"""
Core rendering logic for GreaterTables.
Defines the `GreaterTables` class, which formats and renders pandas DataFrames
to HTML, plain text, or LaTeX output using a validated configuration model.
This is the main entry point for rendering logic. See `gtconfig.py` for configuration schema.
"""
from collections import namedtuple
from decimal import InvalidOperation
from io import StringIO
from itertools import groupby
import logging
import os
from pathlib import Path
import re
import tempfile
from typing import Optional, Union, Literal
import warnings
import yaml
from bs4 import BeautifulSoup
from cachetools import LRUCache
import numpy as np
import pandas as pd
from pandas.errors import IntCastingNaNError
from pandas.api.types import is_datetime64_any_dtype, is_integer_dtype, \
is_float_dtype, is_extension_array_dtype
from pydantic import ValidationError
from rich import box
from IPython.display import display, SVG
from . enums import Breakability
from . config import Configurator
from . hasher import df_short_hash
from . etcher import Etcher
from . utilities import *
# turn off this fuck-fest
pd.set_option('future.no_silent_downcasting', True)
# pandas complaining about casting columns eg putting object in float column
warnings.simplefilter(action='ignore', category=FutureWarning)
# GPT recommended approach
logger = logging.getLogger(__name__)
class GT(object):
"""
Create a greater_tables formatting object.
Provides html and latex output in quarto/Jupyter accessible manner.
Wraps AND COPIES the dataframe df. WILL NOT REFLECT CHANGES TO DF.
Recommended usage is to subclass GT (or use functools.partial) and set
defaults suitable to your particular
application. In that way you can maintain a "house-style"
Process
--------
**Input transformation**
* ``namedtuple`` converted to ``DataFrame``
* ``pd.Series`` converted to ``DataFrame``
* ``list`` converted to ``DataFrame``, optionally using row 0 as
``config.header_row``
* A string is assumed to be a pipe-separated markdown table which is
converted to a ``DataFrame`` setting aligners per the alignment row
* All other input types are an error
The input ``df`` must have unique column names. It is then copied into
``self.df`` which will be changed and ``self.raw_df`` for reference.
The copy is hashed for the table name.
**Mangling**
* If show_index, the index is reset and kept, so that all columns are on an
config.equal footing
* The index change levels are computed to determine LaTeX hrules
* ratio year, and raw columns converted to a list (can be input as a single
string name)
* Columns, except raw columns, are cast to floats
* Column types by index determined
* default formatter function set (wrapping input, if any)
* Aligner column input decoded into aligner values
(``grt-left,grt-right,grt-center``); index aligners separated
* Formatters decoded, strings mapped to lambda functions as f-string
formatters, integers as number of decimals
* Tab values expanded into an iterable
* Dataframe at this point (index reset, cast) saved to
``df_pre_applying_formatters``
* Determine formatters (``df_formatters`` property, a list of column index
formatting functions:
* Make the default float formatter if entered (callable, string, number;
wrapped in try/except)
* Determine each column's format type and add function
* Run ``apply_formatters`` to apply all format choices to ``df``. This
function handles index columns slightly differently, but results in the
formatters being applied to each column.
* Sparsify if requested and if multiindex
* Result is a dataframe with all object column types and values that
reflect the formatting choices.
"""
def __init__(
self,
df,
*,
caption='',
label='',
aligners: dict[str, callable] | None = None,
formatters: dict[str, callable] | None = None,
tabs: Optional[Union[list[float], float, int]] | None = None,
unbreakable=None,
ratio_cols=None,
year_cols=None,
date_cols=None,
raw_cols=None,
show_index=True,
#
config: Configurator | None = None,
config_path: Path | None = None,
**overrides,
):
"""
Parameters:
:param df: target DataFrame or list of lists or markdown table string
:param caption: table caption, optional (GT will look for gt_caption
attribute of df and use that)
:param label: TeX label (used in \\label{} command). For markdown
tables with #tbl:... in the caption it is extracted automatically.
:param aligners: None or dict (type or colname) -> left | center |
right
:param formatters: None or dict (type or colname) or callable ->
format function for the column; formatters trump ratio_cols; if
callable passed, it is applied to **all columns**.
:param tabs: None or list of column widths in characters or a common
int or float width. (It is converted into em; one character is about
0.5em on average; digits are exactly 0.5em.) If None, will be calculated.
Default None.
:param unbreakable: None or list of columns to be considered unbreakable
:param ratio_cols: None, or "all" or list of column names treated as
ratios. Set defaults in derived class suitable to application.
:param year_cols: None, or "all" or list of column names treated as
years (no commas, no decimals). Set defaults in derived class suitable
to application.
:param date_cols: None, or "all" or list of column names treated as
dates. Set defaults in derived class suitable to application.
:param raw_cols: None, or "all" or list of column names that are NOT
cast to floats. Set defaults in derived class suitable to application.
:param show_index: if True, show the index columns, default True
Available keyword *overrides:
:param config.default_integer_str: format f-string for integers, default
value '{x:,d}'
:param config.default_float_str: format f-string for floats, default
value '{x:,.3f}'
:param config.default_date_str: format f-string for dates, default '%Y-%m-%d'.
NOTE: no braces or x!
:param config.default_ratio_str: format f-string for ratios, default '{x:.1%}'
:param config.table_float_format: None or format string for floats in the
table format function, applied to entire table, default None
:param config.table_hrule_width: width of the table top, botton and header
hrule, default 1
:param config.table_vrule_width: width of the table vrule, separating the
index from the body, default 1
:param config.hrule_widths: None or tuple of three ints for hrule widths
(for use with multiindexes)
:param config.vrule_widths: None or tuple of three ints for vrule widths
(for use when columns have multiindexes)
:param config.sparsify: if True, config.sparsify the index columns, you almost always
want this to be true!
:param config.sparsify_columns: if True, config.sparsify the columns, default True,
generally a better look, headings centered in colspans
:param config.spacing: 'tight', 'medium', 'wide' to quickly set cell padding.
Medium is default (2, 10, 2, 10).
