def line_generator(ctx, font, maxlines, encoding, normalization, renormalize,
reorder, font_size, font_weight, language, max_length,
strip, disable_degradation, alpha, beta, distort,
distortion_sigma, legacy, output, text):
"""
Generates artificial text line training data.
"""
import errno
import numpy as np
from kraken import linegen
from kraken.lib.util import make_printable
lines: Set[str] = set()
if not text:
return
with log.progressbar(text, label='Reading texts') as bar:
for t in text:
with click.open_file(t, encoding=encoding) as fp:
logger.info('Reading {}'.format(t))
for l in fp:
lines.add(l.rstrip('\r\n'))
if normalization:
lines = set(
[unicodedata.normalize(normalization, line) for line in lines])
if strip:
lines = set([line.strip() for line in lines])
if max_length:
lines = set([line for line in lines if len(line) < max_length])
logger.info('Read {} lines'.format(len(lines)))
message('Read {} unique lines'.format(len(lines)))
if maxlines and maxlines < len(lines):
message('Sampling {} lines\t'.format(maxlines), nl=False)
llist = list(lines)
lines = set(llist[idx]
for idx in np.random.randint(0, len(llist), maxlines))
message('\u2713', fg='green')
try:
os.makedirs(output)
except OSError as e:
if e.errno != errno.EEXIST:
raise
# calculate the alphabet and print it for verification purposes
alphabet: Set[str] = set()
for line in lines:
alphabet.update(line)
chars = []
combining = []
for char in sorted(alphabet):
k = make_printable(char)
if k != char:
combining.append(k)
else:
chars.append(k)
message('Σ (len: {})'.format(len(alphabet)))
message('Symbols: {}'.format(''.join(chars)))
if combining:
message('Combining Characters: {}'.format(', '.join(combining)))
lg = linegen.LineGenerator(font, font_size, font_weight, language)
with log.progressbar(lines, label='Writing images') as bar:
for idx, line in enumerate(bar):
logger.info(line)
try:
if renormalize:
im = lg.render_line(
unicodedata.normalize(renormalize, line))
else:
im = lg.render_line(line)
except KrakenCairoSurfaceException as e:
logger.info('{}: {} {}'.format(e.message, e.width, e.height))
continue
if not disable_degradation and not legacy:
im = linegen.degrade_line(im, alpha=alpha, beta=beta)
im = linegen.distort_line(
im, abs(np.random.normal(distort)),
abs(np.random.normal(distortion_sigma)))
elif legacy:
im = linegen.ocropy_degrade(im)
im.save('{}/{:06d}.png'.format(output, idx))
with open('{}/{:06d}.gt.txt'.format(output, idx), 'wb') as fp:
if reorder:
fp.write(get_display(line).encode('utf-8'))
else:
fp.write(line.encode('utf-8'))
def line_generator(ctx, font, maxlines, encoding, normalization, renormalize,
reorder, font_size, font_weight, language, max_length,
strip, disable_degradation, binarize, mean, sigma, density,
distort, distortion_sigma, legacy, output, text):
"""
Generates artificial text line training data.
"""
lines = set()
if not text:
return
st_time = time.time()
for t in text:
with click.open_file(t, encoding=encoding) as fp:
if ctx.meta['verbose'] > 0:
click.echo(u'[{:2.4f}] Reading {}'.format(
time.time() - st_time, t))
else:
spin('Reading texts')
lines.update(fp.readlines())
if normalization:
lines = set(
[unicodedata.normalize(normalization, line) for line in lines])
if strip:
lines = set([line.strip() for line in lines])
if max_length:
lines = set([line for line in lines if len(line) < max_length])
if ctx.meta['verbose'] > 0:
click.echo(u'[{:2.4f}] Read {} lines'.format(time.time() - st_time,
len(lines)))
else:
click.secho(u'\b\u2713', fg='green', nl=False)
click.echo('\033[?25h\n', nl=False)
click.echo('Read {} unique lines'.format(len(lines)))
if maxlines and maxlines < len(lines):
click.echo('Sampling {} lines\t'.format(maxlines), nl=False)
lines = list(lines)
lines = [
lines[idx] for idx in np.random.randint(0, len(lines), maxlines)
]
click.secho(u'\u2713', fg='green')
try:
os.makedirs(output)
except OSError as e:
if e.errno != errno.EEXIST:
raise
lines = [line.strip() for line in lines]
# calculate the alphabet and print it for verification purposes
alphabet = set()
for line in lines:
alphabet.update(line)
chars = []
combining = []
for char in sorted(alphabet):
if unicodedata.combining(char):
combining.append(unicodedata.name(char))
else:
chars.append(char)
click.echo(u'Σ (len: {})'.format(len(alphabet)))
click.echo(u'Symbols: {}'.format(''.join(chars)))
if combining:
click.echo(u'Combining Characters: {}'.format(', '.join(combining)))
lg = linegen.LineGenerator(font, font_size, font_weight, language)
for idx, line in enumerate(lines):
if ctx.meta['verbose'] > 0:
click.echo(u'[{:2.4f}] {}'.format(time.time() - st_time, line))
else:
spin('Writing images')
try:
if renormalize:
im = lg.render_line(unicodedata.normalize(renormalize, line))
else:
im = lg.render_line(line)
except KrakenCairoSurfaceException as e:
if ctx.meta['verbose'] > 0:
click.echo('[{:2.4f}] {}: {} {}'.format(
time.time() - st_time, e.message, e.width, e.height))
else:
click.secho(u'\b\u2717', fg='red')
click.echo('{}: {} {}'.format(e.message, e.width, e.height))
continue
if not disable_degradation and not legacy:
im = linegen.distort_line(im, abs(np.random.normal(distort)),
abs(np.random.normal(distortion_sigma)))
im = linegen.degrade_line(im, abs(np.random.normal(mean)),
abs(np.random.normal(sigma)),
abs(np.random.normal(density)))
elif legacy:
im = linegen.ocropy_degrade(im)
if binarize:
try:
im = binarization.nlbin(im)
except KrakenInputException as e:
click.echo('{}'.format(e.message))
continue
im.save('{}/{:06d}.png'.format(output, idx))
with open('{}/{:06d}.gt.txt'.format(output, idx), 'wb') as fp:
if reorder:
fp.write(get_display(line).encode('utf-8'))
else:
fp.write(line.encode('utf-8'))
if ctx.meta['verbose'] == 0:
click.secho(u'\b\u2713', fg='green', nl=False)
click.echo('\033[?25h\n', nl=False)