Reading tables

The majority of commonly encountered ASCII tables can be easily read with the read() function:

>>> from astropy.io import ascii
>>> data = ascii.read(table)

where table is the name of a file, a string representation of a table, or a list of table lines. The return value (data in this case) is a Table object.

By default read() will try to guess the table format by trying all the supported formats. Guessing the file format is often slow for large files because the reader simply tries parsing the file with every allowed format until one succeeds. For large files it is recommended to disable guessing with guess=False.

If guessing does not work (for unusually formatted tables) then one needs give astropy.io.ascii additional hints about the format, for example:

>>> data = astropy.io.ascii.read('data/nls1_stackinfo.dbout', data_start=2, delimiter='|')  
>>> data = astropy.io.ascii.read('data/simple.txt', quotechar="'")  
>>> data = astropy.io.ascii.read('data/simple4.txt', format='no_header', delimiter='|')  
>>> data = astropy.io.ascii.read('data/tab_and_space.txt', delimiter=r'\s')

The read() function accepts a number of parameters that specify the detailed table format. Different formats can define different defaults, so the descriptions below sometimes mention “typical” default values. This refers to the Basic format reader and other similar character-separated formats.

Parameters for read()

tableinput table

There are four ways to specify the table to be read:

  • Name of a file (string)

  • Single string containing all table lines separated by newlines

  • File-like object with a callable read() method

  • List of strings where each list element is a table line

The first two options are distinguished by the presence of a newline in the string. This assumes that valid file names will not normally contain a newline.

formatfile format (default=’basic’)

This specifies the top-level format of the ASCII table, for example if it is a basic character delimited table, fixed format table, or a CDS-compatible table, etc. The value of this parameter must be one of the Supported formats.

guesstry to guess table format (default=True)

If set to True then read() will try to guess the table format by cycling through a number of possible table format permutations and attempting to read the table in each case. See the Guess table format section for further details.

delimitercolumn delimiter string

A one-character string used to separate fields which typically defaults to the space character. Other common values might be “\s” (whitespace), “,” or “|” or “\t” (tab). A value of “\s” allows any combination of the tab and space characters to delimit columns.

commentregular expression defining a comment line in table

If the comment regular expression matches the beginning of a table line then that line will be discarded from header or data processing. For the basic format this defaults to “\s*#” (any whitespace followed by #).

quotecharone-character string to quote fields containing special characters

This specifies the quote character and will typically be either the single or double quote character. This is can be useful for reading text fields with spaces in a space-delimited table. The default is typically the double quote.

header_startline index for the header line

This includes only significant non-comment lines and counting starts at 0. If set to None this indicates that there is no header line and the column names will be auto-generated. See Specifying header and data location for more details.

data_startline index for the start of data counting

This includes only significant non-comment lines and counting starts at 0. See Specifying header and data location for more details.

data_endline index for the end of data

This includes only significant non-comment line and can be negative to count from end. See Specifying header and data location for more details.

encoding: encoding to read the file (default=`None`)

When None use locale.getpreferredencoding as an encoding. This matches the default behavior of the built-in open when no mode argument is provided.

convertersdict of data type converters

See the Converters section for more information.

nameslist of names corresponding to each data column

Define the complete list of names for each data column. This will override names found in the header (if it exists). If not supplied then use names from the header or auto-generated names if there is no header.

include_nameslist of names to include in output

From the list of column names found from the header or the names parameter, select for output only columns within this list. If not supplied then include all names.

exclude_nameslist of names to exclude from output

Exclude these names from the list of output columns. This is applied after the include_names filtering. If not specified then no columns are excluded.

fill_valueslist of fill value specifiers

Specify input table entries which should be masked in the output table because they are bad or missing. See the Bad or missing values section for more information and examples. The default is that any blank table values are treated as missing.

fill_include_nameslist of column names, which are affected by fill_values.

If not supplied, then fill_values can affect all columns.

fill_exclude_nameslist of column names, which are not affected by fill_values.

If not supplied, then fill_values can affect all columns.

OutputterOutputter class

This converts the raw data tables value into the output object that gets returned by read(). The default is TableOutputter, which returns a Table object (see Data Tables).

