IO Tools (Text, CSV, HDF5, ...)¶
The pandas I/O API is a set of top level reader
functions accessed like pd.read_csv()
that generally return a pandas
object.
- read_csv
- read_excel
- read_hdf
- read_sql
- read_json
- read_msgpack (experimental)
- read_html
- read_gbq (experimental)
- read_stata
- read_sas
- read_clipboard
- read_pickle
The corresponding writer
functions are object methods that are accessed like df.to_csv()
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.
CSV & Text files¶
The two workhorse functions for reading text files (a.k.a. flat files) are
read_csv()
and read_table()
.
They both use the same parsing code to intelligently convert tabular
data into a DataFrame object. See the cookbook
for some advanced strategies
They can take a number of arguments:
filepath_or_buffer
: Either a path to a file (apython:str
,python:pathlib.Path
, orpy:py._path.local.LocalPath
), URL (including http, ftp, and S3 locations), or any object with aread
method (such as an open file orStringIO
).sep
ordelimiter
: A delimiter / separator to split fields on. Withsep=None
,read_csv
will try to infer the delimiter automatically in some cases by “sniffing”. The separator may be specified as a regular expression; for instance you may use ‘|\s*’ to indicate a pipe plus arbitrary whitespace.delim_whitespace
: Parse whitespace-delimited (spaces or tabs) file (much faster than using a regular expression)compression
: decompress'gzip'
and'bz2'
formats on the fly. Set to'infer'
(the default) to guess a format based on the file extension.dialect
: string orpython:csv.Dialect
instance to expose more ways to specify the file formatdtype
: A data type name or a dict of column name to data type. If not specified, data types will be inferred. (Unsupported withengine='python'
)header
: row number(s) to use as the column names, and the start of the data. Defaults to 0 if nonames
passed, otherwiseNone
. Explicitly passheader=0
to be able to replace existing names. The header can be a list of integers that specify row locations for a multi-index on the columns E.g. [0,1,3]. Intervening rows that are not specified will be skipped (e.g. 2 in this example are skipped). Note that this parameter ignores commented lines and empty lines ifskip_blank_lines=True
(the default), so header=0 denotes the first line of data rather than the first line of the file.skip_blank_lines
: whether to skip over blank lines rather than interpreting them as NaN valuesskiprows
: A collection of numbers for rows in the file to skip. Can also be an integer to skip the firstn
rowsindex_col
: column number, column name, or list of column numbers/names, to use as theindex
(row labels) of the resulting DataFrame. By default, it will number the rows without using any column, unless there is one more data column than there are headers, in which case the first column is taken as the index.names
: List of column names to use as column names. To replace header existing in file, explicitly passheader=0
.na_values
: optional string or list of strings to recognize as NaN (missing values), either in addition to or in lieu of the default set.true_values
: list of strings to recognize asTrue
false_values
: list of strings to recognize asFalse
keep_default_na
: whether to include the default set of missing values in addition to the ones specified inna_values
parse_dates
: if True then index will be parsed as dates (False by default). You can specify more complicated options to parse a subset of columns or a combination of columns into a single date column (list of ints or names, list of lists, or dict) [1, 2, 3] -> try parsing columns 1, 2, 3 each as a separate date column [[1, 3]] -> combine columns 1 and 3 and parse as a single date column {‘foo’ : [1, 3]} -> parse columns 1, 3 as date and call result ‘foo’keep_date_col
: if True, then date component columns passed intoparse_dates
will be retained in the output (False by default).date_parser
: function to use to parse strings into datetime objects. Ifparse_dates
is True, it defaults to the very robustdateutil.parser
. Specifying this implicitly setsparse_dates
as True. You can also use functions from community supported date converters from date_converters.pydayfirst
: if True then uses the DD/MM international/European date format (This is False by default)thousands
: specifies the thousands separator. If not None, this character will be stripped from numeric dtypes. However, if it is the first character in a field, that column will be imported as a string. In the PythonParser, if not None, then parser will try to look for it in the output and parse relevant data to numeric dtypes. Because it has to essentially scan through the data again, this causes a significant performance hit so only use if necessary.lineterminator
: string (length 1), defaultNone
, Character to break file into lines. Only valid with C parserquotechar
: string, The character to used to denote the start and end of a quoted item. Quoted items can include the delimiter and it will be ignored.quoting
: int, Controls whether quotes should be recognized. Values are taken from csv.QUOTE_* values. Acceptable values are 0, 1, 2, and 3 for QUOTE_MINIMAL, QUOTE_ALL, QUOTE_NONNUMERIC and QUOTE_NONE, respectively.skipinitialspace
: boolean, defaultFalse
, Skip spaces after delimiterescapechar
: string, to specify how to escape quoted datacomment
: 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, fully commented lines are ignored by the parameter header but not by skiprows. For example, if comment=’#’, parsing ‘#emptyn1,2,3na,b,c’ with header=0 will result in ‘1,2,3’ being treated as the header.nrows
: Number of rows to read out of the file. Useful to only read a small portion of a large fileiterator
: If True, return aTextFileReader
to enable reading a file into memory piece by piecechunksize
: An number of rows to be used to “chunk” a file into pieces. Will cause anTextFileReader
object to be returned. More on this below in the section on iterating and chunkingskip_footer
: number of lines to skip at bottom of file (default 0) (Unsupported withengine='c'
)converters
: a dictionary of functions for converting values in certain columns, where keys are either integers or column labelsencoding
: a string representing the encoding to use for decoding unicode data, e.g.'utf-8`
or'latin-1'
. Full list of Python standard encodingsverbose
: show number of NA values inserted in non-numeric columnssqueeze
: if True then output with only one column is turned into Serieserror_bad_lines
: if False then any lines causing an error will be skipped bad linesusecols
: a subset of columns to return, results in much faster parsing time and lower memory usage.mangle_dupe_cols
: boolean, default True, then duplicate columns will be specified as ‘X.0’...’X.N’, rather than ‘X’...’X’tupleize_cols
: boolean, default False, if False, convert a list of tuples to a multi-index of columns, otherwise, leave the column index as a list of tuplesfloat_precision
: string, default None. Specifies which converter the C engine should use for floating-point values. The options are None for the ordinary converter, ‘high’ for the high-precision converter, and ‘round_trip’ for the round-trip converter.
Consider a typical CSV file containing, in this case, some time series data:
In [1]: print(open('foo.csv').read())
date,A,B,C
20090101,a,1,2
20090102,b,3,4
20090103,c,4,5
The default for read_csv is to create a DataFrame with simple numbered rows:
In [2]: pd.read_csv('foo.csv')
Out[2]:
date A B C
0 20090101 a 1 2
1 20090102 b 3 4
2 20090103 c 4 5
In the case of indexed data, you can pass the column number or column name you wish to use as the index:
In [3]: pd.read_csv('foo.csv', index_col=0)
Out[3]:
A B C
date
20090101 a 1 2
20090102 b 3 4
20090103 c 4 5
In [4]: pd.read_csv('foo.csv', index_col='date')
Out[4]:
A B C
date
20090101 a 1 2
20090102 b 3 4
20090103 c 4 5
You can also use a list of columns to create a hierarchical index:
In [5]: pd.read_csv('foo.csv', index_col=[0, 'A'])
Out[5]:
B C
date A
20090101 a 1 2
20090102 b 3 4
20090103 c 4 5
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 python:csv.Dialect
instance.
Suppose you had data with unenclosed quotes:
In [6]: 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 [7]: dia = csv.excel()
In [8]: dia.quoting = csv.QUOTE_NONE
In [9]: pd.read_csv(StringIO(data), dialect=dia)
Out[9]:
label1 label2 label3
index1 "a c e
index2 b d f
All of the dialect options can be specified separately by keyword arguments:
In [10]: data = 'a,b,c~1,2,3~4,5,6'
In [11]: pd.read_csv(StringIO(data), lineterminator='~')
Out[11]:
a b c
0 1 2 3
1 4 5 6
Another common dialect option is skipinitialspace
, to skip any whitespace
after a delimiter:
In [12]: data = 'a, b, c\n1, 2, 3\n4, 5, 6'
In [13]: print(data)
a, b, c
1, 2, 3
4, 5, 6
In [14]: pd.read_csv(StringIO(data), skipinitialspace=True)
Out[14]:
a b c
0 1 2 3
1 4 5 6
The parsers make every attempt to “do the right thing” and not be very fragile. Type inference is a pretty big deal. So if a column can be coerced to integer dtype without altering the contents, it will do so. Any non-numeric columns will come through as object dtype as with the rest of pandas objects.
Specifying column data types¶
Starting with v0.10, you can indicate the data type for the whole DataFrame or individual columns:
In [15]: data = 'a,b,c\n1,2,3\n4,5,6\n7,8,9'
In [16]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [17]: df = pd.read_csv(StringIO(data), dtype=object)
In [18]: df
Out[18]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
In [19]: df['a'][0]
Out[19]: '1'
In [20]: df = pd.read_csv(StringIO(data), dtype={'b': object, 'c': np.float64})
In [21]: df.dtypes
Out[21]:
a int64
b object
c float64
dtype: object
Note
The dtype
option is currently only supported by the C engine.
Specifying dtype
with engine
other than ‘c’ raises a
ValueError
.
Naming and Using Columns¶
Handling column names¶
A file may or may not have a header row. pandas assumes the first row should be used as the column names:
In [22]: data = 'a,b,c\n1,2,3\n4,5,6\n7,8,9'
In [23]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [24]: pd.read_csv(StringIO(data))
Out[24]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
By specifying the names
argument in conjunction with header
you can
indicate other names to use and whether or not to throw away the header row (if
any):
In [25]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [26]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=0)
Out[26]:
foo bar baz
0 1 2 3
1 4 5 6
2 7 8 9
In [27]: pd.read_csv(StringIO(data), names=['foo', 'bar', 'baz'], header=None)
Out[27]:
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 [28]: data = 'skip this skip it\na,b,c\n1,2,3\n4,5,6\n7,8,9'
In [29]: pd.read_csv(StringIO(data), header=1)
Out[29]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
Filtering columns (usecols
)¶
The usecols
argument allows you to select any subset of the columns in a
file, either using the column names or position numbers:
In [30]: data = 'a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz'
In [31]: pd.read_csv(StringIO(data))
Out[31]:
a b c d
0 1 2 3 foo
1 4 5 6 bar
2 7 8 9 baz
In [32]: pd.read_csv(StringIO(data), usecols=['b', 'd'])
Out[32]:
b d
0 2 foo
1 5 bar
2 8 baz
In [33]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])
Out[33]:
a c d
0 1 3 foo
1 4 6 bar
2 7 9 baz
Comments and Empty Lines¶
Ignoring line comments and empty lines¶
If the comment
parameter is specified, then completely commented lines will
be ignored. By default, completely blank lines will be ignored as well. Both of
these are API changes introduced in version 0.15.
In [34]: data = '\na,b,c\n \n# commented line\n1,2,3\n\n4,5,6'
In [35]: print(data)
a,b,c
1,2,3
4,5,6
# commented line
In [36]: pd.read_csv(StringIO(data), comment='#')
Out[36]:
a b c
0 1 2 3
1 4 5 6
If skip_blank_lines=False
, then read_csv
will not ignore blank lines:
In [37]: data = 'a,b,c\n\n1,2,3\n\n\n4,5,6'
In [38]: pd.read_csv(StringIO(data), skip_blank_lines=False)
Out[38]:
a b c
0 NaN NaN NaN
1 1 2 3
2 NaN NaN NaN
3 NaN NaN NaN
4 4 5 6
Warning
The presence of ignored lines might create ambiguities involving line numbers;
the parameter header
uses row numbers (ignoring commented/empty
lines), while skiprows
uses line numbers (including commented/empty lines):
In [39]: data = '#comment\na,b,c\nA,B,C\n1,2,3'
In [40]: pd.read_csv(StringIO(data), comment='#', header=1)
Out[40]:
A B C
0 1 2 3
In [41]: data = 'A,B,C\n#comment\na,b,c\n1,2,3'
In [42]: pd.read_csv(StringIO(data), comment='#', skiprows=2)
Out[42]:
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 [43]: data = '# empty\n# second empty line\n# third empty' \
In [43]: 'line\nX,Y,Z\n1,2,3\nA,B,C\n1,2.,4.\n5.,NaN,10.0'
In [44]: 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 [45]: pd.read_csv(StringIO(data), comment='#', skiprows=4, header=1)
Out[45]:
A B C
0 1 2 4
1 5 NaN 10
Comments¶
Sometimes comments or meta data may be included in a file:
In [46]: 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 [47]: df = pd.read_csv('tmp.csv')
In [48]: df
Out[48]:
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 [49]: df = pd.read_csv('tmp.csv', comment='#')
In [50]: df
Out[50]:
ID level category
0 Patient1 123000 x
1 Patient2 23000 y
2 Patient3 1234018 z
Dealing with Unicode Data¶
The encoding
argument should be used for encoded unicode data, which will
result in byte strings being decoded to unicode in the result:
In [51]: data = b'word,length\nTr\xc3\xa4umen,7\nGr\xc3\xbc\xc3\x9fe,5'.decode('utf8').encode('latin-1')
In [52]: df = pd.read_csv(BytesIO(data), encoding='latin-1')
In [53]: df
Out[53]:
word length
0 Tr?umen 7
1 Gr??e 5
In [54]: df['word'][1]
Out[54]: u'Gr\xfc\xdfe'
Some formats which encode all characters as multiple bytes, like UTF-16, won’t parse correctly at all without specifying the encoding. Full list of Python standard encodings
Index columns and trailing delimiters¶
If a file has one more column of data than the number of column names, the first column will be used as the DataFrame’s row names:
In [55]: data = 'a,b,c\n4,apple,bat,5.7\n8,orange,cow,10'
In [56]: pd.read_csv(StringIO(data))
Out[56]:
a b c
4 apple bat 5.7
8 orange cow 10.0
In [57]: data = 'index,a,b,c\n4,apple,bat,5.7\n8,orange,cow,10'
In [58]: pd.read_csv(StringIO(data), index_col=0)
Out[58]:
a b c
index
4 apple bat 5.7
8 orange cow 10.0
Ordinarily, you can achieve this behavior using the index_col
option.
There are some exception cases when a file has been prepared with delimiters at
the end of each data line, confusing the parser. To explicitly disable the
index column inference and discard the last column, pass index_col=False
:
In [59]: data = 'a,b,c\n4,apple,bat,\n8,orange,cow,'
In [60]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,
In [61]: pd.read_csv(StringIO(data))
Out[61]:
a b c
4 apple bat NaN
8 orange cow NaN
In [62]: pd.read_csv(StringIO(data), index_col=False)
Out[62]:
a b c
0 4 apple bat
1 8 orange cow
Date Handling¶
Specifying Date Columns¶
To better facilitate working with datetime data,
read_csv()
and read_table()
uses the keyword arguments parse_dates
and date_parser
to allow users
to specify a variety of columns and date/time formats to turn the input text
data into datetime
objects.
The simplest case is to just pass in parse_dates=True
:
# Use a column as an index, and parse it as dates.
In [63]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True)
In [64]: df
Out[64]:
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 [65]: df.index
Out[65]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', name=u'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 [66]: 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 [67]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]])
In [68]: df
Out[68]:
1_2 1_3 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59
By default the parser removes the component date columns, but you can choose
to retain them via the keep_date_col
keyword:
In [69]: df = pd.read_csv('tmp.csv', header=None, parse_dates=[[1, 2], [1, 3]],
....: keep_date_col=True)
....:
In [70]: df
Out[70]:
1_2 1_3 0 1 2 \
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 19990127 19:00:00
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 19990127 20:00:00
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD 19990127 21:00:00
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD 19990127 21:00:00
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD 19990127 22:00:00
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD 19990127 23:00:00
3 4
0 18:56:00 0.81
1 19:56:00 0.01
2 20:56:00 -0.59
3 21:18:00 -0.99
4 21:56:00 -0.59
5 22:56:00 -0.59
Note that if you wish to combine multiple columns into a single date column, a
nested list must be used. In other words, parse_dates=[1, 2]
indicates that
the second and third columns should each be parsed as separate date columns
while parse_dates=[[1, 2]]
means the two columns should be parsed into a
single column.
