IO tools (text, CSV, HDF5, …)¶
The pandas I/O API is a set of top level reader functions accessed like
pandas.read_csv() that generally return a pandas object. The corresponding
writer functions are object methods that are accessed like
DataFrame.to_csv(). Below is a table containing available readers and
writers.
Format Type |
Data Description |
Reader |
Writer |
|---|---|---|---|
text |
|||
text |
Fixed-Width Text File |
||
text |
|||
text |
|||
text |
|||
text |
|||
text |
Local clipboard |
||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
binary |
|||
SQL |
|||
SQL |
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
with from io import StringIO for Python 3.
CSV & text files¶
The workhorse function for reading text files (a.k.a. flat files) is
read_csv(). See the cookbook for some advanced strategies.
Parsing options¶
read_csv() accepts the following common arguments:
Basic¶
- filepath_or_buffervarious
Either a path to a file (a
str,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).- sepstr, defaults to
','forread_csv(),\tforread_table() Delimiter to use. If sep is
None, the C engine cannot automatically detect the separator, but the Python parsing engine can, meaning the latter will be used and automatically detect the separator by Python’s builtin sniffer tool,csv.Sniffer. In addition, separators longer than 1 character and different from'\s+'will be interpreted as regular expressions and will also force the use of the Python parsing engine. Note that regex delimiters are prone to ignoring quoted data. Regex example:'\\r\\t'.- delimiterstr, default
None Alternative argument name for sep.
- delim_whitespaceboolean, default False
Specifies whether or not whitespace (e.g.
' 'or'\t') will be used as the delimiter. Equivalent to settingsep='\s+'. If this option is set toTrue, nothing should be passed in for thedelimiterparameter.
Column and index locations and names¶
- headerint or list of ints, default
'infer' Row number(s) to use as the column names, and the start of the data. Default behavior is to infer the column names: if no names are passed the behavior is identical to
header=0and column names are inferred from the first line of the file, if column names are passed explicitly then the behavior is identical toheader=None. Explicitly passheader=0to be able to replace existing names.The header can be a list of ints that specify row locations for a MultiIndex on the columns e.g.
[0,1,3]. Intervening rows that are not specified will be skipped (e.g. 2 in this example is skipped). Note that this parameter ignores commented lines and empty lines ifskip_blank_lines=True, so header=0 denotes the first line of data rather than the first line of the file.- namesarray-like, default
None List of column names to use. If file contains no header row, then you should explicitly pass
header=None. Duplicates in this list are not allowed.- index_colint, str, sequence of int / str, or False, optional, default
None Column(s) to use as the row labels of the
DataFrame, either given as string name or column index. If a sequence of int / str is given, a MultiIndex is used.Note:
index_col=Falsecan be used to force pandas to not use the first column as the index, e.g. when you have a malformed file with delimiters at the end of each line.The default value of
Noneinstructs pandas to guess. If the number of fields in the column header row is equal to the number of fields in the body of the data file, then a default index is used. If it is larger, then the first columns are used as index so that the remaining number of fields in the body are equal to the number of fields in the header.The first row after the header is used to determine the number of columns, which will go into the index. If the subsequent rows contain less columns than the first row, they are filled with
NaN.This can be avoided through
usecols. This ensures that the columns are taken as is and the trailing data are ignored.- usecolslist-like or callable, default
None Return a subset of the columns. If list-like, all elements must either be positional (i.e. integer indices into the document columns) or strings that correspond to column names provided either by the user in
namesor inferred from the document header row(s). Ifnamesare given, the document header row(s) are not taken into account. For example, a valid list-likeusecolsparameter would be[0, 1, 2]or['foo', 'bar', 'baz'].Element order is ignored, so
usecols=[0, 1]is the same as[1, 0]. To instantiate a DataFrame fromdatawith element order preserved usepd.read_csv(data, usecols=['foo', 'bar'])[['foo', 'bar']]for columns in['foo', 'bar']order orpd.read_csv(data, usecols=['foo', 'bar'])[['bar', 'foo']]for['bar', 'foo']order.If callable, the callable function will be evaluated against the column names, returning names where the callable function evaluates to True:
In [1]: import pandas as pd In [2]: from io import StringIO In [3]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3" In [4]: pd.read_csv(StringIO(data)) Out[4]: col1 col2 col3 0 a b 1 1 a b 2 2 c d 3 In [5]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["COL1", "COL3"]) Out[5]: col1 col3 0 a 1 1 a 2 2 c 3
Using this parameter results in much faster parsing time and lower memory usage when using the c engine. The Python engine loads the data first before deciding which columns to drop.
- squeezeboolean, default
False If the parsed data only contains one column then return a
Series.Deprecated since version 1.4.0: Append
.squeeze("columns")to the call to{func_name}to squeeze the data.- prefixstr, default
None Prefix to add to column numbers when no header, e.g. ‘X’ for X0, X1, …
Deprecated since version 1.4.0: Use a list comprehension on the DataFrame’s columns after calling
read_csv.In [6]: data = "col1,col2,col3\na,b,1" In [7]: df = pd.read_csv(StringIO(data)) In [8]: df.columns = [f"pre_{col}" for col in df.columns] In [9]: df Out[9]: pre_col1 pre_col2 pre_col3 0 a b 1
- mangle_dupe_colsboolean, default
True Duplicate columns will be specified as ‘X’, ‘X.1’…’X.N’, rather than ‘X’…’X’. Passing in
Falsewill cause data to be overwritten if there are duplicate names in the columns.
General parsing configuration¶
- dtypeType name or dict of column -> type, default
None Data type for data or columns. E.g.
{'a': np.float64, 'b': np.int32}(unsupported withengine='python'). Usestrorobjecttogether with suitablena_valuessettings to preserve and not interpret dtype.- engine{
'c','python','pyarrow'} Parser engine to use. The C and pyarrow engines are faster, while the python engine is currently more feature-complete. Multithreading is currently only supported by the pyarrow engine.
New in version 1.4.0: The “pyarrow” engine was added as an experimental engine, and some features are unsupported, or may not work correctly, with this engine.
- convertersdict, default
None Dict of functions for converting values in certain columns. Keys can either be integers or column labels.
- true_valueslist, default
None Values to consider as
True.- false_valueslist, default
None Values to consider as
False.- skipinitialspaceboolean, default
False Skip spaces after delimiter.
- skiprowslist-like or integer, default
None Line numbers to skip (0-indexed) or number of lines to skip (int) at the start of the file.
If callable, the callable function will be evaluated against the row indices, returning True if the row should be skipped and False otherwise:
In [10]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3" In [11]: pd.read_csv(StringIO(data)) Out[11]: col1 col2 col3 0 a b 1 1 a b 2 2 c d 3 In [12]: pd.read_csv(StringIO(data), skiprows=lambda x: x % 2 != 0) Out[12]: col1 col2 col3 0 a b 2
- skipfooterint, default
0 Number of lines at bottom of file to skip (unsupported with engine=’c’).
- nrowsint, default
None Number of rows of file to read. Useful for reading pieces of large files.
- low_memoryboolean, default
True Internally process the file in chunks, resulting in lower memory use while parsing, but possibly mixed type inference. To ensure no mixed types either set
False, or specify the type with thedtypeparameter. Note that the entire file is read into a singleDataFrameregardless, use thechunksizeoriteratorparameter to return the data in chunks. (Only valid with C parser)- memory_mapboolean, default False
If a filepath is provided for
filepath_or_buffer, map the file object directly onto memory and access the data directly from there. Using this option can improve performance because there is no longer any I/O overhead.
NA and missing data handling¶
- na_valuesscalar, str, list-like, or dict, default
None Additional strings to recognize as NA/NaN. If dict passed, specific per-column NA values. See na values const below for a list of the values interpreted as NaN by default.
- keep_default_naboolean, default
True Whether or not to include the default NaN values when parsing the data. Depending on whether
na_valuesis passed in, the behavior is as follows:If
keep_default_naisTrue, andna_valuesare specified,na_valuesis appended to the default NaN values used for parsing.If
keep_default_naisTrue, andna_valuesare not specified, only the default NaN values are used for parsing.If
keep_default_naisFalse, andna_valuesare specified, only the NaN values specifiedna_valuesare used for parsing.If
keep_default_naisFalse, andna_valuesare not specified, no strings will be parsed as NaN.
Note that if
na_filteris passed in asFalse, thekeep_default_naandna_valuesparameters will be ignored.- na_filterboolean, default
True Detect missing value markers (empty strings and the value of na_values). In data without any NAs, passing
na_filter=Falsecan improve the performance of reading a large file.- verboseboolean, default
False Indicate number of NA values placed in non-numeric columns.
- skip_blank_linesboolean, default
True If
True, skip over blank lines rather than interpreting as NaN values.
Datetime handling¶
- parse_datesboolean or list of ints or names or list of lists or dict, default
False. If
True-> try parsing the index.If
[1, 2, 3]-> try parsing columns 1, 2, 3 each as a separate date column.If
[[1, 3]]-> combine columns 1 and 3 and parse as a single date column.If
{'foo': [1, 3]}-> parse columns 1, 3 as date and call result ‘foo’. A fast-path exists for iso8601-formatted dates.
- infer_datetime_formatboolean, default
False If
Trueand parse_dates is enabled for a column, attempt to infer the datetime format to speed up the processing.- keep_date_colboolean, default
False If
Trueand parse_dates specifies combining multiple columns then keep the original columns.- date_parserfunction, default
None Function to use for converting a sequence of string columns to an array of datetime instances. The default uses
dateutil.parser.parserto do the conversion. pandas will try to call date_parser in three different ways, advancing to the next if an exception occurs: 1) Pass one or more arrays (as defined by parse_dates) as arguments; 2) concatenate (row-wise) the string values from the columns defined by parse_dates into a single array and pass that; and 3) call date_parser once for each row using one or more strings (corresponding to the columns defined by parse_dates) as arguments.- dayfirstboolean, default
False DD/MM format dates, international and European format.
- cache_datesboolean, default True
If True, use a cache of unique, converted dates to apply the datetime conversion. May produce significant speed-up when parsing duplicate date strings, especially ones with timezone offsets.
New in version 0.25.0.
Iteration¶
- iteratorboolean, default
False Return
TextFileReaderobject for iteration or getting chunks withget_chunk().- chunksizeint, default
None Return
TextFileReaderobject for iteration. See iterating and chunking below.
Quoting, compression, and file format¶
- compression{
'infer','gzip','bz2','zip','xz','zstd',None,dict}, default'infer' For on-the-fly decompression of on-disk data. If ‘infer’, then use gzip, bz2, zip, xz, or zstandard if
filepath_or_bufferis path-like ending in ‘.gz’, ‘.bz2’, ‘.zip’, ‘.xz’, ‘.zst’, respectively, and no decompression otherwise. If using ‘zip’, the ZIP file must contain only one data file to be read in. Set toNonefor no decompression. Can also be a dict with key'method'set to one of {'zip','gzip','bz2','zstd'} and other key-value pairs are forwarded tozipfile.ZipFile,gzip.GzipFile,bz2.BZ2File, orzstandard.ZstdDecompressor. As an example, the following could be passed for faster compression and to create a reproducible gzip archive:compression={'method': 'gzip', 'compresslevel': 1, 'mtime': 1}.Changed in version 1.1.0: dict option extended to support
gzipandbz2.Changed in version 1.2.0: Previous versions forwarded dict entries for ‘gzip’ to
gzip.open.- thousandsstr, default
None Thousands separator.
- decimalstr, default
'.' Character to recognize as decimal point. E.g. use
','for European data.- float_precisionstring, default None
Specifies which converter the C engine should use for floating-point values. The options are
Nonefor the ordinary converter,highfor the high-precision converter, andround_tripfor the round-trip converter.- lineterminatorstr (length 1), default
None Character to break file into lines. Only valid with C parser.
- quotecharstr (length 1)
The character used to denote the start and end of a quoted item. Quoted items can include the delimiter and it will be ignored.
- quotingint or
csv.QUOTE_*instance, default0 Control field quoting behavior per
csv.QUOTE_*constants. Use one ofQUOTE_MINIMAL(0),QUOTE_ALL(1),QUOTE_NONNUMERIC(2) orQUOTE_NONE(3).- doublequoteboolean, default
True When
quotecharis specified andquotingis notQUOTE_NONE, indicate whether or not to interpret two consecutivequotecharelements inside a field as a singlequotecharelement.- escapecharstr (length 1), default
None One-character string used to escape delimiter when quoting is
QUOTE_NONE.- commentstr, default
None Indicates remainder of line should not be parsed. If found at the beginning of a line, the line will be ignored altogether. This parameter must be a single character. Like empty lines (as long as
skip_blank_lines=True), fully commented lines are ignored by the parameterheaderbut not byskiprows. For example, ifcomment='#', parsing ‘#empty\na,b,c\n1,2,3’ withheader=0will result in ‘a,b,c’ being treated as the header.- encodingstr, default
None Encoding to use for UTF when reading/writing (e.g.
'utf-8'). List of Python standard encodings.- dialectstr or
csv.Dialectinstance, defaultNone If provided, this parameter will override values (default or not) for the following parameters:
delimiter,doublequote,escapechar,skipinitialspace,quotechar, andquoting. If it is necessary to override values, a ParserWarning will be issued. Seecsv.Dialectdocumentation for more details.
Error handling¶
- error_bad_linesboolean, optional, default
None Lines with too many fields (e.g. a csv line with too many commas) will by default cause an exception to be raised, and no
DataFramewill be returned. IfFalse, then these “bad lines” will dropped from theDataFramethat is returned. See bad lines below.Deprecated since version 1.3.0: The
on_bad_linesparameter should be used instead to specify behavior upon encountering a bad line instead.- warn_bad_linesboolean, optional, default
None If error_bad_lines is
False, and warn_bad_lines isTrue, a warning for each “bad line” will be output.Deprecated since version 1.3.0: The
on_bad_linesparameter should be used instead to specify behavior upon encountering a bad line instead.- on_bad_lines(‘error’, ‘warn’, ‘skip’), default ‘error’
Specifies what to do upon encountering a bad line (a line with too many fields). Allowed values are :
‘error’, raise an ParserError when a bad line is encountered.
‘warn’, print a warning when a bad line is encountered and skip that line.
‘skip’, skip bad lines without raising or warning when they are encountered.
New in version 1.3.0.
Specifying column data types¶
You can indicate the data type for the whole DataFrame or individual
columns:
In [13]: import numpy as np
In [14]: data = "a,b,c,d\n1,2,3,4\n5,6,7,8\n9,10,11"
In [15]: print(data)
a,b,c,d
1,2,3,4
5,6,7,8
9,10,11
In [16]: df = pd.read_csv(StringIO(data), dtype=object)
In [17]: df
Out[17]:
a b c d
0 1 2 3 4
1 5 6 7 8
2 9 10 11 NaN
In [18]: df["a"][0]
Out[18]: '1'
In [19]: df = pd.read_csv(StringIO(data), dtype={"b": object, "c": np.float64, "d": "Int64"})
In [20]: df.dtypes
Out[20]:
a int64
b object
c float64
d Int64
dtype: object
Fortunately, pandas offers more than one way to ensure that your column(s)
contain only one dtype. If you’re unfamiliar with these concepts, you can
see here to learn more about dtypes, and
here to learn more about object conversion in
pandas.
For instance, you can use the converters argument
of read_csv():
In [21]: data = "col_1\n1\n2\n'A'\n4.22"
In [22]: df = pd.read_csv(StringIO(data), converters={"col_1": str})
In [23]: df
Out[23]:
col_1
0 1
1 2
2 'A'
3 4.22
In [24]: df["col_1"].apply(type).value_counts()
Out[24]:
<class 'str'> 4
Name: col_1, dtype: int64
Or you can use the to_numeric() function to coerce the
dtypes after reading in the data,
In [25]: df2 = pd.read_csv(StringIO(data))
In [26]: df2["col_1"] = pd.to_numeric(df2["col_1"], errors="coerce")
In [27]: df2
Out[27]:
col_1
0 1.00
1 2.00
2 NaN
3 4.22
In [28]: df2["col_1"].apply(type).value_counts()
Out[28]:
<class 'float'> 4
Name: col_1, dtype: int64
which will convert all valid parsing to floats, leaving the invalid parsing
as NaN.
Ultimately, how you deal with reading in columns containing mixed dtypes
depends on your specific needs. In the case above, if you wanted to NaN out
the data anomalies, then to_numeric() is probably your best option.
However, if you wanted for all the data to be coerced, no matter the type, then
using the converters argument of read_csv() would certainly be
worth trying.
Note
In some cases, reading in abnormal data with columns containing mixed dtypes will result in an inconsistent dataset. If you rely on pandas to infer the dtypes of your columns, the parsing engine will go and infer the dtypes for different chunks of the data, rather than the whole dataset at once. Consequently, you can end up with column(s) with mixed dtypes. For example,
In [29]: col_1 = list(range(500000)) + ["a", "b"] + list(range(500000))
In [30]: df = pd.DataFrame({"col_1": col_1})
In [31]: df.to_csv("foo.csv")
In [32]: mixed_df = pd.read_csv("foo.csv")
In [33]: mixed_df["col_1"].apply(type).value_counts()
Out[33]:
<class 'int'> 737858
<class 'str'> 262144
Name: col_1, dtype: int64
In [34]: mixed_df["col_1"].dtype
Out[34]: dtype('O')
will result with mixed_df containing an int dtype for certain chunks
of the column, and str for others due to the mixed dtypes from the
data that was read in. It is important to note that the overall column will be
marked with a dtype of object, which is used for columns with mixed dtypes.
Specifying categorical dtype¶
Categorical columns can be parsed directly by specifying dtype='category' or
dtype=CategoricalDtype(categories, ordered).
In [35]: data = "col1,col2,col3\na,b,1\na,b,2\nc,d,3"
In [36]: pd.read_csv(StringIO(data))
Out[36]:
col1 col2 col3
0 a b 1
1 a b 2
2 c d 3
In [37]: pd.read_csv(StringIO(data)).dtypes
Out[37]:
col1 object
col2 object
col3 int64
dtype: object
In [38]: pd.read_csv(StringIO(data), dtype="category").dtypes
Out[38]:
col1 category
col2 category
col3 category
dtype: object
Individual columns can be parsed as a Categorical using a dict
specification:
In [39]: pd.read_csv(StringIO(data), dtype={"col1": "category"}).dtypes
Out[39]:
col1 category
col2 object
col3 int64
dtype: object
Specifying dtype='category' will result in an unordered Categorical
whose categories are the unique values observed in the data. For more
control on the categories and order, create a
CategoricalDtype ahead of time, and pass that for
that column’s dtype.
In [40]: from pandas.api.types import CategoricalDtype
In [41]: dtype = CategoricalDtype(["d", "c", "b", "a"], ordered=True)
In [42]: pd.read_csv(StringIO(data), dtype={"col1": dtype}).dtypes
Out[42]:
col1 category
col2 object
col3 int64
dtype: object
When using dtype=CategoricalDtype, “unexpected” values outside of
dtype.categories are treated as missing values.
In [43]: dtype = CategoricalDtype(["a", "b", "d"]) # No 'c'
In [44]: pd.read_csv(StringIO(data), dtype={"col1": dtype}).col1
Out[44]:
0 a
1 a
2 NaN
Name: col1, dtype: category
Categories (3, object): ['a', 'b', 'd']
This matches the behavior of Categorical.set_categories().
Note
With dtype='category', the resulting categories will always be parsed
as strings (object dtype). If the categories are numeric they can be
converted using the to_numeric() function, or as appropriate, another
converter such as to_datetime().
When dtype is a CategoricalDtype with homogeneous categories (
all numeric, all datetimes, etc.), the conversion is done automatically.
In [45]: df = pd.read_csv(StringIO(data), dtype="category")
In [46]: df.dtypes
Out[46]:
col1 category
col2 category
col3 category
dtype: object
In [47]: df["col3"]
Out[47]:
0 1
1 2
2 3
Name: col3, dtype: category
Categories (3, object): ['1', '2', '3']
In [48]: df["col3"].cat.categories = pd.to_numeric(df["col3"].cat.categories)
In [49]: df["col3"]
Out[49]:
0 1
1 2
2 3
Name: col3, dtype: category
Categories (3, int64): [1, 2, 3]
Naming and using columns¶
Handling column names¶
A file may or may not have a header row. pandas assumes the first row should be used as the column names:
In [50]: data = "a,b,c\n1,2,3\n4,5,6\n7,8,9"
In [51]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [52]: pd.read_csv(StringIO(data))
Out[52]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
By specifying the names argument in conjunction with header you can
indicate other names to use and whether or not to throw away the header row (if
any):
In [53]: print(data)
a,b,c
1,2,3
4,5,6
7,8,9
In [54]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=0)
Out[54]:
foo bar baz
0 1 2 3
1 4 5 6
2 7 8 9
In [55]: pd.read_csv(StringIO(data), names=["foo", "bar", "baz"], header=None)
Out[55]:
foo bar baz
0 a b c
1 1 2 3
2 4 5 6
3 7 8 9
If the header is in a row other than the first, pass the row number to
header. This will skip the preceding rows:
In [56]: data = "skip this skip it\na,b,c\n1,2,3\n4,5,6\n7,8,9"
In [57]: pd.read_csv(StringIO(data), header=1)
Out[57]:
a b c
0 1 2 3
1 4 5 6
2 7 8 9
Note
Default behavior is to infer the column names: if no names are
passed the behavior is identical to header=0 and column names
are inferred from the first non-blank line of the file, if column
names are passed explicitly then the behavior is identical to
header=None.
Duplicate names parsing¶
If the file or header contains duplicate names, pandas will by default distinguish between them so as to prevent overwriting data:
In [58]: data = "a,b,a\n0,1,2\n3,4,5"
In [59]: pd.read_csv(StringIO(data))
Out[59]:
a b a.1
0 0 1 2
1 3 4 5
There is no more duplicate data because mangle_dupe_cols=True by default,
which modifies a series of duplicate columns ‘X’, …, ‘X’ to become
‘X’, ‘X.1’, …, ‘X.N’. If mangle_dupe_cols=False, duplicate data can
arise:
In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
Out[3]:
a b a
0 2 1 2
1 5 4 5
To prevent users from encountering this problem with duplicate data, a ValueError
exception is raised if mangle_dupe_cols != True:
In [2]: data = 'a,b,a\n0,1,2\n3,4,5'
In [3]: pd.read_csv(StringIO(data), mangle_dupe_cols=False)
...
ValueError: Setting mangle_dupe_cols=False is not supported yet
Filtering columns (usecols)¶
The usecols argument allows you to select any subset of the columns in a
file, either using the column names, position numbers or a callable:
In [60]: data = "a,b,c,d\n1,2,3,foo\n4,5,6,bar\n7,8,9,baz"
In [61]: pd.read_csv(StringIO(data))
Out[61]:
a b c d
0 1 2 3 foo
1 4 5 6 bar
2 7 8 9 baz
In [62]: pd.read_csv(StringIO(data), usecols=["b", "d"])
Out[62]:
b d
0 2 foo
1 5 bar
2 8 baz
In [63]: pd.read_csv(StringIO(data), usecols=[0, 2, 3])
Out[63]:
a c d
0 1 3 foo
1 4 6 bar
2 7 9 baz
In [64]: pd.read_csv(StringIO(data), usecols=lambda x: x.upper() in ["A", "C"])
Out[64]:
a c
0 1 3
1 4 6
2 7 9
The usecols argument can also be used to specify which columns not to
use in the final result:
In [65]: pd.read_csv(StringIO(data), usecols=lambda x: x not in ["a", "c"])
Out[65]:
b d
0 2 foo
1 5 bar
2 8 baz
In this case, the callable is specifying that we exclude the “a” and “c” columns from the output.