:param config.padding_trbl: None or tuple of four ints for padding, in order
top, right, bottom, left.
:param config.tikz_scale: scale factor applied to tikz LaTeX tables.
:param config.font_body: font size for body text, default 0.9. Units in em.
:param config.font_head: font size for header text, default 1.0. Units in em.
:param config.font_caption: font size for caption text, default 1.1.
Units in em.
:param config.font_bold_index: if True, make the index columns bold,
default False.
:param config.pef_precision: precision (digits after period) for pandas
engineering format, default 3.
:param config.pef_lower: apply engineering format to floats with absolute
value < 10**config.pef_lower; default -3.
:param config.pef_upper: apply engineering format to floats with absolute
value > 10**config.pef_upper; default 6.
:param config.cast_to_floats: if True, try to cast all non-integer, non-date
columns to floats
:param config.header_row: True: use first row as headers; False no headings.
Default True
:param config.equal: if True, set all column widths config.equal. Default False. Maybe
ignored, depending on computed ideal column widths.
:param config.caption_align: for the caption
:param config.large_ok: signal that you are intentionally applying to a large
dataframe. Sub-classes may restrict or apply .head() to df.
:param config.max_str_length: maximum displayed length of object types, that
are cast to strings. Eg if you have nested DataFrames!
:param str_table_fmt: table border format used for string output
(markdown), default mixed_grid DEPRECATED??
:param config.table_width_mode:
'explicit': set using config.max_table_width_em
'natural': each cell on one line (can be very wide with long strings)
'breakable': wrap breakable cells (text strings) at word boundaries
to fit longest word
'minimum': wrap breakable and ok-to-break (dates) cells
:param config.table_width_header_adjust: additional proportion of table width
used to balance header columns.
:param config.table_width_header_relax: extra spaces allowed per column heading
to facilitate better column header wrapping.
:param config.max_table_width_em: max table width used for markdown string output,
default 200; width is never less than minimum width. Padding (3 chars
per row plus 1) consumed out of config.max_table_width_em in string output mode.
:param config.debug: if True, add id to caption and use colored lines in table,
default False.
"""
if config and config_path:
raise ValueError(
"Pass either 'config' or 'config_path', not both.")
if config:
base_config = config
elif config_path:
config_path = Path(config_path)
try:
raw = yaml.safe_load(config_path.read_text(encoding="utf-8"))
base_config = Configurator.model_validate(raw)
except (ValidationError, OSError) as e:
raise ValueError(
f"Failed to load config from {config_path}") from e
else:
base_config = Configurator()
# access through config
# update and validate; need to merge to avoid repeated args
# merged = dict(base_config.model_dump(), **overrides)
merged = base_config.model_dump() | overrides
self.config = Configurator(**merged)
# no validation
# self.config = base_config.model_copy(update=overrides)
# deal with alternative input modes for df: None, DataFrame, Series, markdown text table
if df is None:
# don't want None to fail
df = pd.DataFrame([])
if isinstance(df, pd.DataFrame):
# usual use case
pass
elif isinstance(df, pd.Series):
df = df.to_frame()
elif isinstance(df, list):
df = pd.DataFrame(df)
# override this selection come what may
show_index = False
if self.config.header_row:
# Set first row as column names
df.columns = df.iloc[0]
# Drop first row and reset index
df = df[1:].reset_index(drop=True)
elif isinstance(df, str):
df = df.strip()
if df == '':
df = pd.DataFrame([])
else:
df, aligners, caption, label = MD2DF.md_to_df(df)
show_index = False
elif GT._is_namedtuple_instance(df):
df = GT._ntdf(df)
else:
raise ValueError(
'df must be a DataFrame, a list of lists, or a markdown table string')
if len(df) > self.config.large_warning and not self.config.large_ok:
raise ValueError(
f'Large dataframe (>{self.config.large_warning} rows) and config.large_ok not set to true. '
'Set large_ok=True or increase value of large_warning.')
if not df.columns.is_unique:
raise ValueError('df column names are not unique')
# extract value BEFORE copying, copying does not carry these attributes over
if caption != '':
self.caption = caption
else:
# used by querex etc.
self.caption = getattr(df, 'gt_caption', '')
self.label = label
self.df = df.copy(deep=True) # the object being formatted
self.raw_df = df.copy(deep=True)
# if not column_names:
# get rid of column names
# self.df.columns.names = [None] * self.df.columns.nlevels
self.df_id = df_short_hash(self.df)
if self.caption != '' and self.config.debug:
self.caption += f' (id: {self.df_id})'
# self.max_str_length = max_str_length
# before messing
self.show_index = show_index
self.nindex = self.df.index.nlevels if self.show_index else 0
self.ncolumns = self.df.columns.nlevels
self.ncols = self.df.shape[1]
self.dt = self.df.dtypes
# reset index to put all columns on an config.equal footing, but note number ofindex cols
with warnings.catch_warnings():
if self.show_index:
warnings.simplefilter(
"ignore", category=pd.errors.PerformanceWarning)
self.df = self.df.reset_index(
drop=False, col_level=self.df.columns.nlevels - 1)
# want the new index to be ints - that is not default if old was multiindex
self.df.index = np.arange(self.df.shape[0], dtype=int)
self.index_change_level = Indexing.changed_column(
self.df.iloc[:, :self.nindex])
if self.ncolumns > 1:
# will be empty rows above the index headers
self.index_change_level = pd.Series(
[i[-1] for i in self.index_change_level])
self.column_change_level = Indexing.changed_level(self.raw_df.columns)
# determine ratio columns
if ratio_cols is not None and not self.df.columns.is_unique:
logger.warning(
'Ratio cols specified with non-unique column names: ignoring request.')
self.ratio_cols = []
else:
if ratio_cols is None:
self.ratio_cols = []
elif ratio_cols == 'all':
self.ratio_cols = [i for i in self.df.columns]
elif ratio_cols is not None and not isinstance(ratio_cols, (tuple, list)):
self.ratio_cols = self.cols_from_regex(
ratio_cols) # [ratio_cols]
else:
self.ratio_cols = ratio_cols
# determine year columns
if year_cols is not None and not self.df.columns.is_unique:
logger.warning(
'Year cols specified with non-unique column names: ignoring request.')