You can also use a dict to specify custom name columns:
In [71]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}
In [72]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec)
In [73]: df
Out[73]:
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 [74]: date_spec = {'nominal': [1, 2], 'actual': [1, 3]}
In [75]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,
....: index_col=0) #index is the nominal column
....:
In [76]: df
Out[76]:
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
Note
read_csv has a fast_path for parsing datetime strings in iso8601 format, e.g “2000-01-01T00:01:02+00:00” and similar variations. If you can arrange for your data to store datetimes in this format, load times will be significantly faster, ~20x has been observed.
Note
When passing a dict as the parse_dates argument, the order of the columns prepended is not guaranteed, because dict objects do not impose an ordering on their keys. On Python 2.7+ you may use collections.OrderedDict instead of a regular dict if this matters to you. Because of this, when using a dict for ‘parse_dates’ in conjunction with the index_col argument, it’s best to specify index_col as a column label rather then as an index on the resulting frame.
Date Parsing Functions¶
Finally, the parser allows you to specify a custom date_parser
function to
take full advantage of the flexibility of the date parsing API:
In [77]: import pandas.io.date_converters as conv
In [78]: df = pd.read_csv('tmp.csv', header=None, parse_dates=date_spec,
....: date_parser=conv.parse_date_time)
....:
In [79]: df
Out[79]:
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'])
)- 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'])
) - If #2 fails,
date_parser
is called once for every row with one or more string arguments from the columns indicated with parse_dates (e.g.,date_parser('2013', '1')
for the first row,date_parser('2013', '2')
for the second, etc.)
Note that performance-wise, you should try these methods of parsing dates in order:
- Try to infer the format using
infer_datetime_format=True
(see section below) - If you know the format, use
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.
Inferring Datetime Format¶
If you have parse_dates
enabled for some or all of your columns, and your
datetime strings are all formatted the same way, you may get a large speed
up by setting infer_datetime_format=True
. If set, pandas will attempt
to guess the format of your datetime strings, and then use a faster means
of parsing the strings. 5-10x parsing speeds have been observed. pandas
will fallback to the usual parsing if either the format cannot be guessed
or the format that was guessed cannot properly parse the entire column
of strings. So in general, infer_datetime_format
should not have any
negative consequences if enabled.
Here are some examples of datetime strings that can be guessed (All representing December 30th, 2011 at 00:00:00)
- “20111230”
- “2011/12/30”
- “20111230 00:00:00”
- “12/30/2011 00:00:00”
- “30/Dec/2011 00:00:00”
- “30/December/2011 00:00:00”
infer_datetime_format
is sensitive to dayfirst
. With
dayfirst=True
, it will guess “01/12/2011” to be December 1st. With
dayfirst=False
(default) it will guess “01/12/2011” to be January 12th.
# Try to infer the format for the index column
In [80]: df = pd.read_csv('foo.csv', index_col=0, parse_dates=True,
....: infer_datetime_format=True)
....:
In [81]: df
Out[81]:
A B C
date
2009-01-01 a 1 2
2009-01-02 b 3 4
2009-01-03 c 4 5
International Date Formats¶
While US date formats tend to be MM/DD/YYYY, many international formats use
DD/MM/YYYY instead. For convenience, a dayfirst
keyword is provided:
In [82]: print(open('tmp.csv').read())
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c
In [83]: pd.read_csv('tmp.csv', parse_dates=[0])
Out[83]:
date value cat
0 2000-01-06 5 a
1 2000-02-06 10 b
2 2000-03-06 15 c
In [84]: pd.read_csv('tmp.csv', dayfirst=True, parse_dates=[0])
Out[84]:
date value cat
0 2000-06-01 5 a
1 2000-06-02 10 b
2 2000-06-03 15 c
Specifying method for floating-point conversion¶
The parameter float_precision
can be specified in order to use
a specific floating-point converter during parsing with the C engine.
The options are the ordinary converter, the high-precision converter, and
the round-trip converter (which is guaranteed to round-trip values after
writing to a file). For example:
In [85]: val = '0.3066101993807095471566981359501369297504425048828125'
In [86]: data = 'a,b,c\n1,2,{0}'.format(val)
In [87]: abs(pd.read_csv(StringIO(data), engine='c', float_precision=None)['c'][0] - float(val))
Out[87]: 1.1102230246251565e-16
In [88]: abs(pd.read_csv(StringIO(data), engine='c', float_precision='high')['c'][0] - float(val))
Out[88]: 5.5511151231257827e-17
In [89]: abs(pd.read_csv(StringIO(data), engine='c', float_precision='round_trip')['c'][0] - float(val))
Out[89]: 0.0
Thousand Separators¶
For large numbers that have been written with a thousands separator, you can
set the thousands
keyword to a string of length 1 so that integers will be parsed
correctly:
By default, numbers with a thousands separator will be parsed as strings
In [90]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y
Patient3|1,234,018|z
In [91]: df = pd.read_csv('tmp.csv', sep='|')
In [92]: df
Out[92]:
ID level category
0 Patient1 123,000 x
1 Patient2 23,000 y
2 Patient3 1,234,018 z
In [93]: df.level.dtype
Out[93]: dtype('O')
The thousands
keyword allows integers to be parsed correctly
In [94]: print(open('tmp.csv').read())
ID|level|category
Patient1|123,000|x
Patient2|23,000|y
Patient3|1,234,018|z
In [95]: df = pd.read_csv('tmp.csv', sep='|', thousands=',')
In [96]: df
Out[96]:
ID level category
0 Patient1 123000 x
1 Patient2 23000 y
2 Patient3 1234018 z
In [97]: df.level.dtype
Out[97]: dtype('int64')
NA Values¶
To control which values are parsed as missing values (which are signified by NaN
), specifiy a
string in na_values
. If you specify a list of strings, then all values in
it are considered to be missing values. If you specify a number (a float
, like 5.0
or an integer
like 5
),
the corresponding equivalent values will also imply a missing value (in this case effectively
[5.0,5]
are recognized as NaN
.
To completely override the default values that are recognized as missing, specify keep_default_na=False
.
The default NaN
recognized values are ['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A','N/A', 'NA',
'#NA', 'NULL', 'NaN', '-NaN', 'nan', '-nan']
. Although a 0-length string
''
is not included in the default NaN
values list, it is still treated
as a missing value.
read_csv(path, na_values=[5])
the default values, in addition to 5
, 5.0
when interpreted as numbers are recognized as NaN
read_csv(path, keep_default_na=False, na_values=[""])
only an empty field will be NaN
read_csv(path, keep_default_na=False, na_values=["NA", "0"])
only NA
and 0
as strings are NaN
read_csv(path, na_values=["Nope"])
the default values, in addition to the string "Nope"
are recognized as NaN
Infinity¶
inf
like values will be parsed as np.inf
(positive infinity), and -inf
as -np.inf
(negative infinity).
These will ignore the case of the value, meaning Inf
, will also be parsed as np.inf
.
Returning Series¶
Using the squeeze
keyword, the parser will return output with a single column
as a Series
:
In [98]: print(open('tmp.csv').read())
level
Patient1,123000
Patient2,23000
Patient3,1234018
In [99]: output = pd.read_csv('tmp.csv', squeeze=True)
In [100]: output
Out[100]:
Patient1 123000
Patient2 23000
Patient3 1234018
Name: level, dtype: int64
In [101]: type(output)
Out[101]: pandas.core.series.Series
Boolean values¶
The common values True
, False
, TRUE
, and FALSE
are all
recognized as boolean. Sometime you would want to recognize some other values
as being boolean. To do this use the true_values
and false_values
options:
In [102]: data= 'a,b,c\n1,Yes,2\n3,No,4'
In [103]: print(data)
a,b,c
1,Yes,2
3,No,4
In [104]: pd.read_csv(StringIO(data))
Out[104]:
a b c
0 1 Yes 2
1 3 No 4
In [105]: pd.read_csv(StringIO(data), true_values=['Yes'], false_values=['No'])
Out[105]:
a b c
0 1 True 2
1 3 False 4
Handling “bad” lines¶
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 will cause an error by default:
In [27]: data = 'a,b,c\n1,2,3\n4,5,6,7\n8,9,10'
In [28]: pd.read_csv(StringIO(data))
---------------------------------------------------------------------------
CParserError Traceback (most recent call last)
CParserError: 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
Quoting and Escape Characters¶
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:
In [106]: data = 'a,b\n"hello, \\"Bob\\", nice to see you",5'
In [107]: print(data)
a,b
"hello, \"Bob\", nice to see you",5
In [108]: pd.read_csv(StringIO(data), escapechar='\\')
Out[108]:
a b
0 hello, "Bob", nice to see you 5
Files with Fixed Width Columns¶
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:
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 behaviour, if not specified, is to infer.widths
: A list of field widths which can be used instead of ‘colspecs’ if the intervals are contiguous.
Consider a typical fixed-width data file:
In [109]: 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 [110]: colspecs = [(0, 6), (8, 20), (21, 33), (34, 43)]
In [111]: df = pd.read_fwf('bar.csv', colspecs=colspecs, header=None, index_col=0)
In [112]: df
Out[112]:
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 [113]: widths = [6, 14, 13, 10]
In [114]: df = pd.read_fwf('bar.csv', widths=widths, header=None)
In [115]: df
Out[115]:
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.
New in version 0.13.0.
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 [116]: df = pd.read_fwf('bar.csv', header=None, index_col=0)
In [117]: df
Out[117]:
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
Indexes¶
Files with an “implicit” index column¶
Consider a file with one less entry in the header than the number of data column:
In [118]: 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 [119]: pd.read_csv('foo.csv')
Out[119]:
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 [120]: df = pd.read_csv('foo.csv', parse_dates=True)
In [121]: df.index
Out[121]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', freq=None)
Reading an index with a MultiIndex
¶
Suppose you have data indexed by two columns:
In [122]: 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
and read_table
can take a list of
column numbers to turn multiple columns into a MultiIndex
for the index of the
returned object:
In [123]: df = pd.read_csv("data/mindex_ex.csv", index_col=[0,1])
In [124]: df
Out[124]:
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 [125]: df.ix[1978]
Out[125]:
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
Reading columns with a MultiIndex
¶
By specifying list of row locations for the header
argument, you
can read in a MultiIndex
for the columns. Specifying non-consecutive
rows will skip the intervening rows. In order to have the pre-0.13 behavior
of tupleizing columns, specify tupleize_cols=True
.
In [126]: from pandas.util.testing import makeCustomDataframe as mkdf
In [127]: df = mkdf(5,3,r_idx_nlevels=2,c_idx_nlevels=4)
In [128]: df.to_csv('mi.csv')
In [129]: 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 [130]: pd.read_csv('mi.csv',header=[0,1,2,3],index_col=[0,1])
Out[130]:
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
Starting in 0.13.0, read_csv
will be able to interpret a more common format
of multi-columns indices.
In [131]: 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 [132]: pd.read_csv('mi2.csv',header=[0,1],index_col=0)
Out[132]:
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.
Automatically “sniffing” the delimiter¶
read_csv
is capable of inferring delimited (not necessarily
comma-separated) files, as pandas uses the python:csv.Sniffer
class of the csv module. For this, you have to specify sep=None
.
In [133]: print(open('tmp2.sv').read())
:0:1:2:3
0:0.469112299907:-0.282863344329:-1.50905850317:-1.13563237102
1:1.21211202502:-0.173214649053:0.119208711297:-1.04423596628
2:-0.861848963348:-2.10456921889:-0.494929274069:1.07180380704
3:0.721555162244:-0.70677113363:-1.03957498511:0.271859885543
4:-0.424972329789:0.567020349794:0.276232019278:-1.08740069129
5:-0.673689708088:0.113648409689:-1.47842655244:0.524987667115
6:0.40470521868:0.57704598592:-1.71500201611:-1.03926848351
7:-0.370646858236:-1.15789225064:-1.34431181273:0.844885141425
8:1.07576978372:-0.10904997528:1.64356307036:-1.46938795954
9:0.357020564133:-0.67460010373:-1.77690371697:-0.968913812447
In [134]: pd.read_csv('tmp2.sv', sep=None, engine='python')
Out[134]:
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
Iterating through files chunk by chunk¶
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 [135]: print(open('tmp.sv').read())
|0|1|2|3
0|0.469112299907|-0.282863344329|-1.50905850317|-1.13563237102
1|1.21211202502|-0.173214649053|0.119208711297|-1.04423596628
2|-0.861848963348|-2.10456921889|-0.494929274069|1.07180380704
3|0.721555162244|-0.70677113363|-1.03957498511|0.271859885543
4|-0.424972329789|0.567020349794|0.276232019278|-1.08740069129
5|-0.673689708088|0.113648409689|-1.47842655244|0.524987667115
6|0.40470521868|0.57704598592|-1.71500201611|-1.03926848351
7|-0.370646858236|-1.15789225064|-1.34431181273|0.844885141425
8|1.07576978372|-0.10904997528|1.64356307036|-1.46938795954
9|0.357020564133|-0.67460010373|-1.77690371697|-0.968913812447
In [136]: table = pd.read_table('tmp.sv', sep='|')
In [137]: table
Out[137]:
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
or read_table
, the return
value will be an iterable object of type TextFileReader
:
In [138]: reader = pd.read_table('tmp.sv', sep='|', chunksize=4)
In [139]: reader
Out[139]: <pandas.io.parsers.TextFileReader at 0x7f81cb343dd0>
In [140]: 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
0 4 -0.424972 0.567020 0.276232 -1.087401
1 5 -0.673690 0.113648 -1.478427 0.524988
2 6 0.404705 0.577046 -1.715002 -1.039268
3 7 -0.370647 -1.157892 -1.344312 0.844885
Unnamed: 0 0 1 2 3
0 8 1.075770 -0.10905 1.643563 -1.469388
1 9 0.357021 -0.67460 -1.776904 -0.968914
Specifying iterator=True
will also return the TextFileReader
object:
In [141]: reader = pd.read_table('tmp.sv', sep='|', iterator=True)
In [142]: reader.get_chunk(5)
Out[142]:
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
Specifying the parser engine¶
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:
sep
other than a single character (e.g. regex separators)skip_footer
sep=None
withdelim_whitespace=False
Specifying any of the above options will produce a ParserWarning
unless the
python engine is selected explicitly using engine='python'
.
Writing out Data¶
Writing to CSV format¶
The Series and DataFrame objects have an instance method to_csv
which
allows storing the contents of the object as a comma-separated-values file. The
function takes a number of arguments. Only the first is required.
path_or_buf
: A string path to the file to write or a StringIOsep
: Field delimiter for the output file (default ”,”)na_rep
: A string representation of a missing value (default ‘’)float_format
: Format string for floating point numberscols
: Columns to write (default None)header
: Whether to write out the column names (default True)index
: whether to write row (index) names (default True)index_label
: Column label(s) for index column(s) if desired. If None (default), and header and index are True, then the index names are used. (A sequence should be given if the DataFrame uses MultiIndex).mode
: Python write mode, default ‘w’encoding
: a string representing the encoding to use if the contents are non-ASCII, for python versions prior to 3line_terminator
: Character sequence denoting line end (default ‘\n’)quoting
: Set quoting rules as in csv module (default csv.QUOTE_MINIMAL)quotechar
: Character used to quote fields (default ‘”’)doublequote
: Control quoting ofquotechar
in fields (default True)escapechar
: Character used to escapesep
andquotechar
when appropriate (default None)chunksize
: Number of rows to write at a timetupleize_cols
: If False (default), write as a list of tuples, otherwise write in an expanded line format suitable forread_csv
date_format
: Format string for datetime objects
Writing a formatted string¶
The DataFrame object has an instance method to_string
which allows control
over the string representation of the object. All arguments are optional:
buf
default None, for example a StringIO objectcolumns
default None, which columns to writecol_space
default None, minimum width of each column.na_rep
defaultNaN
, representation of NA valueformatters
default None, a dictionary (by column) of functions each of which takes a single argument and returns a formatted stringfloat_format
default None, a function which takes a single (float) argument and returns a formatted string; to be applied to floats in the DataFrame.sparsify
default True, set to False for a DataFrame with a hierarchical index to print every multiindex key at each row.index_names
default True, will print the names of the indicesindex
default True, will print the index (ie, row labels)header
default True, will print the column labelsjustify
defaultleft
, will print column headers left- or right-justified
The Series object also has a to_string
method, but with only the buf
,
na_rep
, float_format
arguments. There is also a length
argument
which, if set to True
, will additionally output the length of the Series.