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 [83]: from io import BytesIO
In [84]: data = b"word,length\n" b"Tr\xc3\xa4umen,7\n" b"Gr\xc3\xbc\xc3\x9fe,5"
In [85]: data = data.decode("utf8").encode("latin-1")
In [86]: df = pd.read_csv(BytesIO(data), encoding="latin-1")
In [87]: df
Out[87]:
word length
0 Träumen 7
1 Grüße 5
In [88]: df["word"][1]
Out[88]: 'Grüße'
Some formats which encode all characters as multiple bytes, like UTF-16, won’t parse correctly at all without specifying the encoding. Full list of Python standard encodings.
Index columns and trailing delimiters¶
If a file has one more column of data than the number of column names, the
first column will be used as the DataFrame’s row names:
In [89]: data = "a,b,c\n4,apple,bat,5.7\n8,orange,cow,10"
In [90]: pd.read_csv(StringIO(data))
Out[90]:
a b c
4 apple bat 5.7
8 orange cow 10.0
In [91]: data = "index,a,b,c\n4,apple,bat,5.7\n8,orange,cow,10"
In [92]: pd.read_csv(StringIO(data), index_col=0)
Out[92]:
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 [93]: data = "a,b,c\n4,apple,bat,\n8,orange,cow,"
In [94]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,
In [95]: pd.read_csv(StringIO(data))
Out[95]:
a b c
4 apple bat NaN
8 orange cow NaN
In [96]: pd.read_csv(StringIO(data), index_col=False)
Out[96]:
a b c
0 4 apple bat
1 8 orange cow
If a subset of data is being parsed using the usecols option, the
index_col specification is based on that subset, not the original data.
In [97]: data = "a,b,c\n4,apple,bat,\n8,orange,cow,"
In [98]: print(data)
a,b,c
4,apple,bat,
8,orange,cow,
In [99]: pd.read_csv(StringIO(data), usecols=["b", "c"])
Out[99]:
b c
4 bat NaN
8 cow NaN
In [100]: pd.read_csv(StringIO(data), usecols=["b", "c"], index_col=0)
Out[100]:
b c
4 bat NaN
8 cow NaN
Date Handling¶
Specifying date columns¶
To better facilitate working with datetime data, read_csv()
uses the keyword arguments parse_dates and date_parser
to allow users to specify a variety of columns and date/time formats to turn the
input text data into datetime objects.
The simplest case is to just pass in parse_dates=True:
In [101]: f = open("foo.csv", "w")
In [102]: f.write("date,A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5")
Out[102]: 55
In [103]: f.close()
# Use a column as an index, and parse it as dates.
In [104]: df = pd.read_csv("foo.csv", index_col=0, parse_dates=True)
In [105]: df
Out[105]:
A B C
date
2009-01-01 a 1 2
2009-01-02 b 3 4
2009-01-03 c 4 5
# These are Python datetime objects
In [106]: df.index
Out[106]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', name='date', freq=None)
It is often the case that we may want to store date and time data separately,
or store various date fields separately. the parse_dates keyword can be
used to specify a combination of columns to parse the dates and/or times from.
You can specify a list of column lists to parse_dates, the resulting date
columns will be prepended to the output (so as to not affect the existing column
order) and the new column names will be the concatenation of the component
column names:
In [107]: data = (
.....: "KORD,19990127, 19:00:00, 18:56:00, 0.8100\n"
.....: "KORD,19990127, 20:00:00, 19:56:00, 0.0100\n"
.....: "KORD,19990127, 21:00:00, 20:56:00, -0.5900\n"
.....: "KORD,19990127, 21:00:00, 21:18:00, -0.9900\n"
.....: "KORD,19990127, 22:00:00, 21:56:00, -0.5900\n"
.....: "KORD,19990127, 23:00:00, 22:56:00, -0.5900"
.....: )
.....:
In [108]: with open("tmp.csv", "w") as fh:
.....: fh.write(data)
.....:
In [109]: df = pd.read_csv("tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]])
In [110]: df
Out[110]:
1_2 1_3 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59
By default the parser removes the component date columns, but you can choose
to retain them via the keep_date_col keyword:
In [111]: df = pd.read_csv(
.....: "tmp.csv", header=None, parse_dates=[[1, 2], [1, 3]], keep_date_col=True
.....: )
.....:
In [112]: df
Out[112]:
1_2 1_3 0 1 2 3 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 19990127 19:00:00 18:56:00 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 19990127 20:00:00 19:56:00 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD 19990127 21:00:00 20:56:00 -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD 19990127 21:00:00 21:18:00 -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD 19990127 22:00:00 21:56:00 -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD 19990127 23:00:00 22:56:00 -0.59
Note that if you wish to combine multiple columns into a single date column, a
nested list must be used. In other words, parse_dates=[1, 2] indicates that
the second and third columns should each be parsed as separate date columns
while parse_dates=[[1, 2]] means the two columns should be parsed into a
single column.
You can also use a dict to specify custom name columns:
In [113]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}
In [114]: df = pd.read_csv("tmp.csv", header=None, parse_dates=date_spec)
In [115]: df
Out[115]:
nominal actual 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59
It is important to remember that if multiple text columns are to be parsed into
a single date column, then a new column is prepended to the data. The index_col
specification is based off of this new set of columns rather than the original
data columns:
In [116]: date_spec = {"nominal": [1, 2], "actual": [1, 3]}
In [117]: df = pd.read_csv(
.....: "tmp.csv", header=None, parse_dates=date_spec, index_col=0
.....: ) # index is the nominal column
.....:
In [118]: df
Out[118]:
actual 0 4
nominal
1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59
Note
If a column or index contains an unparsable date, the entire column or
index will be returned unaltered as an object data type. For non-standard
datetime parsing, use to_datetime() after pd.read_csv.
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.
Date parsing functions¶
Finally, the parser allows you to specify a custom date_parser function to
take full advantage of the flexibility of the date parsing API:
In [119]: df = pd.read_csv(
.....: "tmp.csv", header=None, parse_dates=date_spec, date_parser=pd.to_datetime
.....: )
.....:
In [120]: df
Out[120]:
nominal actual 0 4
0 1999-01-27 19:00:00 1999-01-27 18:56:00 KORD 0.81
1 1999-01-27 20:00:00 1999-01-27 19:56:00 KORD 0.01
2 1999-01-27 21:00:00 1999-01-27 20:56:00 KORD -0.59
3 1999-01-27 21:00:00 1999-01-27 21:18:00 KORD -0.99
4 1999-01-27 22:00:00 1999-01-27 21:56:00 KORD -0.59
5 1999-01-27 23:00:00 1999-01-27 22:56:00 KORD -0.59
pandas will try to call the date_parser function in three different ways. If
an exception is raised, the next one is tried:
date_parseris first called with one or more arrays as arguments, as defined usingparse_dates(e.g.,date_parser(['2013', '2013'], ['1', '2'])).If #1 fails,
date_parseris called with all the columns concatenated row-wise into a single array (e.g.,date_parser(['2013 1', '2013 2'])).
Note that performance-wise, you should try these methods of parsing dates in order:
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_parserfunction. For optimal performance, this should be vectorized, i.e., it should accept arrays as arguments.
Parsing a CSV with mixed timezones¶
pandas cannot natively represent a column or index with mixed timezones. If your CSV
file contains columns with a mixture of timezones, the default result will be
an object-dtype column with strings, even with parse_dates.
In [121]: content = """\
.....: a
.....: 2000-01-01T00:00:00+05:00
.....: 2000-01-01T00:00:00+06:00"""
.....:
In [122]: df = pd.read_csv(StringIO(content), parse_dates=["a"])
In [123]: df["a"]
Out[123]:
0 2000-01-01 00:00:00+05:00
1 2000-01-01 00:00:00+06:00
Name: a, dtype: object
To parse the mixed-timezone values as a datetime column, pass a partially-applied
to_datetime() with utc=True as the date_parser.
In [124]: df = pd.read_csv(
.....: StringIO(content),
.....: parse_dates=["a"],
.....: date_parser=lambda col: pd.to_datetime(col, utc=True),
.....: )
.....:
In [125]: df["a"]
Out[125]:
0 1999-12-31 19:00:00+00:00
1 1999-12-31 18:00:00+00:00
Name: a, dtype: datetime64[ns, UTC]
Inferring datetime format¶
If you have parse_dates enabled for some or all of your columns, and your
datetime strings are all formatted the same way, you may get a large speed
up by setting infer_datetime_format=True. If set, pandas will attempt
to guess the format of your datetime strings, and then use a faster means
of parsing the strings. 5-10x parsing speeds have been observed. pandas
will fallback to the usual parsing if either the format cannot be guessed
or the format that was guessed cannot properly parse the entire column
of strings. So in general, infer_datetime_format should not have any
negative consequences if enabled.
Here are some examples of datetime strings that can be guessed (All representing December 30th, 2011 at 00:00:00):
“20111230”
“2011/12/30”
“20111230 00:00:00”
“12/30/2011 00:00:00”
“30/Dec/2011 00:00:00”
“30/December/2011 00:00:00”
Note that infer_datetime_format is sensitive to dayfirst. With
dayfirst=True, it will guess “01/12/2011” to be December 1st. With
dayfirst=False (default) it will guess “01/12/2011” to be January 12th.
# Try to infer the format for the index column
In [126]: df = pd.read_csv(
.....: "foo.csv",
.....: index_col=0,
.....: parse_dates=True,
.....: infer_datetime_format=True,
.....: )
.....:
In [127]: df
Out[127]:
A B C
date
2009-01-01 a 1 2
2009-01-02 b 3 4
2009-01-03 c 4 5
International date formats¶
While US date formats tend to be MM/DD/YYYY, many international formats use
DD/MM/YYYY instead. For convenience, a dayfirst keyword is provided:
In [128]: data = "date,value,cat\n1/6/2000,5,a\n2/6/2000,10,b\n3/6/2000,15,c"
In [129]: print(data)
date,value,cat
1/6/2000,5,a
2/6/2000,10,b
3/6/2000,15,c
In [130]: with open("tmp.csv", "w") as fh:
.....: fh.write(data)
.....:
In [131]: pd.read_csv("tmp.csv", parse_dates=[0])
Out[131]:
date value cat
0 2000-01-06 5 a
1 2000-02-06 10 b
2 2000-03-06 15 c
In [132]: pd.read_csv("tmp.csv", dayfirst=True, parse_dates=[0])
Out[132]:
date value cat
0 2000-06-01 5 a
1 2000-06-02 10 b
2 2000-06-03 15 c
Writing CSVs to binary file objects¶
New in version 1.2.0.
df.to_csv(..., mode="wb") allows writing a CSV to a file object
opened binary mode. In most cases, it is not necessary to specify
mode as Pandas will auto-detect whether the file object is
opened in text or binary mode.
In [133]: import io
In [134]: data = pd.DataFrame([0, 1, 2])
In [135]: buffer = io.BytesIO()
In [136]: data.to_csv(buffer, encoding="utf-8", compression="gzip")
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 [137]: val = "0.3066101993807095471566981359501369297504425048828125"
In [138]: data = "a,b,c\n1,2,{0}".format(val)
In [139]: abs(
.....: pd.read_csv(
.....: StringIO(data),
.....: engine="c",
.....: float_precision=None,
.....: )["c"][0] - float(val)
.....: )
.....:
Out[139]: 5.551115123125783e-17
In [140]: abs(
.....: pd.read_csv(
.....: StringIO(data),
.....: engine="c",
.....: float_precision="high",
.....: )["c"][0] - float(val)
.....: )
.....:
Out[140]: 5.551115123125783e-17
In [141]: abs(
.....: pd.read_csv(StringIO(data), engine="c", float_precision="round_trip")["c"][0]
.....: - float(val)
.....: )
.....:
Out[141]: 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 [142]: data = (
.....: "ID|level|category\n"
.....: "Patient1|123,000|x\n"
.....: "Patient2|23,000|y\n"
.....: "Patient3|1,234,018|z"
.....: )
.....:
In [143]: with open("tmp.csv", "w") as fh:
.....: fh.write(data)
.....:
In [144]: df = pd.read_csv("tmp.csv", sep="|")
In [145]: df
Out[145]:
ID level category
0 Patient1 123,000 x
1 Patient2 23,000 y
2 Patient3 1,234,018 z
In [146]: df.level.dtype
Out[146]: dtype('O')
The thousands keyword allows integers to be parsed correctly:
In [147]: df = pd.read_csv("tmp.csv", sep="|", thousands=",")
In [148]: df
Out[148]:
ID level category
0 Patient1 123000 x
1 Patient2 23000 y
2 Patient3 1234018 z
In [149]: df.level.dtype
Out[149]: dtype('int64')
NA values¶
To control which values are parsed as missing values (which are signified by
NaN), specify a string in na_values. If you specify a list of strings,
then all values in it are considered to be missing values. If you specify a
number (a float, like 5.0 or an integer like 5), the
corresponding equivalent values will also imply a missing value (in this case
effectively [5.0, 5] are recognized as NaN).
To completely override the default values that are recognized as missing, specify keep_default_na=False.
Let us consider some examples:
pd.read_csv("path_to_file.csv", na_values=[5])
In the example above 5 and 5.0 will be recognized as NaN, in
addition to the defaults. A string will first be interpreted as a numerical
5, then as a NaN.
pd.read_csv("path_to_file.csv", keep_default_na=False, na_values=[""])
Above, only an empty field will be recognized as NaN.
pd.read_csv("path_to_file.csv", keep_default_na=False, na_values=["NA", "0"])
Above, both NA and 0 as strings are NaN.
pd.read_csv("path_to_file.csv", na_values=["Nope"])
The default values, in addition to the string "Nope" are recognized as
NaN.
Infinity¶
inf like values will be parsed as np.inf (positive infinity), and -inf as -np.inf (negative infinity).
These will ignore the case of the value, meaning Inf, will also be parsed as np.inf.
Returning Series¶
Using the squeeze keyword, the parser will return output with a single column
as a Series:
Deprecated since version 1.4.0: Users should append .squeeze("columns") to the DataFrame returned by
read_csv instead.
In [150]: data = "level\nPatient1,123000\nPatient2,23000\nPatient3,1234018"
In [151]: with open("tmp.csv", "w") as fh:
.....: fh.write(data)
.....:
In [152]: print(open("tmp.csv").read())
level
Patient1,123000
Patient2,23000
Patient3,1234018
In [153]: output = pd.read_csv("tmp.csv", squeeze=True)
In [154]: output
Out[154]:
Patient1 123000
Patient2 23000
Patient3 1234018
Name: level, dtype: int64
In [155]: type(output)
Out[155]: pandas.core.series.Series
Boolean values¶
The common values True, False, TRUE, and FALSE are all
recognized as boolean. Occasionally you might want to recognize other values
as being boolean. To do this, use the true_values and false_values
options as follows:
In [156]: data = "a,b,c\n1,Yes,2\n3,No,4"
In [157]: print(data)
a,b,c
1,Yes,2
3,No,4
In [158]: pd.read_csv(StringIO(data))
Out[158]:
a b c
0 1 Yes 2
1 3 No 4
In [159]: pd.read_csv(StringIO(data), true_values=["Yes"], false_values=["No"])
Out[159]:
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 fields will raise an error by default:
In [160]: data = "a,b,c\n1,2,3\n4,5,6,7\n8,9,10"
In [161]: pd.read_csv(StringIO(data))
---------------------------------------------------------------------------
ParserError Traceback (most recent call last)
Input In [161], in <cell line: 1>()
----> 1 pd.read_csv(StringIO(data))
File /pandas/pandas/util/_decorators.py:311, in deprecate_nonkeyword_arguments.<locals>.decorate.<locals>.wrapper(*args, **kwargs)
305 if len(args) > num_allow_args:
306 warnings.warn(
307 msg.format(arguments=arguments),
308 FutureWarning,
309 stacklevel=stacklevel,
310 )
--> 311 return func(*args, **kwargs)
File /pandas/pandas/io/parsers/readers.py:678, in read_csv(filepath_or_buffer, sep, delimiter, header, names, index_col, usecols, squeeze, prefix, mangle_dupe_cols, dtype, engine, converters, true_values, false_values, skipinitialspace, skiprows, skipfooter, nrows, na_values, keep_default_na, na_filter, verbose, skip_blank_lines, parse_dates, infer_datetime_format, keep_date_col, date_parser, dayfirst, cache_dates, iterator, chunksize, compression, thousands, decimal, lineterminator, quotechar, quoting, doublequote, escapechar, comment, encoding, encoding_errors, dialect, error_bad_lines, warn_bad_lines, on_bad_lines, delim_whitespace, low_memory, memory_map, float_precision, storage_options)
663 kwds_defaults = _refine_defaults_read(
664 dialect,
665 delimiter,
(...)
674 defaults={"delimiter": ","},
675 )
676 kwds.update(kwds_defaults)
--> 678 return _read(filepath_or_buffer, kwds)
File /pandas/pandas/io/parsers/readers.py:581, in _read(filepath_or_buffer, kwds)
578 return parser
580 with parser:
--> 581 return parser.read(nrows)
File /pandas/pandas/io/parsers/readers.py:1253, in TextFileReader.read(self, nrows)
1251 nrows = validate_integer("nrows", nrows)
1252 try:
-> 1253 index, columns, col_dict = self._engine.read(nrows)
1254 except Exception:
1255 self.close()
File /pandas/pandas/io/parsers/c_parser_wrapper.py:225, in CParserWrapper.read(self, nrows)
223 try:
224 if self.low_memory:
--> 225 chunks = self._reader.read_low_memory(nrows)
226 # destructive to chunks
227 data = _concatenate_chunks(chunks)
File /pandas/pandas/_libs/parsers.pyx:805, in pandas._libs.parsers.TextReader.read_low_memory()
File /pandas/pandas/_libs/parsers.pyx:861, in pandas._libs.parsers.TextReader._read_rows()
File /pandas/pandas/_libs/parsers.pyx:847, in pandas._libs.parsers.TextReader._tokenize_rows()
File /pandas/pandas/_libs/parsers.pyx:1960, in pandas._libs.parsers.raise_parser_error()
ParserError: Error tokenizing data. C error: Expected 3 fields in line 3, saw 4
You can elect to skip bad lines:
In [29]: pd.read_csv(StringIO(data), on_bad_lines="warn")
Skipping line 3: expected 3 fields, saw 4
Out[29]:
a b c
0 1 2 3
1 8 9 10
Or pass a callable function to handle the bad line if engine="python".
The bad line will be a list of strings that was split by the sep:
In [29]: external_list = []
In [30]: def bad_lines_func(line):
...: external_list.append(line)
...: return line[-3:]
In [31]: pd.read_csv(StringIO(data), on_bad_lines=bad_lines_func, engine="python")
Out[31]:
a b c
0 1 2 3
1 5 6 7
2 8 9 10
In [32]: external_list
Out[32]: [4, 5, 6, 7]
.. versionadded:: 1.4.0
You can also use the usecols parameter to eliminate extraneous column
data that appear in some lines but not others:
In [33]: pd.read_csv(StringIO(data), usecols=[0, 1, 2])
Out[33]:
a b c
0 1 2 3
1 4 5 6
2 8 9 10
In case you want to keep all data including the lines with too many fields, you can
specify a sufficient number of names. This ensures that lines with not enough
fields are filled with NaN.
In [34]: pd.read_csv(StringIO(data), names=['a', 'b', 'c', 'd'])
Out[34]:
a b c d
0 1 2 3 NaN
1 4 5 6 7
2 8 9 10 NaN
Dialect¶
The dialect keyword gives greater flexibility in specifying the file format.
By default it uses the Excel dialect but you can specify either the dialect name
or a csv.Dialect instance.
Suppose you had data with unenclosed quotes:
In [162]: data = "label1,label2,label3\n" 'index1,"a,c,e\n' "index2,b,d,f"
In [163]: 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 [164]: import csv
In [165]: dia = csv.excel()
In [166]: dia.quoting = csv.QUOTE_NONE
In [167]: pd.read_csv(StringIO(data), dialect=dia)
Out[167]:
label1 label2 label3
index1 "a c e
index2 b d f
All of the dialect options can be specified separately by keyword arguments:
In [168]: data = "a,b,c~1,2,3~4,5,6"
In [169]: pd.read_csv(StringIO(data), lineterminator="~")
Out[169]:
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 [170]: data = "a, b, c\n1, 2, 3\n4, 5, 6"
In [171]: print(data)
a, b, c
1, 2, 3
4, 5, 6
In [172]: pd.read_csv(StringIO(data), skipinitialspace=True)
Out[172]:
a b c
0 1 2 3
1 4 5 6
The parsers make every attempt to “do the right thing” and not be fragile. Type inference is a pretty big deal. If a column can be coerced to integer dtype without altering the contents, the parser will do so. Any non-numeric columns will come through as object dtype as with the rest of pandas objects.
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 [173]: data = 'a,b\n"hello, \\"Bob\\", nice to see you",5'
In [174]: print(data)
a,b
"hello, \"Bob\", nice to see you",5
In [175]: pd.read_csv(StringIO(data), escapechar="\\")
Out[175]:
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, and
a different usage of the delimiter parameter:
colspecs: A list of pairs (tuples) giving the extents of the fixed-width fields of each line as half-open intervals (i.e., [from, to[ ). String value ‘infer’ can be used to instruct the parser to try detecting the column specifications from the first 100 rows of the data. Default behavior, if not specified, is to infer.widths: A list of field widths which can be used instead of ‘colspecs’ if the intervals are contiguous.delimiter: Characters to consider as filler characters in the fixed-width file. Can be used to specify the filler character of the fields if it is not spaces (e.g., ‘~’).
Consider a typical fixed-width data file:
In [176]: f = open("bar.csv", "w")
In [177]: data1 = (
.....: "id8141 360.242940 149.910199 11950.7\n"
.....: "id1594 444.953632 166.985655 11788.4\n"
.....: "id1849 364.136849 183.628767 11806.2\n"
.....: "id1230 413.836124 184.375703 11916.8\n"
.....: "id1948 502.953953 173.237159 12468.3"
.....: )
.....:
In [178]: f.write(data1)
Out[178]: 219
In [179]: f.close()
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 [180]: colspecs = [(0, 6), (8, 20), (21, 33), (34, 43)]
In [181]: df = pd.read_fwf("bar.csv", colspecs=colspecs, header=None, index_col=0)
In [182]: df
Out[182]:
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 [183]: widths = [6, 14, 13, 10]
In [184]: df = pd.read_fwf("bar.csv", widths=widths, header=None)
In [185]: df
Out[185]:
0 1 2 3
0 id8141 360.242940 149.910199 11950.7
1 id1594 444.953632 166.985655 11788.4
2 id1849 364.136849 183.628767 11806.2
3 id1230 413.836124 184.375703 11916.8
4 id1948 502.953953 173.237159 12468.3
The parser will take care of extra white spaces around the columns so it’s ok to have extra separation between the columns in the file.
