mirror of
https://git.lyx.org/repos/lyx.git
synced 2024-11-14 15:05:56 +00:00
31d64c7395
In cooperation with Thibaut Cuvelier: lib/scripts/spreadsheet_to_docbook.py: Strip the document header and convert some flags lib/xtemplates/gnumeric.xtemplate: use this output to be inserted in docbook5 lib/configure.py: Add needed conversion entries
71 lines
2.5 KiB
Python
71 lines
2.5 KiB
Python
#!/usr/bin/python3
|
|
|
|
# file spreadsheet_to_docbook.py
|
|
# This file is part of LyX, the document processor.
|
|
# Licence details can be found in the file COPYING.
|
|
|
|
# author Thibaut Cuvelier & Kornel Benko
|
|
|
|
# Full author contact details are available in file CREDITS.
|
|
|
|
"""reformat output of ssconvert of a single spreadsheet to match the needs
|
|
of docbook5 table format .
|
|
|
|
Expects to read from file specified by sys.argv[1]
|
|
and output to to file specified by sys.argv[2]
|
|
"""
|
|
|
|
import re
|
|
import sys
|
|
import subprocess
|
|
|
|
|
|
def process_file(contents):
|
|
# Scrap the header and the footer.
|
|
contents = contents.split("<body>")[1]
|
|
contents = contents.split("</body>")[0]
|
|
|
|
# Gnumeric may generate more than one table, just take the first one.
|
|
contents = contents.split("</table>")[0] + "\n</table>"
|
|
|
|
# Convert the rest of the table to DocBook.
|
|
contents = contents.replace("<p></p>", "")
|
|
contents = contents.replace("<i>", "<emphasis>")
|
|
contents = contents.replace("</i>", "</emphasis>")
|
|
contents = contents.replace("<b>", "<emphasis role='bold'>")
|
|
contents = contents.replace("</b>", "</emphasis>")
|
|
contents = contents.replace("<u>", "<emphasis role='underline'>")
|
|
contents = contents.replace("</u>", "</emphasis>")
|
|
|
|
contents = re.sub(r"<font color=\"(.*)\">", "<phrase role='color \\1'>", contents)
|
|
assert '<font' not in contents # If this happens, implement something to catch these cases.
|
|
contents = contents.replace("</font>", "</phrase>") # Generates invalid XML if there are still font tags left...
|
|
|
|
# If the table has a caption, then the right tag is <table>. Otherwise, it's <informaltable>.
|
|
if '<caption>' not in contents:
|
|
contents = contents.replace("<table", "<informaltable")
|
|
contents = contents.replace("</table>", "</informaltable>")
|
|
|
|
# Return the processed string.
|
|
contents = contents.replace("\n\n", "\n")
|
|
return contents
|
|
|
|
|
|
if __name__ == "__main__":
|
|
if len(sys.argv) == 1:
|
|
# Read from stdin, output to stdout.
|
|
contents = sys.stdin.read()
|
|
f = sys.stdout
|
|
else:
|
|
# Read from output of ssconvert
|
|
assert len(sys.argv) == 3 # Script name, file to process, output file.
|
|
proc = subprocess.Popen(["ssconvert", "--export-type=Gnumeric_html:xhtml", sys.argv[1], "fd://1"], stdout=subprocess.PIPE)
|
|
f = open(sys.argv[2], 'w')
|
|
sys.stdout = f # Redirect stdout to the output file.
|
|
contents = proc.stdout.read()
|
|
|
|
# Process and output to stdout.
|
|
print(process_file(contents))
|
|
f.close()
|
|
exit(0)
|