JSON¶
Read and write JSON
format files and strings.
Writing JSON¶
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 beNone
in which case a JSON string is returnedorient
:- Series :
- default is
index
- allowed values are {
split
,records
,index
}
- default is
- DataFrame
- default is
columns
- allowed values are {
split
,records
,index
,columns
,values
}
- default is
The format of the JSON string
split
dict like {index -> [index], columns -> [columns], data -> [values]} records
list like [{column -> value}, ... , {column -> value}] index
dict like {index -> {column -> value}} columns
dict like {column -> {index -> value}} values
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.force_ascii
: force encoded string to be ASCII, default True.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’.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.
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.
In [143]: dfj = DataFrame(randn(5, 2), columns=list('AB'))
In [144]: json = dfj.to_json()
In [145]: json
Out[145]: '{"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}}'
Orient Options¶
There are a number of different options for the format of the resulting JSON file / string. Consider the following DataFrame and Series:
In [146]: dfjo = DataFrame(dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)),
.....: columns=list('ABC'), index=list('xyz'))
.....:
In [147]: dfjo
Out[147]:
A B C
x 1 4 7
y 2 5 8
z 3 6 9
In [148]: sjo = Series(dict(x=15, y=16, z=17), name='D')
In [149]: sjo
Out[149]:
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 [150]: dfjo.to_json(orient="columns")
Out[150]: '{"A":{"x":1,"y":2,"z":3},"B":{"x":4,"y":5,"z":6},"C":{"x":7,"y":8,"z":9}}'
Index oriented (the default for Series
) similar to column oriented
but the index labels are now primary:
In [151]: dfjo.to_json(orient="index")
Out[151]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,"C":9}}'
In [152]: sjo.to_json(orient="index")
Out[152]: '{"x":15,"y":16,"z":17}'
Record oriented serializes the data to a JSON array of column -> value records, index labels are not included. This is useful for passing DataFrame data to plotting libraries, for example the JavaScript library d3.js:
In [153]: dfjo.to_json(orient="records")
Out[153]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]'
In [154]: sjo.to_json(orient="records")
Out[154]: '[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 [155]: dfjo.to_json(orient="values")
Out[155]: '[[1,4,7],[2,5,8],[3,6,9]]'
Split oriented serializes to a JSON object containing separate entries for
values, index and columns. Name is also included for Series
:
In [156]: dfjo.to_json(orient="split")
Out[156]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,5,8],[3,6,9]]}'
In [157]: sjo.to_json(orient="split")
Out[157]: '{"name":"D","index":["x","y","z"],"data":[15,16,17]}'
Note
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.
Date Handling¶
Writing in ISO date format
In [158]: dfd = DataFrame(randn(5, 2), columns=list('AB'))
In [159]: dfd['date'] = Timestamp('20130101')
In [160]: dfd = dfd.sort_index(1, ascending=False)
In [161]: json = dfd.to_json(date_format='iso')
In [162]: json
Out[162]: '{"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 [163]: json = dfd.to_json(date_format='iso', date_unit='us')
In [164]: json
Out[164]: '{"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 [165]: json = dfd.to_json(date_format='epoch', date_unit='s')
In [166]: json
Out[166]: '{"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 [167]: dfj2 = dfj.copy()
In [168]: dfj2['date'] = Timestamp('20130101')
In [169]: dfj2['ints'] = list(range(5))
In [170]: dfj2['bools'] = True
In [171]: dfj2.index = date_range('20130101', periods=5)
In [172]: dfj2.to_json('test.json')
In [173]: open('test.json').read()
Out[173]: '{"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}}'
Fallback Behavior¶
If the JSON serializer cannot handle the container contents directly it will fallback in the following manner:
- if a
toDict
method is defined by the unrecognised object then that will be called and its returneddict
will be JSON serialized. - if a
default_handler
has been passed toto_json
that will be called to convert the object. - otherwise an attempt is made to convert the object to a
dict
by parsing its contents. However if the object is complex this will often fail with anOverflowError
.
Your best bet when encountering OverflowError
during serialization
is to specify a default_handler
. For example timedelta
can cause
problems:
In [141]: from datetime import timedelta
In [142]: dftd = DataFrame([timedelta(23), timedelta(seconds=5), 42])
In [143]: dftd.to_json()
---------------------------------------------------------------------------
OverflowError Traceback (most recent call last)
OverflowError: Maximum recursion level reached
which can be dealt with by specifying a simple default_handler
:
In [174]: dftd.to_json(default_handler=str)
Out[174]: '{"0":{"0":1987200000,"1":5000,"2":42}}'
In [175]: def my_handler(obj):
.....: return obj.total_seconds()
.....:
Reading JSON¶
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
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.jsontyp
: type of object to recover (series or frame), default ‘frame’orient
:- Series :
- default is
index
- allowed values are {
split
,records
,index
}
- default is
- DataFrame
- default is
columns
- allowed values are {
split
,records
,index
,columns
,values
}
- default is
The format of the JSON string
split
dict like {index -> [index], columns -> [columns], data -> [values]} records
list like [{column -> value}, ... , {column -> value}] index
dict like {index -> {column -> value}} columns
dict like {column -> {index -> value}} values
just the values array 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 dataconvert_axes
: boolean, try to convert the axes to the proper dtypes, default is Trueconvert_dates
: a list of columns to parse for dates; If True, then try to parse date-like columns, default is Truekeep_default_dates
: boolean, default True. If parsing dates, then parse the default date-like columnsnumpy
: 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 ifnumpy=True
precise_float
: boolean, defaultFalse
. 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 functionalitydate_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.
The parser will raise one of ValueError/TypeError/AssertionError
if the JSON is not parseable.
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.
Data Conversion¶
The default of convert_axes=True
, dtype=True
, and convert_dates=True
will try to parse the axes, and all of the data
into appropriate types, including dates. If you need to override specific dtypes, pass a dict to dtype
. convert_axes
should only
be set to False
if you need to preserve string-like numbers (e.g. ‘1’, ‘2’) in an axes.
Note
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'
Warning
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 tointeger
if it can be done safely, e.g. a column of1.
- 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 [176]: pd.read_json(json)
Out[176]:
A B date
0 -1.206412 2.565646 2013-01-01
1 1.431256 1.340309 2013-01-01
2 -1.170299 -0.226169 2013-01-01
3 0.410835 0.813850 2013-01-01
4 0.132003 -0.827317 2013-01-01
Reading from a file:
In [177]: pd.read_json('test.json')
Out[177]:
A B bools date ints
2013-01-01 -1.294524 0.413738 True 2013-01-01 0
2013-01-02 0.276662 -0.472035 True 2013-01-01 1
2013-01-03 -0.013960 -0.362543 True 2013-01-01 2
2013-01-04 -0.006154 -0.923061 True 2013-01-01 3
2013-01-05 0.895717 0.805244 True 2013-01-01 4
Don’t convert any data (but still convert axes and dates):
In [178]: pd.read_json('test.json', dtype=object).dtypes
Out[178]:
A object
B object
bools object
date object
ints object
dtype: object
Specify dtypes for conversion:
In [179]: pd.read_json('test.json', dtype={'A' : 'float32', 'bools' : 'int8'}).dtypes
Out[179]:
A float32
B float64
bools int8
date datetime64[ns]
ints int64
dtype: object
Preserve string indices:
In [180]: si = DataFrame(np.zeros((4, 4)),
.....: columns=list(range(4)),
.....: index=[str(i) for i in range(4)])
.....:
In [181]: si
Out[181]:
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 [182]: si.index
Out[182]: Index([u'0', u'1', u'2', u'3'], dtype='object')
In [183]: si.columns
Out[183]: Int64Index([0, 1, 2, 3], dtype='int64')
In [184]: json = si.to_json()
In [185]: sij = pd.read_json(json, convert_axes=False)
In [186]: sij
Out[186]:
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 [187]: sij.index
Out[187]: Index([u'0', u'1', u'2', u'3'], dtype='object')
In [188]: sij.columns
Out[188]: Index([u'0', u'1', u'2', u'3'], dtype='object')
Dates written in nanoseconds need to be read back in nanoseconds:
In [189]: json = dfj2.to_json(date_unit='ns')
# Try to parse timestamps as millseconds -> Won't Work
In [190]: dfju = pd.read_json(json, date_unit='ms')
In [191]: dfju
Out[191]:
A B bools date ints
1.356998e+18 -1.294524 0.413738 True 1356998400000000000 0
1.357085e+18 0.276662 -0.472035 True 1356998400000000000 1
1.357171e+18 -0.013960 -0.362543 True 1356998400000000000 2
1.357258e+18 -0.006154 -0.923061 True 1356998400000000000 3
1.357344e+18 0.895717 0.805244 True 1356998400000000000 4
# Let pandas detect the correct precision
In [192]: dfju = pd.read_json(json)
In [193]: dfju
Out[193]:
A B bools date ints
2013-01-01 -1.294524 0.413738 True 2013-01-01 0
2013-01-02 0.276662 -0.472035 True 2013-01-01 1
2013-01-03 -0.013960 -0.362543 True 2013-01-01 2
2013-01-04 -0.006154 -0.923061 True 2013-01-01 3
2013-01-05 0.895717 0.805244 True 2013-01-01 4
# Or specify that all timestamps are in nanoseconds
In [194]: dfju = pd.read_json(json, date_unit='ns')
In [195]: dfju
Out[195]:
A B bools date ints
2013-01-01 -1.294524 0.413738 True 2013-01-01 0
2013-01-02 0.276662 -0.472035 True 2013-01-01 1
2013-01-03 -0.013960 -0.362543 True 2013-01-01 2
2013-01-04 -0.006154 -0.923061 True 2013-01-01 3
2013-01-05 0.895717 0.805244 True 2013-01-01 4
The Numpy Parameter¶
Note
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 [196]: randfloats = np.random.uniform(-100, 1000, 10000)
In [197]: randfloats.shape = (1000, 10)
In [198]: dffloats = DataFrame(randfloats, columns=list('ABCDEFGHIJ'))
In [199]: jsonfloats = dffloats.to_json()
In [200]: timeit read_json(jsonfloats)
100 loops, best of 3: 12.9 ms per loop
In [201]: timeit read_json(jsonfloats, numpy=True)
100 loops, best of 3: 12.1 ms per loop
The speedup is less noticeable for smaller datasets:
In [202]: jsonfloats = dffloats.head(100).to_json()
In [203]: timeit read_json(jsonfloats)
100 loops, best of 3: 5.57 ms per loop
In [204]: timeit read_json(jsonfloats, numpy=True)
100 loops, best of 3: 5.93 ms per loop
Warning
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.
Normalization¶
New in version 0.13.0.
pandas provides a utility function to take a dict or list of dicts and normalize this semi-structured data into a flat table.
In [205]: from pandas.io.json import json_normalize
In [206]: data = [{'state': 'Florida',
.....: 'shortname': 'FL',
.....: 'info': {
.....: 'governor': 'Rick Scott'
.....: },
.....: 'counties': [{'name': 'Dade', 'population': 12345},
.....: {'name': 'Broward', 'population': 40000},
.....: {'name': 'Palm Beach', 'population': 60000}]},
.....: {'state': 'Ohio',
.....: 'shortname': 'OH',
.....: 'info': {
.....: 'governor': 'John Kasich'
.....: },
.....: 'counties': [{'name': 'Summit', 'population': 1234},
.....: {'name': 'Cuyahoga', 'population': 1337}]}]
.....:
In [207]: json_normalize(data, 'counties', ['state', 'shortname', ['info', 'governor']])
Out[207]:
name population info.governor state shortname
0 Dade 12345 Rick Scott Florida FL
1 Broward 40000 Rick Scott Florida FL
2 Palm Beach 60000 Rick Scott Florida FL
3 Summit 1234 John Kasich Ohio OH
4 Cuyahoga 1337 John Kasich Ohio OH
HTML¶
Reading HTML Content¶
Warning
We highly encourage you to read the HTML parsing gotchas regarding the issues surrounding the BeautifulSoup4/html5lib/lxml parsers.
New in version 0.12.0.
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.
Note
read_html
returns a list
of DataFrame
objects, even if there is
only a single table contained in the HTML content
Read a URL with no options
In [208]: url = 'http://www.fdic.gov/bank/individual/failed/banklist.html'
In [209]: dfs = read_html(url)
---------------------------------------------------------------------------
URLError Traceback (most recent call last)
<ipython-input-209-15368cbb4893> in <module>()
----> 1 dfs = read_html(url)
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in read_html(io, match, flavor, header, index_col, skiprows, attrs, parse_dates, tupleize_cols, thousands, encoding)
864 _validate_header_arg(header)
865 return _parse(flavor, io, match, header, index_col, skiprows,
--> 866 parse_dates, tupleize_cols, thousands, attrs, encoding)
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in _parse(flavor, io, match, header, index_col, skiprows, parse_dates, tupleize_cols, thousands, attrs, encoding)
726 break
727 else:
--> 728 raise_with_traceback(retained)
729
730 ret = []
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in _parse(flavor, io, match, header, index_col, skiprows, parse_dates, tupleize_cols, thousands, attrs, encoding)
720
721 try:
--> 722 tables = p.parse_tables()
723 except Exception as caught:
724 retained = caught
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in parse_tables(self)
191
192 def parse_tables(self):
--> 193 tables = self._parse_tables(self._build_doc(), self.match, self.attrs)
194 return (self._build_table(table) for table in tables)
195
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in _build_doc(self)
442 def _build_doc(self):
443 from bs4 import BeautifulSoup
--> 444 return BeautifulSoup(self._setup_build_doc(), features='html5lib',
445 from_encoding=self.encoding)
446
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in _setup_build_doc(self)
435
436 def _setup_build_doc(self):
--> 437 raw_text = _read(self.io)
438 if not raw_text:
439 raise ValueError('No text parsed from document: %s' % self.io)
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/html.pyc in _read(obj)
126 """
127 if _is_url(obj):
--> 128 with urlopen(obj) as url:
129 text = url.read()
130 elif hasattr(obj, 'read'):
/usr/lib/python2.7/contextlib.pyc in __enter__(self)
15 def __enter__(self):
16 try:
---> 17 return self.gen.next()
18 except StopIteration:
19 raise RuntimeError("generator didn't yield")
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/common.pyc in urlopen(*args, **kwargs)
49 @contextmanager
50 def urlopen(*args, **kwargs):
---> 51 with closing(_urlopen(*args, **kwargs)) as f:
52 yield f
53
/usr/lib/python2.7/urllib2.pyc in urlopen(url, data, timeout, cafile, capath, cadefault, context)
152 else:
153 opener = _opener
--> 154 return opener.open(url, data, timeout)
155
156 def install_opener(opener):
/usr/lib/python2.7/urllib2.pyc in open(self, fullurl, data, timeout)
429 req = meth(req)
430
--> 431 response = self._open(req, data)
432
433 # post-process response
/usr/lib/python2.7/urllib2.pyc in _open(self, req, data)
447 protocol = req.get_type()
448 result = self._call_chain(self.handle_open, protocol, protocol +
--> 449 '_open', req)
450 if result:
451 return result
/usr/lib/python2.7/urllib2.pyc in _call_chain(self, chain, kind, meth_name, *args)
407 func = getattr(handler, meth_name)
408
--> 409 result = func(*args)
410 if result is not None:
411 return result
/usr/lib/python2.7/urllib2.pyc in http_open(self, req)
1225
1226 def http_open(self, req):
-> 1227 return self.do_open(httplib.HTTPConnection, req)
1228
1229 http_request = AbstractHTTPHandler.do_request_
/usr/lib/python2.7/urllib2.pyc in do_open(self, http_class, req, **http_conn_args)
1195 except socket.error, err: # XXX what error?