By default, read_fwf will try to infer the file’s colspecs by using the
first 100 rows of the file. It can do it only in cases when the columns are
aligned and correctly separated by the provided delimiter (default delimiter
is whitespace).
In [186]: df = pd.read_fwf("bar.csv", header=None, index_col=0)
In [187]: df
Out[187]:
1 2 3
0
id8141 360.242940 149.910199 11950.7
id1594 444.953632 166.985655 11788.4
id1849 364.136849 183.628767 11806.2
id1230 413.836124 184.375703 11916.8
id1948 502.953953 173.237159 12468.3
read_fwf supports the dtype parameter for specifying the types of
parsed columns to be different from the inferred type.
In [188]: pd.read_fwf("bar.csv", header=None, index_col=0).dtypes
Out[188]:
1 float64
2 float64
3 float64
dtype: object
In [189]: pd.read_fwf("bar.csv", header=None, dtype={2: "object"}).dtypes
Out[189]:
0 object
1 float64
2 object
3 float64
dtype: object
Indexes¶
Files with an “implicit” index column¶
Consider a file with one less entry in the header than the number of data column:
In [190]: data = "A,B,C\n20090101,a,1,2\n20090102,b,3,4\n20090103,c,4,5"
In [191]: print(data)
A,B,C
20090101,a,1,2
20090102,b,3,4
20090103,c,4,5
In [192]: with open("foo.csv", "w") as f:
.....: f.write(data)
.....:
In this special case, read_csv assumes that the first column is to be used
as the index of the DataFrame:
In [193]: pd.read_csv("foo.csv")
Out[193]:
A B C
20090101 a 1 2
20090102 b 3 4
20090103 c 4 5
Note that the dates weren’t automatically parsed. In that case you would need to do as before:
In [194]: df = pd.read_csv("foo.csv", parse_dates=True)
In [195]: df.index
Out[195]: DatetimeIndex(['2009-01-01', '2009-01-02', '2009-01-03'], dtype='datetime64[ns]', freq=None)
Reading an index with a MultiIndex¶
Suppose you have data indexed by two columns:
In [196]: data = 'year,indiv,zit,xit\n1977,"A",1.2,.6\n1977,"B",1.5,.5'
In [197]: print(data)
year,indiv,zit,xit
1977,"A",1.2,.6
1977,"B",1.5,.5
In [198]: with open("mindex_ex.csv", mode="w") as f:
.....: f.write(data)
.....:
The index_col argument to read_csv can take a list of
column numbers to turn multiple columns into a MultiIndex for the index of the
returned object:
In [199]: df = pd.read_csv("mindex_ex.csv", index_col=[0, 1])
In [200]: df
Out[200]:
zit xit
year indiv
1977 A 1.2 0.6
B 1.5 0.5
In [201]: df.loc[1977]
Out[201]:
zit xit
indiv
A 1.2 0.6
B 1.5 0.5
Reading columns with a MultiIndex¶
By specifying list of row locations for the header argument, you
can read in a MultiIndex for the columns. Specifying non-consecutive
rows will skip the intervening rows.
In [202]: from pandas._testing import makeCustomDataframe as mkdf
In [203]: df = mkdf(5, 3, r_idx_nlevels=2, c_idx_nlevels=4)
In [204]: df.to_csv("mi.csv")
In [205]: print(open("mi.csv").read())
C0,,C_l0_g0,C_l0_g1,C_l0_g2
C1,,C_l1_g0,C_l1_g1,C_l1_g2
C2,,C_l2_g0,C_l2_g1,C_l2_g2
C3,,C_l3_g0,C_l3_g1,C_l3_g2
R0,R1,,,
R_l0_g0,R_l1_g0,R0C0,R0C1,R0C2
R_l0_g1,R_l1_g1,R1C0,R1C1,R1C2
R_l0_g2,R_l1_g2,R2C0,R2C1,R2C2
R_l0_g3,R_l1_g3,R3C0,R3C1,R3C2
R_l0_g4,R_l1_g4,R4C0,R4C1,R4C2
In [206]: pd.read_csv("mi.csv", header=[0, 1, 2, 3], index_col=[0, 1])
Out[206]:
C0 C_l0_g0 C_l0_g1 C_l0_g2
C1 C_l1_g0 C_l1_g1 C_l1_g2
C2 C_l2_g0 C_l2_g1 C_l2_g2
C3 C_l3_g0 C_l3_g1 C_l3_g2
R0 R1
R_l0_g0 R_l1_g0 R0C0 R0C1 R0C2
R_l0_g1 R_l1_g1 R1C0 R1C1 R1C2
R_l0_g2 R_l1_g2 R2C0 R2C1 R2C2
R_l0_g3 R_l1_g3 R3C0 R3C1 R3C2
R_l0_g4 R_l1_g4 R4C0 R4C1 R4C2
read_csv is also able to interpret a more common format
of multi-columns indices.
In [207]: data = ",a,a,a,b,c,c\n,q,r,s,t,u,v\none,1,2,3,4,5,6\ntwo,7,8,9,10,11,12"
In [208]: print(data)
,a,a,a,b,c,c
,q,r,s,t,u,v
one,1,2,3,4,5,6
two,7,8,9,10,11,12
In [209]: with open("mi2.csv", "w") as fh:
.....: fh.write(data)
.....:
In [210]: pd.read_csv("mi2.csv", header=[0, 1], index_col=0)
Out[210]:
a b c
q r s t u v
one 1 2 3 4 5 6
two 7 8 9 10 11 12
Note: If an index_col is not specified (e.g. you don’t have an index, or wrote it
with df.to_csv(..., index=False), then any names on the columns index will be lost.
Automatically “sniffing” the delimiter¶
read_csv is capable of inferring delimited (not necessarily
comma-separated) files, as pandas uses the csv.Sniffer
class of the csv module. For this, you have to specify sep=None.
In [211]: df = pd.DataFrame(np.random.randn(10, 4))
In [212]: df.to_csv("tmp.csv", sep="|")
In [213]: df.to_csv("tmp2.csv", sep=":")
In [214]: pd.read_csv("tmp2.csv", sep=None, engine="python")
Out[214]:
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
Reading multiple files to create a single DataFrame¶
It’s best to use concat() to combine multiple files.
See the cookbook for an example.
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 [215]: df = pd.DataFrame(np.random.randn(10, 4))
In [216]: df.to_csv("tmp.csv", sep="|")
In [217]: table = pd.read_csv("tmp.csv", sep="|")
In [218]: table
Out[218]:
Unnamed: 0 0 1 2 3
0 0 -1.294524 0.413738 0.276662 -0.472035
1 1 -0.013960 -0.362543 -0.006154 -0.923061
2 2 0.895717 0.805244 -1.206412 2.565646
3 3 1.431256 1.340309 -1.170299 -0.226169
4 4 0.410835 0.813850 0.132003 -0.827317
5 5 -0.076467 -1.187678 1.130127 -1.436737
6 6 -1.413681 1.607920 1.024180 0.569605
7 7 0.875906 -2.211372 0.974466 -2.006747
8 8 -0.410001 -0.078638 0.545952 -1.219217
9 9 -1.226825 0.769804 -1.281247 -0.727707
By specifying a chunksize to read_csv, the return
value will be an iterable object of type TextFileReader:
In [219]: with pd.read_csv("tmp.csv", sep="|", chunksize=4) as reader:
.....: reader
.....: for chunk in reader:
.....: print(chunk)
.....:
Unnamed: 0 0 1 2 3
0 0 -1.294524 0.413738 0.276662 -0.472035
1 1 -0.013960 -0.362543 -0.006154 -0.923061
2 2 0.895717 0.805244 -1.206412 2.565646
3 3 1.431256 1.340309 -1.170299 -0.226169
Unnamed: 0 0 1 2 3
4 4 0.410835 0.813850 0.132003 -0.827317
5 5 -0.076467 -1.187678 1.130127 -1.436737
6 6 -1.413681 1.607920 1.024180 0.569605
7 7 0.875906 -2.211372 0.974466 -2.006747
Unnamed: 0 0 1 2 3
8 8 -0.410001 -0.078638 0.545952 -1.219217
9 9 -1.226825 0.769804 -1.281247 -0.727707
Changed in version 1.2: read_csv/json/sas return a context-manager when iterating through a file.
Specifying iterator=True will also return the TextFileReader object:
In [220]: with pd.read_csv("tmp.csv", sep="|", iterator=True) as reader:
.....: reader.get_chunk(5)
.....:
Specifying the parser engine¶
Pandas currently supports three engines, the C engine, the python engine, and an experimental
pyarrow engine (requires the pyarrow package). In general, the pyarrow engine is fastest
on larger workloads and is equivalent in speed to the C engine on most other workloads.
The python engine tends to be slower than the pyarrow and C engines on most workloads. However,
the pyarrow engine is much less robust than the C engine, which lacks a few features compared to the
Python engine.
Where possible, pandas uses the C parser (specified as engine='c'), but it may fall
back to Python if C-unsupported options are specified.
Currently, options unsupported by the C and pyarrow engines include:
sepother than a single character (e.g. regex separators)skipfootersep=Nonewithdelim_whitespace=False
Specifying any of the above options will produce a ParserWarning unless the
python engine is selected explicitly using engine='python'.
Options that are unsupported by the pyarrow engine which are not covered by the list above include:
float_precisionchunksizecommentnrowsthousandsmemory_mapdialectwarn_bad_lineserror_bad_lineson_bad_linesdelim_whitespacequotinglineterminatorconvertersdecimaliteratordayfirstinfer_datetime_formatverboseskipinitialspacelow_memory
Specifying these options with engine='pyarrow' will raise a ValueError.
Reading/writing remote files¶
You can pass in a URL to read or write remote files to many of pandas’ IO functions - the following example shows reading a CSV file:
df = pd.read_csv("https://download.bls.gov/pub/time.series/cu/cu.item", sep="\t")
New in version 1.3.0.
A custom header can be sent alongside HTTP(s) requests by passing a dictionary
of header key value mappings to the storage_options keyword argument as shown below:
headers = {"User-Agent": "pandas"}
df = pd.read_csv(
"https://download.bls.gov/pub/time.series/cu/cu.item",
sep="\t",
storage_options=headers
)
All URLs which are not local files or HTTP(s) are handled by fsspec, if installed, and its various filesystem implementations (including Amazon S3, Google Cloud, SSH, FTP, webHDFS…). Some of these implementations will require additional packages to be installed, for example S3 URLs require the s3fs library:
df = pd.read_json("s3://pandas-test/adatafile.json")
When dealing with remote storage systems, you might need
extra configuration with environment variables or config files in
special locations. For example, to access data in your S3 bucket,
you will need to define credentials in one of the several ways listed in
the S3Fs documentation. The same is true
for several of the storage backends, and you should follow the links
at fsimpl1 for implementations built into fsspec and fsimpl2
for those not included in the main fsspec
distribution.
You can also pass parameters directly to the backend driver. For example, if you do not have S3 credentials, you can still access public data by specifying an anonymous connection, such as
New in version 1.2.0.
pd.read_csv(
"s3://ncei-wcsd-archive/data/processed/SH1305/18kHz/SaKe2013"
"-D20130523-T080854_to_SaKe2013-D20130523-T085643.csv",
storage_options={"anon": True},
)
fsspec also allows complex URLs, for accessing data in compressed
archives, local caching of files, and more. To locally cache the above
example, you would modify the call to
pd.read_csv(
"simplecache::s3://ncei-wcsd-archive/data/processed/SH1305/18kHz/"
"SaKe2013-D20130523-T080854_to_SaKe2013-D20130523-T085643.csv",
storage_options={"s3": {"anon": True}},
)
where we specify that the “anon” parameter is meant for the “s3” part of the implementation, not to the caching implementation. Note that this caches to a temporary directory for the duration of the session only, but you can also specify a permanent store.
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 file object. If a file object it must be opened withnewline=''sep: Field delimiter for the output file (default “,”)na_rep: A string representation of a missing value (default ‘’)float_format: Format string for floating point numberscolumns: 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), andheaderandindexare True, then the index names are used. (A sequence should be given if theDataFrameuses 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 (defaultos.linesep)quoting: Set quoting rules as in csv module (default csv.QUOTE_MINIMAL). Note that if you have set afloat_formatthen floats are converted to strings and csv.QUOTE_NONNUMERIC will treat them as non-numericquotechar: Character used to quote fields (default ‘”’)doublequote: Control quoting ofquotecharin fields (default True)escapechar: Character used to escapesepandquotecharwhen appropriate (default None)chunksize: Number of rows to write at a timedate_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:
bufdefault None, for example a StringIO objectcolumnsdefault None, which columns to writecol_spacedefault None, minimum width of each column.na_repdefaultNaN, representation of NA valueformattersdefault None, a dictionary (by column) of functions each of which takes a single argument and returns a formatted stringfloat_formatdefault None, a function which takes a single (float) argument and returns a formatted string; to be applied to floats in theDataFrame.sparsifydefault True, set to False for aDataFramewith a hierarchical index to print every MultiIndex key at each row.index_namesdefault True, will print the names of the indicesindexdefault True, will print the index (ie, row labels)headerdefault True, will print the column labelsjustifydefaultleft, 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 beNonein which case a JSON string is returnedorient:Series:default is
indexallowed values are {
split,records,index}
DataFrame:default is
columnsallowed values are {
split,records,index,columns,values,table}
The format of the JSON string
splitdict like {index -> [index], columns -> [columns], data -> [values]}
recordslist like [{column -> value}, … , {column -> value}]
indexdict like {index -> {column -> value}}
columnsdict like {column -> {index -> value}}
valuesjust the values array
tableadhering to the JSON Table Schema
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.lines: Ifrecordsorient, then will write each record per line as json.
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 [221]: dfj = pd.DataFrame(np.random.randn(5, 2), columns=list("AB"))
In [222]: json = dfj.to_json()
In [223]: json
Out[223]: '{"A":{"0":-0.1213062281,"1":0.6957746499,"2":0.9597255933,"3":-0.6199759194,"4":-0.7323393705},"B":{"0":-0.0978826728,"1":0.3417343559,"2":-1.1103361029,"3":0.1497483186,"4":0.6877383895}}'
Orient options¶
There are a number of different options for the format of the resulting JSON
file / string. Consider the following DataFrame and Series:
In [224]: dfjo = pd.DataFrame(
.....: dict(A=range(1, 4), B=range(4, 7), C=range(7, 10)),
.....: columns=list("ABC"),
.....: index=list("xyz"),
.....: )
.....:
In [225]: dfjo
Out[225]:
A B C
x 1 4 7
y 2 5 8
z 3 6 9
In [226]: sjo = pd.Series(dict(x=15, y=16, z=17), name="D")
In [227]: sjo
Out[227]:
x 15
y 16
z 17
Name: D, dtype: int64
Column oriented (the default for DataFrame) serializes the data as
nested JSON objects with column labels acting as the primary index:
In [228]: dfjo.to_json(orient="columns")
Out[228]: '{"A":{"x":1,"y":2,"z":3},"B":{"x":4,"y":5,"z":6},"C":{"x":7,"y":8,"z":9}}'
# Not available for Series
Index oriented (the default for Series) similar to column oriented
but the index labels are now primary:
In [229]: dfjo.to_json(orient="index")
Out[229]: '{"x":{"A":1,"B":4,"C":7},"y":{"A":2,"B":5,"C":8},"z":{"A":3,"B":6,"C":9}}'
In [230]: sjo.to_json(orient="index")
Out[230]: '{"x":15,"y":16,"z":17}'
Record oriented serializes the data to a JSON array of column -> value records,
index labels are not included. This is useful for passing DataFrame data to plotting
libraries, for example the JavaScript library d3.js:
In [231]: dfjo.to_json(orient="records")
Out[231]: '[{"A":1,"B":4,"C":7},{"A":2,"B":5,"C":8},{"A":3,"B":6,"C":9}]'
In [232]: sjo.to_json(orient="records")
Out[232]: '[15,16,17]'
Value oriented is a bare-bones option which serializes to nested JSON arrays of values only, column and index labels are not included:
In [233]: dfjo.to_json(orient="values")
Out[233]: '[[1,4,7],[2,5,8],[3,6,9]]'
# Not available for Series
Split oriented serializes to a JSON object containing separate entries for
values, index and columns. Name is also included for Series:
In [234]: dfjo.to_json(orient="split")
Out[234]: '{"columns":["A","B","C"],"index":["x","y","z"],"data":[[1,4,7],[2,5,8],[3,6,9]]}'
In [235]: sjo.to_json(orient="split")
Out[235]: '{"name":"D","index":["x","y","z"],"data":[15,16,17]}'
Table oriented serializes to the JSON Table Schema, allowing for the preservation of metadata including but not limited to dtypes and index names.
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 [236]: dfd = pd.DataFrame(np.random.randn(5, 2), columns=list("AB"))
In [237]: dfd["date"] = pd.Timestamp("20130101")
In [238]: dfd = dfd.sort_index(axis=1, ascending=False)
In [239]: json = dfd.to_json(date_format="iso")
In [240]: json
Out[240]: '{"date":{"0":"2013-01-01T00:00:00.000Z","1":"2013-01-01T00:00:00.000Z","2":"2013-01-01T00:00:00.000Z","3":"2013-01-01T00:00:00.000Z","4":"2013-01-01T00:00:00.000Z"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'
Writing in ISO date format, with microseconds:
In [241]: json = dfd.to_json(date_format="iso", date_unit="us")
In [242]: json
Out[242]: '{"date":{"0":"2013-01-01T00:00:00.000000Z","1":"2013-01-01T00:00:00.000000Z","2":"2013-01-01T00:00:00.000000Z","3":"2013-01-01T00:00:00.000000Z","4":"2013-01-01T00:00:00.000000Z"},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'
Epoch timestamps, in seconds:
In [243]: json = dfd.to_json(date_format="epoch", date_unit="s")
In [244]: json
Out[244]: '{"date":{"0":1356998400,"1":1356998400,"2":1356998400,"3":1356998400,"4":1356998400},"B":{"0":0.403309524,"1":0.3016244523,"2":-1.3698493577,"3":1.4626960492,"4":-0.8265909164},"A":{"0":0.1764443426,"1":-0.1549507744,"2":-2.1798606054,"3":-0.9542078401,"4":-1.7431609117}}'
Writing to a file, with a date index and a date column:
In [245]: dfj2 = dfj.copy()
In [246]: dfj2["date"] = pd.Timestamp("20130101")
In [247]: dfj2["ints"] = list(range(5))
In [248]: dfj2["bools"] = True
In [249]: dfj2.index = pd.date_range("20130101", periods=5)
In [250]: dfj2.to_json("test.json")
In [251]: with open("test.json") as fh:
.....: print(fh.read())
.....:
{"A":{"1356998400000":-0.1213062281,"1357084800000":0.6957746499,"1357171200000":0.9597255933,"1357257600000":-0.6199759194,"1357344000000":-0.7323393705},"B":{"1356998400000":-0.0978826728,"1357084800000":0.3417343559,"1357171200000":-1.1103361029,"1357257600000":0.1497483186,"1357344000000":0.6877383895},"date":{"1356998400000":1356998400000,"1357084800000":1356998400000,"1357171200000":1356998400000,"1357257600000":1356998400000,"1357344000000":1356998400000},"ints":{"1356998400000":0,"1357084800000":1,"1357171200000":2,"1357257600000":3,"1357344000000":4},"bools":{"1356998400000":true,"1357084800000":true,"1357171200000":true,"1357257600000":true,"1357344000000":true}}
Fallback behavior¶
If the JSON serializer cannot handle the container contents directly it will fall back in the following manner:
if the dtype is unsupported (e.g.
np.complex_) then thedefault_handler, if provided, will be called for each value, otherwise an exception is raised.if an object is unsupported it will attempt the following:
check if the object has defined a
toDictmethod and call it. AtoDictmethod should return adictwhich will then be JSON serialized.invoke the
default_handlerif one was provided.convert the object to a
dictby traversing its contents. However this will often fail with anOverflowErroror give unexpected results.
In general the best approach for unsupported objects or dtypes is to provide a default_handler.
For example:
>>> DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json() # raises
RuntimeError: Unhandled numpy dtype 15
can be dealt with by specifying a simple default_handler:
In [252]: pd.DataFrame([1.0, 2.0, complex(1.0, 2.0)]).to_json(default_handler=str)
Out[252]: '{"0":{"0":"(1+0j)","1":"(2+0j)","2":"(1+2j)"}}'
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
indexallowed values are {
split,records,index}
- DataFrame
default is
columnsallowed values are {
split,records,index,columns,values,table}
The format of the JSON string
splitdict like {index -> [index], columns -> [columns], data -> [values]}
recordslist like [{column -> value}, … , {column -> value}]
indexdict like {index -> {column -> value}}
columnsdict like {column -> {index -> value}}
valuesjust the values array
tableadhering to the JSON Table Schema
dtype: if True, infer dtypes, if a dict of column to dtype, then use those, ifFalse, then don’t infer dtypes at all, default is True, apply only to the data.convert_axes: boolean, try to convert the axes to the proper dtypes, default isTrueconvert_dates: a list of columns to parse for dates; IfTrue, then try to parse date-like columns, default isTrue.keep_default_dates: boolean, defaultTrue. If parsing dates, then parse the default date-like columns.numpy: direct decoding to NumPy arrays. default isFalse; 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 functionality.date_unit: string, the timestamp unit to detect if converting dates. Default None. By default the timestamp precision will be detected, if this is not desired then pass one of ‘s’, ‘ms’, ‘us’ or ‘ns’ to force timestamp precision to seconds, milliseconds, microseconds or nanoseconds respectively.lines: reads file as one json object per line.encoding: The encoding to use to decode py3 bytes.chunksize: when used in combination withlines=True, return a JsonReader which reads inchunksizelines per iteration.
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
floatdata will be converted tointegerif it can be done safely, e.g. a column of1.bool columns will be converted to
integeron reconstruction
Thus there are times where you may want to specify specific dtypes via the dtype keyword argument.