InputterInputter class

This is generally not specified.

data_Splitter : Splitter class to split data columns

header_Splitter : Splitter class to split header columns

fast_readerwhether to use the C engine

This can be True or False, and also be a dict with options. (see Fast ASCII I/O)

ReaderReader class (deprecated in favor of format)

This specifies the top-level format of the ASCII table, for example if it is a basic character delimited table, fixed format table, or a CDS-compatible table, etc. The value of this parameter must be a Reader class. For basic usage this means one of the built-in Extension Reader classes.

Specifying header and data location

The three parameters header_start, data_start and data_end make it possible to read a table file that has extraneous non-table data included. This is a case where you need to help out io.ascii and tell it where to find the header and data.

When processing of a file into a header and data components any blank lines (which might have whitespace characters) and commented lines (starting with the comment character, typically #) are stripped out before the header and data parsing code sees the table content. For example imagine you have the file below. The column on the left is not part of the file but instead shows how io.ascii is viewing each line and the line count index.

Index    Table content
------ ----------------------------------------------------------------
   -  | # This is the start of my data file
   -  |
   0  | Automatically generated by my_script.py at 2012-01-01T12:13:14
   1  | Run parameters: None
   2  | Column header line:
   -  |
   3  | x y z
   -  |
   4  | Data values section:
   -  |
   5  | 1 2 3
   6  | 4 5 6
   -  |
   7  | Run completed at 2012:01-01T12:14:01

In this case you would have header_start=3, data_start=5, and data_end=7. The convention for data_end follows the normal Python slicing convention where to select data rows 5 and 6 you would do rows[5:7]. For data_end you can also supply a negative index to count backward from the end, so data_end=-1 (like rows[5:-1]) would work in this case.

Note

Prior to astropy v1.1 there was a bug in which a blank line that had one or more whitespace characters was mistakenly counted for header_start but was (correctly) not counted for data_start and data_end. If you have code that was depending on the incorrect pre-1.1 behavior then it needs to be modified.

Bad or missing values

ASCII data tables can contain bad or missing values. A common case is when a table contains blank entries with no available data, for example:

>>> weather_data = """
...   day,precip,type
...   Mon,1.5,rain
...   Tues,,
...   Wed,1.1,snow
...   """

By default read() will interpret blank entries as being bad/missing and output a masked Table with those entries masked out by setting the corresponding mask value set to True:

>>> dat = ascii.read(weather_data)
>>> print(dat)
day  precip type
---- ------ ----
 Mon    1.5 rain
Tues     --   --
 Wed    1.1 snow

If you want to replace the masked (missing) values with particular values, set the masked column fill_value attribute and then get the “filled” version of the table. This looks like the following:

>>> dat['precip'].fill_value = -999
>>> dat['type'].fill_value = 'N/A'
>>> print(dat.filled())
day  precip type
---- ------ ----
 Mon    1.5 rain
Tues -999.0  N/A
 Wed    1.1 snow

ASCII tables may also have other indicators of bad or missing data. For example a table may contain string values that are not a valid representation of a number, e.g. "...", or a table may have special values like -999 that are chosen to indicate missing data. The read() function has a flexible system to accommodate these cases by marking specified character sequences in the input data as “missing data” during the conversion process. Whenever missing data is found then the output will be a masked table.

This is done with the fill_values keyword argument, which can be set to a single missing-value specification <missing_spec> or a list of <missing_spec> tuples:

fill_values = <missing_spec> | [<missing_spec1>, <missing_spec2>, ...]
<missing_spec> = (<match_string>, '0', <optional col name 1>, <optional col name 2>, ...)

When reading a table the second element of a <missing_spec> should always be the string '0', otherwise you may get unexpected behavior 1. By default the <missing_spec> is applied to all columns unless column name strings are supplied. An alternate way to limit the columns is via the fill_include_names and fill_exclude_names keyword arguments in read().