1196 h.close()
-> 1197 raise URLError(err)
1198 else:
1199 try:
URLError: <urlopen error [Errno 111] Connection refused>
In [210]: dfs
Out[210]:
( Case Data
0 A 0.174068
1 A -0.439461
2 A -0.741343
3 B -0.079673, Case Data
4 A -0.922875
5 A 0.303638
6 B -0.917368, Case Data
7 A -1.624062
8 A -0.758514)
Note
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 [211]: with open(file_path, 'r') as f:
.....: dfs = read_html(f.read())
.....:
In [212]: dfs
Out[212]:
[ Bank Name City ST CERT \
0 Banks of Wisconsin d/b/a Bank of Kenosha Kenosha WI 35386
1 Central Arizona Bank Scottsdale AZ 34527
2 Sunrise Bank Valdosta GA 58185
3 Pisgah Community Bank Asheville NC 58701
4 Douglas County Bank Douglasville GA 21649
5 Parkway Bank Lenoir NC 57158
6 Chipola Community Bank Marianna FL 58034
.. ... ... .. ...
499 Hamilton Bank, NAEn Espanol Miami FL 24382
500 Sinclair National Bank Gravette AR 34248
501 Superior Bank, FSB Hinsdale IL 32646
502 Malta National Bank Malta OH 6629
503 First Alliance Bank & Trust Co. Manchester NH 34264
504 National State Bank of Metropolis Metropolis IL 3815
505 Bank of Honolulu Honolulu HI 21029
Acquiring Institution Closing Date Updated Date
0 North Shore Bank, FSB May 31, 2013 May 31, 2013
1 Western State Bank May 14, 2013 May 20, 2013
2 Synovus Bank May 10, 2013 May 21, 2013
3 Capital Bank, N.A. May 10, 2013 May 14, 2013
4 Hamilton State Bank April 26, 2013 May 16, 2013
5 CertusBank, National Association April 26, 2013 May 17, 2013
6 First Federal Bank of Florida April 19, 2013 May 16, 2013
.. ... ... ...
499 Israel Discount Bank of New York January 11, 2002 June 5, 2012
500 Delta Trust & Bank September 7, 2001 February 10, 2004
501 Superior Federal, FSB July 27, 2001 June 5, 2012
502 North Valley Bank May 3, 2001 November 18, 2002
503 Southern New Hampshire Bank & Trust February 2, 2001 February 18, 2003
504 Banterra Bank of Marion December 14, 2000 March 17, 2005
505 Bank of the Orient October 13, 2000 March 17, 2005
[506 rows x 7 columns]]
You can even pass in an instance of StringIO
if you so desire
In [213]: with open(file_path, 'r') as f:
.....: sio = StringIO(f.read())
.....:
In [214]: dfs = read_html(sio)
In [215]: dfs
Out[215]:
[ Bank Name City ST CERT \
0 Banks of Wisconsin d/b/a Bank of Kenosha Kenosha WI 35386
1 Central Arizona Bank Scottsdale AZ 34527
2 Sunrise Bank Valdosta GA 58185
3 Pisgah Community Bank Asheville NC 58701
4 Douglas County Bank Douglasville GA 21649
5 Parkway Bank Lenoir NC 57158
6 Chipola Community Bank Marianna FL 58034
.. ... ... .. ...
499 Hamilton Bank, NAEn Espanol Miami FL 24382
500 Sinclair National Bank Gravette AR 34248
501 Superior Bank, FSB Hinsdale IL 32646
502 Malta National Bank Malta OH 6629
503 First Alliance Bank & Trust Co. Manchester NH 34264
504 National State Bank of Metropolis Metropolis IL 3815
505 Bank of Honolulu Honolulu HI 21029
Acquiring Institution Closing Date Updated Date
0 North Shore Bank, FSB May 31, 2013 May 31, 2013
1 Western State Bank May 14, 2013 May 20, 2013
2 Synovus Bank May 10, 2013 May 21, 2013
3 Capital Bank, N.A. May 10, 2013 May 14, 2013
4 Hamilton State Bank April 26, 2013 May 16, 2013
5 CertusBank, National Association April 26, 2013 May 17, 2013
6 First Federal Bank of Florida April 19, 2013 May 16, 2013
.. ... ... ...
499 Israel Discount Bank of New York January 11, 2002 June 5, 2012
500 Delta Trust & Bank September 7, 2001 February 10, 2004
501 Superior Federal, FSB July 27, 2001 June 5, 2012
502 North Valley Bank May 3, 2001 November 18, 2002
503 Southern New Hampshire Bank & Trust February 2, 2001 February 18, 2003
504 Banterra Bank of Marion December 14, 2000 March 17, 2005
505 Bank of the Orient October 13, 2000 March 17, 2005
[506 rows x 7 columns]]
Note
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 = read_html(url, match=match)
Specify a header row (by default <th>
elements are used to form the column
index); if specified, the header row is taken from the data minus the parsed
header elements (<th>
elements).
dfs = read_html(url, header=0)
Specify an index column
dfs = read_html(url, index_col=0)
Specify a number of rows to skip
dfs = read_html(url, skiprows=0)
Specify a number of rows to skip using a list (xrange
(Python 2 only) works
as well)
dfs = read_html(url, skiprows=range(2))
Don’t infer numeric and date types
dfs = read_html(url, infer_types=False)
Specify an HTML attribute
dfs1 = read_html(url, attrs={'id': 'table'})
dfs2 = read_html(url, attrs={'class': 'sortable'})
print(np.array_equal(dfs1[0], dfs2[0])) # Should be True
Use some combination of the above
dfs = read_html(url, match='Metcalf Bank', index_col=0)
Read in pandas to_html
output (with some loss of floating point precision)
df = DataFrame(randn(2, 2))
s = df.to_html(float_format='{0:.40g}'.format)
dfin = 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)
dfs = read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml'])
or
dfs = 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.
dfs = read_html(url, 'Metcalf Bank', index_col=0, flavor=['lxml', 'bs4'])
Writing to HTML files¶
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.
Note
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.
In [216]: df = DataFrame(randn(2, 2))
In [217]: df
Out[217]:
0 1
0 -0.184744 0.496971
1 -0.856240 1.857977
In [218]: 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:
0 | 1 | |
---|---|---|
0 | -0.184744 | 0.496971 |
1 | -0.856240 | 1.857977 |
The columns
argument will limit the columns shown
In [219]: 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>
HTML:
0 | |
---|---|
0 | -0.184744 |
1 | -0.856240 |
float_format
takes a Python callable to control the precision of floating
point values
In [220]: 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>
HTML:
0 | 1 | |
---|---|---|
0 | -0.1847438576 | 0.4969711327 |
1 | -0.8562396763 | 1.8579766508 |
bold_rows
will make the row labels bold by default, but you can turn that
off
In [221]: 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>
0 | 1 | |
---|---|---|
0 | -0.184744 | 0.496971 |
1 | -0.856240 | 1.857977 |
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.
In [222]: 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>
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
In [223]: df = DataFrame({'a': list('&<>'), 'b': randn(3)})
Escaped:
In [224]: 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>
a | b | |
---|---|---|
0 | & | -0.474063 |
1 | < | -0.230305 |
2 | > | -0.400654 |
Not escaped:
In [225]: 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>
a | b | |
---|---|---|
0 | & | -0.474063 |
1 | < | -0.230305 |
2 | > | -0.400654 |
Note
Some browsers may not show a difference in the rendering of the previous two HTML tables.
Excel files¶
The read_excel()
method can read Excel 2003 (.xls
) and
Excel 2007+ (.xlsx
) files using the xlrd
Python
module. 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
Reading Excel Files¶
In the most basic use-case, read_excel
takes a path to an Excel
file, and the sheetname
indicating which sheet to parse.
# Returns a DataFrame
read_excel('path_to_file.xls', sheetname='Sheet1')
ExcelFile
class¶
To facilitate working with multiple sheets from the same file, the ExcelFile
class can be used to wrap the file and can be 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.
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'] = read_excel(xls, 'Sheet1', index_col=None, na_values=['NA'])
data['Sheet2'] = read_excel(xls, 'Sheet2', index_col=None, na_values=['NA'])
# equivalent using the read_excel function
data = read_excel('path_to_file.xls', ['Sheet1', 'Sheet2'], index_col=None, na_values=['NA'])
New in version 0.12.
ExcelFile
has been moved to the top level namespace.
New in version 0.17.
read_excel
can take an ExcelFile
object as input
Specifying Sheets¶
Note
The second argument is sheetname
, not to be confused with ExcelFile.sheet_names
Note
An ExcelFile’s attribute sheet_names
provides access to a list of sheets.
- The arguments
sheetname
allows specifying the sheet or sheets to read. - The default value for
sheetname
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
read_excel('path_to_file.xls', 'Sheet1', index_col=None, na_values=['NA'])
Using the sheet index:
# Returns a DataFrame
read_excel('path_to_file.xls', 0, index_col=None, na_values=['NA'])
Using all default values:
# Returns a DataFrame
read_excel('path_to_file.xls')
Using None to get all sheets:
# Returns a dictionary of DataFrames
read_excel('path_to_file.xls',sheetname=None)
Using a list to get multiple sheets:
# Returns the 1st and 4th sheet, as a dictionary of DataFrames.
read_excel('path_to_file.xls',sheetname=['Sheet1',3])
New in version 0.16.
read_excel
can read more than one sheet, by setting sheetname
to either
a list of sheet names, a list of sheet positions, or None
to read all sheets.
New in version 0.13.
Sheets can be specified by sheet index or sheet name, using an integer or string, respectively.
Reading a MultiIndex
¶
New in version 0.17.
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 [226]: df = pd.DataFrame({'a':[1,2,3,4], 'b':[5,6,7,8]},
.....: index=pd.MultiIndex.from_product([['a','b'],['c','d']]))
.....:
In [227]: df.to_excel('path_to_file.xlsx')
In [228]: df = pd.read_excel('path_to_file.xlsx', index_col=[0,1])
In [229]: df
Out[229]:
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 [230]: df.index = df.index.set_names(['lvl1', 'lvl2'])
In [231]: df.to_excel('path_to_file.xlsx')
In [232]: df = pd.read_excel('path_to_file.xlsx', index_col=[0,1])
In [233]: df
Out[233]:
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 [234]: df.columns = pd.MultiIndex.from_product([['a'],['b', 'd']], names=['c1', 'c2'])
In [235]: df.to_excel('path_to_file.xlsx')
In [236]: df = pd.read_excel('path_to_file.xlsx',
.....: index_col=[0,1], header=[0,1])
.....:
In [237]: df
Out[237]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8
Warning
Excel files saved in version 0.16.2 or prior that had index names will still able to be read in,
but the has_index_names
argument must specified to True
.
Parsing Specific Columns¶
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 parse_cols keyword to allow you to specify a subset of columns to parse.
If parse_cols is an integer, then it is assumed to indicate the last column to be parsed.
read_excel('path_to_file.xls', 'Sheet1', parse_cols=2)
If parse_cols is a list of integers, then it is assumed to be the file column indices to be parsed.
read_excel('path_to_file.xls', 'Sheet1', parse_cols=[0, 2, 3])
Cell Converters¶
It is possible to transform the contents of Excel cells via the converters option. For instance, to convert a column to boolean:
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:
cfun = lambda x: int(x) if x else -1
read_excel('path_to_file.xls', 'Sheet1', converters={'MyInts': cfun})
Writing Excel Files¶
Writing Excel Files to Disk¶
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
.
The DataFrame will be written in a way that tries to mimic the REPL output. One
difference from 0.12.0 is that the index_label
will be placed in the second
row instead of the first. You can get the previous behaviour by setting the
merge_cells
option in to_excel()
to False
:
df.to_excel('path_to_file.xlsx', index_label='label', merge_cells=False)
The Panel class also has a to_excel
instance method,
which writes each DataFrame in the Panel to a separate sheet.
In order to write separate DataFrames to separate sheets in a single Excel file,
one can pass an ExcelWriter
.
with ExcelWriter('path_to_file.xlsx') as writer:
df1.to_excel(writer, sheet_name='Sheet1')
df2.to_excel(writer, sheet_name='Sheet2')
Note
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.
Writing Excel Files to Memory¶
New in version 0.17.
Pandas supports writing Excel files to buffer-like objects such as StringIO
or
BytesIO
using ExcelWriter
.
New in version 0.17.
Added support for Openpyxl >= 2.2
# 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 = 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()
Note
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.
Excel writer engines¶
New in version 0.13.
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
and openpyxl
for .xlsm
files 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.
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
: This includes stable support for Openpyxl from 1.6.1. However, it is advised to use version 2.2 and higher, especially when working with styles.xlsxwriter
xlwt
# By setting the 'engine' in the DataFrame and Panel 'to_excel()' methods.
df.to_excel('path_to_file.xlsx', sheet_name='Sheet1', engine='xlsxwriter')
# By setting the 'engine' in the ExcelWriter constructor.
writer = ExcelWriter('path_to_file.xlsx', engine='xlsxwriter')
# Or via pandas configuration.
from pandas import options
options.io.excel.xlsx.writer = 'xlsxwriter'
df.to_excel('path_to_file.xlsx', sheet_name='Sheet1')
Clipboard¶
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_table
method. For instance, you can copy the following
text to the clipboard (CTRL-C on many operating systems):
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()
In [238]: clipdf
Out[238]:
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.
In [239]: df=pd.DataFrame(randn(5,3))
In [240]: df
Out[240]:
0 1 2
0 -0.288267 -0.084905 0.004772
1 1.382989 0.343635 -1.253994
2 -0.124925 0.212244 0.496654
3 0.525417 1.238640 -1.210543
4 -1.175743 -0.172372 -0.734129
In [241]: df.to_clipboard()
In [242]: pd.read_clipboard()
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-242-475cedd2a1b1> in <module>()
----> 1 pd.read_clipboard()
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/clipboard.pyc in read_clipboard(**kwargs)
49 kwargs['sep'] = '\s+'
50
---> 51 return read_table(StringIO(text), **kwargs)
52
53
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/parsers.pyc in parser_f(filepath_or_buffer, sep, dialect, compression, doublequote, escapechar, quotechar, quoting, skipinitialspace, lineterminator, header, index_col, names, prefix, skiprows, skipfooter, skip_footer, na_values, true_values, false_values, delimiter, converters, dtype, usecols, engine, delim_whitespace, as_recarray, na_filter, compact_ints, use_unsigned, low_memory, buffer_lines, warn_bad_lines, error_bad_lines, keep_default_na, thousands, comment, decimal, parse_dates, keep_date_col, dayfirst, date_parser, memory_map, float_precision, nrows, iterator, chunksize, verbose, encoding, squeeze, mangle_dupe_cols, tupleize_cols, infer_datetime_format, skip_blank_lines)
496 skip_blank_lines=skip_blank_lines)
497
--> 498 return _read(filepath_or_buffer, kwds)
499
500 parser_f.__name__ = name
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/parsers.pyc in _read(filepath_or_buffer, kwds)
273
274 # Create the parser.
--> 275 parser = TextFileReader(filepath_or_buffer, **kwds)
276
277 if (nrows is not None) and (chunksize is not None):
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/parsers.pyc in __init__(self, f, engine, **kwds)
588 self.options['has_index_names'] = kwds['has_index_names']
589
--> 590 self._make_engine(self.engine)
591
592 def _get_options_with_defaults(self, engine):
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/parsers.pyc in _make_engine(self, engine)
729 def _make_engine(self, engine='c'):
730 if engine == 'c':
--> 731 self._engine = CParserWrapper(self.f, **self.options)
732 else:
733 if engine == 'python':
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/io/parsers.pyc in __init__(self, src, **kwds)
1101 kwds['allow_leading_cols'] = self.index_col is not False
1102
-> 1103 self._reader = _parser.TextReader(src, **kwds)
1104
1105 # XXX
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/parser.so in pandas.parser.TextReader.__cinit__ (pandas/parser.c:5030)()
ValueError: No columns to parse from file
We can see that we got the same content back, which we had earlier written to the clipboard.
Note
You may need to install xclip or xsel (with gtk or PyQt4 modules) on Linux to use these methods.
Pickling¶
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.