self.year_cols = []
else:
if year_cols is None:
self.year_cols = []
elif year_cols is not None and not isinstance(year_cols, (tuple, list)):
self.year_cols = self.cols_from_regex(year_cols) # [year_cols]
else:
self.year_cols = year_cols
# determine date columns
if date_cols is not None and not self.df.columns.is_unique:
logger.warning(
'Year cols specified with non-unique column names: ignoring request.')
self.date_cols = []
else:
if date_cols is None:
self.date_cols = []
elif date_cols is not None and not isinstance(date_cols, (tuple, list)):
self.date_cols = self.cols_from_regex(date_cols) # [date_cols]
else:
self.date_cols = date_cols
# determine columns NOT to cast to floats
if raw_cols is not None and not self.df.columns.is_unique:
logger.warning(
'Year cols specified with non-unique column names: ignoring request.')
self.raw_cols = []
else:
if raw_cols is None:
self.raw_cols = []
elif raw_cols is not None and not isinstance(raw_cols, (tuple, list)):
self.raw_cols = self.cols_from_regex(raw_cols) # [raw_cols]
else:
self.raw_cols = raw_cols
# figure the default formatter (used in conjunction with raw columns)
if self.config.default_formatter is None:
self.default_formatter = self._default_formatter
else:
assert callable(
self.config.default_formatter), 'config.default_formatter must be callable'
def wrapped_default_formatter(x):
try:
return self.config.default_formatter(x)
except ValueError:
return str(x)
self.default_formatter = wrapped_default_formatter
# cast as much as possible to floats
with warnings.catch_warnings():
warnings.simplefilter(
"ignore", category=pd.errors.PerformanceWarning)
if self.config.cast_to_floats:
for i, c in enumerate(self.df.columns):
if c in self.raw_cols or c in self.date_cols:
continue
old_type = self.df.dtypes[c]
if not np.any((is_integer_dtype(self.df.iloc[:, i]),
is_datetime64_any_dtype(self.df.iloc[:, i]),
# pyarrow classes excluded too
is_extension_array_dtype(self.df.iloc[:, i]))):
try:
self.df.iloc[:, i] = self.df.iloc[: ,i].astype(float)
logger.debug(
f'coerced column {i}={c} from {old_type} to float success')
except (ValueError, TypeError):
logger.debug(
f'coercing {i}={c} from {old_type} to float FAILED')
# massage unbreakable
if unbreakable is None:
unbreakable = []
elif isinstance(unbreakable, str):
unbreakable = [unbreakable]
# now can determine types and infer the break penalties (for column sizes)
self.float_col_indices = []
self.integer_col_indices = []
self.date_col_indices = []
self.object_col_indices = [] # not actually used, but for neatness
self.break_penalties = []
# manage non-unique col names here
logger.debug('FIGURING TYPES')
for i, cn in enumerate(self.df.columns): # range(self.df.shape[1]):
ser = self.df.iloc[:, i]
if cn in self.date_cols:
logger.debug(f'col {i}/{cn} specified as date col')
self.date_col_indices.append(i)
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.DATE)
elif is_datetime64_any_dtype(ser):
logger.debug(f'col {i} = {self.df.columns[i]} is DATE')
self.date_col_indices.append(i)
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.DATE)
elif is_integer_dtype(ser):
logger.debug(f'col {i} = {self.df.columns[i]} is INTEGER')
self.integer_col_indices.append(i)
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.NEVER)
elif is_float_dtype(ser):
logger.debug(f'col {i} = {self.df.columns[i]} is FLOAT')
self.float_col_indices.append(i)
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.NEVER)
else:
logger.debug(f'col {i} = {self.df.columns[i]} is OBJECT')
self.object_col_indices.append(i)
c = ser.name
if c in self.year_cols or c in self.ratio_cols:
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.NEVER)
else:
self.break_penalties.append(
Breakability.NEVER if cn in unbreakable else Breakability.ACCEPTABLE)
# figure out column and index alignment
if aligners is not None and np.any(self.df.columns.duplicated()):
logger.warning(
'aligners specified with non-unique column names: ignoring request.')
aligners = None
if aligners is None:
# not using
aligners = []
elif isinstance(aligners, str):
# lrc for each column
aligners = {c: a for c, a in zip(self.df.columns, aligners)}
self.df_aligners = []
lrc = {'l': 'grt-left', 'r': 'grt-right', 'c': 'grt-center'}
# TODO: index aligners
for i, c in enumerate(self.df.columns):
# test aligners BEFORE index!
if c in aligners:
self.df_aligners.append(lrc.get(aligners[c], 'grt-center'))
elif i < self.nindex:
# index -> left
self.df_aligners.append('grt-left')
elif c in self.year_cols:
self.df_aligners.append('grt-center')
elif c in self.raw_cols:
# these are strings
self.df_aligners.append('grt-left')
elif i in self.date_col_indices:
# center dates, why not!
self.df_aligners.append('grt-center')
elif c in self.ratio_cols or i in self.float_col_indices or i in self.integer_col_indices:
# number -> right
self.df_aligners.append('grt-right')
else:
# all else, left
self.df_aligners.append('grt-left')
self.df_idx_aligners = self.df_aligners[:self.nindex]
self.default_formatters = {}
if formatters is None:
pass
elif callable(formatters):
# apply to all columns
for k in self.df.columns:
self.default_formatters[k] = formatters
else:
for k, v in formatters.items():
if callable(v):
self.default_formatters[k] = v
elif isinstance(v, str):
self.default_formatters[k] = lambda x: v.format(x=x)
elif isinstance(v, int):
fmt = f'{{x:.{v}f}}'
self.default_formatters[k] = lambda x: fmt.format(x=x)
else:
raise ValueError(
'formatters must be dict of callables or ints or format strings {x:...}')
if tabs is None:
self.tabs = None
elif isinstance(tabs, (int, float)):
self.tabs = (tabs,) * (self.nindex + self.ncols)
elif isinstance(tabs, (np.ndarray, pd.Series, list, tuple)):
if len(tabs) == self.nindex + self.ncols:
self.tabs = tabs # Already iterable and right length, self.tabs = as is
else:
logger.error(
f'{self.tabs=} has wrong length. Ignoring.')