In the example below we read back the weather table after filling the missing values in with typical placeholders:

>>> table = ['day   precip  type',
...          ' Mon     1.5  rain',
...          'Tues  -999.0   N/A',
...          ' Wed     1.1  snow']
>>> t = ascii.read(table, fill_values=[('-999.0', '0', 'precip'), ('N/A', '0', 'type')])
>>> print(t)
day  precip type
---- ------ ----
 Mon    1.5 rain
Tues     --   --
 Wed    1.1 snow

Note

The default in read() is fill_values=('','0'). This marks blank entries as being missing for any data type (int, float, or string). If fill_values is explicitly set in the call to read() then the default behavior of marking blank entries as missing no longer applies. For instance setting fill_values=None will disable this auto-masking without setting any other fill values. This can be useful for a string column where one of values happens to be "".

1

The requirement to put the '0' there is the legacy of an old interface which is maintained for backward compatibility and also to match the format of fill_value for reading with the format of fill_value used for writing tables. On reading, the second element of the <missing_spec> tuple can actually be an arbitrary string value which replaces occurrences of the <match_string> string in the input stream prior to type conversion. This ends up being the value “behind the mask”, which should never be directly accessed. Only the value '0' is neutral when attempting to detect the column data type and perform type conversion. For instance if you used 'nan' for the <match_string> value then integer columns would wind up as float.

Guess table format

If the guess parameter in read() is set to True (which is the default) then read() will try to guess the table format by cycling through a number of possible table format permutations and attempting to read the table in each case. The first format which succeeds and will be used to read the table. To succeed the table must be successfully parsed by the Reader and satisfy the following column requirements:

  • At least two table columns

  • No column names are a float or int number

  • No column names begin or end with space, comma, tab, single quote, double quote, or a vertical bar (|).

These requirements reduce the chance for a false positive where a table is successfully parsed with the wrong format. A common situation is a table with numeric columns but no header row, and in this case astropy.io.ascii will auto-assign column names because of the restriction on column names that look like a number.

Guess order

The order of guessing is shown by this Python code, where Reader is the class which actually implements reading the different file formats:

for Reader in (Ecsv, FixedWidthTwoLine, Rst, FastBasic, Basic,
               FastRdb, Rdb, FastTab, Tab, Cds, Daophot, SExtractor,
               Ipac, Latex, AASTex):
    read(Reader=Reader)

for Reader in (CommentedHeader, FastBasic, Basic, FastNoHeader, NoHeader):
    for delimiter in ("|", ",", " ", "\\s"):
        for quotechar in ('"', "'"):
            read(Reader=Reader, delimiter=delimiter, quotechar=quotechar)

Note that the FixedWidth derived-readers are not included in the default guess sequence (this causes problems), so to read such tables one must explicitly specify the format with the format keyword. Also notice that formats compatible with the fast reading engine attempt to use the fast engine before the ordinary reading engine.

If none of the guesses succeed in reading the table (subject to the column requirements) a final try is made using just the user-supplied parameters but without checking the column requirements. In this way a table with only one column or column names that look like a number can still be successfully read.

The guessing process respects any values of the Reader, delimiter, and quotechar parameters as well as options for the fast reader that were supplied to the read() function. Any guesses that would conflict are skipped. For example the call:

>>> data = ascii.read(table, Reader=ascii.NoHeader, quotechar="'")

would only try the four delimiter possibilities, skipping all the conflicting Reader and quotechar combinations. Similarly with any setting of fast_reader that requires use of the fast engine, only the fast variants in the Reader list above will be tried.

Disabling

Guessing can be disabled in two ways:

import astropy.io.ascii
data = astropy.io.ascii.read(table)               # guessing enabled by default
data = astropy.io.ascii.read(table, guess=False)  # disable for this call
astropy.io.ascii.set_guess(False)                 # set default to False globally
data = astropy.io.ascii.read(table)               # guessing disabled

Debugging

In order to get more insight into the guessing process and possibly debug if something isn’t working as expected, use the get_read_trace() function. This returns a traceback of the attempted read formats for the last call to read().