Reading from a JSON string:
In [253]: pd.read_json(json)
Out[253]:
date B A
0 2013-01-01 0.403310 0.176444
1 2013-01-01 0.301624 -0.154951
2 2013-01-01 -1.369849 -2.179861
3 2013-01-01 1.462696 -0.954208
4 2013-01-01 -0.826591 -1.743161
Reading from a file:
In [254]: pd.read_json("test.json")
Out[254]:
A B date ints bools
2013-01-01 -0.121306 -0.097883 2013-01-01 0 True
2013-01-02 0.695775 0.341734 2013-01-01 1 True
2013-01-03 0.959726 -1.110336 2013-01-01 2 True
2013-01-04 -0.619976 0.149748 2013-01-01 3 True
2013-01-05 -0.732339 0.687738 2013-01-01 4 True
Don’t convert any data (but still convert axes and dates):
In [255]: pd.read_json("test.json", dtype=object).dtypes
Out[255]:
A object
B object
date object
ints object
bools object
dtype: object
Specify dtypes for conversion:
In [256]: pd.read_json("test.json", dtype={"A": "float32", "bools": "int8"}).dtypes
Out[256]:
A float32
B float64
date datetime64[ns]
ints int64
bools int8
dtype: object
Preserve string indices:
In [257]: si = pd.DataFrame(
.....: np.zeros((4, 4)), columns=list(range(4)), index=[str(i) for i in range(4)]
.....: )
.....:
In [258]: si
Out[258]:
0 1 2 3
0 0.0 0.0 0.0 0.0
1 0.0 0.0 0.0 0.0
2 0.0 0.0 0.0 0.0
3 0.0 0.0 0.0 0.0
In [259]: si.index
Out[259]: Index(['0', '1', '2', '3'], dtype='object')
In [260]: si.columns
Out[260]: Int64Index([0, 1, 2, 3], dtype='int64')
In [261]: json = si.to_json()
In [262]: sij = pd.read_json(json, convert_axes=False)
In [263]: sij
Out[263]:
0 1 2 3
0 0 0 0 0
1 0 0 0 0
2 0 0 0 0
3 0 0 0 0
In [264]: sij.index
Out[264]: Index(['0', '1', '2', '3'], dtype='object')
In [265]: sij.columns
Out[265]: Index(['0', '1', '2', '3'], dtype='object')
Dates written in nanoseconds need to be read back in nanoseconds:
In [266]: json = dfj2.to_json(date_unit="ns")
# Try to parse timestamps as milliseconds -> Won't Work
In [267]: dfju = pd.read_json(json, date_unit="ms")
In [268]: dfju
Out[268]:
A B date ints bools
1356998400000000000 -0.121306 -0.097883 1356998400000000000 0 True
1357084800000000000 0.695775 0.341734 1356998400000000000 1 True
1357171200000000000 0.959726 -1.110336 1356998400000000000 2 True
1357257600000000000 -0.619976 0.149748 1356998400000000000 3 True
1357344000000000000 -0.732339 0.687738 1356998400000000000 4 True
# Let pandas detect the correct precision
In [269]: dfju = pd.read_json(json)
In [270]: dfju
Out[270]:
A B date ints bools
2013-01-01 -0.121306 -0.097883 2013-01-01 0 True
2013-01-02 0.695775 0.341734 2013-01-01 1 True
2013-01-03 0.959726 -1.110336 2013-01-01 2 True
2013-01-04 -0.619976 0.149748 2013-01-01 3 True
2013-01-05 -0.732339 0.687738 2013-01-01 4 True
# Or specify that all timestamps are in nanoseconds
In [271]: dfju = pd.read_json(json, date_unit="ns")
In [272]: dfju
Out[272]:
A B date ints bools
2013-01-01 -0.121306 -0.097883 2013-01-01 0 True
2013-01-02 0.695775 0.341734 2013-01-01 1 True
2013-01-03 0.959726 -1.110336 2013-01-01 2 True
2013-01-04 -0.619976 0.149748 2013-01-01 3 True
2013-01-05 -0.732339 0.687738 2013-01-01 4 True
The Numpy parameter¶
Note
This param has been deprecated as of version 1.0.0 and will raise a FutureWarning.
This supports numeric data only. Index and columns labels may be non-numeric, e.g. strings, dates etc.
If numpy=True is passed to read_json an attempt will be made to sniff
an appropriate dtype during deserialization and to subsequently decode directly
to NumPy arrays, bypassing the need for intermediate Python objects.
This can provide speedups if you are deserialising a large amount of numeric data:
In [273]: randfloats = np.random.uniform(-100, 1000, 10000)
In [274]: randfloats.shape = (1000, 10)
In [275]: dffloats = pd.DataFrame(randfloats, columns=list("ABCDEFGHIJ"))
In [276]: jsonfloats = dffloats.to_json()
In [277]: %timeit pd.read_json(jsonfloats)
27.1 ms +- 7.28 ms per loop (mean +- std. dev. of 7 runs, 10 loops each)
In [278]: %timeit pd.read_json(jsonfloats, numpy=True)
22.9 ms +- 3.41 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
The speedup is less noticeable for smaller datasets:
In [279]: jsonfloats = dffloats.head(100).to_json()
In [280]: %timeit pd.read_json(jsonfloats)
12.1 ms +- 3.66 ms per loop (mean +- std. dev. of 7 runs, 100 loops each)
In [281]: %timeit pd.read_json(jsonfloats, numpy=True)
7.65 ms +- 432 us per loop (mean +- std. dev. of 7 runs, 100 loops each)
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
ValueErrormay 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_jsonbut may not be the case if the JSON is from another source.
Normalization¶
pandas provides a utility function to take a dict or list of dicts and normalize this semi-structured data into a flat table.
In [282]: data = [
.....: {"id": 1, "name": {"first": "Coleen", "last": "Volk"}},
.....: {"name": {"given": "Mark", "family": "Regner"}},
.....: {"id": 2, "name": "Faye Raker"},
.....: ]
.....:
In [283]: pd.json_normalize(data)
Out[283]:
id name.first name.last name.given name.family name
0 1.0 Coleen Volk NaN NaN NaN
1 NaN NaN NaN Mark Regner NaN
2 2.0 NaN NaN NaN NaN Faye Raker
In [284]: data = [
.....: {
.....: "state": "Florida",
.....: "shortname": "FL",
.....: "info": {"governor": "Rick Scott"},
.....: "county": [
.....: {"name": "Dade", "population": 12345},
.....: {"name": "Broward", "population": 40000},
.....: {"name": "Palm Beach", "population": 60000},
.....: ],
.....: },
.....: {
.....: "state": "Ohio",
.....: "shortname": "OH",
.....: "info": {"governor": "John Kasich"},
.....: "county": [
.....: {"name": "Summit", "population": 1234},
.....: {"name": "Cuyahoga", "population": 1337},
.....: ],
.....: },
.....: ]
.....:
In [285]: pd.json_normalize(data, "county", ["state", "shortname", ["info", "governor"]])
Out[285]:
name population state shortname info.governor
0 Dade 12345 Florida FL Rick Scott
1 Broward 40000 Florida FL Rick Scott
2 Palm Beach 60000 Florida FL Rick Scott
3 Summit 1234 Ohio OH John Kasich
4 Cuyahoga 1337 Ohio OH John Kasich
The max_level parameter provides more control over which level to end normalization. With max_level=1 the following snippet normalizes until 1st nesting level of the provided dict.
In [286]: data = [
.....: {
.....: "CreatedBy": {"Name": "User001"},
.....: "Lookup": {
.....: "TextField": "Some text",
.....: "UserField": {"Id": "ID001", "Name": "Name001"},
.....: },
.....: "Image": {"a": "b"},
.....: }
.....: ]
.....:
In [287]: pd.json_normalize(data, max_level=1)
Out[287]:
CreatedBy.Name Lookup.TextField Lookup.UserField Image.a
0 User001 Some text {'Id': 'ID001', 'Name': 'Name001'} b
Line delimited json¶
pandas is able to read and write line-delimited json files that are common in data processing pipelines using Hadoop or Spark.
For line-delimited json files, pandas can also return an iterator which reads in chunksize lines at a time. This can be useful for large files or to read from a stream.
In [288]: jsonl = """
.....: {"a": 1, "b": 2}
.....: {"a": 3, "b": 4}
.....: """
.....:
In [289]: df = pd.read_json(jsonl, lines=True)
In [290]: df
Out[290]:
a b
0 1 2
1 3 4
In [291]: df.to_json(orient="records", lines=True)
Out[291]: '{"a":1,"b":2}\n{"a":3,"b":4}\n'
# reader is an iterator that returns ``chunksize`` lines each iteration
In [292]: with pd.read_json(StringIO(jsonl), lines=True, chunksize=1) as reader:
.....: reader
.....: for chunk in reader:
.....: print(chunk)
.....:
Empty DataFrame
Columns: []
Index: []
a b
0 1 2
a b
1 3 4
Table schema¶
Table Schema is a spec for describing tabular datasets as a JSON
object. The JSON includes information on the field names, types, and
other attributes. You can use the orient table to build
a JSON string with two fields, schema and data.
In [293]: df = pd.DataFrame(
.....: {
.....: "A": [1, 2, 3],
.....: "B": ["a", "b", "c"],
.....: "C": pd.date_range("2016-01-01", freq="d", periods=3),
.....: },
.....: index=pd.Index(range(3), name="idx"),
.....: )
.....:
In [294]: df
Out[294]:
A B C
idx
0 1 a 2016-01-01
1 2 b 2016-01-02
2 3 c 2016-01-03
In [295]: df.to_json(orient="table", date_format="iso")
Out[295]: '{"schema":{"fields":[{"name":"idx","type":"integer"},{"name":"A","type":"integer"},{"name":"B","type":"string"},{"name":"C","type":"datetime"}],"primaryKey":["idx"],"pandas_version":"1.4.0"},"data":[{"idx":0,"A":1,"B":"a","C":"2016-01-01T00:00:00.000Z"},{"idx":1,"A":2,"B":"b","C":"2016-01-02T00:00:00.000Z"},{"idx":2,"A":3,"B":"c","C":"2016-01-03T00:00:00.000Z"}]}'
The schema field contains the fields key, which itself contains
a list of column name to type pairs, including the Index or MultiIndex
(see below for a list of types).
The schema field also contains a primaryKey field if the (Multi)index
is unique.
The second field, data, contains the serialized data with the records
orient.
The index is included, and any datetimes are ISO 8601 formatted, as required
by the Table Schema spec.
The full list of types supported are described in the Table Schema spec. This table shows the mapping from pandas types:
pandas type |
Table Schema type |
|---|---|
int64 |
integer |
float64 |
number |
bool |
boolean |
datetime64[ns] |
datetime |
timedelta64[ns] |
duration |
categorical |
any |
object |
str |
A few notes on the generated table schema:
The
schemaobject contains apandas_versionfield. This contains the version of pandas’ dialect of the schema, and will be incremented with each revision.All dates are converted to UTC when serializing. Even timezone naive values, which are treated as UTC with an offset of 0.
In [296]: from pandas.io.json import build_table_schema In [297]: s = pd.Series(pd.date_range("2016", periods=4)) In [298]: build_table_schema(s) Out[298]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'datetime'}], 'primaryKey': ['index'], 'pandas_version': '1.4.0'}
datetimes with a timezone (before serializing), include an additional field
tzwith the time zone name (e.g.'US/Central').In [299]: s_tz = pd.Series(pd.date_range("2016", periods=12, tz="US/Central")) In [300]: build_table_schema(s_tz) Out[300]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'datetime', 'tz': 'US/Central'}], 'primaryKey': ['index'], 'pandas_version': '1.4.0'}
Periods are converted to timestamps before serialization, and so have the same behavior of being converted to UTC. In addition, periods will contain and additional field
freqwith the period’s frequency, e.g.'A-DEC'.In [301]: s_per = pd.Series(1, index=pd.period_range("2016", freq="A-DEC", periods=4)) In [302]: build_table_schema(s_per) Out[302]: {'fields': [{'name': 'index', 'type': 'datetime', 'freq': 'A-DEC'}, {'name': 'values', 'type': 'integer'}], 'primaryKey': ['index'], 'pandas_version': '1.4.0'}
Categoricals use the
anytype and anenumconstraint listing the set of possible values. Additionally, anorderedfield is included:In [303]: s_cat = pd.Series(pd.Categorical(["a", "b", "a"])) In [304]: build_table_schema(s_cat) Out[304]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'any', 'constraints': {'enum': ['a', 'b']}, 'ordered': False}], 'primaryKey': ['index'], 'pandas_version': '1.4.0'}
A
primaryKeyfield, containing an array of labels, is included if the index is unique:In [305]: s_dupe = pd.Series([1, 2], index=[1, 1]) In [306]: build_table_schema(s_dupe) Out[306]: {'fields': [{'name': 'index', 'type': 'integer'}, {'name': 'values', 'type': 'integer'}], 'pandas_version': '1.4.0'}
The
primaryKeybehavior is the same with MultiIndexes, but in this case theprimaryKeyis an array:In [307]: s_multi = pd.Series(1, index=pd.MultiIndex.from_product([("a", "b"), (0, 1)])) In [308]: build_table_schema(s_multi) Out[308]: {'fields': [{'name': 'level_0', 'type': 'string'}, {'name': 'level_1', 'type': 'integer'}, {'name': 'values', 'type': 'integer'}], 'primaryKey': FrozenList(['level_0', 'level_1']), 'pandas_version': '1.4.0'}
The default naming roughly follows these rules:
For series, the
object.nameis used. If that’s none, then the name isvaluesFor
DataFrames, the stringified version of the column name is usedFor
Index(notMultiIndex),index.nameis used, with a fallback toindexif that is None.For
MultiIndex,mi.namesis used. If any level has no name, thenlevel_<i>is used.
read_json also accepts orient='table' as an argument. This allows for
the preservation of metadata such as dtypes and index names in a
round-trippable manner.
In [309]: df = pd.DataFrame( .....: { .....: "foo": [1, 2, 3, 4], .....: "bar": ["a", "b", "c", "d"], .....: "baz": pd.date_range("2018-01-01", freq="d", periods=4), .....: "qux": pd.Categorical(["a", "b", "c", "c"]), .....: }, .....: index=pd.Index(range(4), name="idx"), .....: ) .....: In [310]: df Out[310]: foo bar baz qux idx 0 1 a 2018-01-01 a 1 2 b 2018-01-02 b 2 3 c 2018-01-03 c 3 4 d 2018-01-04 c In [311]: df.dtypes Out[311]: foo int64 bar object baz datetime64[ns] qux category dtype: object In [312]: df.to_json("test.json", orient="table") In [313]: new_df = pd.read_json("test.json", orient="table") In [314]: new_df Out[314]: foo bar baz qux idx 0 1 a 2018-01-01 a 1 2 b 2018-01-02 b 2 3 c 2018-01-03 c 3 4 d 2018-01-04 c In [315]: new_df.dtypes Out[315]: foo int64 bar object baz datetime64[ns] qux category dtype: object
Please note that the literal string ‘index’ as the name of an Index
is not round-trippable, nor are any names beginning with 'level_' within a
MultiIndex. These are used by default in DataFrame.to_json() to
indicate missing values and the subsequent read cannot distinguish the intent.
In [316]: df.index.name = "index"
In [317]: df.to_json("test.json", orient="table")
In [318]: new_df = pd.read_json("test.json", orient="table")
In [319]: print(new_df.index.name)
None
When using orient='table' along with user-defined ExtensionArray,
the generated schema will contain an additional extDtype key in the respective
fields element. This extra key is not standard but does enable JSON roundtrips
for extension types (e.g. read_json(df.to_json(orient="table"), orient="table")).
The extDtype key carries the name of the extension, if you have properly registered
the ExtensionDtype, pandas will use said name to perform a lookup into the registry
and re-convert the serialized data into your custom dtype.
HTML¶
Reading HTML content¶
Warning
We highly encourage you to read the HTML Table Parsing gotchas below regarding the issues surrounding the BeautifulSoup4/html5lib/lxml parsers.
The top-level read_html() function can accept an HTML
string/file/URL and will parse HTML tables into list of pandas DataFrames.
Let’s look at a few examples.
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 [320]: url = "https://www.fdic.gov/resources/resolutions/bank-failures/failed-bank-list"
In [321]: dfs = pd.read_html(url)
In [322]: dfs
Out[322]:
[ Bank NameBank CityCity StateSt ... Acquiring InstitutionAI Closing DateClosing FundFund
0 Almena State Bank Almena KS ... Equity Bank October 23, 2020 10538
1 First City Bank of Florida Fort Walton Beach FL ... United Fidelity Bank, fsb October 16, 2020 10537
2 The First State Bank Barboursville WV ... MVB Bank, Inc. April 3, 2020 10536
3 Ericson State Bank Ericson NE ... Farmers and Merchants Bank February 14, 2020 10535
4 City National Bank of New Jersey Newark NJ ... Industrial Bank November 1, 2019 10534
.. ... ... ... ... ... ... ...
558 Superior Bank, FSB Hinsdale IL ... Superior Federal, FSB July 27, 2001 6004
559 Malta National Bank Malta OH ... North Valley Bank May 3, 2001 4648
560 First Alliance Bank & Trust Co. Manchester NH ... Southern New Hampshire Bank & Trust February 2, 2001 4647
561 National State Bank of Metropolis Metropolis IL ... Banterra Bank of Marion December 14, 2000 4646
562 Bank of Honolulu Honolulu HI ... Bank of the Orient October 13, 2000 4645
[563 rows x 7 columns]]
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 [323]: html_str = """
.....: <table>
.....: <tr>
.....: <th>A</th>
.....: <th colspan="1">B</th>
.....: <th rowspan="1">C</th>
.....: </tr>
.....: <tr>
.....: <td>a</td>
.....: <td>b</td>
.....: <td>c</td>
.....: </tr>
.....: </table>
.....: """
.....:
In [324]: with open("tmp.html", "w") as f:
.....: f.write(html_str)
.....:
In [325]: df = pd.read_html("tmp.html")
In [326]: df[0]
Out[326]:
A B C
0 a b c
You can even pass in an instance of StringIO if you so desire:
In [327]: dfs = pd.read_html(StringIO(html_str))
In [328]: dfs[0]
Out[328]:
A B C
0 a b c
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 = pd.read_html(url, match=match)
Specify a header row (by default <th> or <td> elements located within a
<thead> are used to form the column index, if multiple rows are contained within
<thead> then a MultiIndex is created); if specified, the header row is taken
from the data minus the parsed header elements (<th> elements).
dfs = pd.read_html(url, header=0)
Specify an index column:
dfs = pd.read_html(url, index_col=0)
Specify a number of rows to skip:
dfs = pd.read_html(url, skiprows=0)
Specify a number of rows to skip using a list (range works
as well):
dfs = pd.read_html(url, skiprows=range(2))
Specify an HTML attribute:
dfs1 = pd.read_html(url, attrs={"id": "table"})
dfs2 = pd.read_html(url, attrs={"class": "sortable"})
print(np.array_equal(dfs1[0], dfs2[0])) # Should be True
Specify values that should be converted to NaN:
dfs = pd.read_html(url, na_values=["No Acquirer"])
Specify whether to keep the default set of NaN values:
dfs = pd.read_html(url, keep_default_na=False)
Specify converters for columns. This is useful for numerical text data that has leading zeros. By default columns that are numerical are cast to numeric types and the leading zeros are lost. To avoid this, we can convert these columns to strings.
url_mcc = "https://en.wikipedia.org/wiki/Mobile_country_code"
dfs = pd.read_html(
url_mcc,
match="Telekom Albania",
header=0,
converters={"MNC": str},
)
Use some combination of the above:
dfs = pd.read_html(url, match="Metcalf Bank", index_col=0)
Read in pandas to_html output (with some loss of floating point precision):
df = pd.DataFrame(np.random.randn(2, 2))
s = df.to_html(float_format="{0:.40g}".format)
dfin = pd.read_html(s, index_col=0)
The lxml backend will raise an error on a failed parse if that is the only
parser you provide. If you only have a single parser you can provide just a
string, but it is considered good practice to pass a list with one string if,
for example, the function expects a sequence of strings. You may use:
dfs = pd.read_html(url, "Metcalf Bank", index_col=0, flavor=["lxml"])
Or you could pass flavor='lxml' without a list:
dfs = pd.read_html(url, "Metcalf Bank", index_col=0, flavor="lxml")
However, if you have bs4 and html5lib installed and pass None or ['lxml',
'bs4'] then the parse will most likely succeed. Note that as soon as a parse
succeeds, the function will return.
dfs = pd.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.
Note
In an HTML-rendering supported environment like a Jupyter Notebook, display(HTML(...))`
will render the raw HTML into the environment.
In [329]: from IPython.display import display, HTML
In [330]: df = pd.DataFrame(np.random.randn(2, 2))
In [331]: df
Out[331]:
0 1
0 0.070319 1.773907
1 0.253908 0.414581
In [332]: html = df.to_html()
In [333]: print(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>0.070319</td>
<td>1.773907</td>
</tr>
<tr>
<th>1</th>
<td>0.253908</td>
<td>0.414581</td>
</tr>
</tbody>
</table>
In [334]: display(HTML(html))
<IPython.core.display.HTML object>
The columns argument will limit the columns shown:
In [335]: html = df.to_html(columns=[0])
In [336]: print(html)
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>0</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>0.070319</td>
</tr>
<tr>
<th>1</th>
<td>0.253908</td>
</tr>
</tbody>
</table>
In [337]: display(HTML(html))
<IPython.core.display.HTML object>
float_format takes a Python callable to control the precision of floating
point values:
In [338]: html = df.to_html(float_format="{0:.10f}".format)
In [339]: print(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>0.0703192665</td>
<td>1.7739074228</td>
</tr>
<tr>
<th>1</th>
<td>0.2539083433</td>
<td>0.4145805920</td>
</tr>
</tbody>
</table>
In [340]: display(HTML(html))
<IPython.core.display.HTML object>
bold_rows will make the row labels bold by default, but you can turn that
off:
In [341]: html = df.to_html(bold_rows=False)
In [342]: print(html)
<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>0.070319</td>
<td>1.773907</td>
</tr>
<tr>
<td>1</td>
<td>0.253908</td>
<td>0.414581</td>
</tr>
</tbody>
</table>
In [343]: display(HTML(html))
<IPython.core.display.HTML object>
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 [344]: 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>0.070319</td>
<td>1.773907</td>
</tr>
<tr>
<th>1</th>
<td>0.253908</td>
<td>0.414581</td>
</tr>
</tbody>
</table>
The render_links argument provides the ability to add hyperlinks to cells
that contain URLs.
In [345]: url_df = pd.DataFrame(
.....: {
.....: "name": ["Python", "pandas"],
.....: "url": ["https://www.python.org/", "https://pandas.pydata.org"],
.....: }
.....: )
.....:
In [346]: html = url_df.to_html(render_links=True)
In [347]: print(html)
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>name</th>
<th>url</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>Python</td>
<td><a href="https://www.python.org/" target="_blank">https://www.python.org/</a></td>
</tr>
<tr>
<th>1</th>
<td>pandas</td>
<td><a href="https://pandas.pydata.org" target="_blank">https://pandas.pydata.org</a></td>
</tr>
</tbody>
</table>
In [348]: display(HTML(html))
<IPython.core.display.HTML object>
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 [349]: df = pd.DataFrame({"a": list("&<>"), "b": np.random.randn(3)})
Escaped:
In [350]: html = df.to_html()
In [351]: print(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>0.842321</td>
</tr>
<tr>
<th>1</th>
<td><</td>
<td>0.211337</td>
</tr>
<tr>
<th>2</th>
<td>></td>
<td>-1.055427</td>
</tr>
</tbody>
</table>
In [352]: display(HTML(html))
<IPython.core.display.HTML object>
Not escaped:
In [353]: html = df.to_html(escape=False)
In [354]: print(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>0.842321</td>
</tr>
<tr>
<th>1</th>
<td><</td>
<td>0.211337</td>
</tr>
<tr>
<th>2</th>
<td>></td>
<td>-1.055427</td>
</tr>
</tbody>
</table>
In [355]: display(HTML(html))
<IPython.core.display.HTML object>
Note
Some browsers may not show a difference in the rendering of the previous two HTML tables.
HTML Table Parsing Gotchas¶
There are some versioning issues surrounding the libraries that are used to
parse HTML tables in the top-level pandas io function read_html.
Issues with lxml
Benefits
Drawbacks
lxml does not make any guarantees about the results of its parse unless it is given strictly valid markup.
In light of the above, we have chosen to allow you, the user, to use the lxml backend, but this backend will use html5lib if lxml fails to parse
It is therefore highly recommended that you install both BeautifulSoup4 and html5lib, so that you will still get a valid result (provided everything else is valid) even if lxml fails.