self.tabs = None
else:
logger.error(
f'{self.tabs=} must be None, a single number, or a list of '
'numbers of the correct length. Ignoring.')
self.tabs = None
if self.config.padding_trbl is not None:
padding_trbl = self.config.padding_trbl
elif self.config.padding_trbl is None:
if self.config.spacing == 'tight':
padding_trbl = (0, 5, 0, 5)
elif self.config.spacing == 'medium':
padding_trbl = (2, 10, 2, 10)
elif self.config.spacing == 'wide':
padding_trbl = (4, 15, 4, 15)
else:
raise ValueError(
'config.spacing must be tight, medium, or wide or tuple of four ints.')
# pydantic will see to it this is OK
self.padt, self.padr, self.padb, self.padl = padding_trbl
# because of the problem of non-unique indexes use a list and
# not a dict to pass the formatters to to_html
self.max_table_width_em = self.config.max_table_inch_width * 72 / self.config.table_font_pt_size
self._pef = None
self._df_formatters = None
self.df_style = ''
self.df_html = ''
self._clean_html = ''
self._clean_tex = ''
self._rich_table = None
self._string = ''
self._df_html_text = ""
self._df_style_text = ""
self._cache = LRUCache(20)
self._text_knowledge_df = None
self._html_knowledge_df = None
self._tex_knowledge_df = None
self._knowledge_dfs = None
# finally config.sparsify and then apply formaters
# this radically alters the df, so keep a copy for now...
self.df_pre_applying_formatters = self.df.copy()
self.df = self.apply_formatters(self.df)
# config.sparsify
if self.config.sparsify and self.nindex > 1:
self.df = Sparsify.sparsify(self.df, self.df.columns[:self.nindex])
# for c in self.df.columns[:self.nindex]:
# # config.sparsify returns some other stuff...
# self.df[c], _ = GT.config.sparsify(self.df[c])
# make final tex and html versions
if self.config.tex_to_html is not None:
# NEED TO WORK ON INDEXES TOO
self.df_html = self.df.map(self.config.tex_to_html)
else:
self.df_html = self.df
if self.config.tikz_escape_tex:
self.df_tex = Escaping.escape_df_tex(self.df)
else:
self.df_tex = self.df
def __repr__(self):
"""Basic representation."""
return f"GT(df_id={self.df_id})"
def __str__(self):
"""String representation, for print()."""
return self.make_string()
def _repr_html_(self):
"""
Apply format to self.df.
ratio cols like in constructor
"""
return self.html
def _repr_latex_(self):
"""Generate a LaTeX tabular representation."""
# return ''
# latex = self.df.to_latex(caption=self.caption, formatters=self._df_formatters)
if self._clean_tex == '':
self._clean_tex = self.make_tikz()
logger.info('CREATED LATEX')
return self._clean_tex
def cache_get(self, key):
"""Retrieve item from cache."""
return self._cache.get(key, None)
def cache_set(self, key, value):
"""Add item to cache."""
self._cache[key] = value
def cols_from_regex(self, regex):
"""
Return columns matching a regex.
For Index and MultiIndex. Operates on ``self.df`` and includes
index (if ``show_index``) and columns of input dataframe. Search
applies to any level of the index. Case sensitive.
"""
pattern = re.compile(regex)
matching_cols = [
col for col in self.df.columns
if any(pattern.search(str(level))
for level in (col if isinstance(col, tuple) else (col,)))
]
return matching_cols
# return [col for col in self.df.columns if isinstance(col, str) and re.search(regex, col)]
# define the default and easy formatters ===================================================
def default_ratio_formatter(self, x):
"""Ratio formatter."""
try:
return self.config.default_ratio_str.format(x=x)
except ValueError:
return str(x)
def default_date_formatter(self, x):
"""Date formatter that works for strings too."""
if pd.isna(x):
return ""
try:
dt = pd.to_datetime(x, errors='coerce')
if pd.isna(dt):
return str(x)
return dt.strftime(self.config.default_date_str)
except Exception:
logger.error("date error with %s", x)
return str(x)
def default_integer_formatter(self, x):
"""Integer formatter."""
try:
return self.config.default_integer_str.format(x=x)
except ValueError:
return str(x)
def default_year_formatter(self, x):
"""Year formatter."""
try:
return f'{int(x):d}'
except (TypeError, ValueError):
return str(x)
def default_raw_formatter(self, x):
"""Formatter for columns flagged as raw."""
return str(x)
@staticmethod
def default_float_format(x, neng=3):
"""
the endless quest for the perfect float formatter...
NOT USED AT THE MINUTE.
tester::
for x in 1.123123982398324723947 * 10.**np.arange(-23, 23):
print(default_float_format(x))
:param x:
:return:
"""
ef = pd.io.formats.format.EngFormatter(neng, True) # noqa
try:
if x == 0:
ans = '0'
elif 1e-3 <= abs(x) < 1e6:
if abs(x) <= 10:
ans = f'{x:.3g}'
elif abs(x) < 100:
ans = f'{x:,.2f}'
elif abs(x) < 1000:
ans = f'{x:,.1f}'
else:
ans = f'{x:,.0f}'
else:
ans = ef(x)
return ans
except ValueError as e:
logger.debug(f'ValueError {e}')
return str(x)
except TypeError as e:
logger.debug(f'TypeError {e}')
return str(x)
except AttributeError as e:
logger.debug(f'AttributeError {e}')
return str(x)
def _default_formatter(self, x):
"""Default universal formatter for other types."""
try:
f = float(x)
except (TypeError, ValueError):
s = str(x)
return s if self.config.max_str_length < 0 else s[:self.config.max_str_length]
if self.default_float_formatter:
return self.default_float_formatter(f)
if np.isinf(f) or np.isnan(f): # clearer handling of weird float cases
return str(x)
if f.is_integer():
return self.config.default_integer_str.format(x=int(f))
else:
return self.config.default_float_str.format(x=f)
def pef(self, x):
"""Pandas engineering format."""
if self._pef is None:
self._pef = pd.io.formats.format.EngFormatter(accuracy=self.config.pef_precision, use_eng_prefix=True) # noqa
return self._pef(x)
def make_float_formatter(self, ser):
"""
Make a float formatter suitable for the Series ser.