Comments and metadata

Any comment lines detected during reading are inserted into the output table via the comments key in the table’s .meta dictionary. For example:

>>> table='''# TELESCOPE = 30 inch
...          # TARGET = PV Ceph
...          # BAND = V
...          MJD mag
...          55555 12.3
...          55556 12.4'''
>>> dat = ascii.read(table)
>>> print(dat.meta['comments'])
['TELESCOPE = 30 inch', 'TARGET = PV Ceph', 'BAND = V']

While astropy.io.ascii will not do any post-processing on comment lines, custom post-processing can be accomplished by re-reading with the metadata line comments. Here is one example, where comments are of the form “# KEY = VALUE”:

>>> header = ascii.read(dat.meta['comments'], delimiter='=',
...                     format='no_header', names=['key', 'val'])
>>> print(header)
   key      val
--------- -------
TELESCOPE 30 inch
   TARGET PV Ceph
     BAND       V

Converters

astropy.io.ascii converts the raw string values from the table into numeric data types by using converter functions such as the Python int and float functions. For example int("5.0") will fail while float(“5.0”) will succeed and return 5.0 as a Python float.

The default converters are:

default_converters = [astropy.io.ascii.convert_numpy(numpy.int),
                      astropy.io.ascii.convert_numpy(numpy.float),
                      astropy.io.ascii.convert_numpy(numpy.str)]

These take advantage of the convert_numpy() function which returns a 2-element tuple (converter_func, converter_type) as described in the previous section. The type provided to convert_numpy() must be a valid numpy type, for example numpy.int, numpy.uint, numpy.int8, numpy.int64, numpy.float, numpy.float64, numpy.str.

The default converters for each column can be overridden with the converters keyword:

>>> import numpy as np
>>> converters = {'col1': [ascii.convert_numpy(np.uint)],
...               'col2': [ascii.convert_numpy(np.float32)]}
>>> ascii.read('file.dat', converters=converters)

Fortran-style exponents

The fast converter available with the C input parser provides an exponent_style option to define a custom character instead of the standard 'e' for exponential formats in the input file, to read for example Fortran-style double precision numbers like '1.495978707D+13':

>>> ascii.read('double.dat', format='basic', guess=False,
...            fast_reader={'exponent_style': 'D'})

The special setting 'fortran' is provided to allow for the auto-detection of any valid Fortran exponent character ('E', 'D', 'Q'), as well as of triple-digit exponents prefixed with no character at all (e.g. '2.1127123261674622-107'). All values and exponent characters in the input data are case-insensitive; any value other than the default 'E' implies the automatic setting of 'use_fast_converter': True.

Advanced customization

Here we provide a few examples that demonstrate how to extend the base functionality to handle special cases. To go beyond these simple examples the best reference is to read the code for the existing Extension Reader classes.

Define custom readers by class inheritance

The most useful way to define a new reader class is by inheritance. This is the way all the build-in readers are defined, so there are plenty of examples in the code.

In most cases, you will define one class to handle the header, one class that handles the data and a reader class that ties it all together. Here is a simple example from the code that defines a reader that is just like the basic reader, but header and data start in different lines of the file:

# Note: NoHeader is already included in astropy.io.ascii for convenience.
class NoHeaderHeader(BasicHeader):
    '''Reader for table header without a header

    Set the start of header line number to `None`, which tells the basic
    reader there is no header line.
    '''
    start_line = None

class NoHeaderData(BasicData):
    '''Reader for table data without a header

    Data starts at first uncommented line since there is no header line.
    '''
    start_line = 0

class NoHeader(Basic):
    """Read a table with no header line.  Columns are autonamed using
    header.auto_format which defaults to "col%d".  Otherwise this reader
    the same as the :class:`Basic` class from which it is derived.  Example::

      # Table data
      1 2 "hello there"
      3 4 world
    """
    _format_name = 'no_header'
    _description = 'Basic table with no headers'
    header_class = NoHeaderHeader
    data_class = NoHeaderData

In a slightly more involved case, the implementation can also override some of the methods in the base class:

