can be used as an iterator.
In [548]: reader = pd.read_stata('stata.dta', chunksize=3) In [549]: for df in reader: .....: print(df.shape) .....: (3, 3) (3, 3) (3, 3) (1, 3)
For more fine-grained control, use iterator=True and specify chunksize with each call to read().
In [550]: reader = pd.read_stata('stata.dta', iterator=True) In [551]: chunk1 = reader.read(5) In [552]: chunk2 = reader.read(5)
Currently the index is retrieved as a column.
The parameter convert_categoricals indicates whether value labels should be read and used to create a Categorical variable from them. Value labels can also be retrieved by the function value_labels, which requires read() to be called before use.
convert_categoricals
value_labels
The parameter convert_missing indicates whether missing value representations in Stata should be preserved. If False (the default), missing values are represented as np.nan. If True, missing values are represented using StataMissingValue objects, and columns containing missing values will have object data type.
convert_missing
StataMissingValue
read_stata() and StataReader support .dta formats 113-115 (Stata 10-12), 117 (Stata 13), and 118 (Stata 14).
read_stata()
Setting preserve_dtypes=False will upcast to the standard pandas data types: int64 for all integer types and float64 for floating point data. By default, the Stata data types are preserved when importing.
preserve_dtypes=False
Categorical data can be exported to Stata data files as value labeled data. The exported data consists of the underlying category codes as integer data values and the categories as value labels. Stata does not have an explicit equivalent to a Categorical and information about whether the variable is ordered is lost when exporting.
Stata only supports string value labels, and so str is called on the categories when exporting data. Exporting Categorical variables with non-string categories produces a warning, and can result a loss of information if the str representations of the categories are not unique.
Labeled data can similarly be imported from Stata data files as Categorical variables using the keyword argument convert_categoricals (True by default). The keyword argument order_categoricals (True by default) determines whether imported Categorical variables are ordered.
order_categoricals
When importing categorical data, the values of the variables in the Stata data file are not preserved since Categorical variables always use integer data types between -1 and n-1 where n is the number of categories. If the original values in the Stata data file are required, these can be imported by setting convert_categoricals=False, which will import original data (but not the variable labels). The original values can be matched to the imported categorical data since there is a simple mapping between the original Stata data values and the category codes of imported Categorical variables: missing values are assigned code -1, and the smallest original value is assigned 0, the second smallest is assigned 1 and so on until the largest original value is assigned the code n-1.
-1
n-1
convert_categoricals=False
1
Stata supports partially labeled series. These series have value labels for some but not all data values. Importing a partially labeled series will produce a Categorical with string categories for the values that are labeled and numeric categories for values with no label.
The top-level function read_sas() can read (but not write) SAS xport (.XPT) and (since v0.18.0) SAS7BDAT (.sas7bdat) format files.
read_sas()
SAS files only contain two value types: ASCII text and floating point values (usually 8 bytes but sometimes truncated). For xport files, there is no automatic type conversion to integers, dates, or categoricals. For SAS7BDAT files, the format codes may allow date variables to be automatically converted to dates. By default the whole file is read and returned as a DataFrame.
Specify a chunksize or use iterator=True to obtain reader objects (XportReader or SAS7BDATReader) for incrementally reading the file. The reader objects also have attributes that contain additional information about the file and its variables.
XportReader
SAS7BDATReader
Read a SAS7BDAT file:
df = pd.read_sas('sas_data.sas7bdat')
Obtain an iterator and read an XPORT file 100,000 lines at a time:
def do_something(chunk): pass rdr = pd.read_sas('sas_xport.xpt', chunk=100000) for chunk in rdr: do_something(chunk)
The specification for the xport file format is available from the SAS web site.
No official documentation is available for the SAS7BDAT format.
The top-level function read_spss() can read (but not write) SPSS sav (.sav) and zsav (.zsav) format files.
read_spss()
SPSS files contain column names. By default the whole file is read, categorical columns are converted into pd.Categorical, and a DataFrame with all columns is returned.
pd.Categorical
Specify the usecols parameter to obtain a subset of columns. Specify convert_categoricals=False to avoid converting categorical columns into pd.Categorical.
Read an SPSS file:
df = pd.read_spss('spss_data.sav')
Extract a subset of columns contained in usecols from an SPSS file and avoid converting categorical columns into pd.Categorical:
df = pd.read_spss('spss_data.sav', usecols=['foo', 'bar'], convert_categoricals=False)
More information about the sav and zsav file format is available here.
pandas itself only supports IO with a limited set of file formats that map cleanly to its tabular data model. For reading and writing other file formats into and from pandas, we recommend these packages from the broader community.
xarray provides data structures inspired by the pandas DataFrame for working with multi-dimensional datasets, with a focus on the netCDF file format and easy conversion to and from pandas.
This is an informal comparison of various IO methods, using pandas 0.24.2. Timings are machine dependent and small differences should be ignored.