Obeys these rules:
* All elements in the column are formatted consistently
* ...
TODO flesh out... at some point shd use pef?!
"""
amean = ser.abs().mean()
# mean = ser.mean()
amn = ser.abs().min()
amx = ser.abs().max()
# smallest = ser.abs().min()
# sd = ser.sd()
# p10, p50, p90 = np.quantile(ser, [0.1, .5, 0.9], method='inverted_cdf')
# pl = 10. ** self.config.pef_lower
# pu = 10. ** self.config.pef_upper
pl, pu = 10. ** self.config.pef_lower, 10. ** self.config.pef_upper
if amean < 1:
precision = 5
elif amean < 10:
precision = 3
elif amean < 20000:
precision = 2
else:
precision = 0
fmt = f'{{x:,.{precision}f}}'
logger.debug(f'{ser.name=}, {amean=}, {fmt=}')
if amean < pl or amean > pu or amx / max(1, amn) > pu:
# go with eng
def ff(x):
try:
return self.pef(x)
except (ValueError, TypeError, InvalidOperation):
return str(x)
else:
def ff(x):
try:
return fmt.format(x=x)
# well and good but results in ugly differences
# by entries in a column
# if x == int(x) and np.abs(x) < pu:
# return f'{x:,.0f}.'
# else:
# return fmt.format(x=x)
except (ValueError, TypeError):
return str(x)
return ff
@ property
def df_formatters(self):
"""
Make and return the list of formatters.
Created one per column. Int, date, objects use defaults, but
for float cols the formatter is created custom to the details of
each column.
"""
if self._df_formatters is None:
# because of non-unique indexes, index by position not name
if self.config.table_float_format is not None:
if callable(self.config.table_float_format):
# wrap in error protections
def ff(x):
try:
return self.config.table_float_format(x=x)
except ValueError:
return str(x)
except Exception as e:
logger.error(f'Custom float function raised {e=}')
self.default_float_formatter = ff
else:
if type(self.config.table_float_format) != str:
raise ValueError(
'config.table_float_format must be a string or a function')
fmt = self.config.table_float_format
def ff(x):
try:
return fmt.format(x=x)
except ValueError:
return str(x)
except Exception as e:
logger.error(
f'Custom float format string raised {e=}')
self.default_float_formatter = ff
else:
self.default_float_formatter = False
self._df_formatters = []
for i, c in enumerate(self.df.columns):
# set a default, note here can have
# non-unique index so work with position i
if c in self.default_formatters:
self._df_formatters.append(self.default_formatters[c])
elif c in self.ratio_cols:
# print(f'{i} ratio')
self._df_formatters.append(self.default_ratio_formatter)
elif c in self.year_cols:
self._df_formatters.append(self.default_year_formatter)
elif c in self.raw_cols:
self._df_formatters.append(self.default_raw_formatter)
elif i in self.date_col_indices:
self._df_formatters.append(self.default_date_formatter)
elif i in self.integer_col_indices:
# print(f'{i} int')
self._df_formatters.append(self.default_integer_formatter)
elif i in self.float_col_indices:
# trickier approach...
self._df_formatters.append(
self.default_float_formatter or self.make_float_formatter(self.df.iloc[:, i]))
else:
# print(f'{i} default')
self._df_formatters.append(self.default_formatter)
# self._df_formatters is now a list of length config.equal to cols in df
if len(self._df_formatters) != self.df.shape[1]:
raise ValueError(
f'Something wrong: {len(self._df_formatters)=} != {self.df.shape=}')
return self._df_formatters
@staticmethod
def apply_formatters_work(df, formatters):
"""Apply formatters to a DataFrame."""
try:
# very surprising bug: if df is empty new_df will have type float!!
new_df = pd.DataFrame({i: map(f, df.iloc[:, i])
for i, f in enumerate(formatters)}, dtype=object)
except TypeError:
print('NASTY TYPE ERROR')
raise
new_df.columns = df.columns
return new_df
def apply_formatters(self, df, mode='adjusted'):
"""
Replace df (the raw df) with formatted df, including the index.
If mode is 'adjusted' operates on columns only, does not touch the
index. Otherwise, called from tikz and operating on raw_df
"""
if mode == 'adjusted':
# apply to df where the index has been reset
# number of columns = len(self.df_formatters)
return GT.apply_formatters_work(df, self.df_formatters)
elif mode == 'raw':
# work on raw_df where the index has not been reset
# because of non-unique indexes, index by position not name
# create the df and the index separately
data_formatters = self.df_formatters[self.nindex:]
new_body = GT.apply_formatters_work(df, data_formatters)
if not self.show_index:
return new_body
# else have to handle the index
index_formatters = self.df_formatters[:self.nindex]
df_index = df.reset_index(
drop=False, col_level=self.df.columns.nlevels - 1).iloc[:, :self.nindex]
new_index = GT.apply_formatters_work(df_index, index_formatters)
# put them back together
new_df = pd.concat([new_index, new_body], axis=1)
new_df = new_df.set_index(list(df_index.columns))
new_df.index.names = df.index.names
return new_df
else:
raise ValueError(f'unknown mode {mode}')
@property
def text_knowledge_df(self):
"""Uber source of information for text formatting."""
if self._text_knowledge_df is None:
self._text_knowledge_df = self.estimate_column_widths_by_mode('text')
return self._text_knowledge_df
@property
def html_knowledge_df(self):
"""Uber source of information for html formatting."""