Issues with BeautifulSoup4 using lxml as a backend
The above issues hold here as well since BeautifulSoup4 is essentially just a wrapper around a parser backend.
Issues with BeautifulSoup4 using html5lib as a backend
Benefits
html5lib is far more lenient than lxml and consequently deals with real-life markup in a much saner way rather than just, e.g., dropping an element without notifying you.
html5lib generates valid HTML5 markup from invalid markup automatically. This is extremely important for parsing HTML tables, since it guarantees a valid document. However, that does NOT mean that it is “correct”, since the process of fixing markup does not have a single definition.
html5lib is pure Python and requires no additional build steps beyond its own installation.
Drawbacks
The biggest drawback to using html5lib is that it is slow as molasses. However consider the fact that many tables on the web are not big enough for the parsing algorithm runtime to matter. It is more likely that the bottleneck will be in the process of reading the raw text from the URL over the web, i.e., IO (input-output). For very large tables, this might not be true.
LaTeX¶
New in version 1.3.0.
Currently there are no methods to read from LaTeX, only output methods.
Writing to LaTeX files¶
Note
DataFrame and Styler objects currently have a to_latex method. We recommend
using the Styler.to_latex() method
over DataFrame.to_latex() due to the former’s greater flexibility with
conditional styling, and the latter’s possible future deprecation.
Review the documentation for Styler.to_latex, which gives examples of conditional styling and explains the operation of its keyword arguments.
For simple application the following pattern is sufficient.
In [356]: df = pd.DataFrame([[1, 2], [3, 4]], index=["a", "b"], columns=["c", "d"])
In [357]: print(df.style.to_latex())
\begin{tabular}{lrr}
& c & d \\
a & 1 & 2 \\
b & 3 & 4 \\
\end{tabular}
To format values before output, chain the Styler.format method.
In [358]: print(df.style.format("€ {}").to_latex())
\begin{tabular}{lrr}
& c & d \\
a & € 1 & € 2 \\
b & € 3 & € 4 \\
\end{tabular}
XML¶
Reading XML¶
New in version 1.3.0.
The top-level read_xml() function can accept an XML
string/file/URL and will parse nodes and attributes into a pandas DataFrame.
Note
Since there is no standard XML structure where design types can vary in
many ways, read_xml works best with flatter, shallow versions. If
an XML document is deeply nested, use the stylesheet feature to
transform XML into a flatter version.
Let’s look at a few examples.
Read an XML string:
In [359]: xml = """<?xml version="1.0" encoding="UTF-8"?>
.....: <bookstore>
.....: <book category="cooking">
.....: <title lang="en">Everyday Italian</title>
.....: <author>Giada De Laurentiis</author>
.....: <year>2005</year>
.....: <price>30.00</price>
.....: </book>
.....: <book category="children">
.....: <title lang="en">Harry Potter</title>
.....: <author>J K. Rowling</author>
.....: <year>2005</year>
.....: <price>29.99</price>
.....: </book>
.....: <book category="web">
.....: <title lang="en">Learning XML</title>
.....: <author>Erik T. Ray</author>
.....: <year>2003</year>
.....: <price>39.95</price>
.....: </book>
.....: </bookstore>"""
.....:
In [360]: df = pd.read_xml(xml)
In [361]: df
Out[361]:
category title author year price
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00
1 children Harry Potter J K. Rowling 2005 29.99
2 web Learning XML Erik T. Ray 2003 39.95
Read a URL with no options:
In [362]: df = pd.read_xml("https://www.w3schools.com/xml/books.xml")
In [363]: df
Out[363]:
category title author year price cover
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00 None
1 children Harry Potter J K. Rowling 2005 29.99 None
2 web XQuery Kick Start Vaidyanathan Nagarajan 2003 49.99 None
3 web Learning XML Erik T. Ray 2003 39.95 paperback
Read in the content of the “books.xml” file and pass it to read_xml
as a string:
In [364]: file_path = "books.xml"
In [365]: with open(file_path, "w") as f:
.....: f.write(xml)
.....:
In [366]: with open(file_path, "r") as f:
.....: df = pd.read_xml(f.read())
.....:
In [367]: df
Out[367]:
category title author year price
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00
1 children Harry Potter J K. Rowling 2005 29.99
2 web Learning XML Erik T. Ray 2003 39.95
Read in the content of the “books.xml” as instance of StringIO or
BytesIO and pass it to read_xml:
In [368]: with open(file_path, "r") as f:
.....: sio = StringIO(f.read())
.....:
In [369]: df = pd.read_xml(sio)
In [370]: df
Out[370]:
category title author year price
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00
1 children Harry Potter J K. Rowling 2005 29.99
2 web Learning XML Erik T. Ray 2003 39.95
In [371]: with open(file_path, "rb") as f:
.....: bio = BytesIO(f.read())
.....:
In [372]: df = pd.read_xml(bio)
In [373]: df
Out[373]:
category title author year price
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00
1 children Harry Potter J K. Rowling 2005 29.99
2 web Learning XML Erik T. Ray 2003 39.95
Even read XML from AWS S3 buckets such as NIH NCBI PMC Article Datasets providing Biomedical and Life Science Jorurnals:
In [374]: df = pd.read_xml(
.....: "s3://pmc-oa-opendata/oa_comm/xml/all/PMC1236943.xml",
.....: xpath=".//journal-meta",
.....: )
.....:
In [375]: df
Out[375]:
journal-id journal-title issn publisher
0 Cardiovasc Ultrasound Cardiovascular Ultrasound 1476-7120 NaN
With lxml as default parser, you access the full-featured XML library
that extends Python’s ElementTree API. One powerful tool is ability to query
nodes selectively or conditionally with more expressive XPath:
In [376]: df = pd.read_xml(file_path, xpath="//book[year=2005]")
In [377]: df
Out[377]:
category title author year price
0 cooking Everyday Italian Giada De Laurentiis 2005 30.00
1 children Harry Potter J K. Rowling 2005 29.99
Specify only elements or only attributes to parse:
In [378]: df = pd.read_xml(file_path, elems_only=True)
In [379]: df
Out[379]:
title author year price
0 Everyday Italian Giada De Laurentiis 2005 30.00
1 Harry Potter J K. Rowling 2005 29.99
2 Learning XML Erik T. Ray 2003 39.95
In [380]: df = pd.read_xml(file_path, attrs_only=True)
In [381]: df
Out[381]:
category
0 cooking
1 children
2 web
XML documents can have namespaces with prefixes and default namespaces without
prefixes both of which are denoted with a special attribute xmlns. In order
to parse by node under a namespace context, xpath must reference a prefix.
For example, below XML contains a namespace with prefix, doc, and URI at
https://example.com. In order to parse doc:row nodes,
namespaces must be used.
In [382]: xml = """<?xml version='1.0' encoding='utf-8'?>
.....: <doc:data xmlns:doc="https://example.com">
.....: <doc:row>
.....: <doc:shape>square</doc:shape>
.....: <doc:degrees>360</doc:degrees>
.....: <doc:sides>4.0</doc:sides>
.....: </doc:row>
.....: <doc:row>
.....: <doc:shape>circle</doc:shape>
.....: <doc:degrees>360</doc:degrees>
.....: <doc:sides/>
.....: </doc:row>
.....: <doc:row>
.....: <doc:shape>triangle</doc:shape>
.....: <doc:degrees>180</doc:degrees>
.....: <doc:sides>3.0</doc:sides>
.....: </doc:row>
.....: </doc:data>"""
.....:
In [383]: df = pd.read_xml(xml,
.....: xpath="//doc:row",
.....: namespaces={"doc": "https://example.com"})
.....:
In [384]: df
Out[384]:
shape degrees sides
0 square 360 4.0
1 circle 360 NaN
2 triangle 180 3.0
Similarly, an XML document can have a default namespace without prefix. Failing
to assign a temporary prefix will return no nodes and raise a ValueError.
But assigning any temporary name to correct URI allows parsing by nodes.
In [385]: xml = """<?xml version='1.0' encoding='utf-8'?>
.....: <data xmlns="https://example.com">
.....: <row>
.....: <shape>square</shape>
.....: <degrees>360</degrees>
.....: <sides>4.0</sides>
.....: </row>
.....: <row>
.....: <shape>circle</shape>
.....: <degrees>360</degrees>
.....: <sides/>
.....: </row>
.....: <row>
.....: <shape>triangle</shape>
.....: <degrees>180</degrees>
.....: <sides>3.0</sides>
.....: </row>
.....: </data>"""
.....:
In [386]: df = pd.read_xml(xml,
.....: xpath="//pandas:row",
.....: namespaces={"pandas": "https://example.com"})
.....:
In [387]: df
Out[387]:
shape degrees sides
0 square 360 4.0
1 circle 360 NaN
2 triangle 180 3.0
However, if XPath does not reference node names such as default, /*, then
namespaces is not required.
With lxml as parser, you can flatten nested XML documents with an XSLT script which also can be string/file/URL types. As background, XSLT is a special-purpose language written in a special XML file that can transform original XML documents into other XML, HTML, even text (CSV, JSON, etc.) using an XSLT processor.
For example, consider this somewhat nested structure of Chicago “L” Rides
where station and rides elements encapsulate data in their own sections.
With below XSLT, lxml can transform original nested document into a flatter
output (as shown below for demonstration) for easier parse into DataFrame:
In [388]: xml = """<?xml version='1.0' encoding='utf-8'?>
.....: <response>
.....: <row>
.....: <station id="40850" name="Library"/>
.....: <month>2020-09-01T00:00:00</month>
.....: <rides>
.....: <avg_weekday_rides>864.2</avg_weekday_rides>
.....: <avg_saturday_rides>534</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>417.2</avg_sunday_holiday_rides>
.....: </rides>
.....: </row>
.....: <row>
.....: <station id="41700" name="Washington/Wabash"/>
.....: <month>2020-09-01T00:00:00</month>
.....: <rides>
.....: <avg_weekday_rides>2707.4</avg_weekday_rides>
.....: <avg_saturday_rides>1909.8</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>1438.6</avg_sunday_holiday_rides>
.....: </rides>
.....: </row>
.....: <row>
.....: <station id="40380" name="Clark/Lake"/>
.....: <month>2020-09-01T00:00:00</month>
.....: <rides>
.....: <avg_weekday_rides>2949.6</avg_weekday_rides>
.....: <avg_saturday_rides>1657</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>1453.8</avg_sunday_holiday_rides>
.....: </rides>
.....: </row>
.....: </response>"""
.....:
In [389]: xsl = """<xsl:stylesheet version="1.0" xmlns:xsl="http://www.w3.org/1999/XSL/Transform">
.....: <xsl:output method="xml" omit-xml-declaration="no" indent="yes"/>
.....: <xsl:strip-space elements="*"/>
.....: <xsl:template match="/response">
.....: <xsl:copy>
.....: <xsl:apply-templates select="row"/>
.....: </xsl:copy>
.....: </xsl:template>
.....: <xsl:template match="row">
.....: <xsl:copy>
.....: <station_id><xsl:value-of select="station/@id"/></station_id>
.....: <station_name><xsl:value-of select="station/@name"/></station_name>
.....: <xsl:copy-of select="month|rides/*"/>
.....: </xsl:copy>
.....: </xsl:template>
.....: </xsl:stylesheet>"""
.....:
In [390]: output = """<?xml version='1.0' encoding='utf-8'?>
.....: <response>
.....: <row>
.....: <station_id>40850</station_id>
.....: <station_name>Library</station_name>
.....: <month>2020-09-01T00:00:00</month>
.....: <avg_weekday_rides>864.2</avg_weekday_rides>
.....: <avg_saturday_rides>534</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>417.2</avg_sunday_holiday_rides>
.....: </row>
.....: <row>
.....: <station_id>41700</station_id>
.....: <station_name>Washington/Wabash</station_name>
.....: <month>2020-09-01T00:00:00</month>
.....: <avg_weekday_rides>2707.4</avg_weekday_rides>
.....: <avg_saturday_rides>1909.8</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>1438.6</avg_sunday_holiday_rides>
.....: </row>
.....: <row>
.....: <station_id>40380</station_id>
.....: <station_name>Clark/Lake</station_name>
.....: <month>2020-09-01T00:00:00</month>
.....: <avg_weekday_rides>2949.6</avg_weekday_rides>
.....: <avg_saturday_rides>1657</avg_saturday_rides>
.....: <avg_sunday_holiday_rides>1453.8</avg_sunday_holiday_rides>
.....: </row>
.....: </response>"""
.....:
In [391]: df = pd.read_xml(xml, stylesheet=xsl)
In [392]: df
Out[392]:
station_id station_name month avg_weekday_rides avg_saturday_rides avg_sunday_holiday_rides
0 40850 Library 2020-09-01T00:00:00 864.2 534.0 417.2
1 41700 Washington/Wabash 2020-09-01T00:00:00 2707.4 1909.8 1438.6
2 40380 Clark/Lake 2020-09-01T00:00:00 2949.6 1657.0 1453.8
Writing XML¶
New in version 1.3.0.
DataFrame objects have an instance method to_xml which renders the
contents of the DataFrame as an XML document.
Note
This method does not support special properties of XML including DTD,
CData, XSD schemas, processing instructions, comments, and others.
Only namespaces at the root level is supported. However, stylesheet
allows design changes after initial output.
Let’s look at a few examples.
Write an XML without options:
In [393]: geom_df = pd.DataFrame(
.....: {
.....: "shape": ["square", "circle", "triangle"],
.....: "degrees": [360, 360, 180],
.....: "sides": [4, np.nan, 3],
.....: }
.....: )
.....:
In [394]: print(geom_df.to_xml())
<?xml version='1.0' encoding='utf-8'?>
<data>
<row>
<index>0</index>
<shape>square</shape>
<degrees>360</degrees>
<sides>4.0</sides>
</row>
<row>
<index>1</index>
<shape>circle</shape>
<degrees>360</degrees>
<sides/>
</row>
<row>
<index>2</index>
<shape>triangle</shape>
<degrees>180</degrees>
<sides>3.0</sides>
</row>
</data>
Write an XML with new root and row name:
In [395]: print(geom_df.to_xml(root_name="geometry", row_name="objects"))
<?xml version='1.0' encoding='utf-8'?>
<geometry>
<objects>
<index>0</index>
<shape>square</shape>
<degrees>360</degrees>
<sides>4.0</sides>
</objects>
<objects>
<index>1</index>
<shape>circle</shape>
<degrees>360</degrees>
<sides/>
</objects>
<objects>
<index>2</index>
<shape>triangle</shape>
<degrees>180</degrees>
<sides>3.0</sides>
</objects>
</geometry>
Write an attribute-centric XML:
In [396]: print(geom_df.to_xml(attr_cols=geom_df.columns.tolist()))
<?xml version='1.0' encoding='utf-8'?>
<data>
<row index="0" shape="square" degrees="360" sides="4.0"/>
<row index="1" shape="circle" degrees="360"/>
<row index="2" shape="triangle" degrees="180" sides="3.0"/>
</data>
Write a mix of elements and attributes:
In [397]: print(
.....: geom_df.to_xml(
.....: index=False,
.....: attr_cols=['shape'],
.....: elem_cols=['degrees', 'sides'])
.....: )
.....:
<?xml version='1.0' encoding='utf-8'?>
<data>
<row shape="square">
<degrees>360</degrees>
<sides>4.0</sides>
</row>
<row shape="circle">
<degrees>360</degrees>
<sides/>
</row>
<row shape="triangle">
<degrees>180</degrees>
<sides>3.0</sides>
</row>
</data>
Any DataFrames with hierarchical columns will be flattened for XML element names
with levels delimited by underscores:
In [398]: ext_geom_df = pd.DataFrame(
.....: {
.....: "type": ["polygon", "other", "polygon"],
.....: "shape": ["square", "circle", "triangle"],
.....: "degrees": [360, 360, 180],
.....: "sides": [4, np.nan, 3],
.....: }
.....: )
.....:
In [399]: pvt_df = ext_geom_df.pivot_table(index='shape',
.....: columns='type',
.....: values=['degrees', 'sides'],
.....: aggfunc='sum')
.....:
In [400]: pvt_df
Out[400]:
degrees sides
type other polygon other polygon
shape
circle 360.0 NaN 0.0 NaN
square NaN 360.0 NaN 4.0
triangle NaN 180.0 NaN 3.0
In [401]: print(pvt_df.to_xml())
<?xml version='1.0' encoding='utf-8'?>
<data>
<row>
<shape>circle</shape>
<degrees_other>360.0</degrees_other>
<degrees_polygon/>
<sides_other>0.0</sides_other>
<sides_polygon/>
</row>
<row>
<shape>square</shape>
<degrees_other/>
<degrees_polygon>360.0</degrees_polygon>
<sides_other/>
<sides_polygon>4.0</sides_polygon>
</row>
<row>
<shape>triangle</shape>
<degrees_other/>
<degrees_polygon>180.0</degrees_polygon>
<sides_other/>
<sides_polygon>3.0</sides_polygon>
</row>
</data>
Write an XML with default namespace:
In [402]: print(geom_df.to_xml(namespaces={"": "https://example.com"}))
<?xml version='1.0' encoding='utf-8'?>
<data xmlns="https://example.com">
<row>
<index>0</index>
<shape>square</shape>
<degrees>360</degrees>
<sides>4.0</sides>
</row>
<row>
<index>1</index>
<shape>circle</shape>
<degrees>360</degrees>
<sides/>
</row>
<row>
<index>2</index>
<shape>triangle</shape>
<degrees>180</degrees>
<sides>3.0</sides>
</row>
</data>
Write an XML with namespace prefix:
In [403]: print(
.....: geom_df.to_xml(namespaces={"doc": "https://example.com"},
.....: prefix="doc")
.....: )
.....:
<?xml version='1.0' encoding='utf-8'?>
<doc:data xmlns:doc="https://example.com">
<doc:row>
<doc:index>0</doc:index>
<doc:shape>square</doc:shape>
<doc:degrees>360</doc:degrees>
<doc:sides>4.0</doc:sides>
</doc:row>
<doc:row>
<doc:index>1</doc:index>
<doc:shape>circle</doc:shape>
<doc:degrees>360</doc:degrees>
<doc:sides/>
</doc:row>
<doc:row>
<doc:index>2</doc:index>
<doc:shape>triangle</doc:shape>
<doc:degrees>180</doc:degrees>
<doc:sides>3.0</doc:sides>
</doc:row>
</doc:data>
Write an XML without declaration or pretty print:
In [404]: print(
.....: geom_df.to_xml(xml_declaration=False,
.....: pretty_print=False)
.....: )
.....:
<data><row><index>0</index><shape>square</shape><degrees>360</degrees><sides>4.0</sides></row><row><index>1</index><shape>circle</shape><degrees>360</degrees><sides/></row><row><index>2</index><shape>triangle</shape><degrees>180</degrees><sides>3.0</sides></row></data>
Write an XML and transform with stylesheet:
In [405]: xsl = """<xsl:stylesheet version="1.0" xmlns:xsl="http://www.w3.org/1999/XSL/Transform">
.....: <xsl:output method="xml" omit-xml-declaration="no" indent="yes"/>
.....: <xsl:strip-space elements="*"/>
.....: <xsl:template match="/data">
.....: <geometry>
.....: <xsl:apply-templates select="row"/>
.....: </geometry>
.....: </xsl:template>
.....: <xsl:template match="row">
.....: <object index="{index}">
.....: <xsl:if test="shape!='circle'">
.....: <xsl:attribute name="type">polygon</xsl:attribute>
.....: </xsl:if>
.....: <xsl:copy-of select="shape"/>
.....: <property>
.....: <xsl:copy-of select="degrees|sides"/>
.....: </property>
.....: </object>
.....: </xsl:template>
.....: </xsl:stylesheet>"""
.....:
In [406]: print(geom_df.to_xml(stylesheet=xsl))
<?xml version="1.0"?>
<geometry>
<object index="0" type="polygon">
<shape>square</shape>
<property>
<degrees>360</degrees>
<sides>4.0</sides>
</property>
</object>
<object index="1">
<shape>circle</shape>
<property>
<degrees>360</degrees>
<sides/>
</property>
</object>
<object index="2" type="polygon">
<shape>triangle</shape>
<property>
<degrees>180</degrees>
<sides>3.0</sides>
</property>
</object>
</geometry>
XML Final Notes¶
All XML documents adhere to W3C specifications. Both
etreeandlxmlparsers will fail to parse any markup document that is not well-formed or follows XML syntax rules. Do be aware HTML is not an XML document unless it follows XHTML specs. However, other popular markup types including KML, XAML, RSS, MusicML, MathML are compliant XML schemas.For above reason, if your application builds XML prior to pandas operations, use appropriate DOM libraries like
etreeandlxmlto build the necessary document and not by string concatenation or regex adjustments. Always remember XML is a special text file with markup rules.With very large XML files (several hundred MBs to GBs), XPath and XSLT can become memory-intensive operations. Be sure to have enough available RAM for reading and writing to large XML files (roughly about 5 times the size of text).
Because XSLT is a programming language, use it with caution since such scripts can pose a security risk in your environment and can run large or infinite recursive operations. Always test scripts on small fragments before full run.
The etree parser supports all functionality of both
read_xmlandto_xmlexcept for complex XPath and any XSLT. Though limited in features,etreeis still a reliable and capable parser and tree builder. Its performance may traillxmlto a certain degree for larger files but relatively unnoticeable on small to medium size files.
Excel files¶
The read_excel() method can read Excel 2007+ (.xlsx) files
using the openpyxl Python module. Excel 2003 (.xls) files
can be read using xlrd. Binary Excel (.xlsb)
files can be read using pyxlsb.
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.
Warning
The xlwt package for writing old-style .xls
excel files is no longer maintained.
The xlrd package is now only for reading
old-style .xls files.
Before pandas 1.3.0, the default argument engine=None to read_excel()
would result in using the xlrd engine in many cases, including new
Excel 2007+ (.xlsx) files. pandas will now default to using the
openpyxl engine.
It is strongly encouraged to install openpyxl to read Excel 2007+
(.xlsx) files.
Please do not report issues when using ``xlrd`` to read ``.xlsx`` files.
This is no longer supported, switch to using openpyxl instead.
Attempting to use the the xlwt engine will raise a FutureWarning
unless the option io.excel.xls.writer is set to "xlwt".
While this option is now deprecated and will also raise a FutureWarning,
it can be globally set and the warning suppressed. Users are recommended to
write .xlsx files using the openpyxl engine instead.
Reading Excel files¶
In the most basic use-case, read_excel takes a path to an Excel
file, and the sheet_name indicating which sheet to parse.
# Returns a DataFrame
pd.read_excel("path_to_file.xls", sheet_name="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 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"] = pd.read_excel(xls, "Sheet1", index_col=None, na_values=["NA"])
data["Sheet2"] = pd.read_excel(xls, "Sheet2", index_col=None, na_values=["NA"])
# equivalent using the read_excel function
data = pd.read_excel(
"path_to_file.xls", ["Sheet1", "Sheet2"], index_col=None, na_values=["NA"]
)
ExcelFile can also be called with a xlrd.book.Book object
as a parameter. This allows the user to control how the excel file is read.
For example, sheets can be loaded on demand by calling xlrd.open_workbook()
with on_demand=True.
import xlrd
xlrd_book = xlrd.open_workbook("path_to_file.xls", on_demand=True)
with pd.ExcelFile(xlrd_book) as xls:
df1 = pd.read_excel(xls, "Sheet1")
df2 = pd.read_excel(xls, "Sheet2")
Specifying sheets¶
Note
The second argument is sheet_name, not to be confused with ExcelFile.sheet_names.