In [243]: df
Out[243]:
0 1 2
0 -0.288267 -0.084905 0.004772
1 1.382989 0.343635 -1.253994
2 -0.124925 0.212244 0.496654
3 0.525417 1.238640 -1.210543
4 -1.175743 -0.172372 -0.734129
In [244]: 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:
In [245]: read_pickle('foo.pkl')
Out[245]:
0 1 2
0 -0.288267 -0.084905 0.004772
1 1.382989 0.343635 -1.253994
2 -0.124925 0.212244 0.496654
3 0.525417 1.238640 -1.210543
4 -1.175743 -0.172372 -0.734129
Warning
Loading pickled data received from untrusted sources can be unsafe.
Warning
Several internal refactorings, 0.13 (Series Refactoring), and 0.15 (Index Refactoring),
preserve compatibility with pickles created prior to these versions. However, these must
be read with pd.read_pickle
, rather than the default python pickle.load
.
See this question
for a detailed explanation.
Note
These methods were previously pd.save
and pd.load
, prior to 0.12.0, and are now deprecated.
msgpack (experimental)¶
New in version 0.13.0.
Starting in 0.13.0, pandas is supporting the msgpack
format for
object serialization. This is a lightweight portable binary format, similar
to binary JSON, that is highly space efficient, and provides good performance
both on the writing (serialization), and reading (deserialization).
Warning
This is a very new feature of pandas. We intend to provide certain
optimizations in the io of the msgpack
data. Since this is marked
as an EXPERIMENTAL LIBRARY, the storage format may not be stable until a future release.
In [246]: df = DataFrame(np.random.rand(5,2),columns=list('AB'))
In [247]: df.to_msgpack('foo.msg')
In [248]: pd.read_msgpack('foo.msg')
Out[248]:
A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106
In [249]: s = Series(np.random.rand(5),index=date_range('20130101',periods=5))
You can pass a list of objects and you will receive them back on deserialization.
In [250]: pd.to_msgpack('foo.msg', df, 'foo', np.array([1,2,3]), s)
In [251]: pd.read_msgpack('foo.msg')
Out[251]:
[ A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106, 'foo', array([1, 2, 3]), 2013-01-01 0.690810
2013-01-02 0.235907
2013-01-03 0.712756
2013-01-04 0.119599
2013-01-05 0.023493
Freq: D, dtype: float64]
You can pass iterator=True
to iterate over the unpacked results
In [252]: for o in pd.read_msgpack('foo.msg',iterator=True):
.....: print o
.....:
A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106
foo
[1 2 3]
2013-01-01 0.690810
2013-01-02 0.235907
2013-01-03 0.712756
2013-01-04 0.119599
2013-01-05 0.023493
Freq: D, dtype: float64
You can pass append=True
to the writer to append to an existing pack
In [253]: df.to_msgpack('foo.msg',append=True)
In [254]: pd.read_msgpack('foo.msg')
Out[254]:
[ A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106, 'foo', array([1, 2, 3]), 2013-01-01 0.690810
2013-01-02 0.235907
2013-01-03 0.712756
2013-01-04 0.119599
2013-01-05 0.023493
Freq: D, dtype: float64, A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106]
Unlike other io methods, to_msgpack
is available on both a per-object basis,
df.to_msgpack()
and using the top-level pd.to_msgpack(...)
where you
can pack arbitrary collections of python lists, dicts, scalars, while intermixing
pandas objects.
In [255]: pd.to_msgpack('foo2.msg', { 'dict' : [ { 'df' : df }, { 'string' : 'foo' }, { 'scalar' : 1. }, { 's' : s } ] })
In [256]: pd.read_msgpack('foo2.msg')
Out[256]:
{'dict': ({'df': A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106},
{'string': 'foo'},
{'scalar': 1.0},
{'s': 2013-01-01 0.690810
2013-01-02 0.235907
2013-01-03 0.712756
2013-01-04 0.119599
2013-01-05 0.023493
Freq: D, dtype: float64})}
Read/Write API¶
Msgpacks can also be read from and written to strings.
In [257]: df.to_msgpack()
Out[257]: '\x84\xc4\x06blocks\x91\x87\xc4\x05items\x86\xc4\x04name\xc0\xc4\x05dtype\xc4\x06object\xc4\x08compress\xc0\xc4\x04data\x92\xc4\x01A\xc4\x01B\xc4\x05klass\xc4\x05Index\xc4\x03typ\xc4\x05index\xc4\x08compress\xc0\xc4\x04locs\x86\xc4\x04ndim\x01\xc4\x05dtype\xc4\x05int64\xc4\x08compress\xc0\xc4\x04data\xd8\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x00\x00\x00\x00\x00\xc4\x05shape\x91\x02\xc4\x03typ\xc4\x07ndarray\xc4\x05shape\x92\x02\x05\xc4\x06values\xc7P\x00\xa0\xab\xfb6H\xc1\xc3?\x98(oMgz\xd9?\x17\xaed\\\xa5\xc6\xe2?\xdc\xd0\x1bd(\x94\xd2?\xb5\xe8\xf5\x0e\x8d\xa2\xef?\x02D\xebO\x80\xc0\xe6?\x16\xbddQ\xae|\xe8?\x10?Ya[\xc1\xd2?\xa8\xfd\xcf\xa0\xbc\xbe\xe6? Z\xe1\ti\xcc\xaf?\xc4\x05klass\xc4\nFloatBlock\xc4\x05dtype\xc4\x07float64\xc4\x04axes\x92\x86\xc4\x04name\xc0\xc4\x05dtype\xc4\x06object\xc4\x08compress\xc0\xc4\x04data\x92\xc4\x01A\xc4\x01B\xc4\x05klass\xc4\x05Index\xc4\x03typ\xc4\x05index\x86\xc4\x04name\xc0\xc4\x05dtype\xc4\x05int64\xc4\x08compress\xc0\xc4\x04data\xc7(\x00\x00\x00\x00\x00\x00\x00\x00\x00\x01\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x04\x00\x00\x00\x00\x00\x00\x00\xc4\x05klass\xc4\nInt64Index\xc4\x03typ\xc4\x05index\xc4\x03typ\xc4\rblock_manager\xc4\x05klass\xc4\tDataFrame'
Furthermore you can concatenate the strings to produce a list of the original objects.
In [258]: pd.read_msgpack(df.to_msgpack() + s.to_msgpack())
Out[258]:
[ A B
0 0.154336 0.710999
1 0.398096 0.765220
2 0.586749 0.293052
3 0.290293 0.710783
4 0.988593 0.062106, 2013-01-01 0.690810
2013-01-02 0.235907
2013-01-03 0.712756
2013-01-04 0.119599
2013-01-05 0.023493
Freq: D, dtype: float64]
HDF5 (PyTables)¶
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
Warning
As of version 0.15.0, pandas requires PyTables
>= 3.0.0. Stores written with prior versions of pandas / PyTables
>= 2.3 are fully compatible (this was the previous minimum PyTables
required version).
Warning
There is a PyTables
indexing bug 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.
Warning
As of version 0.17.0, HDFStore
will not drop rows that have all missing values by default. Previously, if all values (except the index) were missing, HDFStore
would not write those rows to disk.
In [259]: store = HDFStore('store.h5')
In [260]: print(store)
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
Empty
Objects can be written to the file just like adding key-value pairs to a dict:
In [261]: np.random.seed(1234)
In [262]: index = date_range('1/1/2000', periods=8)
In [263]: s = Series(randn(5), index=['a', 'b', 'c', 'd', 'e'])
In [264]: df = DataFrame(randn(8, 3), index=index,
.....: columns=['A', 'B', 'C'])
.....:
In [265]: wp = Panel(randn(2, 5, 4), items=['Item1', 'Item2'],
.....: major_axis=date_range('1/1/2000', periods=5),
.....: minor_axis=['A', 'B', 'C', 'D'])
.....:
# store.put('s', s) is an equivalent method
In [266]: store['s'] = s
In [267]: store['df'] = df
In [268]: store['wp'] = wp
# the type of stored data
In [269]: store.root.wp._v_attrs.pandas_type
Out[269]: 'wide'
In [270]: store
Out[270]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame (shape->[8,3])
/s series (shape->[5])
/wp wide (shape->[2,5,4])
In a current or later Python session, you can retrieve stored objects:
# store.get('df') is an equivalent method
In [271]: store['df']
Out[271]:
A B C
2000-01-01 0.887163 0.859588 -0.636524
2000-01-02 0.015696 -2.242685 1.150036
2000-01-03 0.991946 0.953324 -2.021255
2000-01-04 -0.334077 0.002118 0.405453
2000-01-05 0.289092 1.321158 -1.546906
2000-01-06 -0.202646 -0.655969 0.193421
2000-01-07 0.553439 1.318152 -0.469305
2000-01-08 0.675554 -1.817027 -0.183109
# dotted (attribute) access provides get as well
In [272]: store.df
Out[272]:
A B C
2000-01-01 0.887163 0.859588 -0.636524
2000-01-02 0.015696 -2.242685 1.150036
2000-01-03 0.991946 0.953324 -2.021255
2000-01-04 -0.334077 0.002118 0.405453
2000-01-05 0.289092 1.321158 -1.546906
2000-01-06 -0.202646 -0.655969 0.193421
2000-01-07 0.553439 1.318152 -0.469305
2000-01-08 0.675554 -1.817027 -0.183109
Deletion of the object specified by the key
# store.remove('wp') is an equivalent method
In [273]: del store['wp']
In [274]: store
Out[274]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame (shape->[8,3])
/s series (shape->[5])
Closing a Store, Context Manager
In [275]: store.close()
In [276]: store
Out[276]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
File is CLOSED
In [277]: store.is_open
Out[277]: False
# Working with, and automatically closing the store with the context
# manager
In [278]: with HDFStore('store.h5') as store:
.....: store.keys()
.....:
Read/Write API¶
HDFStore
supports an top-level API using read_hdf
for reading and to_hdf
for writing,
similar to how read_csv
and to_csv
work. (new in 0.11.0)
In [279]: df_tl = DataFrame(dict(A=list(range(5)), B=list(range(5))))
In [280]: df_tl.to_hdf('store_tl.h5','table',append=True)
In [281]: read_hdf('store_tl.h5', 'table', where = ['index>2'])
Out[281]:
A B
3 3 3
4 4 4
As of version 0.17.0, HDFStore will no longer drop rows that are all missing by default. This behavior can be enabled by setting dropna=True
.
In [282]: df_with_missing = pd.DataFrame({'col1':[0, np.nan, 2],
.....: 'col2':[1, np.nan, np.nan]})
.....:
In [283]: df_with_missing
Out[283]:
col1 col2
0 0 1
1 NaN NaN
2 2 NaN
In [284]: df_with_missing.to_hdf('file.h5', 'df_with_missing',
.....: format = 'table', mode='w')
.....:
In [285]: pd.read_hdf('file.h5', 'df_with_missing')
Out[285]:
col1 col2
0 0 1
1 NaN NaN
2 2 NaN
In [286]: df_with_missing.to_hdf('file.h5', 'df_with_missing',
.....: format = 'table', mode='w', dropna=True)
.....:
In [287]: pd.read_hdf('file.h5', 'df_with_missing')
Out[287]:
col1 col2
0 0 1
2 2 NaN
This is also true for the major axis of a Panel
:
In [288]: matrix = [[[np.nan, np.nan, np.nan],[1,np.nan,np.nan]],
.....: [[np.nan, np.nan, np.nan], [np.nan,5,6]],
.....: [[np.nan, np.nan, np.nan],[np.nan,3,np.nan]]]
.....:
In [289]: panel_with_major_axis_all_missing = Panel(matrix,
.....: items=['Item1', 'Item2','Item3'],
.....: major_axis=[1,2],
.....: minor_axis=['A', 'B', 'C'])
.....:
In [290]: panel_with_major_axis_all_missing
Out[290]:
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 2 (major_axis) x 3 (minor_axis)
Items axis: Item1 to Item3
Major_axis axis: 1 to 2
Minor_axis axis: A to C
In [291]: panel_with_major_axis_all_missing.to_hdf('file.h5', 'panel',
.....: dropna = True,
.....: format='table',
.....: mode='w')
.....:
In [292]: reloaded = read_hdf('file.h5', 'panel')
In [293]: reloaded
Out[293]:
<class 'pandas.core.panel.Panel'>
Dimensions: 3 (items) x 1 (major_axis) x 3 (minor_axis)
Items axis: Item1 to Item3
Major_axis axis: 2 to 2
Minor_axis axis: A to C
Fixed Format¶
Note
This was prior to 0.13.0 the Storer
format.
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 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'
Warning
A fixed
format will raise a TypeError
if you try to retrieve using a where
.
DataFrame(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
Table Format¶
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 & query type operations are
supported. This format is specified by format='table'
or format='t'
to append
or put
or to_hdf
New in version 0.13.
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.
In [294]: store = HDFStore('store.h5')
In [295]: df1 = df[0:4]
In [296]: df2 = df[4:]
# append data (creates a table automatically)
In [297]: store.append('df', df1)
In [298]: store.append('df', df2)
In [299]: store
Out[299]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
# select the entire object
In [300]: store.select('df')
Out[300]:
A B C
2000-01-01 0.887163 0.859588 -0.636524
2000-01-02 0.015696 -2.242685 1.150036
2000-01-03 0.991946 0.953324 -2.021255
2000-01-04 -0.334077 0.002118 0.405453
2000-01-05 0.289092 1.321158 -1.546906
2000-01-06 -0.202646 -0.655969 0.193421
2000-01-07 0.553439 1.318152 -0.469305
2000-01-08 0.675554 -1.817027 -0.183109
# the type of stored data
In [301]: store.root.df._v_attrs.pandas_type
Out[301]: 'frame_table'
Note
You can also create a table
by passing format='table'
or format='t'
to a put
operation.
Hierarchical Keys¶
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 with out 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.
In [302]: store.put('foo/bar/bah', df)
In [303]: store.append('food/orange', df)
In [304]: store.append('food/apple', df)
In [305]: store
Out[305]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/foo/bar/bah frame (shape->[8,3])
/food/apple frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/food/orange frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
# a list of keys are returned
In [306]: store.keys()
Out[306]: ['/df', '/food/apple', '/food/orange', '/foo/bar/bah']
# remove all nodes under this level
In [307]: store.remove('food')
In [308]: store
Out[308]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/foo/bar/bah frame (shape->[8,3])
Storing Types¶
Storing Mixed Types in a Table¶
Storing mixed-dtype data is supported. Strings are stored as a fixed-width using the maximum size of the appended column. Subsequent appends will truncate strings at this length.
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.
In [309]: df_mixed = DataFrame({ 'A' : randn(8),
.....: 'B' : randn(8),
.....: 'C' : np.array(randn(8),dtype='float32'),
.....: 'string' :'string',
.....: 'int' : 1,
.....: 'bool' : True,
.....: 'datetime64' : Timestamp('20010102')},
.....: index=list(range(8)))
.....:
In [310]: df_mixed.ix[3:5,['A', 'B', 'string', 'datetime64']] = np.nan
In [311]: store.append('df_mixed', df_mixed, min_itemsize = {'values': 50})
In [312]: df_mixed1 = store.select('df_mixed')
In [313]: df_mixed1
Out[313]:
A B C bool datetime64 int string
0 0.704721 -1.152659 -0.430096 True 2001-01-02 1 string
1 -0.785435 0.631979 0.767369 True 2001-01-02 1 string
2 0.462060 0.039513 0.984920 True 2001-01-02 1 string
3 NaN NaN 0.270836 True NaT 1 NaN
4 NaN NaN 1.391986 True NaT 1 NaN
5 NaN NaN 0.079842 True NaT 1 NaN
6 2.007843 0.152631 -0.399965 True 2001-01-02 1 string
7 0.226963 0.164530 -1.027851 True 2001-01-02 1 string
In [314]: df_mixed1.get_dtype_counts()
Out[314]:
bool 1
datetime64[ns] 1
float32 1
float64 2
int64 1
object 1
dtype: int64
# we have provided a minimum string column size
In [315]: store.root.df_mixed.table
Out[315]:
/df_mixed/table (Table(8,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(2,), dflt=0.0, pos=1),
"values_block_1": Float32Col(shape=(1,), dflt=0.0, pos=2),
"values_block_2": Int64Col(shape=(1,), dflt=0, pos=3),
"values_block_3": Int64Col(shape=(1,), dflt=0, pos=4),
"values_block_4": BoolCol(shape=(1,), dflt=False, pos=5),
"values_block_5": StringCol(itemsize=50, shape=(1,), dflt='', pos=6)}
byteorder := 'little'
chunkshape := (689,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False}
Storing Multi-Index DataFrames¶
Storing multi-index dataframes as tables is very similar to storing/selecting from homogeneous index DataFrames.