if self._html_knowledge_df is None:
self._html_knowledge_df = self.estimate_column_widths_by_mode('html')
return self._html_knowledge_df
@property
def tex_knowledge_df(self):
"""Uber source of information for tex formatting."""
if self._tex_knowledge_df is None:
# seems this is unlikely to be a good idea!
# if (all(self.df_tex.index == self.df_html.index)
# and all(self.df_tex.columns == self.df_html.columns)
# and all(self.df_tex == self.df_html)):
# self._tex_knowledge_df = self.html_knowledge_df
# else:
if not self.config.tikz:
# just repeat html so you have something
self._tex_knowledge_df = self.html_knowledge_df
else:
self._tex_knowledge_df = self.estimate_column_widths_by_mode('tex')
return self._tex_knowledge_df
@property
def knowledge_dfs(self):
if self._knowledge_dfs is None:
self._knowledge_dfs = pd.concat((self.text_knowledge_df.T,
self.html_knowledge_df.T, self.tex_knowledge_df.T),
keys=['text','html', 'tex'], names=['mode', 'measure'])
self._knowledge_dfs['Total'] = self._knowledge_dfs.fillna(0.).apply(
lambda row: sum(x for x in row if pd.api.types.is_number(x)), axis=1)
idx = self._knowledge_dfs.query('Total == 0').index
self._knowledge_dfs.loc[idx, 'Total'] = ''
self._knowledge_dfs = self._knowledge_dfs.fillna('')
return self._knowledge_dfs
def width_report(self):
"""Return a report summarizing the width information."""
natural = self.text_knowledge_df.natural_width.sum()
minimum = self.text_knowledge_df.minimum_width.sum()
text = self.text_knowledge_df.recommended.sum()
h = self.html_knowledge_df.recommended.sum()
tex = self.tex_knowledge_df.recommended.sum()
tikz = self.tex_knowledge_df.tikz_colw.sum()
mtw = self.max_table_width_em
mtiw = self.config.max_table_inch_width
pts = self.config.table_font_pt_size
bit = pd.DataFrame({
'text natural': self.text_knowledge_df.natural_width,
'text minimum': self.text_knowledge_df.minimum_width,
'text recommended': self.text_knowledge_df.recommended,
'html recommended': self.html_knowledge_df.recommended,
'tex recommended': self.tex_knowledge_df.recommended,
'tikz recommended': self.tex_knowledge_df.tikz_colw,
}).fillna(0)
ser = pd.Series({
'text natural': natural,
'text minimum': minimum,
'text recommended': text,
'html recommended': h,
'tex recommended': tex,
'tikz recommended': tikz,
})
bit.loc['total', :] = ser
print(f"requested width = {mtw} em\n"
f"max tbl inch w = {mtiw} inches\n"
f"font pts = {pts} pts\n"
f"width in em chk = {mtiw * 72 / pts} em\n"
f"width mode = {self.config.table_width_mode}\n"
f"header relax = {self.config.table_width_header_adjust}\n"
f"header chars = {self.config.table_width_header_relax}")
return bit
def estimate_column_widths_by_mode(self, mode):
r"""
Return dataframe of width information: three modes for text, html, and tex.
Mode adjusts which df is used and how widths are estimated
* text -> self.df and len = str.len
* html -> self.df_html and len =
* tex -> self.df_tex and len =
Returned dataframe has columns named mode_xxx, where xxx can be
* natural: max len by col
* minimum width = max length given breaks
* acceptable = allowing for break type by column
* head_natural, head_min, head_acceptable for the heading
* raw_recommended
* header_adjustment
* recommended
pat and iso_date_split regex explanation:
# re.split(r'(?<=[\s.,:;!?()\[\]{}\-\\/|])\s*', text)
# (?<=...) is a lookbehind to preserve the break character with the left-hand fragment.
# [\s.,:;!?()\[\]{}\-\\/|] matches common punctuation and separators:
# \s = whitespace
# . , : ; ! ? = terminal punctuation
# () [] {} = brackets
# \- = dash
# \\/| = slash, backslash, pipe
"""
assert mode in ('text', 'html', 'tex'), 'Only html, text and tex modes valid.'
if mode == 'text':
df = self.df
len_function = len
# no bold in text mode
bold_adjustment = 1.0
elif mode == 'html':
df = self.df_html
len_function = TextLength.text_display_len
bold_adjustment = 1.1
else: # mode == 'tex':
df = self.df_tex
len_function = TextLength.text_display_len
bold_adjustment = 1.1
n_row, n_col = df.shape
# The width if content didn't wrap (single line)
# Series=dict colname->max width of cells in column
natural_width = df.map(lambda x: len_function(x.strip())).max(axis=0).to_dict()
# in text mode: figure out where you can break; pat breaks after punctuation or at -
pat = r'(?<=[.,;:!?)\]}\u2014\u2013])\s+|--*\s+|\s+'
iso_date_split = r'(?<=\b\d{4})-(?=\d{2}-\d{2})'
pat = f'{pat}|{iso_date_split}'
# Calculate ideal (no wrap) and minimum possible widths for all columns
# The absolute minimum width each column can take (e.g., longest word for text)
minimum_width = {}
header_natural = {}
header_minimum = {}
for col_name in df.columns:
minimum_width[col_name] = (
df[col_name].str
.split(pat=pat, regex=True, expand=True)
.fillna('')
.map(len_function)
.max(axis=1)
.max()
)
# ensure is a tuple
ctuple = col_name if isinstance(col_name, tuple) else (col_name, )
header_natural[col_name] = bold_adjustment * max(map(len_function, ctuple))
header_minimum[col_name] = bold_adjustment * min(len_function(part) for i in ctuple for part in re.split(pat, str(i)))
# begin to assemble the parts
# ans will be the col_width_df; break_penalties needed by all methods
ans = pd.DataFrame({
'alignment': [i[4:] for i in self.df_aligners],
'break_penalties': self.break_penalties,
'breakability': [x.name for x in self.break_penalties],
'natural_width': natural_width.values(),
'minimum_width': minimum_width.values(),
}, index=df.columns)
ans['acceptable_width'] = np.where(
ans.break_penalties == Breakability.ACCEPTABLE, ans.minimum_width, ans.natural_width)
ans['header_natural'] = header_natural