In [1]: sz = 1000000 In [2]: df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz}) In [3]: df.info() <class 'pandas.core.frame.DataFrame'> RangeIndex: 1000000 entries, 0 to 999999 Data columns (total 2 columns): A 1000000 non-null float64 B 1000000 non-null int64 dtypes: float64(1), int64(1) memory usage: 15.3 MB
Given the next test set:
import numpy as np import os sz = 1000000 df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz}) sz = 1000000 np.random.seed(42) df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz}) def test_sql_write(df): if os.path.exists('test.sql'): os.remove('test.sql') sql_db = sqlite3.connect('test.sql') df.to_sql(name='test_table', con=sql_db) sql_db.close() def test_sql_read(): sql_db = sqlite3.connect('test.sql') pd.read_sql_query("select * from test_table", sql_db) sql_db.close() def test_hdf_fixed_write(df): df.to_hdf('test_fixed.hdf', 'test', mode='w') def test_hdf_fixed_read(): pd.read_hdf('test_fixed.hdf', 'test') def test_hdf_fixed_write_compress(df): df.to_hdf('test_fixed_compress.hdf', 'test', mode='w', complib='blosc') def test_hdf_fixed_read_compress(): pd.read_hdf('test_fixed_compress.hdf', 'test') def test_hdf_table_write(df): df.to_hdf('test_table.hdf', 'test', mode='w', format='table') def test_hdf_table_read(): pd.read_hdf('test_table.hdf', 'test') def test_hdf_table_write_compress(df): df.to_hdf('test_table_compress.hdf', 'test', mode='w', complib='blosc', format='table') def test_hdf_table_read_compress(): pd.read_hdf('test_table_compress.hdf', 'test') def test_csv_write(df): df.to_csv('test.csv', mode='w') def test_csv_read(): pd.read_csv('test.csv', index_col=0) def test_feather_write(df): df.to_feather('test.feather') def test_feather_read(): pd.read_feather('test.feather') def test_pickle_write(df): df.to_pickle('test.pkl') def test_pickle_read(): pd.read_pickle('test.pkl') def test_pickle_write_compress(df): df.to_pickle('test.pkl.compress', compression='xz') def test_pickle_read_compress(): pd.read_pickle('test.pkl.compress', compression='xz') def test_parquet_write(df): df.to_parquet('test.parquet') def test_parquet_read(): pd.read_parquet('test.parquet')
When writing, the top-three functions in terms of speed are test_feather_write, test_hdf_fixed_write and test_hdf_fixed_write_compress.
test_feather_write
test_hdf_fixed_write
test_hdf_fixed_write_compress
In [4]: %timeit test_sql_write(df) 3.29 s ± 43.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [5]: %timeit test_hdf_fixed_write(df) 19.4 ms ± 560 µs per loop (mean ± std. dev. of 7 runs, 1 loop each) In [6]: %timeit test_hdf_fixed_write_compress(df) 19.6 ms ± 308 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) In [7]: %timeit test_hdf_table_write(df) 449 ms ± 5.61 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [8]: %timeit test_hdf_table_write_compress(df) 448 ms ± 11.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [9]: %timeit test_csv_write(df) 3.66 s ± 26.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [10]: %timeit test_feather_write(df) 9.75 ms ± 117 µs per loop (mean ± std. dev. of 7 runs, 100 loops each) In [11]: %timeit test_pickle_write(df) 30.1 ms ± 229 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) In [12]: %timeit test_pickle_write_compress(df) 4.29 s ± 15.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [13]: %timeit test_parquet_write(df) 67.6 ms ± 706 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
When reading, the top three are test_feather_read, test_pickle_read and test_hdf_fixed_read.
test_feather_read
test_pickle_read
test_hdf_fixed_read
In [14]: %timeit test_sql_read() 1.77 s ± 17.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [15]: %timeit test_hdf_fixed_read() 19.4 ms ± 436 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) In [16]: %timeit test_hdf_fixed_read_compress() 19.5 ms ± 222 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) In [17]: %timeit test_hdf_table_read() 38.6 ms ± 857 µs per loop (mean ± std. dev. of 7 runs, 10 loops each) In [18]: %timeit test_hdf_table_read_compress() 38.8 ms ± 1.49 ms per loop (mean ± std. dev. of 7 runs, 10 loops each) In [19]: %timeit test_csv_read() 452 ms ± 9.04 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [20]: %timeit test_feather_read() 12.4 ms ± 99.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each) In [21]: %timeit test_pickle_read() 18.4 ms ± 191 µs per loop (mean ± std. dev. of 7 runs, 100 loops each) In [22]: %timeit test_pickle_read_compress() 915 ms ± 7.48 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) In [23]: %timeit test_parquet_read() 24.4 ms ± 146 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
For this test case test.pkl.compress, test.parquet and test.feather took the least space on disk. Space on disk (in bytes)
test.pkl.compress
test.parquet
test.feather
29519500 Oct 10 06:45 test.csv 16000248 Oct 10 06:45 test.feather 8281983 Oct 10 06:49 test.parquet 16000857 Oct 10 06:47 test.pkl 7552144 Oct 10 06:48 test.pkl.compress 34816000 Oct 10 06:42 test.sql 24009288 Oct 10 06:43 test_fixed.hdf 24009288 Oct 10 06:43 test_fixed_compress.hdf 24458940 Oct 10 06:44 test_table.hdf 24458940 Oct 10 06:44 test_table_compress.hdf