In [316]: index = MultiIndex(levels=[['foo', 'bar', 'baz', 'qux'],
.....: ['one', 'two', 'three']],
.....: labels=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3],
.....: [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
.....: names=['foo', 'bar'])
.....:
In [317]: df_mi = DataFrame(np.random.randn(10, 3), index=index,
.....: columns=['A', 'B', 'C'])
.....:
In [318]: df_mi
Out[318]:
A B C
foo bar
foo one -0.584718 0.816594 -0.081947
two -0.344766 0.528288 -1.068989
three -0.511881 0.291205 0.566534
bar one 0.503592 0.285296 0.484288
two 1.363482 -0.781105 -0.468018
baz two 1.224574 -1.281108 0.875476
three -1.710715 -0.450765 0.749164
qux one -0.203933 -0.182175 0.680656
two -1.818499 0.047072 0.394844
three -0.248432 -0.617707 -0.682884
In [319]: store.append('df_mi',df_mi)
In [320]: store.select('df_mi')
Out[320]:
A B C
foo bar
foo one -0.584718 0.816594 -0.081947
two -0.344766 0.528288 -1.068989
three -0.511881 0.291205 0.566534
bar one 0.503592 0.285296 0.484288
two 1.363482 -0.781105 -0.468018
baz two 1.224574 -1.281108 0.875476
three -1.710715 -0.450765 0.749164
qux one -0.203933 -0.182175 0.680656
two -1.818499 0.047072 0.394844
three -0.248432 -0.617707 -0.682884
# the levels are automatically included as data columns
In [321]: store.select('df_mi', 'foo=bar')
Out[321]:
A B C
foo bar
bar one 0.503592 0.285296 0.484288
two 1.363482 -0.781105 -0.468018
Querying¶
Querying a Table¶
Warning
This query capabilities have changed substantially starting in 0.13.0
.
Queries from prior version are accepted (with a DeprecationWarning
) printed
if its not string-like.
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.
A query is specified using the Term
class under the hood, as a boolean expression.
index
andcolumns
are supported indexers of a DataFramemajor_axis
,minor_axis
, anditems
are supported indexers of the Panel- if
data_columns
are specified, these can be used as additional indexers
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.
Note
=
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:
columns
, 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')
- strings, e.g.
"bar"
- date-like, e.g.
20130101
, or"20130101"
- lists, e.g.
"['A','B']"
- variables that are defined in the local names space, e.g.
date
Note
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.
If you must interpolate, use the '%r'
format specifier
store.select('df', 'index == %r' % string)
which will quote string
.
Here are some examples:
In [322]: dfq = DataFrame(randn(10,4),columns=list('ABCD'),index=date_range('20130101',periods=10))
In [323]: store.append('dfq',dfq,format='table',data_columns=True)
Use boolean expressions, with in-line function evaluation.
In [324]: store.select('dfq',"index>Timestamp('20130104') & columns=['A', 'B']")
Out[324]:
A B
2013-01-05 1.210384 0.797435
2013-01-06 -0.850346 1.176812
2013-01-07 0.984188 -0.121728
2013-01-08 0.796595 -0.474021
2013-01-09 -0.804834 -2.123620
2013-01-10 0.334198 0.536784
Use and inline column reference
In [325]: store.select('dfq',where="A>0 or C>0")
Out[325]:
A B C D
2013-01-01 0.436258 -1.703013 0.393711 -0.479324
2013-01-02 -0.299016 0.694103 0.678630 0.239556
2013-01-03 0.151227 0.816127 1.893534 0.639633
2013-01-04 -0.962029 -2.085266 1.930247 -1.735349
2013-01-05 1.210384 0.797435 -0.379811 0.702562
2013-01-07 0.984188 -0.121728 2.365769 0.496143
2013-01-08 0.796595 -0.474021 -0.056696 1.357797
2013-01-10 0.334198 0.536784 -0.743830 -0.320204
Works with a Panel as well.
In [326]: store.append('wp',wp)
In [327]: store
Out[327]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/df_mi frame_table (typ->appendable_multi,nrows->10,ncols->5,indexers->[index],dc->[bar,foo])
/df_mixed frame_table (typ->appendable,nrows->8,ncols->7,indexers->[index])
/dfq frame_table (typ->appendable,nrows->10,ncols->4,indexers->[index],dc->[A,B,C,D])
/foo/bar/bah frame (shape->[8,3])
/wp wide_table (typ->appendable,nrows->20,ncols->2,indexers->[major_axis,minor_axis])
In [328]: store.select('wp', "major_axis>Timestamp('20000102') & minor_axis=['A', 'B']")
Out[328]:
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 3 (major_axis) x 2 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to B
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'
:
In [329]: store.select('df', "columns=['A', 'B']")
Out[329]:
A B
2000-01-01 0.887163 0.859588
2000-01-02 0.015696 -2.242685
2000-01-03 0.991946 0.953324
2000-01-04 -0.334077 0.002118
2000-01-05 0.289092 1.321158
2000-01-06 -0.202646 -0.655969
2000-01-07 0.553439 1.318152
2000-01-08 0.675554 -1.817027
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.
# this is effectively what the storage of a Panel looks like
In [330]: wp.to_frame()
Out[330]:
Item1 Item2
major minor
2000-01-01 A 1.058969 0.215269
B -0.397840 0.841009
C 0.337438 -1.445810
D 1.047579 -1.401973
2000-01-02 A 1.045938 -0.100918
B 0.863717 -0.548242
C -0.122092 -0.144620
... ... ...
2000-01-04 B 0.036142 0.307969
C -2.074978 -0.208499
D 0.247792 1.033801
2000-01-05 A -0.897157 -2.400454
B -0.136795 2.030604
C 0.018289 -1.142631
D 0.755414 0.211883
[20 rows x 2 columns]
# limiting the search
In [331]: store.select('wp',"major_axis>20000102 & minor_axis=['A','B']",
.....: start=0, stop=10)
.....:
Out[331]:
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 1 (major_axis) x 2 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 2000-01-03 00:00:00 to 2000-01-03 00:00:00
Minor_axis axis: A to B
Note
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.
Using timedelta64[ns]¶
New in version 0.13.
Beginning in 0.13.0, 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:
In [332]: from datetime import timedelta
In [333]: dftd = DataFrame(dict(A = Timestamp('20130101'), B = [ Timestamp('20130101') + timedelta(days=i,seconds=10) for i in range(10) ]))
In [334]: dftd['C'] = dftd['A']-dftd['B']
In [335]: dftd
Out[335]:
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 [336]: store.append('dftd',dftd,data_columns=True)
In [337]: store.select('dftd',"C<'-3.5D'")
Out[337]:
A B C
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
Indexing¶
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
.
Note
Indexes are automagically created (starting 0.10.1
) on the indexables
and any data columns you specify. This behavior can be turned off by passing
index=False
to append
.
# we have automagically already created an index (in the first section)
In [338]: i = store.root.df.table.cols.index.index
In [339]: i.optlevel, i.kind
Out[339]: (6, 'medium')
# change an index by passing new parameters
In [340]: store.create_table_index('df', optlevel=9, kind='full')
In [341]: i = store.root.df.table.cols.index.index
In [342]: i.optlevel, i.kind
Out[342]: (9, 'full')
Ofentimes 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 [343]: df_1 = DataFrame(randn(10,2),columns=list('AB'))
In [344]: df_2 = DataFrame(randn(10,2),columns=list('AB'))
In [345]: st = pd.HDFStore('appends.h5',mode='w')
In [346]: st.append('df', df_1, data_columns=['B'], index=False)
In [347]: st.append('df', df_2, data_columns=['B'], index=False)
In [348]: st.get_storer('df').table
Out[348]:
/df/table (Table(20,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2)}
byteorder := 'little'
chunkshape := (2730,)
Then create the index when finished appending.
In [349]: st.create_table_index('df', columns=['B'], optlevel=9, kind='full')
In [350]: st.get_storer('df').table
Out[350]:
/df/table (Table(20,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2)}
byteorder := 'little'
chunkshape := (2730,)
autoindex := True
colindexes := {
"B": Index(9, full, shuffle, zlib(1)).is_csi=True}
In [351]: st.close()
See here for how to create a completely-sorted-index (CSI) on an existing store.
Query via Data Columns¶
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
In [352]: df_dc = df.copy()
In [353]: df_dc['string'] = 'foo'
In [354]: df_dc.ix[4:6,'string'] = np.nan
In [355]: df_dc.ix[7:9,'string'] = 'bar'
In [356]: df_dc['string2'] = 'cool'
In [357]: df_dc.ix[1:3,['B','C']] = 1.0
In [358]: df_dc
Out[358]:
A B C string string2
2000-01-01 0.887163 0.859588 -0.636524 foo cool
2000-01-02 0.015696 1.000000 1.000000 foo cool
2000-01-03 0.991946 1.000000 1.000000 foo cool
2000-01-04 -0.334077 0.002118 0.405453 foo cool
2000-01-05 0.289092 1.321158 -1.546906 NaN cool
2000-01-06 -0.202646 -0.655969 0.193421 NaN cool
2000-01-07 0.553439 1.318152 -0.469305 foo cool
2000-01-08 0.675554 -1.817027 -0.183109 bar cool
# on-disk operations
In [359]: store.append('df_dc', df_dc, data_columns = ['B', 'C', 'string', 'string2'])
In [360]: store.select('df_dc', [ Term('B>0') ])
Out[360]:
A B C string string2
2000-01-01 0.887163 0.859588 -0.636524 foo cool
2000-01-02 0.015696 1.000000 1.000000 foo cool
2000-01-03 0.991946 1.000000 1.000000 foo cool
2000-01-04 -0.334077 0.002118 0.405453 foo cool
2000-01-05 0.289092 1.321158 -1.546906 NaN cool
2000-01-07 0.553439 1.318152 -0.469305 foo cool
# getting creative
In [361]: store.select('df_dc', 'B > 0 & C > 0 & string == foo')
Out[361]:
A B C string string2
2000-01-02 0.015696 1.000000 1.000000 foo cool
2000-01-03 0.991946 1.000000 1.000000 foo cool
2000-01-04 -0.334077 0.002118 0.405453 foo cool
# this is in-memory version of this type of selection
In [362]: df_dc[(df_dc.B > 0) & (df_dc.C > 0) & (df_dc.string == 'foo')]
Out[362]:
A B C string string2
2000-01-02 0.015696 1.000000 1.000000 foo cool
2000-01-03 0.991946 1.000000 1.000000 foo cool
2000-01-04 -0.334077 0.002118 0.405453 foo cool
# we have automagically created this index and the B/C/string/string2
# columns are stored separately as ``PyTables`` columns
In [363]: store.root.df_dc.table
Out[363]:
/df_dc/table (Table(8,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": Float64Col(shape=(1,), dflt=0.0, pos=1),
"B": Float64Col(shape=(), dflt=0.0, pos=2),
"C": Float64Col(shape=(), dflt=0.0, pos=3),
"string": StringCol(itemsize=3, shape=(), dflt='', pos=4),
"string2": StringCol(itemsize=4, shape=(), dflt='', pos=5)}
byteorder := 'little'
chunkshape := (1680,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"C": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"B": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"string2": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"string": Index(6, medium, shuffle, zlib(1)).is_csi=False}
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!)
Iterator¶
Starting in 0.11.0
, 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.
In [364]: for df in store.select('df', chunksize=3):
.....: print(df)
.....:
A B C
2000-01-01 0.887163 0.859588 -0.636524
2000-01-02 0.015696 -2.242685 1.150036
2000-01-03 0.991946 0.953324 -2.021255
A B C
2000-01-04 -0.334077 0.002118 0.405453
2000-01-05 0.289092 1.321158 -1.546906
2000-01-06 -0.202646 -0.655969 0.193421
A B C
2000-01-07 0.553439 1.318152 -0.469305
2000-01-08 0.675554 -1.817027 -0.183109
Note
New in version 0.12.0.
You can also use the iterator with read_hdf
which will open, then
automatically close the store when finished iterating.
for df in 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 [365]: dfeq = DataFrame({'number': np.arange(1,11)})
In [366]: dfeq
Out[366]:
number
0 1
1 2
2 3
3 4
4 5
5 6
6 7
7 8
8 9
9 10
In [367]: store.append('dfeq', dfeq, data_columns=['number'])
In [368]: def chunks(l, n):
.....: return [l[i:i+n] for i in range(0, len(l), n)]
.....:
In [369]: evens = [2,4,6,8,10]
In [370]: coordinates = store.select_as_coordinates('dfeq','number=evens')
In [371]: for c in chunks(coordinates, 2):
.....: print store.select('dfeq',where=c)
.....:
number
1 2
3 4
number
5 6
7 8
number
9 10
Advanced Queries¶
Select a Single Column¶
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.
In [372]: store.select_column('df_dc', 'index')
Out[372]:
0 2000-01-01
1 2000-01-02
2 2000-01-03
3 2000-01-04
4 2000-01-05
5 2000-01-06
6 2000-01-07
7 2000-01-08
Name: index, dtype: datetime64[ns]
In [373]: store.select_column('df_dc', 'string')
Out[373]:
0 foo
1 foo
2 foo
3 foo
4 NaN
5 NaN
6 foo
7 bar
Name: string, dtype: object
Selecting coordinates¶
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.
In [374]: df_coord = DataFrame(np.random.randn(1000,2),index=date_range('20000101',periods=1000))
In [375]: store.append('df_coord',df_coord)
In [376]: c = store.select_as_coordinates('df_coord','index>20020101')
In [377]: c.summary()
Out[377]: u'Int64Index: 268 entries, 732 to 999'
In [378]: store.select('df_coord',where=c)
Out[378]:
0 1
2002-01-02 -0.178266 -0.064638
2002-01-03 -1.204956 -3.880898
2002-01-04 0.974470 0.415160
2002-01-05 1.751967 0.485011
2002-01-06 -0.170894 0.748870
2002-01-07 0.629793 0.811053
2002-01-08 2.133776 0.238459
... ... ...
2002-09-20 -0.181434 0.612399
2002-09-21 -0.763324 -0.354962
2002-09-22 -0.261776 0.812126
2002-09-23 0.482615 -0.886512
2002-09-24 -0.037757 -0.562953
2002-09-25 0.897706 0.383232
2002-09-26 -1.324806 1.139269
[268 rows x 2 columns]
Selecting using a where mask¶
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.
In [379]: df_mask = DataFrame(np.random.randn(1000,2),index=date_range('20000101',periods=1000))
In [380]: store.append('df_mask',df_mask)
In [381]: c = store.select_column('df_mask','index')
In [382]: where = c[DatetimeIndex(c).month==5].index
In [383]: store.select('df_mask',where=where)
Out[383]:
0 1
2000-05-01 -1.006245 -0.616759
2000-05-02 0.218940 0.717838
2000-05-03 0.013333 1.348060
2000-05-04 0.662176 -1.050645
2000-05-05 -1.034870 -0.243242
2000-05-06 -0.753366 -1.454329
2000-05-07 -1.022920 -0.476989
... ... ...
2002-05-25 -0.509090 -0.389376
2002-05-26 0.150674 1.164337
2002-05-27 -0.332944 0.115181
2002-05-28 -1.048127 -0.605733
2002-05-29 1.418754 -0.442835
2002-05-30 -0.433200 0.835001
2002-05-31 -1.041278 1.401811
[93 rows x 2 columns]
Storer Object¶
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.
In [384]: store.get_storer('df_dc').nrows
Out[384]: 8
Multiple Table Queries¶
New in 0.10.1 are the methods append_to_multiple
and
select_as_multiple
, that 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.
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.
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.