ans['header_minimum'] = header_minimum
if mode in ('html', 'tex'):
# put in some padding TODO KLUDGE
ans['natural_width'] += 1
ans['minimum_width'] += 1
ans['header_natural'] += 1
ans['header_minimum'] += 1
# adjustments and recommendations - these are keyed to text output with padding
natural, acceptable, minimum = ans.iloc[:, 3:6].sum()
head_natural, head_minimum = ans.iloc[:, 6:8].sum()
if mode == 'text':
# +1 for the pipe | symbol
PADDING = 2 # per column TODO enhance
pad_adjustment = (PADDING + 1) * n_col - 1
else:
PADDING = 1 # per column TODO enhance
pad_adjustment = PADDING * n_col
if self.config.table_width_mode == 'explicit':
# target width INCLUDES padding and column marks |
target_width = self.max_table_width_em - pad_adjustment
elif self.config.table_width_mode == 'natural':
target_width = natural + pad_adjustment
elif self.config.table_width_mode == 'breakable':
target_width = acceptable + pad_adjustment
elif self.config.table_width_mode == 'minimum':
target_width = minimum + pad_adjustment
logger.info('table_width_mode = %s', self.config.table_width_mode)
logger.info('config self.max_table_width_em %s', self.max_table_width_em)
logger.info('target width after column spacer adjustment %s', target_width)
# extra space for the headers to relax, if useful
if self.config.table_width_header_adjust > 0:
max_extra = int(self.config.table_width_header_adjust * target_width)
else:
max_extra = 0
if target_width > natural:
# everything gets its natural width
ans['recommended'] = ans['natural_width']
space = target_width - natural
logger.info('Space for NATURAL! Spare space = %s', space)
elif target_width > acceptable:
# strings wrap
ans['recommended'] = ans['acceptable_width']
# use up extra on the ACCEPTABLE cols
space = target_width - acceptable
logger.info(
'Using "breaks acceptable" (dates not wrapped), spare space = %s', space)
elif target_width > minimum:
# strings and dates wrap
ans['recommended'] = ans['minimum_width']
# use up extra on dates first, then strings
space = target_width - minimum
logger.info(
'Using "minimum" (all breakable incl dates), spare space = %s', space)
else:
# OK severely too small
ans['recommended'] = ans['minimum_width']
space = target_width - minimum
# hard to shut this up...
logger.info(
'Mode %s, desired width too small for pleasant formatting, table will be too wide by spare space %s em < 0.',
mode, space)
logger.info(f'{mode=} {target_width=}, {natural=}, {acceptable=}, {minimum=}, {max_extra=}, {space=}')
# this section tweaks the widths for column headers -> text output only.
# trust tex and html output to naturally make better decisions about line breaks in the heading.
if mode == "text" and space > 0 and df.columns.nlevels == 1:
# text mode only: see if some header tweaks are in order (Index only for now, TODO)
# Step 1: baseline comes in from code above
ans['raw_recommended'] = ans['recommended']
# Step 2: optimize to get rid of intra-line breaks
if max_extra > 0:
adj = Width.header_adjustment(df, ans['recommended'], space, max_extra)
# create new col and populate per GPT
ans['header_tweak'] = pd.Series(adj)
else:
ans['header_tweak'] = 0
ans['recommended'] = ans['recommended'] + ans['header_tweak']
# in this case zero out impact of header_natural and header_minimum cos don't want to use them below
ans['header_natural'] = ans['recommended']
ans['header_minimum'] = ans['recommended']
# Step 3 (all modes): distribute remaining shortfall proportionally
# account for
# obvs remaining == space if mode is not text
remaining = target_width - ans['recommended'].sum()
ans['pre_shortfall_recommended'] = ans['recommended']
if remaining > 0:
shortfall = ans[['natural_width', 'header_natural']].max(axis=1) - ans['recommended']
total_shortfall = shortfall.clip(lower=0).sum()
if total_shortfall > 0:
logger.info('total shortfall to allocate after header adjustments = %s', total_shortfall)
fractions = shortfall.clip(lower=0) / total_shortfall
ans['proto_recommended'] = ans['recommended'] + np.floor(fractions * remaining).astype(int)
ans['recommended'] = np.minimum(ans[['natural_width', 'header_natural']].max(axis=1),
ans['proto_recommended'])
else:
logger.info('no shortfall to allocate after header adjustments')
if mode == 'tex':
# tex mode only need tikz raw size for tex code layout
tikz_colw = dict.fromkeys(df.columns, 0)
tikz_headw = dict.fromkeys(df.columns, 0)
for i, c in enumerate(df.columns):
# figure width of the column labels
c0 = c # before we mess around with it, for setting dict values
if not isinstance(c, tuple):
# make it one: now index and multi index on same footing
c = (c,)
# convert to strings
c = [str(i) for i in c]
tikz_headw[c0] = max(map(len, c))
# now figure the width of the elements in the column
tikz_colw[c0] = df.iloc[:, i].map(lambda x: len(str(x))).max()
# needed tikz width is greater of two
for c in df.columns:
tikz_colw[c] = max(tikz_colw[c], tikz_headw[c])
# distribute any overage using the measures already done
ans['tikz_colw'] = tikz_colw
ans['tikz_colw'] += 2 # for \I
# in all cases...assemble the answer with relevant information
return_columns = [
'alignment',
'break_penalties',
'breakability',
'natural_width',
'acceptable_width',
'minimum_width',
'header_natural',
'header_minimum',
'raw_recommended',
'header_tweak',
'pre_space_share_recommended',
'proto_recommended',
'recommended',
'tikz_colw',
]
ans = ans[[i for i in return_columns if i in ans.columns]]
# need recommended to be > 0
ans['recommended'] = np.maximum(ans['recommended'], 1)
return ans
def make_style(self, tabs):
"""Write out custom CSS for the table."""
if self.config.debug:
head_tb = '#0ff'
body_b = '#f0f'
h0 = '#f00'
h1 = '#b00'
h2 = '#900'
bh0 = '#f00'
bh1 = '#b00'
v0 = '#0f0'
v1 = '#0a0'
v2 = '#090'
else:
head_tb = '#000'
body_b = '#000'
h0 = '#000'
h1 = '#000'
h2 = '#000'
bh0 = '#000'
bh1 = '#000'
v0 = '#000'
v1 = '#000'
v2 = '#000'
table_hrule = self.config.table_hrule_width
table_vrule = self.config.table_vrule_width
# for local use
padt, padr, padb, padl = self.padt, self.padr, self.padb, self.padl
style = [f'''
')
logger.info('CREATED CSS')
return '\n'.join(style)
def make_html(self):
"""Convert a pandas DataFrame to an HTML table."""
index_name_to_level = dict(
zip(self.raw_df.index.names, range(self.nindex)))
index_change_level = self.index_change_level.map(index_name_to_level)
# this is easier and computed in the init
column_change_level = self.column_change_level
# Start table
html = [f'']
if self.label != "":
pass
# TODO put in achor tag somehow!!
if self.caption != '':
html.append(f'{self.caption}')
# Process header: allow_duplicates=True means can create cols with the same name
bit = self.df_html.T.reset_index(drop=False, allow_duplicates=True)
idx_header = bit.iloc[:self.nindex, :self.ncolumns]
columns = bit.iloc[self.nindex:, :self.ncolumns]
# figure appropriate widths
tabs = self.html_knowledge_df['recommended'].map(lambda x: np.round(x, 3))
# set column widths; tabs returns lengths of strings in each column
tabs = np.array(tabs) + (self.padl + self.padr) / 12
# this gets stripped out by quarto, so make part of style
html.append('')
for w in tabs:
html.append(f'')
html.append('')
# TODO Add header aligners
# this is TRANSPOSED!!
if self.config.sparsify_columns:
html.append("")
for i in range(self.ncolumns):
# one per row of columns m index, usually only 1
html.append("")
if self.show_index:
for j, r in enumerate(idx_header.iloc[:, i]):
# columns one per level of index
html.append(f'| {r} | ')
# if not for col span issue you could just to this:
# for j in range(self.ncols):
# hrule = f'grt-bhrule-{i}' if i < self.ncolumns - 1 else ''
# if j == 0:
# # start with the first column come what may
# vrule = f'grt-vrule-index'
# elif j >= self.column_change_level[i]:
# vrule = f'grt-vrule-{column_change_level[cum_col]}'
# else:
# vrule = ''
# html.append(f'{nm} | ')
# here, the groupby needs to consider all levels at and above i
# this concats all the levels
# need :i+1 to get down to the ith level
cum_col = 0 # keep track of where we are up to
for j, (nm, g) in enumerate(groupby(columns.iloc[:, :i + 1].
apply(lambda x: ':::'.join(str(i) for i in x), axis=1))):
# ::: needs to be something that does not appear in the col names
# need to combine for groupby but be able to split off the last level
# picks off the name of the bottom level
nm = nm.split(':::')[-1]
hrule = f'grt-bhrule-{i}' if i < self.ncolumns - 1 else ''
colspan = sum(1 for _ in g)
if 0 < j:
vrule = f'grt-vrule-{column_change_level[cum_col]}'
elif j == 0 and self.show_index:
# start with the first column if showing index
vrule = f'grt-vrule-index'
else:
vrule = ''
if j == 0 and not self.show_index:
# first column, no index, left align label
html.append(
f'{nm} | ')
else:
html.append(
f'{nm} | ')
cum_col += colspan
html.append("
")
html.append("")
else:
html.append("")
for i in range(self.ncolumns):
# one per row of columns m index, usually only 1
html.append("")
if self.show_index:
for j, r in enumerate(idx_header.iloc[:, i]):
# columns one per level of index
html.append(f'| {r} | ')
for j, r in enumerate(columns.iloc[:, i]):
# one per column of dataframe
# figure how high up mindex the vrules go
# all headings get hrules, it's the vrules that are tricky
hrule = f'grt-bhrule-{i}' if i < self.ncolumns - 1 else ''
if 0 < j < self.ncols and i >= column_change_level[j]:
vrule = f'grt-vrule-{column_change_level[j]}'
elif j == 0 and self.show_index:
# start with the first column come what may
vrule = f'grt-vrule-index'
else:
vrule = ''
html.append(
f'{r} | ')
html.append("
")
html.append("")
bold_idx = 'grt-bold' if self.config.font_bold_index else ''
html.append("")
for i, (n, r) in enumerate(self.df_html.iterrows()):
# one per row of dataframe
html.append("")
hrule = ''
if self.show_index:
for j, c in enumerate(r.iloc[:self.nindex]):
# dx = data in index
# if this is the level that changes for this row
# will use a top rule hence omit i = 0 which already has an hrule
# appears in the index change level. But if it DOES NOT appear then
# it isn't a change level so no rule required
if i > 0 and hrule == '' and i in index_change_level and j == index_change_level[i]:
hrule = f'grt-hrule-{j}'
# html.append(f'| {c} | ')
col_id = f'grt-c-{j}'
html.append(
f'{c} | ')
for j, c in enumerate(r.iloc[self.nindex:]):
# first col left handled by index/body divider
if 0 < j < self.ncols:
vrule = f'grt-vrule-{column_change_level[j]}'
elif j == 0 and self.show_index:
# start with the first column come what may
vrule = f'grt-vrule-index'
else:
vrule = ''
# html.append(f'{c} | ')
col_id = f'grt-c-{j+self.nindex}'
html.append(
f'{c} | ')
html.append("
")
html.append("")
text = '\n'.join(html)
self._df_html_text = Escaping.clean_html_tex(text)
logger.info('CREATED HTML')
self._df_style_text = self.make_style(tabs)
def clean_style(self, soup):
"""Minify CSS inside
Rendered Table
{table_html}