Note
An ExcelFile’s attribute sheet_names provides access to a list of sheets.
The arguments
sheet_nameallows specifying the sheet or sheets to read.The default value for
sheet_nameis 0, indicating to read the first sheetPass 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
Noneto return a dictionary of all available sheets.
# Returns a DataFrame
pd.read_excel("path_to_file.xls", "Sheet1", index_col=None, na_values=["NA"])
Using the sheet index:
# Returns a DataFrame
pd.read_excel("path_to_file.xls", 0, index_col=None, na_values=["NA"])
Using all default values:
# Returns a DataFrame
pd.read_excel("path_to_file.xls")
Using None to get all sheets:
# Returns a dictionary of DataFrames
pd.read_excel("path_to_file.xls", sheet_name=None)
Using a list to get multiple sheets:
# Returns the 1st and 4th sheet, as a dictionary of DataFrames.
pd.read_excel("path_to_file.xls", sheet_name=["Sheet1", 3])
read_excel can read more than one sheet, by setting sheet_name to either
a list of sheet names, a list of sheet positions, or None to read all sheets.
Sheets can be specified by sheet index or sheet name, using an integer or string,
respectively.
Reading a MultiIndex¶
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 [407]: df = pd.DataFrame(
.....: {"a": [1, 2, 3, 4], "b": [5, 6, 7, 8]},
.....: index=pd.MultiIndex.from_product([["a", "b"], ["c", "d"]]),
.....: )
.....:
In [408]: df.to_excel("path_to_file.xlsx")
In [409]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1])
In [410]: df
Out[410]:
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 [411]: df.index = df.index.set_names(["lvl1", "lvl2"])
In [412]: df.to_excel("path_to_file.xlsx")
In [413]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1])
In [414]: df
Out[414]:
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 [415]: df.columns = pd.MultiIndex.from_product([["a"], ["b", "d"]], names=["c1", "c2"])
In [416]: df.to_excel("path_to_file.xlsx")
In [417]: df = pd.read_excel("path_to_file.xlsx", index_col=[0, 1], header=[0, 1])
In [418]: df
Out[418]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8
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 usecols keyword to allow you to specify a subset of columns to parse.
Changed in version 1.0.0.
Passing in an integer for usecols will no longer work. Please pass in a list
of ints from 0 to usecols inclusive instead.
You can specify a comma-delimited set of Excel columns and ranges as a string:
pd.read_excel("path_to_file.xls", "Sheet1", usecols="A,C:E")
If usecols is a list of integers, then it is assumed to be the file column
indices to be parsed.
pd.read_excel("path_to_file.xls", "Sheet1", usecols=[0, 2, 3])
Element order is ignored, so usecols=[0, 1] is the same as [1, 0].
If usecols is a list of strings, it is assumed that each string corresponds
to a column name provided either by the user in names or inferred from the
document header row(s). Those strings define which columns will be parsed:
pd.read_excel("path_to_file.xls", "Sheet1", usecols=["foo", "bar"])
Element order is ignored, so usecols=['baz', 'joe'] is the same as ['joe', 'baz'].
If usecols is callable, the callable function will be evaluated against
the column names, returning names where the callable function evaluates to True.
pd.read_excel("path_to_file.xls", "Sheet1", usecols=lambda x: x.isalpha())
Parsing dates¶
Datetime-like values are normally automatically converted to the appropriate
dtype when reading the excel file. But if you have a column of strings that
look like dates (but are not actually formatted as dates in excel), you can
use the parse_dates keyword to parse those strings to datetimes:
pd.read_excel("path_to_file.xls", "Sheet1", parse_dates=["date_strings"])
Cell converters¶
It is possible to transform the contents of Excel cells via the converters
option. For instance, to convert a column to boolean:
pd.read_excel("path_to_file.xls", "Sheet1", converters={"MyBools": bool})
This options handles missing values and treats exceptions in the converters as missing data. Transformations are applied cell by cell rather than to the column as a whole, so the array dtype is not guaranteed. For instance, a column of integers with missing values cannot be transformed to an array with integer dtype, because NaN is strictly a float. You can manually mask missing data to recover integer dtype:
def cfun(x):
return int(x) if x else -1
pd.read_excel("path_to_file.xls", "Sheet1", converters={"MyInts": cfun})
Dtype specifications¶
As an alternative to converters, the type for an entire column can
be specified using the dtype keyword, which takes a dictionary
mapping column names to types. To interpret data with
no type inference, use the type str or object.
pd.read_excel("path_to_file.xls", dtype={"MyInts": "int64", "MyText": str})
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.
The index_label will be placed in the second
row instead of the first. You can place it in the first row by setting the
merge_cells option in to_excel() to False:
df.to_excel("path_to_file.xlsx", index_label="label", merge_cells=False)
In order to write separate DataFrames to separate sheets in a single Excel file,
one can pass an ExcelWriter.
with pd.ExcelWriter("path_to_file.xlsx") as writer:
df1.to_excel(writer, sheet_name="Sheet1")
df2.to_excel(writer, sheet_name="Sheet2")
Writing Excel files to memory¶
pandas supports writing Excel files to buffer-like objects such as StringIO or
BytesIO using ExcelWriter.
from io import BytesIO
bio = BytesIO()
# By setting the 'engine' in the ExcelWriter constructor.
writer = pd.ExcelWriter(bio, engine="xlsxwriter")
df.to_excel(writer, sheet_name="Sheet1")
# Save the workbook
writer.save()
# Seek to the beginning and read to copy the workbook to a variable in memory
bio.seek(0)
workbook = bio.read()
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¶
Deprecated since version 1.2.0: As the xlwt package is no longer
maintained, the xlwt engine will be removed from a future version
of pandas. This is the only engine in pandas that supports writing to
.xls files.
pandas chooses an Excel writer via two methods:
the
enginekeyword argumentthe filename extension (via the default specified in config options)
By default, pandas uses the XlsxWriter for .xlsx, openpyxl
for .xlsm, and xlwt for .xls files. If you have multiple
engines installed, you can set the default engine through setting the
config options io.excel.xlsx.writer and
io.excel.xls.writer. pandas will fall back on openpyxl for .xlsx
files if Xlsxwriter is not available.
To specify which writer you want to use, you can pass an engine keyword
argument to to_excel and to ExcelWriter. The built-in engines are:
openpyxl: version 2.4 or higher is requiredxlsxwriterxlwt
# By setting the 'engine' in the DataFrame 'to_excel()' methods.
df.to_excel("path_to_file.xlsx", sheet_name="Sheet1", engine="xlsxwriter")
# By setting the 'engine' in the ExcelWriter constructor.
writer = pd.ExcelWriter("path_to_file.xlsx", engine="xlsxwriter")
# Or via pandas configuration.
from pandas import options # noqa: E402
options.io.excel.xlsx.writer = "xlsxwriter"
df.to_excel("path_to_file.xlsx", sheet_name="Sheet1")
Style and formatting¶
The look and feel of Excel worksheets created from pandas can be modified using the following parameters on the DataFrame’s to_excel method.
float_format: Format string for floating point numbers (defaultNone).freeze_panes: A tuple of two integers representing the bottommost row and rightmost column to freeze. Each of these parameters is one-based, so (1, 1) will freeze the first row and first column (defaultNone).
Using the Xlsxwriter engine provides many options for controlling the
format of an Excel worksheet created with the to_excel method. Excellent examples can be found in the
Xlsxwriter documentation here: https://xlsxwriter.readthedocs.io/working_with_pandas.html
OpenDocument Spreadsheets¶
New in version 0.25.
The read_excel() method can also read OpenDocument spreadsheets
using the odfpy module. The semantics and features for reading
OpenDocument spreadsheets match what can be done for Excel files using
engine='odf'.
# Returns a DataFrame
pd.read_excel("path_to_file.ods", engine="odf")
Note
Currently pandas only supports reading OpenDocument spreadsheets. Writing is not implemented.
Binary Excel (.xlsb) files¶
New in version 1.0.0.
The read_excel() method can also read binary Excel files
using the pyxlsb module. The semantics and features for reading
binary Excel files mostly match what can be done for Excel files using
engine='pyxlsb'. pyxlsb does not recognize datetime types
in files and will return floats instead.
# Returns a DataFrame
pd.read_excel("path_to_file.xlsb", engine="pyxlsb")
Note
Currently pandas only supports reading binary Excel files. Writing is not implemented.
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_csv 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()
>>> clipdf
A B C
x 1 4 p
y 2 5 q
z 3 6 r
The to_clipboard method can be used to write the contents of a DataFrame to
the clipboard. Following which you can paste the clipboard contents into other
applications (CTRL-V on many operating systems). Here we illustrate writing a
DataFrame into clipboard and reading it back.
>>> df = pd.DataFrame(
... {"A": [1, 2, 3], "B": [4, 5, 6], "C": ["p", "q", "r"]}, index=["x", "y", "z"]
... )
>>> df
A B C
x 1 4 p
y 2 5 q
z 3 6 r
>>> df.to_clipboard()
>>> pd.read_clipboard()
A B C
x 1 4 p
y 2 5 q
z 3 6 r
We can see that we got the same content back, which we had earlier written to the clipboard.
Note
You may need to install xclip or xsel (with PyQt5, PyQt4 or qtpy) 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 [419]: df
Out[419]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8
In [420]: 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 [421]: pd.read_pickle("foo.pkl")
Out[421]:
c1 a
c2 b d
lvl1 lvl2
a c 1 5
d 2 6
b c 3 7
d 4 8
Warning
Loading pickled data received from untrusted sources can be unsafe.
Warning
read_pickle() is only guaranteed backwards compatible back to pandas version 0.20.3
Compressed pickle files¶
read_pickle(), DataFrame.to_pickle() and Series.to_pickle() can read
and write compressed pickle files. The compression types of gzip, bz2, xz, zstd are supported for reading and writing.
The zip file format only supports reading and must contain only one data file
to be read.
The compression type can be an explicit parameter or be inferred from the file extension.
If ‘infer’, then use gzip, bz2, zip, xz, zstd if filename ends in '.gz', '.bz2', '.zip',
'.xz', or '.zst', respectively.
The compression parameter can also be a dict in order to pass options to the
compression protocol. It must have a 'method' key set to the name
of the compression protocol, which must be one of
{'zip', 'gzip', 'bz2', 'xz', 'zstd'}. All other key-value pairs are passed to
the underlying compression library.
In [422]: df = pd.DataFrame(
.....: {
.....: "A": np.random.randn(1000),
.....: "B": "foo",
.....: "C": pd.date_range("20130101", periods=1000, freq="s"),
.....: }
.....: )
.....:
In [423]: df
Out[423]:
A B C
0 -0.828876 foo 2013-01-01 00:00:00
1 -0.110383 foo 2013-01-01 00:00:01
2 2.357598 foo 2013-01-01 00:00:02
3 -1.620073 foo 2013-01-01 00:00:03
4 0.440903 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.177365 foo 2013-01-01 00:16:35
996 1.236988 foo 2013-01-01 00:16:36
997 0.743946 foo 2013-01-01 00:16:37
998 -0.533097 foo 2013-01-01 00:16:38
999 -0.140850 foo 2013-01-01 00:16:39
[1000 rows x 3 columns]
Using an explicit compression type:
In [424]: df.to_pickle("data.pkl.compress", compression="gzip")
In [425]: rt = pd.read_pickle("data.pkl.compress", compression="gzip")
In [426]: rt
Out[426]:
A B C
0 -0.828876 foo 2013-01-01 00:00:00
1 -0.110383 foo 2013-01-01 00:00:01
2 2.357598 foo 2013-01-01 00:00:02
3 -1.620073 foo 2013-01-01 00:00:03
4 0.440903 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.177365 foo 2013-01-01 00:16:35
996 1.236988 foo 2013-01-01 00:16:36
997 0.743946 foo 2013-01-01 00:16:37
998 -0.533097 foo 2013-01-01 00:16:38
999 -0.140850 foo 2013-01-01 00:16:39
[1000 rows x 3 columns]
Inferring compression type from the extension:
In [427]: df.to_pickle("data.pkl.xz", compression="infer")
In [428]: rt = pd.read_pickle("data.pkl.xz", compression="infer")
In [429]: rt
Out[429]:
A B C
0 -0.828876 foo 2013-01-01 00:00:00
1 -0.110383 foo 2013-01-01 00:00:01
2 2.357598 foo 2013-01-01 00:00:02
3 -1.620073 foo 2013-01-01 00:00:03
4 0.440903 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.177365 foo 2013-01-01 00:16:35
996 1.236988 foo 2013-01-01 00:16:36
997 0.743946 foo 2013-01-01 00:16:37
998 -0.533097 foo 2013-01-01 00:16:38
999 -0.140850 foo 2013-01-01 00:16:39
[1000 rows x 3 columns]
The default is to ‘infer’:
In [430]: df.to_pickle("data.pkl.gz")
In [431]: rt = pd.read_pickle("data.pkl.gz")
In [432]: rt
Out[432]:
A B C
0 -0.828876 foo 2013-01-01 00:00:00
1 -0.110383 foo 2013-01-01 00:00:01
2 2.357598 foo 2013-01-01 00:00:02
3 -1.620073 foo 2013-01-01 00:00:03
4 0.440903 foo 2013-01-01 00:00:04
.. ... ... ...
995 -1.177365 foo 2013-01-01 00:16:35
996 1.236988 foo 2013-01-01 00:16:36
997 0.743946 foo 2013-01-01 00:16:37
998 -0.533097 foo 2013-01-01 00:16:38
999 -0.140850 foo 2013-01-01 00:16:39
[1000 rows x 3 columns]
In [433]: df["A"].to_pickle("s1.pkl.bz2")
In [434]: rt = pd.read_pickle("s1.pkl.bz2")
In [435]: rt
Out[435]:
0 -0.828876
1 -0.110383
2 2.357598
3 -1.620073
4 0.440903
...
995 -1.177365
996 1.236988
997 0.743946
998 -0.533097
999 -0.140850
Name: A, Length: 1000, dtype: float64
Passing options to the compression protocol in order to speed up compression:
In [436]: df.to_pickle("data.pkl.gz", compression={"method": "gzip", "compresslevel": 1})
msgpack¶
pandas support for msgpack has been removed in version 1.0.0. It is
recommended to use pickle instead.
Alternatively, you can also the Arrow IPC serialization format for on-the-wire transmission of pandas objects. For documentation on pyarrow, see here.
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
pandas uses PyTables for reading and writing HDF5 files, which allows serializing object-dtype data with pickle. Loading pickled data received from untrusted sources can be unsafe.
See: https://docs.python.org/3/library/pickle.html for more.
In [437]: store = pd.HDFStore("store.h5")
In [438]: print(store)
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
Objects can be written to the file just like adding key-value pairs to a dict:
In [439]: index = pd.date_range("1/1/2000", periods=8)
In [440]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [441]: df = pd.DataFrame(np.random.randn(8, 3), index=index, columns=["A", "B", "C"])
# store.put('s', s) is an equivalent method
In [442]: store["s"] = s
In [443]: store["df"] = df
In [444]: store
Out[444]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
In a current or later Python session, you can retrieve stored objects:
# store.get('df') is an equivalent method
In [445]: store["df"]
Out[445]:
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
# dotted (attribute) access provides get as well
In [446]: store.df
Out[446]:
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
Deletion of the object specified by the key:
# store.remove('df') is an equivalent method
In [447]: del store["df"]
In [448]: store
Out[448]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
Closing a Store and using a context manager:
In [449]: store.close()
In [450]: store
Out[450]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
In [451]: store.is_open
Out[451]: False
# Working with, and automatically closing the store using a context manager
In [452]: with pd.HDFStore("store.h5") as store:
.....: store.keys()
.....:
Read/write API¶
HDFStore supports a top-level API using read_hdf for reading and to_hdf for writing,
similar to how read_csv and to_csv work.
In [453]: df_tl = pd.DataFrame({"A": list(range(5)), "B": list(range(5))})
In [454]: df_tl.to_hdf("store_tl.h5", "table", append=True)
In [455]: pd.read_hdf("store_tl.h5", "table", where=["index>2"])
Out[455]:
A B
3 3 3
4 4 4
HDFStore will by default not drop rows that are all missing. This behavior can be changed by setting dropna=True.
In [456]: df_with_missing = pd.DataFrame(
.....: {
.....: "col1": [0, np.nan, 2],
.....: "col2": [1, np.nan, np.nan],
.....: }
.....: )
.....:
In [457]: df_with_missing
Out[457]:
col1 col2
0 0.0 1.0
1 NaN NaN
2 2.0 NaN
In [458]: df_with_missing.to_hdf("file.h5", "df_with_missing", format="table", mode="w")
In [459]: pd.read_hdf("file.h5", "df_with_missing")
Out[459]:
col1 col2
0 0.0 1.0
1 NaN NaN
2 2.0 NaN
In [460]: df_with_missing.to_hdf(
.....: "file.h5", "df_with_missing", format="table", mode="w", dropna=True
.....: )
.....:
In [461]: pd.read_hdf("file.h5", "df_with_missing")
Out[461]:
col1 col2
0 0.0 1.0
2 2.0 NaN
Fixed 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 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:
>>> pd.DataFrame(np.random.randn(10, 2)).to_hdf("test_fixed.h5", "df")
>>> pd.read_hdf("test_fixed.h5", "df", where="index>5")
TypeError: cannot pass a where specification when reading a fixed format.
this store must be selected in its entirety
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 and query type operations are
supported. This format is specified by format='table' or format='t'
to append or put or to_hdf.
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 [462]: store = pd.HDFStore("store.h5")
In [463]: df1 = df[0:4]
In [464]: df2 = df[4:]
# append data (creates a table automatically)
In [465]: store.append("df", df1)
In [466]: store.append("df", df2)
In [467]: store
Out[467]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
# select the entire object
In [468]: store.select("df")
Out[468]:
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
# the type of stored data
In [469]: store.root.df._v_attrs.pandas_type
Out[469]: '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 without the leading ‘/’ and are always
absolute (e.g. ‘foo’ refers to ‘/foo’). Removal operations can remove
everything in the sub-store and below, so be careful.
In [470]: store.put("foo/bar/bah", df)
In [471]: store.append("food/orange", df)
In [472]: store.append("food/apple", df)
In [473]: store
Out[473]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
# a list of keys are returned
In [474]: store.keys()
Out[474]: ['/df', '/food/apple', '/food/orange', '/foo/bar/bah']
# remove all nodes under this level
In [475]: store.remove("food")
In [476]: store
Out[476]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
You can walk through the group hierarchy using the walk method which
will yield a tuple for each group key along with the relative keys of its contents.
In [477]: for (path, subgroups, subkeys) in store.walk():
.....: for subgroup in subgroups:
.....: print("GROUP: {}/{}".format(path, subgroup))
.....: for subkey in subkeys:
.....: key = "/".join([path, subkey])
.....: print("KEY: {}".format(key))
.....: print(store.get(key))
.....:
GROUP: /foo
KEY: /df
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
GROUP: /foo/bar
KEY: /foo/bar/bah
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
Warning
Hierarchical keys cannot be retrieved as dotted (attribute) access as described above for items stored under the root node.
In [8]: store.foo.bar.bah
AttributeError: 'HDFStore' object has no attribute 'foo'
# you can directly access the actual PyTables node but using the root node
In [9]: store.root.foo.bar.bah
Out[9]:
/foo/bar/bah (Group) ''
children := ['block0_items' (Array), 'block0_values' (Array), 'axis0' (Array), 'axis1' (Array)]
Instead, use explicit string based keys:
In [478]: store["foo/bar/bah"]
Out[478]:
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
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 attempts
at appending longer strings will raise a ValueError.
Passing min_itemsize={`values`: size} as a parameter to append
will set a larger minimum for the string columns. Storing floats,
strings, ints, bools, datetime64 are currently supported. For string
columns, passing nan_rep = 'nan' to append will change the default
nan representation on disk (which converts to/from np.nan), this
defaults to nan.
In [479]: df_mixed = pd.DataFrame(
.....: {
.....: "A": np.random.randn(8),
.....: "B": np.random.randn(8),
.....: "C": np.array(np.random.randn(8), dtype="float32"),
.....: "string": "string",
.....: "int": 1,
.....: "bool": True,
.....: "datetime64": pd.Timestamp("20010102"),
.....: },
.....: index=list(range(8)),
.....: )
.....:
In [480]: df_mixed.loc[df_mixed.index[3:5], ["A", "B", "string", "datetime64"]] = np.nan
In [481]: store.append("df_mixed", df_mixed, min_itemsize={"values": 50})
In [482]: df_mixed1 = store.select("df_mixed")
In [483]: df_mixed1
Out[483]:
A B C string int bool datetime64
0 1.778161 -0.898283 -0.263043 string 1 True 2001-01-02
1 -0.913867 -0.218499 -0.639244 string 1 True 2001-01-02
2 -0.030004 1.408028 -0.866305 string 1 True 2001-01-02
3 NaN NaN -0.225250 NaN 1 True NaT
4 NaN NaN -0.890978 NaN 1 True NaT
5 0.081323 0.520995 -0.553839 string 1 True 2001-01-02
6 -0.268494 0.620028 -2.762875 string 1 True 2001-01-02
7 0.168016 0.159416 -1.244763 string 1 True 2001-01-02
In [484]: df_mixed1.dtypes.value_counts()
Out[484]:
float64 2
float32 1
object 1
int64 1
bool 1
datetime64[ns] 1
dtype: int64
# we have provided a minimum string column size
In [485]: store.root.df_mixed.table
Out[485]:
/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": StringCol(itemsize=50, shape=(1,), dflt=b'', 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": Int64Col(shape=(1,), dflt=0, pos=6)}
byteorder := 'little'
chunkshape := (689,)
autoindex := True
colindexes := {
"index": Index(6, mediumshuffle, zlib(1)).is_csi=False}
Storing MultiIndex DataFrames¶
Storing MultiIndex DataFrames as tables is very similar to
storing/selecting from homogeneous index DataFrames.
In [486]: index = pd.MultiIndex(
.....: levels=[["foo", "bar", "baz", "qux"], ["one", "two", "three"]],
.....: codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
.....: names=["foo", "bar"],
.....: )
.....:
In [487]: df_mi = pd.DataFrame(np.random.randn(10, 3), index=index, columns=["A", "B", "C"])
In [488]: df_mi
Out[488]:
A B C
foo bar
foo one -1.280289 0.692545 -0.536722
two 1.005707 0.296917 0.139796
three -1.083889 0.811865 1.648435
bar one -0.164377 -0.402227 1.618922
two -1.424723 -0.023232 0.948196
baz two 0.183573 0.145277 0.308146
three -1.043530 -0.708145 1.430905
qux one -0.850136 0.813949 1.508891
two -1.556154 0.187597 1.176488
three -1.246093 -0.002726 -0.444249
In [489]: store.append("df_mi", df_mi)
In [490]: store.select("df_mi")
Out[490]:
A B C
foo bar
foo one -1.280289 0.692545 -0.536722
two 1.005707 0.296917 0.139796
three -1.083889 0.811865 1.648435
bar one -0.164377 -0.402227 1.618922
two -1.424723 -0.023232 0.948196
baz two 0.183573 0.145277 0.308146
three -1.043530 -0.708145 1.430905
qux one -0.850136 0.813949 1.508891
two -1.556154 0.187597 1.176488
three -1.246093 -0.002726 -0.444249
# the levels are automatically included as data columns
In [491]: store.select("df_mi", "foo=bar")
Out[491]:
A B C
foo bar
bar one -0.164377 -0.402227 1.618922
two -1.424723 -0.023232 0.948196
Note
The index keyword is reserved and cannot be use as a level name.