In [385]: df_mt = DataFrame(randn(8, 6), index=date_range('1/1/2000', periods=8),
.....: columns=['A', 'B', 'C', 'D', 'E', 'F'])
.....:
In [386]: df_mt['foo'] = 'bar'
In [387]: df_mt.ix[1, ('A', 'B')] = np.nan
# you can also create the tables individually
In [388]: store.append_to_multiple({'df1_mt': ['A', 'B'], 'df2_mt': None },
.....: df_mt, selector='df1_mt')
.....:
In [389]: store
Out[389]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/df1_mt frame_table (typ->appendable,nrows->8,ncols->2,indexers->[index],dc->[A,B])
/df2_mt frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index])
/df_coord frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_dc frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index],dc->[B,C,string,string2])
/df_mask frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_mi frame_table (typ->appendable_multi,nrows->10,ncols->5,indexers->[index],dc->[bar,foo])
/df_mixed frame_table (typ->appendable,nrows->8,ncols->7,indexers->[index])
/dfeq frame_table (typ->appendable,nrows->10,ncols->1,indexers->[index],dc->[number])
/dfq frame_table (typ->appendable,nrows->10,ncols->4,indexers->[index],dc->[A,B,C,D])
/dftd frame_table (typ->appendable,nrows->10,ncols->3,indexers->[index],dc->[A,B,C])
/foo/bar/bah frame (shape->[8,3])
/wp wide_table (typ->appendable,nrows->20,ncols->2,indexers->[major_axis,minor_axis])
# individual tables were created
In [390]: store.select('df1_mt')
Out[390]:
A B
2000-01-01 0.714697 0.318215
2000-01-02 NaN NaN
2000-01-03 -0.086919 0.416905
2000-01-04 0.489131 -0.253340
2000-01-05 -0.382952 -0.397373
2000-01-06 0.538116 0.226388
2000-01-07 -2.073479 -0.115926
2000-01-08 -0.695400 0.402493
In [391]: store.select('df2_mt')
Out[391]:
C D E F foo
2000-01-01 0.607460 0.790907 0.852225 0.096696 bar
2000-01-02 0.811031 -0.356817 1.047085 0.664705 bar
2000-01-03 -0.764381 -0.287229 -0.089351 -1.035115 bar
2000-01-04 -1.948100 -0.116556 0.800597 -0.796154 bar
2000-01-05 -0.717627 0.156995 -0.344718 -0.171208 bar
2000-01-06 1.541729 0.205256 1.998065 0.953591 bar
2000-01-07 1.391070 0.303013 1.093347 -0.101000 bar
2000-01-08 -1.507639 0.089575 0.658822 -1.037627 bar
# as a multiple
In [392]: store.select_as_multiple(['df1_mt', 'df2_mt'], where=['A>0', 'B>0'],
.....: selector = 'df1_mt')
.....:
Out[392]:
A B C D E F foo
2000-01-01 0.714697 0.318215 0.607460 0.790907 0.852225 0.096696 bar
2000-01-06 0.538116 0.226388 1.541729 0.205256 1.998065 0.953591 bar
Delete from a Table¶
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. This is especially true in higher dimensional
objects (Panel
and Panel4D
). To get optimal performance, it’s
worthwhile to have the dimension you are deleting be the first of the
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:
- date_1 - id_1 - id_2 - . - id_n
- date_2 - id_1 - . - 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.
# returns the number of rows deleted
In [393]: store.remove('wp', 'major_axis>20000102' )
Out[393]: 12
In [394]: store.select('wp')
Out[394]:
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 2 (major_axis) x 4 (minor_axis)
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-02 00:00:00
Minor_axis axis: A to D
Warning
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
Notes & Caveats¶
Compression¶
PyTables
allows the stored data to be compressed. This applies to
all kinds of stores, not just tables.
- Pass
complevel=int
for a compression level (1-9, with 0 being no compression, and the default) - Pass
complib=lib
where lib is any ofzlib, bzip2, lzo, blosc
for whichever compression library you prefer.
HDFStore
will use the file based compression scheme if no overriding
complib
or complevel
options are provided. blosc
offers very
fast compression, and is my most used. Note that lzo
and bzip2
may not be installed (by Python) by default.
Compression for all objects within the file
store_compressed = HDFStore('store_compressed.h5', complevel=9, complib='blosc')
Or on-the-fly compression (this only applies to tables). You can turn
off file compression for a specific table by passing complevel=0
store.append('df', df, complib='zlib', complevel=5)
ptrepack¶
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 --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.
Caveats¶
Warning
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 usestore.flush(fsync=True)
to do this for you. - Once a
table
is created its items (Panel) / 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 usetz_convert
with the updated timezone definition.
Warning
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.
DataTypes¶
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 |
np.nan |
integer : int64, int32, int8, uint64,uint32, uint8 |
|
boolean | |
datetime64[ns] |
NaT |
timedelta64[ns] |
NaT |
categorical : see the section below | |
object : strings |
np.nan |
unicode
columns are not supported, and WILL FAIL.
Categorical Data¶
New in version 0.15.2.
Writing data to a HDFStore
that contains a category
dtype was implemented
in 0.15.2. Queries work the same as if it was an object array. However, the category
dtyped data is
stored in a more efficient manner.
In [395]: dfcat = DataFrame({ 'A' : Series(list('aabbcdba')).astype('category'),
.....: 'B' : np.random.randn(8) })
.....:
In [396]: dfcat
Out[396]:
A B
0 a 0.603273
1 a 0.262554
2 b -0.979586
3 b 2.132387
4 c 0.892485
5 d 1.996474
6 b 0.231425
7 a 0.980070
In [397]: dfcat.dtypes
Out[397]:
A category
B float64
dtype: object
In [398]: cstore = pd.HDFStore('cats.h5', mode='w')
In [399]: cstore.append('dfcat', dfcat, format='table', data_columns=['A'])
In [400]: result = cstore.select('dfcat', where="A in ['b','c']")
In [401]: result
Out[401]:
A B
2 b -0.979586
3 b 2.132387
4 c 0.892485
6 b 0.231425
In [402]: result.dtypes
Out[402]:
A category
B float64
dtype: object
Warning
The format of the Categorical
is readable by prior versions of pandas (< 0.15.2), but will retrieve
the data as an integer based column (e.g. the codes
). However, the categories
can be retrieved
but require the user to select them manually using the explicit meta path.
The data is stored like so:
In [403]: cstore
Out[403]:
<class 'pandas.io.pytables.HDFStore'>
File path: cats.h5
/dfcat frame_table (typ->appendable,nrows->8,ncols->2,indexers->[index],dc->[A])
/dfcat/meta/A/meta series_table (typ->appendable,nrows->4,ncols->1,indexers->[index],dc->[values])
# to get the categories
In [404]: cstore.select('dfcat/meta/A/meta')
Out[404]:
0 a
1 b
2 c
3 d
dtype: object
String Columns¶
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.
Starting in 0.11.0, passing a min_itemsize
dict will cause all passed columns to be created as data_columns automatically.
Note
If you are not passing any data_columns
, then the min_itemsize
will be the maximum of the length of any string passed
In [405]: dfs = DataFrame(dict(A = 'foo', B = 'bar'),index=list(range(5)))
In [406]: dfs
Out[406]:
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 [407]: store.append('dfs', dfs, min_itemsize = 30)
In [408]: store.get_storer('dfs').table
Out[408]:
/dfs/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=30, shape=(2,), dflt='', pos=1)}
byteorder := 'little'
chunkshape := (963,)
autoindex := True
colindexes := {
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False}
# A is created as a data_column with a size of 30
# B is size is calculated
In [409]: store.append('dfs2', dfs, min_itemsize = { 'A' : 30 })
In [410]: store.get_storer('dfs2').table
Out[410]:
/dfs2/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=3, shape=(1,), dflt='', pos=1),
"A": StringCol(itemsize=30, shape=(), dflt='', pos=2)}
byteorder := 'little'
chunkshape := (1598,)
autoindex := True
colindexes := {
"A": Index(6, medium, shuffle, zlib(1)).is_csi=False,
"index": Index(6, medium, shuffle, zlib(1)).is_csi=False}
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.
In [411]: dfss = DataFrame(dict(A = ['foo','bar','nan']))
In [412]: dfss
Out[412]:
A
0 foo
1 bar
2 nan
In [413]: store.append('dfss', dfss)
In [414]: store.select('dfss')
Out[414]:
A
0 foo
1 bar
2 NaN
# here you need to specify a different nan rep
In [415]: store.append('dfss2', dfss, nan_rep='_nan_')
In [416]: store.select('dfss2')
Out[416]:
A
0 foo
1 bar
2 nan
External Compatibility¶
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:
In [417]: np.random.seed(1)
In [418]: 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 [419]: df_for_r.head()
Out[419]:
class first second
0 0 0.417022 0.326645
1 0 0.720324 0.527058
2 1 0.000114 0.885942
3 1 0.302333 0.357270
4 1 0.146756 0.908535
In [420]: store_export = HDFStore('export.h5')
In [421]: store_export.append('df_for_r', df_for_r, data_columns=df_dc.columns)
In [422]: store_export
Out[422]:
<class 'pandas.io.pytables.HDFStore'>
File path: export.h5
/df_for_r frame_table (typ->appendable,nrows->100,ncols->3,indexers->[index])
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
:
# 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 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
Note
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.
Backwards Compatibility¶
0.10.1 of HDFStore
can read tables created in a prior version of pandas,
however query terms using the
prior (undocumented) methodology are unsupported. HDFStore
will
issue a warning if you try to use a legacy-format file. You must
read in the entire file and write it out using the new format, using the
method copy
to take advantage of the updates. The group attribute
pandas_version
contains the version information. copy
takes a
number of options, please see the docstring.
# a legacy store
In [423]: legacy_store = HDFStore(legacy_file_path,'r')
In [424]: legacy_store
Out[424]:
<class 'pandas.io.pytables.HDFStore'>
File path: /build/pandas-8apiR_/pandas-0.17.1/doc/source/_static/legacy_0.10.h5
/a series (shape->[30])
/b frame (shape->[30,4])
/df1_mixed frame_table [0.10.0] (typ->appendable,nrows->30,ncols->11,indexers->[index])
/foo/bar wide (shape->[3,30,4])
/p1_mixed wide_table [0.10.0] (typ->appendable,nrows->120,ncols->9,indexers->[major_axis,minor_axis])
/p4d_mixed ndim_table [0.10.0] (typ->appendable,nrows->360,ncols->9,indexers->[items,major_axis,minor_axis])
# copy (and return the new handle)
In [425]: new_store = legacy_store.copy('store_new.h5')
In [426]: new_store
Out[426]:
<class 'pandas.io.pytables.HDFStore'>
File path: store_new.h5
/a series (shape->[30])
/b frame (shape->[30,4])
/df1_mixed frame_table (typ->appendable,nrows->30,ncols->11,indexers->[index])
/foo/bar wide (shape->[3,30,4])
/p1_mixed wide_table (typ->appendable,nrows->120,ncols->9,indexers->[major_axis,minor_axis])
/p4d_mixed wide_table (typ->appendable,nrows->360,ncols->9,indexers->[items,major_axis,minor_axis])
In [427]: new_store.close()
Performance¶
tables
format come with a writing performance penalty as compared tofixed
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.- You can pass
chunksize=<int>
toappend
, specifying the write chunksize (default is 50000). This will significantly lower your memory usage on writing. - You can pass
expectedrows=<int>
to the firstappend
, to set the TOTAL number of expected rows thatPyTables
will expected. This will optimize read/write performance. - 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.
Experimental¶
HDFStore supports Panel4D
storage.
In [428]: p4d = Panel4D({ 'l1' : wp })
In [429]: p4d
Out[429]:
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 1 (labels) x 2 (items) x 5 (major_axis) x 4 (minor_axis)
Labels axis: l1 to l1
Items axis: Item1 to Item2
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-05 00:00:00
Minor_axis axis: A to D
In [430]: store.append('p4d', p4d)
In [431]: store
Out[431]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/df1_mt frame_table (typ->appendable,nrows->8,ncols->2,indexers->[index],dc->[A,B])
/df2_mt frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index])
/df_coord frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_dc frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index],dc->[B,C,string,string2])
/df_mask frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_mi frame_table (typ->appendable_multi,nrows->10,ncols->5,indexers->[index],dc->[bar,foo])
/df_mixed frame_table (typ->appendable,nrows->8,ncols->7,indexers->[index])
/dfeq frame_table (typ->appendable,nrows->10,ncols->1,indexers->[index],dc->[number])
/dfq frame_table (typ->appendable,nrows->10,ncols->4,indexers->[index],dc->[A,B,C,D])
/dfs frame_table (typ->appendable,nrows->5,ncols->2,indexers->[index])
/dfs2 frame_table (typ->appendable,nrows->5,ncols->2,indexers->[index],dc->[A])
/dfss frame_table (typ->appendable,nrows->3,ncols->1,indexers->[index])
/dfss2 frame_table (typ->appendable,nrows->3,ncols->1,indexers->[index])
/dftd frame_table (typ->appendable,nrows->10,ncols->3,indexers->[index],dc->[A,B,C])
/foo/bar/bah frame (shape->[8,3])
/p4d wide_table (typ->appendable,nrows->40,ncols->1,indexers->[items,major_axis,minor_axis])
/wp wide_table (typ->appendable,nrows->8,ncols->2,indexers->[major_axis,minor_axis])
These, by default, index the three axes items, major_axis,
minor_axis
. On an AppendableTable
it is possible to setup with the
first append a different indexing scheme, depending on how you want to
store your data. Pass the axes
keyword with a list of dimensions
(currently must by exactly 1 less than the total dimensions of the
object). This cannot be changed after table creation.
In [432]: store.append('p4d2', p4d, axes=['labels', 'major_axis', 'minor_axis'])
In [433]: store
Out[433]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
/df frame_table (typ->appendable,nrows->8,ncols->3,indexers->[index])
/df1_mt frame_table (typ->appendable,nrows->8,ncols->2,indexers->[index],dc->[A,B])
/df2_mt frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index])
/df_coord frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_dc frame_table (typ->appendable,nrows->8,ncols->5,indexers->[index],dc->[B,C,string,string2])
/df_mask frame_table (typ->appendable,nrows->1000,ncols->2,indexers->[index])
/df_mi frame_table (typ->appendable_multi,nrows->10,ncols->5,indexers->[index],dc->[bar,foo])
/df_mixed frame_table (typ->appendable,nrows->8,ncols->7,indexers->[index])
/dfeq frame_table (typ->appendable,nrows->10,ncols->1,indexers->[index],dc->[number])
/dfq frame_table (typ->appendable,nrows->10,ncols->4,indexers->[index],dc->[A,B,C,D])
/dfs frame_table (typ->appendable,nrows->5,ncols->2,indexers->[index])
/dfs2 frame_table (typ->appendable,nrows->5,ncols->2,indexers->[index],dc->[A])
/dfss frame_table (typ->appendable,nrows->3,ncols->1,indexers->[index])
/dfss2 frame_table (typ->appendable,nrows->3,ncols->1,indexers->[index])
/dftd frame_table (typ->appendable,nrows->10,ncols->3,indexers->[index],dc->[A,B,C])
/foo/bar/bah frame (shape->[8,3])
/p4d wide_table (typ->appendable,nrows->40,ncols->1,indexers->[items,major_axis,minor_axis])
/p4d2 wide_table (typ->appendable,nrows->20,ncols->2,indexers->[labels,major_axis,minor_axis])
/wp wide_table (typ->appendable,nrows->8,ncols->2,indexers->[major_axis,minor_axis])
In [434]: store.select('p4d2', [ Term('labels=l1'), Term('items=Item1'), Term('minor_axis=A_big_strings') ])
Out[434]:
<class 'pandas.core.panelnd.Panel4D'>
Dimensions: 0 (labels) x 1 (items) x 0 (major_axis) x 0 (minor_axis)
Labels axis: None
Items axis: Item1 to Item1
Major_axis axis: None
Minor_axis axis: None
SQL Queries¶
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.
New in version 0.14.0.
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 database table into a DataFrame. |
read_sql_query (sql, con[, index_col, ...]) |
Read SQL query into a DataFrame. |
read_sql (sql, con[, index_col, ...]) |
Read SQL query or database table into a DataFrame. |
DataFrame.to_sql (name, con[, flavor, ...]) |
Write records stored in a DataFrame to a SQL database. |
Note
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.