# Note: CommentedHeader is already included in astropy.io.ascii for convenience.
class CommentedHeaderHeader(BasicHeader):
    """Header class for which the column definition line starts with the
    comment character.  See the :class:`CommentedHeader` class  for an example.
    """
    def process_lines(self, lines):
        """Return only lines that start with the comment regexp.  For these
        lines strip out the matching characters."""
        re_comment = re.compile(self.comment)
        for line in lines:
            match = re_comment.match(line)
            if match:
                yield line[match.end():]

    def write(self, lines):
        lines.append(self.write_comment + self.splitter.join(self.colnames))


class CommentedHeader(Basic):
    """Read a file where the column names are given in a line that begins with
    the header comment character. ``header_start`` can be used to specify the
    line index of column names, and it can be a negative index (for example -1
    for the last commented line).  The default delimiter is the <space>
    character.::

      # col1 col2 col3
      # Comment line
      1 2 3
      4 5 6
    """
    _format_name = 'commented_header'
    _description = 'Column names in a commented line'

    header_class = CommentedHeaderHeader
    data_class = NoHeaderData

Define a custom reader functionally Instead of defining a new class, it is also possible to obtain an instance of a reader and then to modify the properties of this one reader instance in a function:

def read_rdb_table(table):
    reader = astropy.io.ascii.Basic()
    reader.header.splitter.delimiter = '\t'
    reader.data.splitter.delimiter = '\t'
    reader.header.splitter.process_line = None
    reader.data.splitter.process_line = None
    reader.data.start_line = 2

    return reader.read(table)

Create a custom splitter.process_val function

# The default process_val() normally just strips whitespace.
# In addition have it replace empty fields with -999.
def process_val(x):
    """Custom splitter process_val function: Remove whitespace at the beginning
    or end of value and substitute -999 for any blank entries."""
    x = x.strip()
    if x == '':
        x = '-999'
    return x

# Create an RDB reader and override the splitter.process_val function
rdb_reader = astropy.io.ascii.get_reader(Reader=astropy.io.ascii.Rdb)
rdb_reader.data.splitter.process_val = process_val

Reading large tables in chunks

The default process for reading ASCII tables is not memory efficient and may temporarily require much more memory than the size of the file (up to a factor of 5 to 10). In cases where the temporary memory requirement exceeds available memory this can cause significant slowdown when disk cache gets used.

In this situation there is a way to read the table in smaller chunks which are limited in size. There are two possible ways to do this:

  • Read the table in chunks and aggregate the final table along the way. This uses only somewhat more memory than the final table requires.

  • Use a Python generator function to return a Table object for each chunk of the input table. This allows for scanning through arbitrarily large tables since it never returns the final aggregate table.

The chunk reading functionality is most useful for very large tables, so this is available only for the Fast ASCII I/O readers. The following formats are supported: tab, csv, no_header, rdb, and basic. The commented_header format is not directly supported, but as a workaround one can read using the no_header format and explicitly supply the column names using the names argument.

In order to read a table in chunks one must provide the fast_reader keyword argument with a dict that includes the chunk_size key with the value being the approximate size (in bytes) of each chunk of the input table to read. In addition, if one provides a chunk_generator key which is set to True, then instead of returning a single table for the whole input it returns an iterator that provides a table for each chunk of the input.

Example: Reading an entire table while limiting peak memory usage.

# Read a large CSV table in 100 Mb chunks.

tbl = ascii.read('large_table.csv', format='csv', guess=False,
                 fast_reader={'chunk_size': 100 * 1000000})

Example: Reading the table in chunks with an iterator.

Here we iterate over a CSV table and select all rows where the Vmag column is less than 8.0 (e.g. all stars in table brighter than 8.0 mag). We collect all these sub-tables and then stack them at the end.

from astropy.table import vstack

# tbls is an iterator over the chunks (no actual reading done yet)
tbls = ascii.read('large_table.csv', format='csv', guess=False,
                  fast_reader={'chunk_size': 100 * 1000000,
                               'chunk_generator': True})

out_tbls = []

# At this point the file is actually read in chunks.
for tbl in tbls:
    bright = tbl['Vmag'] < 8.0
    if np.count_nonzero(bright):
        out_tbls.append(tbl[bright])

out_tbl = vstack(out_tbls)

Note

Performance

Specifying the format explicitly and using guess=False is a good idea for large tables. This prevent unnecessary guessing in the typical case where the format is already known.

The chunk_size should generally be set to the largest value that is reasonable given available system memory. There is overhead associated with processing each chunk, so the fewer chunks the better.