Querying¶
Querying a table¶
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.
indexandcolumnsare supported indexers ofDataFrames.if
data_columnsare specified, these can be used as additional indexers.level name in a MultiIndex, with default name
level_0,level_1, … if not provided.
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 circumstancesIf 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', f'index == {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 [492]: dfq = pd.DataFrame(
.....: np.random.randn(10, 4),
.....: columns=list("ABCD"),
.....: index=pd.date_range("20130101", periods=10),
.....: )
.....:
In [493]: store.append("dfq", dfq, format="table", data_columns=True)
Use boolean expressions, with in-line function evaluation.
In [494]: store.select("dfq", "index>pd.Timestamp('20130104') & columns=['A', 'B']")
Out[494]:
A B
2013-01-05 1.366810 1.073372
2013-01-06 2.119746 -2.628174
2013-01-07 0.337920 -0.634027
2013-01-08 1.053434 1.109090
2013-01-09 -0.772942 -0.269415
2013-01-10 0.048562 -0.285920
Use inline column reference.
In [495]: store.select("dfq", where="A>0 or C>0")
Out[495]:
A B C D
2013-01-01 0.856838 1.491776 0.001283 0.701816
2013-01-02 -1.097917 0.102588 0.661740 0.443531
2013-01-03 0.559313 -0.459055 -1.222598 -0.455304
2013-01-05 1.366810 1.073372 -0.994957 0.755314
2013-01-06 2.119746 -2.628174 -0.089460 -0.133636
2013-01-07 0.337920 -0.634027 0.421107 0.604303
2013-01-08 1.053434 1.109090 -0.367891 -0.846206
2013-01-10 0.048562 -0.285920 1.334100 0.194462
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 [496]: store.select("df", "columns=['A', 'B']")
Out[496]:
A B
2000-01-01 -0.398501 -0.677311
2000-01-02 -1.167564 -0.593353
2000-01-03 -0.131959 0.089012
2000-01-04 0.169405 -1.358046
2000-01-05 0.492195 0.076693
2000-01-06 -0.285283 -1.210529
2000-01-07 0.941577 -0.342447
2000-01-08 0.052607 2.093214
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.
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.
Query timedelta64[ns]¶
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 [497]: from datetime import timedelta
In [498]: dftd = pd.DataFrame(
.....: {
.....: "A": pd.Timestamp("20130101"),
.....: "B": [
.....: pd.Timestamp("20130101") + timedelta(days=i, seconds=10)
.....: for i in range(10)
.....: ],
.....: }
.....: )
.....:
In [499]: dftd["C"] = dftd["A"] - dftd["B"]
In [500]: dftd
Out[500]:
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 [501]: store.append("dftd", dftd, data_columns=True)
In [502]: store.select("dftd", "C<'-3.5D'")
Out[502]:
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
Query MultiIndex¶
Selecting from a MultiIndex can be achieved by using the name of the level.
In [503]: df_mi.index.names
Out[503]: FrozenList(['foo', 'bar'])
In [504]: store.select("df_mi", "foo=baz and bar=two")
Out[504]:
A B C
foo bar
baz two 0.183573 0.145277 0.308146
If the MultiIndex levels names are None, the levels are automatically made available via
the level_n keyword with n the level of the MultiIndex you want to select from.
In [505]: index = pd.MultiIndex(
.....: levels=[["foo", "bar", "baz", "qux"], ["one", "two", "three"]],
.....: codes=[[0, 0, 0, 1, 1, 2, 2, 3, 3, 3], [0, 1, 2, 0, 1, 1, 2, 0, 1, 2]],
.....: )
.....:
In [506]: df_mi_2 = pd.DataFrame(np.random.randn(10, 3), index=index, columns=["A", "B", "C"])
In [507]: df_mi_2
Out[507]:
A B C
foo one -0.646538 1.210676 -0.315409
two 1.528366 0.376542 0.174490
three 1.247943 -0.742283 0.710400
bar one 0.434128 -1.246384 1.139595
two 1.388668 -0.413554 -0.666287
baz two 0.010150 -0.163820 -0.115305
three 0.216467 0.633720 0.473945
qux one -0.155446 1.287082 0.320201
two -1.256989 0.874920 0.765944
three 0.025557 -0.729782 -0.127439
In [508]: store.append("df_mi_2", df_mi_2)
# the levels are automatically included as data columns with keyword level_n
In [509]: store.select("df_mi_2", "level_0=foo and level_1=two")
Out[509]:
A B C
foo two 1.528366 0.376542 0.17449
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 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 [510]: i = store.root.df.table.cols.index.index
In [511]: i.optlevel, i.kind
Out[511]: (6, 'medium')
# change an index by passing new parameters
In [512]: store.create_table_index("df", optlevel=9, kind="full")
In [513]: i = store.root.df.table.cols.index.index
In [514]: i.optlevel, i.kind
Out[514]: (9, 'full')
Oftentimes when appending large amounts of data to a store, it is useful to turn off index creation for each append, then recreate at the end.
In [515]: df_1 = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))
In [516]: df_2 = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))
In [517]: st = pd.HDFStore("appends.h5", mode="w")
In [518]: st.append("df", df_1, data_columns=["B"], index=False)
In [519]: st.append("df", df_2, data_columns=["B"], index=False)
In [520]: st.get_storer("df").table
Out[520]:
/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 [521]: st.create_table_index("df", columns=["B"], optlevel=9, kind="full")
In [522]: st.get_storer("df").table
Out[522]:
/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, fullshuffle, zlib(1)).is_csi=True}
In [523]: 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 [524]: df_dc = df.copy()
In [525]: df_dc["string"] = "foo"
In [526]: df_dc.loc[df_dc.index[4:6], "string"] = np.nan
In [527]: df_dc.loc[df_dc.index[7:9], "string"] = "bar"
In [528]: df_dc["string2"] = "cool"
In [529]: df_dc.loc[df_dc.index[1:3], ["B", "C"]] = 1.0
In [530]: df_dc
Out[530]:
A B C string string2
2000-01-01 -0.398501 -0.677311 -0.874991 foo cool
2000-01-02 -1.167564 1.000000 1.000000 foo cool
2000-01-03 -0.131959 1.000000 1.000000 foo cool
2000-01-04 0.169405 -1.358046 -0.105563 foo cool
2000-01-05 0.492195 0.076693 0.213685 NaN cool
2000-01-06 -0.285283 -1.210529 -1.408386 NaN cool
2000-01-07 0.941577 -0.342447 0.222031 foo cool
2000-01-08 0.052607 2.093214 1.064908 bar cool
# on-disk operations
In [531]: store.append("df_dc", df_dc, data_columns=["B", "C", "string", "string2"])
In [532]: store.select("df_dc", where="B > 0")
Out[532]:
A B C string string2
2000-01-02 -1.167564 1.000000 1.000000 foo cool
2000-01-03 -0.131959 1.000000 1.000000 foo cool
2000-01-05 0.492195 0.076693 0.213685 NaN cool
2000-01-08 0.052607 2.093214 1.064908 bar cool
# getting creative
In [533]: store.select("df_dc", "B > 0 & C > 0 & string == foo")
Out[533]:
A B C string string2
2000-01-02 -1.167564 1.0 1.0 foo cool
2000-01-03 -0.131959 1.0 1.0 foo cool
# this is in-memory version of this type of selection
In [534]: df_dc[(df_dc.B > 0) & (df_dc.C > 0) & (df_dc.string == "foo")]
Out[534]:
A B C string string2
2000-01-02 -1.167564 1.0 1.0 foo cool
2000-01-03 -0.131959 1.0 1.0 foo cool
# we have automagically created this index and the B/C/string/string2
# columns are stored separately as ``PyTables`` columns
In [535]: store.root.df_dc.table
Out[535]:
/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=b'', pos=4),
"string2": StringCol(itemsize=4, shape=(), dflt=b'', pos=5)}
byteorder := 'little'
chunkshape := (1680,)
autoindex := True
colindexes := {
"index": Index(6, mediumshuffle, zlib(1)).is_csi=False,
"B": Index(6, mediumshuffle, zlib(1)).is_csi=False,
"C": Index(6, mediumshuffle, zlib(1)).is_csi=False,
"string": Index(6, mediumshuffle, zlib(1)).is_csi=False,
"string2": Index(6, mediumshuffle, 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¶
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 [536]: for df in store.select("df", chunksize=3):
.....: print(df)
.....:
A B C
2000-01-01 -0.398501 -0.677311 -0.874991
2000-01-02 -1.167564 -0.593353 0.146262
2000-01-03 -0.131959 0.089012 0.667450
A B C
2000-01-04 0.169405 -1.358046 -0.105563
2000-01-05 0.492195 0.076693 0.213685
2000-01-06 -0.285283 -1.210529 -1.408386
A B C
2000-01-07 0.941577 -0.342447 0.222031
2000-01-08 0.052607 2.093214 1.064908
Note
You can also use the iterator with read_hdf which will open, then
automatically close the store when finished iterating.
for df in pd.read_hdf("store.h5", "df", chunksize=3):
print(df)
Note, that the chunksize keyword applies to the source rows. So if you are doing a query, then the chunksize will subdivide the total rows in the table and the query applied, returning an iterator on potentially unequal sized chunks.
Here is a recipe for generating a query and using it to create equal sized return chunks.
In [537]: dfeq = pd.DataFrame({"number": np.arange(1, 11)})
In [538]: dfeq
Out[538]:
number
0 1
1 2
2 3
3 4
4 5
5 6
6 7
7 8
8 9
9 10
In [539]: store.append("dfeq", dfeq, data_columns=["number"])
In [540]: def chunks(l, n):
.....: return [l[i: i + n] for i in range(0, len(l), n)]
.....:
In [541]: evens = [2, 4, 6, 8, 10]
In [542]: coordinates = store.select_as_coordinates("dfeq", "number=evens")
In [543]: 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 [544]: store.select_column("df_dc", "index")
Out[544]:
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 [545]: store.select_column("df_dc", "string")
Out[545]:
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 [546]: df_coord = pd.DataFrame(
.....: np.random.randn(1000, 2), index=pd.date_range("20000101", periods=1000)
.....: )
.....:
In [547]: store.append("df_coord", df_coord)
In [548]: c = store.select_as_coordinates("df_coord", "index > 20020101")
In [549]: c
Out[549]:
Int64Index([732, 733, 734, 735, 736, 737, 738, 739, 740, 741,
...
990, 991, 992, 993, 994, 995, 996, 997, 998, 999],
dtype='int64', length=268)
In [550]: store.select("df_coord", where=c)
Out[550]:
0 1
2002-01-02 0.009035 0.921784
2002-01-03 -1.476563 -1.376375
2002-01-04 1.266731 2.173681
2002-01-05 0.147621 0.616468
2002-01-06 0.008611 2.136001
... ... ...
2002-09-22 0.781169 -0.791687
2002-09-23 -0.764810 -2.000933
2002-09-24 -0.345662 0.393915
2002-09-25 -0.116661 0.834638
2002-09-26 -1.341780 0.686366
[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 [551]: df_mask = pd.DataFrame(
.....: np.random.randn(1000, 2), index=pd.date_range("20000101", periods=1000)
.....: )
.....:
In [552]: store.append("df_mask", df_mask)
In [553]: c = store.select_column("df_mask", "index")
In [554]: where = c[pd.DatetimeIndex(c).month == 5].index
In [555]: store.select("df_mask", where=where)
Out[555]:
0 1
2000-05-01 -0.386742 -0.977433
2000-05-02 -0.228819 0.471671
2000-05-03 0.337307 1.840494
2000-05-04 0.050249 0.307149
2000-05-05 -0.802947 -0.946730
... ... ...
2002-05-27 1.605281 1.741415
2002-05-28 -0.804450 -0.715040
2002-05-29 -0.874851 0.037178
2002-05-30 -0.161167 -1.294944
2002-05-31 -0.258463 -0.731969
[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 [556]: store.get_storer("df_dc").nrows
Out[556]: 8
Multiple table queries¶
The methods append_to_multiple and
select_as_multiple can perform appending/selecting from
multiple tables at once. The idea is to have one table (call it the
selector table) that you index most/all of the columns, and perform your
queries. The other table(s) are data tables with an index matching the
selector table’s index. You can then perform a very fast query
on the selector table, yet get lots of data back. This method is similar to
having a very wide table, but enables more efficient queries.
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 [557]: df_mt = pd.DataFrame(
.....: np.random.randn(8, 6),
.....: index=pd.date_range("1/1/2000", periods=8),
.....: columns=["A", "B", "C", "D", "E", "F"],
.....: )
.....:
In [558]: df_mt["foo"] = "bar"
In [559]: df_mt.loc[df_mt.index[1], ("A", "B")] = np.nan
# you can also create the tables individually
In [560]: store.append_to_multiple(
.....: {"df1_mt": ["A", "B"], "df2_mt": None}, df_mt, selector="df1_mt"
.....: )
.....:
In [561]: store
Out[561]:
<class 'pandas.io.pytables.HDFStore'>
File path: store.h5
# individual tables were created
In [562]: store.select("df1_mt")
Out[562]:
A B
2000-01-01 0.079529 -1.459471
2000-01-02 NaN NaN
2000-01-03 -0.423113 2.314361
2000-01-04 0.756744 -0.792372
2000-01-05 -0.184971 0.170852
2000-01-06 0.678830 0.633974
2000-01-07 0.034973 0.974369
2000-01-08 -2.110103 0.243062
In [563]: store.select("df2_mt")
Out[563]:
C D E F foo
2000-01-01 -0.596306 -0.910022 -1.057072 -0.864360 bar
2000-01-02 0.477849 0.283128 -2.045700 -0.338206 bar
2000-01-03 -0.033100 -0.965461 -0.001079 -0.351689 bar
2000-01-04 -0.513555 -1.484776 -0.796280 -0.182321 bar
2000-01-05 -0.872407 -1.751515 0.934334 0.938818 bar
2000-01-06 -1.398256 1.347142 -0.029520 0.082738 bar
2000-01-07 -0.755544 0.380786 -1.634116 1.293610 bar
2000-01-08 1.453064 0.500558 -0.574475 0.694324 bar
# as a multiple
In [564]: store.select_as_multiple(
.....: ["df1_mt", "df2_mt"],
.....: where=["A>0", "B>0"],
.....: selector="df1_mt",
.....: )
.....:
Out[564]:
A B C D E F foo
2000-01-06 0.678830 0.633974 -1.398256 1.347142 -0.029520 0.082738 bar
2000-01-07 0.034973 0.974369 -0.755544 0.380786 -1.634116 1.293610 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. 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.
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. Two parameters are used to
control compression: complevel and complib.
complevelspecifies if and how hard data is to be compressed.complevel=0andcomplevel=Nonedisables compression and0<complevel<10enables compression.complibspecifies which compression library to use. If nothing is specified the default libraryzlibis used. A compression library usually optimizes for either good compression rates or speed and the results will depend on the type of data. Which type of compression to choose depends on your specific needs and data. The list of supported compression libraries:zlib: The default compression library. A classic in terms of compression, achieves good compression rates but is somewhat slow.
lzo: Fast compression and decompression.
bzip2: Good compression rates.
blosc: Fast compression and decompression.
Support for alternative blosc compressors:
blosc:blosclz This is the default compressor for
bloscblosc:lz4: A compact, very popular and fast compressor.
blosc:lz4hc: A tweaked version of LZ4, produces better compression ratios at the expense of speed.
blosc:snappy: A popular compressor used in many places.
blosc:zlib: A classic; somewhat slower than the previous ones, but achieving better compression ratios.
blosc:zstd: An extremely well balanced codec; it provides the best compression ratios among the others above, and at reasonably fast speed.
If
complibis defined as something other than the listed libraries aValueErrorexception is issued.
Note
If the library specified with the complib option is missing on your platform,
compression defaults to zlib without further ado.
Enable compression for all objects within the file:
store_compressed = pd.HDFStore(
"store_compressed.h5", complevel=9, complib="blosc:blosclz"
)
Or on-the-fly compression (this only applies to tables) in stores where compression is not enabled:
store.append("df", df, complib="zlib", complevel=5)
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
tableis created columns (DataFrame) are fixed; only exactly the same columns can be appendedBe 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_convertwith 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 : |
|
integer : |
|
boolean |
|
|
|
|
|
categorical : see the section below |
|
object : |
|
unicode columns are not supported, and WILL FAIL.
Categorical data¶
You can write data that contains category dtypes to a HDFStore.
Queries work the same as if it was an object array. However, the category dtyped data is
stored in a more efficient manner.
In [565]: dfcat = pd.DataFrame(
.....: {"A": pd.Series(list("aabbcdba")).astype("category"), "B": np.random.randn(8)}
.....: )
.....:
In [566]: dfcat
Out[566]:
A B
0 a -1.608059
1 a 0.851060
2 b -0.736931
3 b 0.003538
4 c -1.422611
5 d 2.060901
6 b 0.993899
7 a -1.371768
In [567]: dfcat.dtypes
Out[567]:
A category
B float64
dtype: object
In [568]: cstore = pd.HDFStore("cats.h5", mode="w")
In [569]: cstore.append("dfcat", dfcat, format="table", data_columns=["A"])
In [570]: result = cstore.select("dfcat", where="A in ['b', 'c']")
In [571]: result
Out[571]:
A B
2 b -0.736931
3 b 0.003538
4 c -1.422611
6 b 0.993899
In [572]: result.dtypes
Out[572]:
A category
B float64
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.
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 [573]: dfs = pd.DataFrame({"A": "foo", "B": "bar"}, index=list(range(5)))
In [574]: dfs
Out[574]:
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 [575]: store.append("dfs", dfs, min_itemsize=30)
In [576]: store.get_storer("dfs").table
Out[576]:
/dfs/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=30, shape=(2,), dflt=b'', pos=1)}
byteorder := 'little'
chunkshape := (963,)
autoindex := True
colindexes := {
"index": Index(6, mediumshuffle, zlib(1)).is_csi=False}
# A is created as a data_column with a size of 30
# B is size is calculated
In [577]: store.append("dfs2", dfs, min_itemsize={"A": 30})
In [578]: store.get_storer("dfs2").table
Out[578]:
/dfs2/table (Table(5,)) ''
description := {
"index": Int64Col(shape=(), dflt=0, pos=0),
"values_block_0": StringCol(itemsize=3, shape=(1,), dflt=b'', pos=1),
"A": StringCol(itemsize=30, shape=(), dflt=b'', pos=2)}
byteorder := 'little'
chunkshape := (1598,)
autoindex := True
colindexes := {
"index": Index(6, mediumshuffle, zlib(1)).is_csi=False,
"A": Index(6, mediumshuffle, 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 [579]: dfss = pd.DataFrame({"A": ["foo", "bar", "nan"]})
In [580]: dfss
Out[580]:
A
0 foo
1 bar
2 nan
In [581]: store.append("dfss", dfss)
In [582]: store.select("dfss")
Out[582]:
A
0 foo
1 bar
2 NaN
# here you need to specify a different nan rep
In [583]: store.append("dfss2", dfss, nan_rep="_nan_")
In [584]: store.select("dfss2")
Out[584]:
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 [585]: 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 [586]: df_for_r.head()
Out[586]:
first second class
0 0.013480 0.504941 0
1 0.690984 0.898188 1
2 0.510113 0.618748 1
3 0.357698 0.004972 0
4 0.451658 0.012065 1
In [587]: store_export = pd.HDFStore("export.h5")
In [588]: store_export.append("df_for_r", df_for_r, data_columns=df_dc.columns)
In [589]: store_export
Out[589]:
<class 'pandas.io.pytables.HDFStore'>
File path: export.h5
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 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.
Performance¶
tablesformat come with a writing performance penalty as compared tofixedstores. 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 rows thatPyTableswill expect. 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
PerformanceWarningwill 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.
Feather¶
Feather provides binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy.
Feather is designed to faithfully serialize and de-serialize DataFrames, supporting all of the pandas dtypes, including extension dtypes such as categorical and datetime with tz.
Several caveats:
The format will NOT write an
Index, orMultiIndexfor theDataFrameand will raise an error if a non-default one is provided. You can.reset_index()to store the index or.reset_index(drop=True)to ignore it.Duplicate column names and non-string columns names are not supported
Actual Python objects in object dtype columns are not supported. These will raise a helpful error message on an attempt at serialization.
See the Full Documentation.
In [590]: df = pd.DataFrame(
.....: {
.....: "a": list("abc"),
.....: "b": list(range(1, 4)),
.....: "c": np.arange(3, 6).astype("u1"),
.....: "d": np.arange(4.0, 7.0, dtype="float64"),
.....: "e": [True, False, True],
.....: "f": pd.Categorical(list("abc")),
.....: "g": pd.date_range("20130101", periods=3),
.....: "h": pd.date_range("20130101", periods=3, tz="US/Eastern"),
.....: "i": pd.date_range("20130101", periods=3, freq="ns"),
.....: }
.....: )
.....:
In [591]: df
Out[591]:
a b c d e f g h i
0 a 1 3 4.0 True a 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01 00:00:00.000000000
1 b 2 4 5.0 False b 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01 00:00:00.000000001
2 c 3 5 6.0 True c 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01 00:00:00.000000002
In [592]: df.dtypes
Out[592]:
a object
b int64
c uint8
d float64
e bool
f category
g datetime64[ns]
h datetime64[ns, US/Eastern]
i datetime64[ns]
dtype: object
Write to a feather file.
In [593]: df.to_feather("example.feather")
Read from a feather file.
In [594]: result = pd.read_feather("example.feather")
In [595]: result
Out[595]:
a b c d e f g h i
0 a 1 3 4.0 True a 2013-01-01 2013-01-01 00:00:00-05:00 2013-01-01 00:00:00.000000000
1 b 2 4 5.0 False b 2013-01-02 2013-01-02 00:00:00-05:00 2013-01-01 00:00:00.000000001
2 c 3 5 6.0 True c 2013-01-03 2013-01-03 00:00:00-05:00 2013-01-01 00:00:00.000000002
# we preserve dtypes
In [596]: result.dtypes
Out[596]:
a object
b int64
c uint8
d float64
e bool
f category
g datetime64[ns]
h datetime64[ns, US/Eastern]
i datetime64[ns]
dtype: object
Parquet¶
Apache Parquet provides a partitioned binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy. Parquet can use a variety of compression techniques to shrink the file size as much as possible while still maintaining good read performance.
Parquet is designed to faithfully serialize and de-serialize DataFrame s, supporting all of the pandas
dtypes, including extension dtypes such as datetime with tz.
Several caveats.
Duplicate column names and non-string columns names are not supported.
The
pyarrowengine always writes the index to the output, butfastparquetonly writes non-default indexes. This extra column can cause problems for non-pandas consumers that are not expecting it. You can force including or omitting indexes with theindexargument, regardless of the underlying engine.Index level names, if specified, must be strings.