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
In [435]: from sqlalchemy import create_engine
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-435-f1d47d3a77bf> in <module>()
----> 1 from sqlalchemy import create_engine
ImportError: No module named sqlalchemy
# Create your engine.
In [436]: engine = create_engine('sqlite:///:memory:')
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-436-51912da07a87> 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)
Writing DataFrames¶
Assuming the following data is in a DataFrame data
, we can insert it into
the database using to_sql()
.
id | Date | Col_1 | Col_2 | Col_3 |
---|---|---|---|---|
26 | 2012-10-18 | X | 25.7 | True |
42 | 2012-10-19 | Y | -12.4 | False |
63 | 2012-10-20 | Z | 5.73 | True |
In [437]: data.to_sql('data', engine)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-437-b10ff34c034b> 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 [438]: data.to_sql('data_chunked', engine, chunksize=1000)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-438-e899c31edee9> in <module>()
----> 1 data.to_sql('data_chunked', engine, chunksize=1000)
NameError: name 'engine' is not defined
SQL data types¶
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:
In [439]: from sqlalchemy.types import String
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-439-83ecf8de5e0b> in <module>()
----> 1 from sqlalchemy.types import String
ImportError: No module named sqlalchemy.types
In [440]: data.to_sql('data_dtype', engine, dtype={'Col_1': String})
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-440-96dff3d2d9fe> in <module>()
----> 1 data.to_sql('data_dtype', engine, dtype={'Col_1': String})
NameError: name 'engine' is not defined
Note
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.
Note
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.
Reading Tables¶
read_sql_table()
will read a database table given the
table name and optionally a subset of columns to read.
Note
In order to use read_sql_table()
, you must have the
SQLAlchemy optional dependency installed.
In [441]: pd.read_sql_table('data', engine)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-441-ce9b1988c629> in <module>()
----> 1 pd.read_sql_table('data', engine)
NameError: name 'engine' is not defined
You can also specify the name of the column as the DataFrame index, and specify a subset of columns to be read.
In [442]: pd.read_sql_table('data', engine, index_col='id')
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-442-1614c2f6316d> in <module>()
----> 1 pd.read_sql_table('data', engine, index_col='id')
NameError: name 'engine' is not defined
In [443]: pd.read_sql_table('data', engine, columns=['Col_1', 'Col_2'])
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-443-b4a6dece8ff2> 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 [444]: pd.read_sql_table('data', engine, parse_dates=['Date'])
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-444-0e196c651a7f> 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()
:
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()
Schema support¶
New in version 0.15.0.
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')
Querying¶
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 [445]: pd.read_sql_query('SELECT * FROM data', engine)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-445-0b51156a07a3> 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 [446]: pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;", engine)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-446-638de09f3f40> 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 [447]: df = pd.DataFrame(np.random.randn(20, 3), columns=list('abc'))
In [448]: df.to_sql('data_chunks', engine, index=False)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-448-0b7ccd98cc80> in <module>()
----> 1 df.to_sql('data_chunks', engine, index=False)
NameError: name 'engine' is not defined
In [449]: for chunk in pd.read_sql_query("SELECT * FROM data_chunks", engine, chunksize=5):
.....: print(chunk)
.....:
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-449-e9c157f2270a> in <module>()
----> 1 for chunk in pd.read_sql_query("SELECT * FROM data_chunks", engine, chunksize=5):
2 print(chunk)
3
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.
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)])
Engine connection examples¶
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
Advanced SQLAlchemy queries¶
You can use SQLAlchemy constructs to describe your query.
Use sqlalchemy.text()
to specify query parameters in a backend-neutral way
In [450]: import sqlalchemy as sa
---------------------------------------------------------------------------
ImportError Traceback (most recent call last)
<ipython-input-450-efbf8abb5ae0> in <module>()
----> 1 import sqlalchemy as sa
ImportError: No module named sqlalchemy
In [451]: pd.read_sql(sa.text('SELECT * FROM data where Col_1=:col1'), engine, params={'col1': 'X'})
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-451-5cd50b4746de> in <module>()
----> 1 pd.read_sql(sa.text('SELECT * FROM data where Col_1=:col1'), engine, params={'col1': 'X'})
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/core/generic.pyc in __getattr__(self, name)
2358 return self[name]
2359 raise AttributeError("'%s' object has no attribute '%s'" %
-> 2360 (type(self).__name__, name))
2361
2362 def __setattr__(self, name, value):
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 [452]: metadata = sa.MetaData()
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-452-bea88176fe71> in <module>()
----> 1 metadata = sa.MetaData()
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/core/generic.pyc in __getattr__(self, name)
2358 return self[name]
2359 raise AttributeError("'%s' object has no attribute '%s'" %
-> 2360 (type(self).__name__, name))
2361
2362 def __setattr__(self, name, value):
AttributeError: 'Series' object has no attribute 'MetaData'
In [453]: 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-453-aff99cbc17d9> 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),
/build/pandas-8apiR_/pandas-0.17.1/debian/python-pandas/usr/lib/python2.7/dist-packages/pandas/core/generic.pyc in __getattr__(self, name)
2358 return self[name]
2359 raise AttributeError("'%s' object has no attribute '%s'" %
-> 2360 (type(self).__name__, name))
2361
2362 def __setattr__(self, name, value):
AttributeError: 'Series' object has no attribute 'Table'
In [454]: pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 == True), engine)
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-454-5e38c8249be6> in <module>()
----> 1 pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 == True), engine)
NameError: name 'data_table' is not defined
You can combine SQLAlchemy expressions with parameters passed to read_sql()
using sqlalchemy.bindparam()
In [455]: import datetime as dt
In [456]: expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam('date'))
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-456-8f297d55803c> in <module>()
----> 1 expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam('date'))
NameError: name 'data_table' is not defined
In [457]: pd.read_sql(expr, engine, params={'date': dt.datetime(2010, 10, 18)})
---------------------------------------------------------------------------
NameError Traceback (most recent call last)
<ipython-input-457-ba4c9a9d7e36> in <module>()
----> 1 pd.read_sql(expr, engine, params={'date': dt.datetime(2010, 10, 18)})
NameError: name 'engine' is not defined
Sqlite fallback¶
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', cnx)
pd.read_sql_query("SELECT * FROM data", con)
Google BigQuery (Experimental)¶
New in version 0.13.0.
The pandas.io.gbq
module provides a wrapper for Google’s BigQuery
analytics web service to simplify retrieving results from BigQuery tables
using SQL-like queries. Result sets are parsed into a pandas
DataFrame with a shape and data types derived from the source table.
Additionally, DataFrames can be inserted into new BigQuery tables or appended
to existing tables.
Warning
To use this module, you will need a valid BigQuery account. Refer to the BigQuery Documentation for details on the service itself.
The key functions are:
read_gbq (query[, project_id, index_col, ...]) |
Load data from Google BigQuery. |
to_gbq (dataframe, destination_table, project_id) |
Write a DataFrame to a Google BigQuery table. |
Querying¶
Suppose you want to load all data from an existing BigQuery table : test_dataset.test_table
into a DataFrame using the read_gbq()
function.
# Insert your BigQuery Project ID Here
# Can be found in the Google web console
projectid = "xxxxxxxx"
data_frame = pd.read_gbq('SELECT * FROM test_dataset.test_table', projectid)
You will then be authenticated to the specified BigQuery account via Google’s Oauth2 mechanism. In general, this is as simple as following the prompts in a browser window which will be opened for you. Should the browser not be available, or fail to launch, a code will be provided to complete the process manually. Additional information on the authentication mechanism can be found here.
You can define which column from BigQuery to use as an index in the destination DataFrame as well as a preferred column order as follows:
data_frame = pd.read_gbq('SELECT * FROM test_dataset.test_table',
index_col='index_column_name',
col_order=['col1', 'col2', 'col3'], projectid)
Note
You can find your project id in the BigQuery management console.
Note
You can toggle the verbose output via the verbose
flag which defaults to True
.
Writing DataFrames¶
Assume we want to write a DataFrame df
into a BigQuery table using to_gbq()
.
In [458]: df = pd.DataFrame({'my_string': list('abc'),
.....: 'my_int64': list(range(1, 4)),
.....: 'my_float64': np.arange(4.0, 7.0),
.....: 'my_bool1': [True, False, True],
.....: 'my_bool2': [False, True, False],
.....: 'my_dates': pd.date_range('now', periods=3)})
.....:
In [459]: df
Out[459]:
my_bool1 my_bool2 my_dates my_float64 my_int64 my_string
0 True False 2016-02-09 20:29:47.382064 4 1 a
1 False True 2016-02-10 20:29:47.382064 5 2 b
2 True False 2016-02-11 20:29:47.382064 6 3 c
In [460]: df.dtypes
Out[460]:
my_bool1 bool
my_bool2 bool
my_dates datetime64[ns]
my_float64 float64
my_int64 int64
my_string object
dtype: object
df.to_gbq('my_dataset.my_table', projectid)
Note
The destination table and destination dataset will automatically be created if they do not already exist.
The if_exists
argument can be used to dictate whether to 'fail'
, 'replace'
or 'append'
if the destination table already exists. The default value is 'fail'
.
For example, assume that if_exists
is set to 'fail'
. The following snippet will raise
a TableCreationError
if the destination table already exists.
df.to_gbq('my_dataset.my_table', projectid, if_exists='fail')
Note
If the if_exists
argument is set to 'append'
, the destination dataframe will
be written to the table using the defined table schema and column types. The
dataframe must match the destination table in column order, structure, and
data types.
If the if_exists
argument is set to 'replace'
, and the existing table has a
different schema, a delay of 2 minutes will be forced to ensure that the new schema
has propagated in the Google environment. See
Google BigQuery issue 191.
Writing large DataFrames can result in errors due to size limitations being exceeded.
This can be avoided by setting the chunksize
argument when calling to_gbq()
.
For example, the following writes df
to a BigQuery table in batches of 10000 rows at a time:
df.to_gbq('my_dataset.my_table', projectid, chunksize=10000)
You can also see the progress of your post via the verbose
flag which defaults to True
.
For example:
In [8]: df.to_gbq('my_dataset.my_table', projectid, chunksize=10000, verbose=True)
Streaming Insert is 10% Complete
Streaming Insert is 20% Complete
Streaming Insert is 30% Complete
Streaming Insert is 40% Complete
Streaming Insert is 50% Complete
Streaming Insert is 60% Complete
Streaming Insert is 70% Complete
Streaming Insert is 80% Complete
Streaming Insert is 90% Complete
Streaming Insert is 100% Complete
Note
If an error occurs while streaming data to BigQuery, see Troubleshooting BigQuery Errors.
Note
The BigQuery SQL query language has some oddities, see the BigQuery Query Reference Documentation.
Note
While BigQuery uses SQL-like syntax, it has some important differences from traditional databases both in functionality, API limitations (size and quantity of queries or uploads), and how Google charges for use of the service. You should refer to Google BigQuery documentation often as the service seems to be changing and evolving. BiqQuery is best for analyzing large sets of data quickly, but it is not a direct replacement for a transactional database.
Creating BigQuery Tables¶
Warning
As of 0.17, the function generate_bq_schema()
has been deprecated and will be
removed in a future version.
As of 0.15.2, the gbq module has a function generate_bq_schema()
which will
produce the dictionary representation schema of the specified pandas DataFrame.
In [10]: gbq.generate_bq_schema(df, default_type='STRING')
Out[10]: {'fields': [{'name': 'my_bool1', 'type': 'BOOLEAN'},
{'name': 'my_bool2', 'type': 'BOOLEAN'},
{'name': 'my_dates', 'type': 'TIMESTAMP'},
{'name': 'my_float64', 'type': 'FLOAT'},
{'name': 'my_int64', 'type': 'INTEGER'},
{'name': 'my_string', 'type': 'STRING'}]}
Note
If you delete and re-create a BigQuery table with the same name, but different table schema, you must wait 2 minutes before streaming data into the table. As a workaround, consider creating the new table with a different name. Refer to Google BigQuery issue 191.
Stata Format¶
New in version 0.12.0.
Writing to Stata format¶
The method to_stata()
will write a DataFrame
into a .dta file. The format version of this file is always 115 (Stata 12).
In [461]: df = DataFrame(randn(10, 2), columns=list('AB'))
In [462]: 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).
Note
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
.
Warning
Conversion from int64
to float64
may result in a loss of precision
if int64
values are larger than 2**53.
Warning
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
.
Reading from Stata format¶
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.
In [463]: pd.read_stata('stata.dta')
Out[463]:
index A B
0 0 1.810535 -1.305727
1 1 -0.344987 -0.230840
2 2 -2.793085 1.937529
3 3 0.366332 -1.044589
4 4 2.051173 0.585662
5 5 0.429526 -0.606998
6 6 0.106223 -1.525680
7 7 0.795026 -0.374438
8 8 0.134048 1.202055
9 9 0.284748 0.262467
New in version 0.16.0.
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 [464]: reader = pd.read_stata('stata.dta', chunksize=3)
In [465]: 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 [466]: reader = pd.read_stata('stata.dta', iterator=True)
In [467]: chunk1 = reader.read(5)
In [468]: 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.
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.
Note
read_stata()
and
StataReader
support .dta formats 113-115
(Stata 10-12), 117 (Stata 13), and 118 (Stata 14).
Note
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.
Categorical Data¶
New in version 0.15.2.
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.
Warning
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.
Note
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
.
Note
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 Categorial
with string categories for the values that are labeled and
numeric categories for values with no label.
Other file formats¶
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.
SAS Format¶
New in version 0.17.0.
The top-level function read_sas()
currently can read (but
not write) SAS xport (.XPT) format files. Pandas cannot currently
handle SAS7BDAT files.
XPORT files only contain two value types: ASCII text and double
precision numeric values. There is no automatic type conversion to
integers, dates, or categoricals. By default the whole file is read
and returned as a DataFrame
.
Specify a chunksize
or use iterator=True
to obtain an
XportReader
object for incrementally reading the file. The
XportReader
object also has attributes that contain additional
information about the file and its variables.
Read a SAS XPORT file:
df = pd.read_sas('sas_xport.xpt')
Obtain an iterator and read an XPORT file 100,000 lines at a time:
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.
Performance Considerations¶
This is an informal comparison of various IO methods, using pandas 0.13.1.
In [1]: df = DataFrame(randn(1000000,2),columns=list('AB'))
In [2]: df.info()
<class 'pandas.core.frame.DataFrame'>
Int64Index: 1000000 entries, 0 to 999999
Data columns (total 2 columns):
A 1000000 non-null float64
B 1000000 non-null float64
dtypes: float64(2)
memory usage: 22.9 MB
Writing
In [14]: %timeit test_sql_write(df)
1 loops, best of 3: 6.24 s per loop
In [15]: %timeit test_hdf_fixed_write(df)
1 loops, best of 3: 237 ms per loop
In [26]: %timeit test_hdf_fixed_write_compress(df)
1 loops, best of 3: 245 ms per loop
In [16]: %timeit test_hdf_table_write(df)
1 loops, best of 3: 901 ms per loop
In [27]: %timeit test_hdf_table_write_compress(df)
1 loops, best of 3: 952 ms per loop
In [17]: %timeit test_csv_write(df)
1 loops, best of 3: 3.44 s per loop
Reading
In [18]: %timeit test_sql_read()
1 loops, best of 3: 766 ms per loop
In [19]: %timeit test_hdf_fixed_read()
10 loops, best of 3: 19.1 ms per loop
In [28]: %timeit test_hdf_fixed_read_compress()
10 loops, best of 3: 36.3 ms per loop
In [20]: %timeit test_hdf_table_read()
10 loops, best of 3: 39 ms per loop
In [29]: %timeit test_hdf_table_read_compress()
10 loops, best of 3: 60.6 ms per loop
In [22]: %timeit test_csv_read()
1 loops, best of 3: 620 ms per loop
Space on disk (in bytes)
25843712 Apr 8 14:11 test.sql
24007368 Apr 8 14:11 test_fixed.hdf
15580682 Apr 8 14:11 test_fixed_compress.hdf
24458444 Apr 8 14:11 test_table.hdf
16797283 Apr 8 14:11 test_table_compress.hdf
46152810 Apr 8 14:11 test.csv
And here’s the code
import sqlite3
import os
from pandas.io import sql
df = DataFrame(randn(1000000,2),columns=list('AB'))
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)