def _trim(text, trim_begin=True, trim_end=True):
- """
+ """
Cut off everything before RE_TRIM_BEGIN and after RE_TRIM_END, so
that eg. one big XML file can be compiled from many small XML files.
"""
def compile_text(parts):
- """
+ """
Compiles full text from an iterable of parts,
trimming where applicable.
"""
name_elem = deepcopy(element)
for tag in 'extra', 'motyw', 'pa', 'pe', 'pr', 'pt', 'uwaga':
for a in name_elem.findall('.//' + tag):
- a.text=''
+ a.text = ''
del a[:]
name = etree.tostring(name_elem, method='text', encoding='utf-8').strip()
while parent[0] is not element:
del parent[0]
element, parent = parent, parent.getparent()
- chunks[:0] = [[name,
- unicode(etree.tostring(copied, encoding='utf-8'), 'utf-8')
- ]]
+ chunks[:0] = [[name, unicode(etree.tostring(copied, encoding='utf-8'), 'utf-8')]]
parts = src.findall('.//naglowek_rozdzial')
- chunks[:0] = [[u'początek',
- unicode(etree.tostring(src, encoding='utf-8'), 'utf-8')
- ]]
+ chunks[:0] = [[u'początek', unicode(etree.tostring(src, encoding='utf-8'), 'utf-8')]]
for ch in chunks[1:]:
ch[1] = add_trim_begin(ch[1])