In the
pyarrowengine, categorical dtypes for non-string types can be serialized to parquet, but will de-serialize as their primitive dtype.The
pyarrowengine preserves theorderedflag of categorical dtypes with string types.fastparquetdoes not preserve theorderedflag.Non supported types include
Intervaland actual Python object types. These will raise a helpful error message on an attempt at serialization.Periodtype is supported with pyarrow >= 0.16.0.The
pyarrowengine preserves extension data types such as the nullable integer and string data type (requiring pyarrow >= 0.16.0, and requiring the extension type to implement the needed protocols, see the extension types documentation).
You can specify an engine to direct the serialization. This can be one of pyarrow, or fastparquet, or auto.
If the engine is NOT specified, then the pd.options.io.parquet.engine option is checked; if this is also auto,
then pyarrow is tried, and falling back to fastparquet.
See the documentation for pyarrow and fastparquet.
Note
These engines are very similar and should read/write nearly identical parquet format files.
Currently pyarrow does not support timedelta data, fastparquet>=0.1.4 supports timezone aware datetimes.
These libraries differ by having different underlying dependencies (fastparquet by using numba, while pyarrow uses a c-library).
In [597]: df = pd.DataFrame(
.....: {
.....: "a": list("abc"),
.....: "b": list(range(1, 4)),
.....: "c": np.arange(3, 6).astype("u1"),
.....: "d": np.arange(4.0, 7.0, dtype="float64"),
.....: "e": [True, False, True],
.....: "f": pd.date_range("20130101", periods=3),
.....: "g": pd.date_range("20130101", periods=3, tz="US/Eastern"),
.....: "h": pd.Categorical(list("abc")),
.....: "i": pd.Categorical(list("abc"), ordered=True),
.....: }
.....: )
.....:
In [598]: df
Out[598]:
a b c d e f g h i
0 a 1 3 4.0 True 2013-01-01 2013-01-01 00:00:00-05:00 a a
1 b 2 4 5.0 False 2013-01-02 2013-01-02 00:00:00-05:00 b b
2 c 3 5 6.0 True 2013-01-03 2013-01-03 00:00:00-05:00 c c
In [599]: df.dtypes
Out[599]:
a object
b int64
c uint8
d float64
e bool
f datetime64[ns]
g datetime64[ns, US/Eastern]
h category
i category
dtype: object
Write to a parquet file.
In [600]: df.to_parquet("example_pa.parquet", engine="pyarrow")
In [601]: df.to_parquet("example_fp.parquet", engine="fastparquet")
Read from a parquet file.
In [602]: result = pd.read_parquet("example_fp.parquet", engine="fastparquet")
In [603]: result = pd.read_parquet("example_pa.parquet", engine="pyarrow")
In [604]: result.dtypes
Out[604]:
a object
b int64
c uint8
d float64
e bool
f datetime64[ns]
g datetime64[ns, US/Eastern]
h category
i category
dtype: object
Read only certain columns of a parquet file.
In [605]: result = pd.read_parquet(
.....: "example_fp.parquet",
.....: engine="fastparquet",
.....: columns=["a", "b"],
.....: )
.....:
In [606]: result = pd.read_parquet(
.....: "example_pa.parquet",
.....: engine="pyarrow",
.....: columns=["a", "b"],
.....: )
.....:
In [607]: result.dtypes
Out[607]:
a object
b int64
dtype: object
Handling indexes¶
Serializing a DataFrame to parquet may include the implicit index as one or
more columns in the output file. Thus, this code:
In [608]: df = pd.DataFrame({"a": [1, 2], "b": [3, 4]})
In [609]: df.to_parquet("test.parquet", engine="pyarrow")
creates a parquet file with three columns if you use pyarrow for serialization:
a, b, and __index_level_0__. If you’re using fastparquet, the
index may or may not
be written to the file.
This unexpected extra column causes some databases like Amazon Redshift to reject the file, because that column doesn’t exist in the target table.
If you want to omit a dataframe’s indexes when writing, pass index=False to
to_parquet():
In [610]: df.to_parquet("test.parquet", index=False)
This creates a parquet file with just the two expected columns, a and b.
If your DataFrame has a custom index, you won’t get it back when you load
this file into a DataFrame.
Passing index=True will always write the index, even if that’s not the
underlying engine’s default behavior.
Partitioning Parquet files¶
Parquet supports partitioning of data based on the values of one or more columns.
In [611]: df = pd.DataFrame({"a": [0, 0, 1, 1], "b": [0, 1, 0, 1]})
In [612]: df.to_parquet(path="test", engine="pyarrow", partition_cols=["a"], compression=None)
The path specifies the parent directory to which data will be saved.
The partition_cols are the column names by which the dataset will be partitioned.
Columns are partitioned in the order they are given. The partition splits are
determined by the unique values in the partition columns.
The above example creates a partitioned dataset that may look like:
test
├── a=0
│ ├── 0bac803e32dc42ae83fddfd029cbdebc.parquet
│ └── ...
└── a=1
├── e6ab24a4f45147b49b54a662f0c412a3.parquet
└── ...
ORC¶
New in version 1.0.0.
Similar to the parquet format, the ORC Format is a binary columnar serialization
for data frames. It is designed to make reading data frames efficient. pandas provides only a reader for the
ORC format, read_orc(). This requires the pyarrow library.
Warning
It is highly recommended to install pyarrow using conda due to some issues occurred by pyarrow.
read_orc()is not supported on Windows yet, you can find valid environments on install optional dependencies.
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. Examples of such drivers are psycopg2
for PostgreSQL or pymysql for MySQL.
For SQLite this is
included in Python’s standard library by default.
You can find an overview of supported drivers for each SQL dialect in the
SQLAlchemy docs.
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 database table into a DataFrame. |
|
Read SQL query into a DataFrame. |
|
Read SQL query or database table into a DataFrame. |
|
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 [613]: from sqlalchemy import create_engine
# Create your engine.
In [614]: engine = create_engine("sqlite:///:memory:")
If you want to manage your own connections you can pass one of those instead. The example below opens a connection to the database using a Python context manager that automatically closes the connection after the block has completed. See the SQLAlchemy docs for an explanation of how the database connection is handled.
with engine.connect() as conn, conn.begin():
data = pd.read_sql_table("data", conn)
Warning
When you open a connection to a database you are also responsible for closing it. Side effects of leaving a connection open may include locking the database or other breaking behaviour.
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 [615]: import datetime
In [616]: c = ["id", "Date", "Col_1", "Col_2", "Col_3"]
In [617]: d = [
.....: (26, datetime.datetime(2010, 10, 18), "X", 27.5, True),
.....: (42, datetime.datetime(2010, 10, 19), "Y", -12.5, False),
.....: (63, datetime.datetime(2010, 10, 20), "Z", 5.73, True),
.....: ]
.....:
In [618]: data = pd.DataFrame(d, columns=c)
In [619]: data
Out[619]:
id Date Col_1 Col_2 Col_3
0 26 2010-10-18 X 27.50 True
1 42 2010-10-19 Y -12.50 False
2 63 2010-10-20 Z 5.73 True
In [620]: data.to_sql("data", engine)
Out[620]: 3
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 [621]: data.to_sql("data_chunked", engine, chunksize=1000)
Out[621]: 3
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 [622]: from sqlalchemy.types import String
In [623]: data.to_sql("data_dtype", engine, dtype={"Col_1": String})
Out[623]: 3
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.
Datetime data types¶
Using SQLAlchemy, to_sql() is capable of writing
datetime data that is timezone naive or timezone aware. However, the resulting
data stored in the database ultimately depends on the supported data type
for datetime data of the database system being used.
The following table lists supported data types for datetime data for some common databases. Other database dialects may have different data types for datetime data.
Database |
SQL Datetime Types |
Timezone Support |
|---|---|---|
SQLite |
|
No |
MySQL |
|
No |
PostgreSQL |
|
Yes |
When writing timezone aware data to databases that do not support timezones, the data will be written as timezone naive timestamps that are in local time with respect to the timezone.
read_sql_table() is also capable of reading datetime data that is
timezone aware or naive. When reading TIMESTAMP WITH TIME ZONE types, pandas
will convert the data to UTC.
Insertion method¶
The parameter method controls the SQL insertion clause used.
Possible values are:
None: Uses standard SQLINSERTclause (one per row).'multi': Pass multiple values in a singleINSERTclause. It uses a special SQL syntax not supported by all backends. This usually provides better performance for analytic databases like Presto and Redshift, but has worse performance for traditional SQL backend if the table contains many columns. For more information check the SQLAlchemy documentation.callable with signature
(pd_table, conn, keys, data_iter): This can be used to implement a more performant insertion method based on specific backend dialect features.
Example of a callable using PostgreSQL COPY clause:
# Alternative to_sql() *method* for DBs that support COPY FROM
import csv
from io import StringIO
def psql_insert_copy(table, conn, keys, data_iter):
"""
Execute SQL statement inserting data
Parameters
----------
table : pandas.io.sql.SQLTable
conn : sqlalchemy.engine.Engine or sqlalchemy.engine.Connection
keys : list of str
Column names
data_iter : Iterable that iterates the values to be inserted
"""
# gets a DBAPI connection that can provide a cursor
dbapi_conn = conn.connection
with dbapi_conn.cursor() as cur:
s_buf = StringIO()
writer = csv.writer(s_buf)
writer.writerows(data_iter)
s_buf.seek(0)
columns = ', '.join(['"{}"'.format(k) for k in keys])
if table.schema:
table_name = '{}.{}'.format(table.schema, table.name)
else:
table_name = table.name
sql = 'COPY {} ({}) FROM STDIN WITH CSV'.format(
table_name, columns)
cur.copy_expert(sql=sql, file=s_buf)
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 [624]: pd.read_sql_table("data", engine)
Out[624]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 X 27.50 True
1 1 42 2010-10-19 Y -12.50 False
2 2 63 2010-10-20 Z 5.73 True
Note
Note that pandas infers column dtypes from query outputs, and not by looking
up data types in the physical database schema. For example, assume userid
is an integer column in a table. Then, intuitively, select userid ... will
return integer-valued series, while select cast(userid as text) ... will
return object-valued (str) series. Accordingly, if the query output is empty,
then all resulting columns will be returned as object-valued (since they are
most general). If you foresee that your query will sometimes generate an empty
result, you may want to explicitly typecast afterwards to ensure dtype
integrity.
You can also specify the name of the column as the DataFrame index,
and specify a subset of columns to be read.
In [625]: pd.read_sql_table("data", engine, index_col="id")
Out[625]:
index Date Col_1 Col_2 Col_3
id
26 0 2010-10-18 X 27.50 True
42 1 2010-10-19 Y -12.50 False
63 2 2010-10-20 Z 5.73 True
In [626]: pd.read_sql_table("data", engine, columns=["Col_1", "Col_2"])
Out[626]:
Col_1 Col_2
0 X 27.50
1 Y -12.50
2 Z 5.73
And you can explicitly force columns to be parsed as dates:
In [627]: pd.read_sql_table("data", engine, parse_dates=["Date"])
Out[627]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 X 27.50 True
1 1 42 2010-10-19 Y -12.50 False
2 2 63 2010-10-20 Z 5.73 True
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¶
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 [628]: pd.read_sql_query("SELECT * FROM data", engine)
Out[628]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 00:00:00.000000 X 27.50 1
1 1 42 2010-10-19 00:00:00.000000 Y -12.50 0
2 2 63 2010-10-20 00:00:00.000000 Z 5.73 1
Of course, you can specify a more “complex” query.
In [629]: pd.read_sql_query("SELECT id, Col_1, Col_2 FROM data WHERE id = 42;", engine)
Out[629]:
id Col_1 Col_2
0 42 Y -12.5
The read_sql_query() function supports a chunksize argument.
Specifying this will return an iterator through chunks of the query result:
In [630]: df = pd.DataFrame(np.random.randn(20, 3), columns=list("abc"))
In [631]: df.to_sql("data_chunks", engine, index=False)
Out[631]: 20
In [632]: for chunk in pd.read_sql_query("SELECT * FROM data_chunks", engine, chunksize=5):
.....: print(chunk)
.....:
a b c
0 0.070470 0.901320 0.937577
1 0.295770 1.420548 -0.005283
2 -1.518598 -0.730065 0.226497
3 -2.061465 0.632115 0.853619
4 2.719155 0.139018 0.214557
a b c
0 -1.538924 -0.366973 -0.748801
1 -0.478137 -1.559153 -3.097759
2 -2.320335 -0.221090 0.119763
3 0.608228 1.064810 -0.780506
4 -2.736887 0.143539 1.170191
a b c
0 -1.573076 0.075792 -1.722223
1 -0.774650 0.803627 0.221665
2 0.584637 0.147264 1.057825
3 -0.284136 0.912395 1.552808
4 0.189376 -0.109830 0.539341
a b c
0 0.592591 -0.155407 -1.356475
1 0.833837 1.524249 1.606722
2 -0.029487 -0.051359 1.700152
3 0.921484 -0.926347 0.979818
4 0.182380 -0.186376 0.049820
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:[email protected]: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 [633]: import sqlalchemy as sa
In [634]: pd.read_sql(
.....: sa.text("SELECT * FROM data where Col_1=:col1"), engine, params={"col1": "X"}
.....: )
.....:
Out[634]:
index id Date Col_1 Col_2 Col_3
0 0 26 2010-10-18 00:00:00.000000 X 27.5 1
If you have an SQLAlchemy description of your database you can express where conditions using SQLAlchemy expressions
In [635]: metadata = sa.MetaData()
In [636]: 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),
.....: )
.....:
In [637]: pd.read_sql(sa.select([data_table]).where(data_table.c.Col_3 is True), engine)
Out[637]:
Empty DataFrame
Columns: [index, Date, Col_1, Col_2, Col_3]
Index: []
You can combine SQLAlchemy expressions with parameters passed to read_sql() using sqlalchemy.bindparam()
In [638]: import datetime as dt
In [639]: expr = sa.select([data_table]).where(data_table.c.Date > sa.bindparam("date"))
In [640]: pd.read_sql(expr, engine, params={"date": dt.datetime(2010, 10, 18)})
Out[640]:
index Date Col_1 Col_2 Col_3
0 1 2010-10-19 Y -12.50 False
1 2 2010-10-20 Z 5.73 True
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", con)
pd.read_sql_query("SELECT * FROM data", con)
Google BigQuery¶
Warning
Starting in 0.20.0, pandas has split off Google BigQuery support into the
separate package pandas-gbq. You can pip install pandas-gbq to get it.
The pandas-gbq package provides functionality to read/write from Google BigQuery.
pandas integrates with this external package. if pandas-gbq is installed, you can
use the pandas methods pd.read_gbq and DataFrame.to_gbq, which will call the
respective functions from pandas-gbq.
Full documentation can be found here.
Stata format¶
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 [641]: df = pd.DataFrame(np.random.randn(10, 2), columns=list("AB"))
In [642]: 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 [643]: pd.read_stata("stata.dta")
Out[643]:
index A B
0 0 -1.690072 0.405144
1 1 -1.511309 -1.531396
2 2 0.572698 -1.106845
3 3 -1.185859 0.174564
4 4 0.603797 -1.796129
5 5 -0.791679 1.173795
6 6 -0.277710 1.859988
7 7 -0.258413 1.251808
8 8 1.443262 0.441553
9 9 1.168163 -2.054946
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 [644]: with pd.read_stata("stata.dta", chunksize=3) as reader:
.....: 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 [645]: with pd.read_stata("stata.dta", iterator=True) as reader:
.....: chunk1 = reader.read(5)
.....: 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¶
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 Categorical with string categories for the values that are labeled and
numeric categories for values with no label.
SAS formats¶
The top-level function read_sas() can read (but not write) SAS
XPORT (.xpt) and (since v0.18.0) SAS7BDAT (.sas7bdat) format files.
SAS files only contain two value types: ASCII text and floating point
values (usually 8 bytes but sometimes truncated). For xport files,
there is no automatic type conversion to integers, dates, or
categoricals. For SAS7BDAT files, the format codes may allow date
variables to be automatically converted to dates. By default the
whole file is read and returned as a DataFrame.
Specify a chunksize or use iterator=True to obtain reader
objects (XportReader or SAS7BDATReader) for incrementally
reading the file. The reader objects also have attributes that
contain additional information about the file and its variables.
Read a SAS7BDAT file:
df = pd.read_sas("sas_data.sas7bdat")
Obtain an iterator and read an XPORT file 100,000 lines at a time:
def do_something(chunk):
pass
with pd.read_sas("sas_xport.xpt", chunk=100000) as rdr:
for chunk in rdr:
do_something(chunk)
The specification for the xport file format is available from the SAS web site.
No official documentation is available for the SAS7BDAT format.
SPSS formats¶
New in version 0.25.0.
The top-level function read_spss() can read (but not write) SPSS
SAV (.sav) and ZSAV (.zsav) format files.
SPSS files contain column names. By default the
whole file is read, categorical columns are converted into pd.Categorical,
and a DataFrame with all columns is returned.
Specify the usecols parameter to obtain a subset of columns. Specify convert_categoricals=False
to avoid converting categorical columns into pd.Categorical.
Read an SPSS file:
df = pd.read_spss("spss_data.sav")
Extract a subset of columns contained in usecols from an SPSS file and
avoid converting categorical columns into pd.Categorical:
df = pd.read_spss(
"spss_data.sav",
usecols=["foo", "bar"],
convert_categoricals=False,
)
More information about the SAV and ZSAV file formats is available here.
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.
netCDF¶
xarray provides data structures inspired by the pandas DataFrame for working
with multi-dimensional datasets, with a focus on the netCDF file format and
easy conversion to and from pandas.
Performance considerations¶
This is an informal comparison of various IO methods, using pandas 0.24.2. Timings are machine dependent and small differences should be ignored.
In [1]: sz = 1000000
In [2]: df = pd.DataFrame({'A': np.random.randn(sz), 'B': [1] * sz})
In [3]: df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1000000 entries, 0 to 999999
Data columns (total 2 columns):
A 1000000 non-null float64
B 1000000 non-null int64
dtypes: float64(1), int64(1)
memory usage: 15.3 MB
The following test functions will be used below to compare the performance of several IO methods:
import numpy as np
import os
sz = 1000000
df = pd.DataFrame({"A": np.random.randn(sz), "B": [1] * sz})
sz = 1000000
np.random.seed(42)
df = pd.DataFrame({"A": np.random.randn(sz), "B": [1] * sz})
def test_sql_write(df):
if os.path.exists("test.sql"):
os.remove("test.sql")
sql_db = sqlite3.connect("test.sql")
df.to_sql(name="test_table", con=sql_db)
sql_db.close()
def test_sql_read():
sql_db = sqlite3.connect("test.sql")
pd.read_sql_query("select * from test_table", sql_db)
sql_db.close()
def test_hdf_fixed_write(df):
df.to_hdf("test_fixed.hdf", "test", mode="w")
def test_hdf_fixed_read():
pd.read_hdf("test_fixed.hdf", "test")
def test_hdf_fixed_write_compress(df):
df.to_hdf("test_fixed_compress.hdf", "test", mode="w", complib="blosc")
def test_hdf_fixed_read_compress():
pd.read_hdf("test_fixed_compress.hdf", "test")
def test_hdf_table_write(df):
df.to_hdf("test_table.hdf", "test", mode="w", format="table")
def test_hdf_table_read():
pd.read_hdf("test_table.hdf", "test")
def test_hdf_table_write_compress(df):
df.to_hdf(
"test_table_compress.hdf", "test", mode="w", complib="blosc", format="table"
)
def test_hdf_table_read_compress():
pd.read_hdf("test_table_compress.hdf", "test")
def test_csv_write(df):
df.to_csv("test.csv", mode="w")
def test_csv_read():
pd.read_csv("test.csv", index_col=0)
def test_feather_write(df):
df.to_feather("test.feather")
def test_feather_read():
pd.read_feather("test.feather")
def test_pickle_write(df):
df.to_pickle("test.pkl")
def test_pickle_read():
pd.read_pickle("test.pkl")
def test_pickle_write_compress(df):
df.to_pickle("test.pkl.compress", compression="xz")
def test_pickle_read_compress():
pd.read_pickle("test.pkl.compress", compression="xz")
def test_parquet_write(df):
df.to_parquet("test.parquet")
def test_parquet_read():
pd.read_parquet("test.parquet")
When writing, the top three functions in terms of speed are test_feather_write, test_hdf_fixed_write and test_hdf_fixed_write_compress.
In [4]: %timeit test_sql_write(df)
3.29 s ± 43.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [5]: %timeit test_hdf_fixed_write(df)
19.4 ms ± 560 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [6]: %timeit test_hdf_fixed_write_compress(df)
19.6 ms ± 308 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [7]: %timeit test_hdf_table_write(df)
449 ms ± 5.61 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [8]: %timeit test_hdf_table_write_compress(df)
448 ms ± 11.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [9]: %timeit test_csv_write(df)
3.66 s ± 26.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [10]: %timeit test_feather_write(df)
9.75 ms ± 117 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
In [11]: %timeit test_pickle_write(df)
30.1 ms ± 229 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [12]: %timeit test_pickle_write_compress(df)
4.29 s ± 15.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [13]: %timeit test_parquet_write(df)
67.6 ms ± 706 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
When reading, the top three functions in terms of speed are test_feather_read, test_pickle_read and
test_hdf_fixed_read.
In [14]: %timeit test_sql_read()
1.77 s ± 17.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [15]: %timeit test_hdf_fixed_read()
19.4 ms ± 436 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [16]: %timeit test_hdf_fixed_read_compress()
19.5 ms ± 222 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [17]: %timeit test_hdf_table_read()
38.6 ms ± 857 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [18]: %timeit test_hdf_table_read_compress()
38.8 ms ± 1.49 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)
In [19]: %timeit test_csv_read()
452 ms ± 9.04 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [20]: %timeit test_feather_read()
12.4 ms ± 99.7 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
In [21]: %timeit test_pickle_read()
18.4 ms ± 191 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
In [22]: %timeit test_pickle_read_compress()
915 ms ± 7.48 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [23]: %timeit test_parquet_read()
24.4 ms ± 146 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
The files test.pkl.compress, test.parquet and test.feather took the least space on disk (in bytes).
29519500 Oct 10 06:45 test.csv
16000248 Oct 10 06:45 test.feather
8281983 Oct 10 06:49 test.parquet
16000857 Oct 10 06:47 test.pkl
7552144 Oct 10 06:48 test.pkl.compress
34816000 Oct 10 06:42 test.sql
24009288 Oct 10 06:43 test_fixed.hdf
24009288 Oct 10 06:43 test_fixed_compress.hdf
24458940 Oct 10 06:44 test_table.hdf
24458940 Oct 10 06:44 test_table_compress.hdf
Comments and empty lines¶
Ignoring line comments and empty lines¶
If the
commentparameter is specified, then completely commented lines will be ignored. By default, completely blank lines will be ignored as well.If
skip_blank_lines=False, thenread_csvwill not ignore blank lines:Warning
The presence of ignored lines might create ambiguities involving line numbers; the parameter
headeruses row numbers (ignoring commented/empty lines), whileskiprowsuses line numbers (including commented/empty lines):If both
headerandskiprowsare specified,headerwill be relative to the end ofskiprows. For example:Comments¶
Sometimes comments or meta data may be included in a file:
By default, the parser includes the comments in the output:
We can suppress the comments using the
commentkeyword: