main.py 35.41 KiB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import argparse
import getpass
import os
import random
from collections import Counter, defaultdict
from enum import Enum
from pathlib import Path
from typing import List
import cv2
import numpy as np
import tqdm
from apistar.exceptions import ErrorResponse
from arkindex import options_from_env
from kaldi_data_generator.image_utils import (
determine_rotate_angle,
download_image,
extract_min_area_rect_image,
extract_polygon_image,
resize_transcription_data,
rotate,
trim,
)
from kaldi_data_generator.utils import (
CachedApiClient,
TranscriptionData,
logger,
write_file,
)
SEED = 42
random.seed(SEED)
MANUAL = "manual"
TEXT_LINE = "text_line"
WHITE = 255
DEFAULT_RESCALE = 1.0
ROTATION_CLASSES_TO_ANGLES = {
"rotate_0": 0,
"rotate_left_90": 90,
"rotate_180": 180,
"rotate_right_90": -90,
}
def create_api_client(cache_dir=None):
logger.info("Creating API client")
# return ArkindexClient(**options_from_env())
return CachedApiClient(cache_root=cache_dir, **options_from_env())
class Extraction(Enum):
boundingRect: int = 0
polygon: int = 1
# minimum containing rectangle with an angle (cv2.min_area_rect)
min_area_rect: int = 2
deskew_polygon: int = 3
deskew_min_area_rect: int = 4
skew_polygon: int = 5
skew_min_area_rect: int = 6
class Style(Enum):
handwritten: str = "handwritten"
typewritten: str = "typewritten"
other: str = "other"
STYLE_CLASSES = [s.name for s in [Style.handwritten, Style.typewritten]]
class HTRDataGenerator:
def __init__(
self,
format,
dataset_name="foo",
out_dir_base="/tmp/kaldi_data",
grayscale=True,
extraction=Extraction.boundingRect,
accepted_classes=None,
ignored_classes=None,
style=None,
skip_vertical_lines=False,
accepted_worker_version_ids=None,
transcription_type=TEXT_LINE,
max_deskew_angle=45,
skew_angle=0,
should_rotate=False,
scale_x=None,
scale_y_top=None,
scale_y_bottom=None,
cache_dir=None,
api_client=None,
):
self.format = format
self.out_dir_base = out_dir_base
self.dataset_name = dataset_name
self.grayscale = grayscale
self.extraction_mode = extraction
self.accepted_classes = accepted_classes
self.ignored_classes = ignored_classes
self.should_filter_by_class = bool(self.accepted_classes) or bool(
self.ignored_classes
)
self.accepted_worker_version_ids = accepted_worker_version_ids
self.should_filter_by_worker = bool(self.accepted_worker_version_ids)
self.style = style
self.should_filter_by_style = bool(self.style)
self.transcription_type = transcription_type
self.skip_vertical_lines = skip_vertical_lines
self.skipped_pages_count = 0
self.skipped_vertical_lines_count = 0
self.accepted_lines_count = 0
self.max_deskew_angle = max_deskew_angle
self.skew_angle = skew_angle
self.should_rotate = should_rotate
if scale_x or scale_y_top or scale_y_bottom:
self.should_resize_polygons = True
# use 1.0 as default - no resize, if not specified
self.scale_x = scale_x or DEFAULT_RESCALE
self.scale_y_top = scale_y_top or DEFAULT_RESCALE
self.scale_y_bottom = scale_y_bottom or DEFAULT_RESCALE
else:
self.should_resize_polygons = False
self.api_client = api_client
if MANUAL in self.accepted_worker_version_ids:
self.accepted_worker_version_ids[
self.accepted_worker_version_ids.index(MANUAL)
] = None
if self.format == "kraken":
self.out_line_dir = out_dir_base
os.makedirs(self.out_line_dir, exist_ok=True)
else:
self.out_line_text_dir = os.path.join(
self.out_dir_base, "Transcriptions", self.dataset_name
)
os.makedirs(self.out_line_text_dir, exist_ok=True)
self.out_line_img_dir = os.path.join(
self.out_dir_base, "Lines", self.dataset_name
)
os.makedirs(self.out_line_img_dir, exist_ok=True)
self.cache_dir = cache_dir
logger.info(f"Setting up cache to {self.cache_dir}")
self.img_cache_dir = self.cache_dir / "images"
self.img_cache_dir.mkdir(exist_ok=True, parents=True)
if not any(self.img_cache_dir.iterdir()):
logger.info("Cache is empty, no need to check")
self._cache_is_empty = True
else:
self._cache_is_empty = False
if self.grayscale:
self._color = "grayscale"
self._cv2_flag = cv2.IMREAD_GRAYSCALE
else:
self._color = "rgb"
self._cv2_flag = cv2.IMREAD_COLOR
def get_image(self, image_url: str, page_id: str) -> "np.ndarray":
# id is last part before full/full/0/default.jpg
img_id = image_url.split("/")[-5].replace("%2F", "/")
cached_img_path = self.img_cache_dir / self._color / img_id
if not self._cache_is_empty and cached_img_path.exists():
logger.info(f"Cached image exists: {cached_img_path} - {page_id}")
else:
logger.info(f"Image not in cache: {cached_img_path} - {page_id}")
cached_img_path.parent.mkdir(exist_ok=True, parents=True)
pil_img = download_image(image_url)
if self.grayscale:
pil_img = pil_img.convert("L")
pil_img.save(cached_img_path, format="jpeg")
img = cv2.imread(str(cached_img_path), self._cv2_flag)
return img
def get_accepted_zones(self, page_id: str):
try:
accepted_zones = []
for elt in self.api_client.cached_paginate(
"ListElementChildren", id=page_id, with_classes=True
):
elem_classes = [c for c in elt["classes"] if c["state"] != "rejected"]
should_accept = True
if self.should_filter_by_class:
# at first filter to only have elements with accepted classes
# if accepted classes list is empty then should accept all
# except for ignored classes
should_accept = len(self.accepted_classes) == 0
for classification in elem_classes:
class_name = classification["ml_class"]["name"]
if class_name in self.accepted_classes:
should_accept = True
break
elif class_name in self.ignored_classes:
should_accept = False
break
if not should_accept:
continue
if self.should_filter_by_style:
style_counts = Counter()
for classification in elem_classes:
class_name = classification["ml_class"]["name"]
if class_name in STYLE_CLASSES:
style_counts[class_name] += 1
if len(style_counts) == 0:
# no handwritten or typewritten found, so other
found_class = Style.other
elif len(style_counts) == 1:
found_class = list(style_counts.keys())[0]
found_class = Style(found_class)
else:
raise ValueError(
f"Multiple style classes on the same element! {elt['id']} - {elem_classes}"
)
if found_class == self.style:
accepted_zones.append(elt["zone"]["id"])
else:
accepted_zones.append(elt["zone"]["id"])
logger.info(
"Number of accepted zone for page {} : {}".format(
page_id, len(accepted_zones)
)
)
return accepted_zones
except ErrorResponse as e:
logger.info(
f"ListTranscriptions failed {e.status_code} - {e.title} - {e.content} - {page_id}"
)
raise e
def _validate_transcriptions(self, page_id: str, lines: List[TranscriptionData]):
if not lines:
return
line_elem_counter = Counter([trans.element_id for trans in lines])
most_common = line_elem_counter.most_common(10)
if most_common[0][-1] > 1:
logger.error("Line elements have multiple transcriptions! Showing top 10:")
logger.error(f"{most_common}")
raise ValueError(f"Multiple transcriptions: {most_common[0]}")
worker_version_counter = Counter([trans.worker_version_id for trans in lines])
if len(worker_version_counter) > 1:
logger.warning(
f"There are transcriptions from multiple worker versions on this page: {page_id}:"
)
logger.warning(
f"Top 10 worker versions: {worker_version_counter.most_common(10)}"
)
def _choose_best_transcriptions(
self, lines: List[TranscriptionData]
) -> List[TranscriptionData]:
"""
Get the best transcription based on the order of accepted worker version ids.
:param lines:
:return:
"""
if not lines:
return []
trans_by_element = defaultdict(list)
for line in lines:
trans_by_element[line.element_id].append(line)
best_transcriptions = []
for elem, trans_list in trans_by_element.items():
tmp_dict = {t.worker_version_id: t for t in trans_list}
for wv in self.accepted_worker_version_ids:
if wv in tmp_dict:
best_transcriptions.append(tmp_dict[wv])
break
else:
logger.info(f"No suitable trans found for {elem}")
return best_transcriptions
def get_transcriptions(self, page_id: str, accepted_zones):
lines = []
try:
for res in self.api_client.cached_paginate(
"ListTranscriptions", id=page_id, recursive=True
):
if (
self.should_filter_by_worker
and res["worker_version_id"] not in self.accepted_worker_version_ids
):
continue
if (self.should_filter_by_class or self.should_filter_by_style) and (
res["element"]["zone"]["id"] not in accepted_zones
):
continue
if res["element"]["type"] != self.transcription_type:
continue
text = res["text"]
if not text or not text.strip():
continue
if "zone" in res:
polygon = res["zone"]["polygon"]
elif "element" in res:
polygon = res["element"]["zone"]["polygon"]
else:
raise ValueError(f"Data problem with polygon :: {res}")
trans_data = TranscriptionData(
element_id=res["element"]["id"],
polygon=polygon,
text=text,
trans_id=res["id"],
worker_version_id=res["worker_version_id"],
)
lines.append(trans_data)
if self.accepted_worker_version_ids:
# if accepted worker versions have been defined then use them
lines = self._choose_best_transcriptions(lines)
else:
# if no accepted worker versions have been defined
# then check that there aren't multiple transcriptions
# on the same text line
self._validate_transcriptions(page_id, lines)
if self.should_rotate:
classes_by_elem = self.get_children_classes(page_id)
for trans in lines:
rotation_classes = [
c
for c in classes_by_elem[trans.element_id]
if c in ROTATION_CLASSES_TO_ANGLES
]
if len(rotation_classes) > 0:
if len(rotation_classes) > 1:
logger.warning(
f"Several rotation classes = {len(rotation_classes)} - {trans.element_id}"
)
trans.rotation_class = rotation_classes[0]
else:
logger.warning(f"No rotation classes on {trans.element_id}")
count_skipped = 0
if self.skip_vertical_lines:
filtered_lines = []
for line in lines:
if line.is_vertical:
count_skipped += 1
continue
filtered_lines.append(line)
lines = filtered_lines
count = len(lines)
return lines, count, count_skipped
except ErrorResponse as e:
logger.info(
f"ListTranscriptions failed {e.status_code} - {e.title} - {e.content} - {page_id}"
)
raise e
def get_children_classes(self, page_id):
return {
elem["id"]: [
best_class["ml_class"]["name"]
for best_class in elem["classes"]
if best_class["state"] != "rejected"
]
for elem in self.api_client.cached_paginate(
"ListElementChildren",
id=page_id,
recursive=True,
type=TEXT_LINE,
with_classes=True,
)
}
def _save_line_image(
self, page_id, i, line_img, manifest_fp=None, trans: TranscriptionData = None
):
if self.should_rotate:
if trans.rotation_class:
rotate_angle = ROTATION_CLASSES_TO_ANGLES[trans.rotation_class]
line_img = self.rotate_and_trim(line_img, rotate_angle)
if self.format == "kraken":
cv2.imwrite(f"{self.out_line_dir}/{page_id}_{i}.png", line_img)
manifest_fp.write(f"{page_id}_{i}.png\n")
else:
cv2.imwrite(f"{self.out_line_img_dir}/{page_id}_{i}.jpg", line_img)
def extract_lines(self, page_id: str, image_data: dict):
if self.should_filter_by_class or self.should_filter_by_style:
accepted_zones = self.get_accepted_zones(page_id)
else:
accepted_zones = []
lines, count, count_skipped = self.get_transcriptions(page_id, accepted_zones)
if count == 0:
self.skipped_pages_count += 1
logger.info(f"Page {page_id} skipped, because it has no lines")
return
logger.debug(f"Total num of lines {count + count_skipped}")
logger.debug(f"Num of accepted lines {count}")
logger.debug(f"Num of skipped lines {count_skipped}")
self.skipped_vertical_lines_count += count_skipped
self.accepted_lines_count += count
full_image_url = image_data["s3_url"]
if full_image_url is None:
full_image_url = image_data["url"] + "/full/full/0/default.jpg"
img = self.get_image(full_image_url, page_id=page_id)
# sort vertically then horizontally
sorted_lines = sorted(lines, key=lambda key: (key.rect.y, key.rect.x))
if self.should_resize_polygons:
sorted_lines = [
resize_transcription_data(
line,
image_data["width"],
image_data["height"],
self.scale_x,
self.scale_y_top,
self.scale_y_bottom,
)
for line in sorted_lines
]
if self.format == "kraken":
manifest_fp = open(f"{self.out_line_dir}/manifest.txt", "a")
# append to file, not re-write it
else:
# not needed for kaldi
manifest_fp = None
if self.extraction_mode == Extraction.boundingRect:
for i, trans in enumerate(sorted_lines):
(x, y, w, h) = trans.rect
cropped = img[y : y + h, x : x + w].copy()
self._save_line_image(page_id, i, cropped, manifest_fp, trans)
elif self.extraction_mode == Extraction.polygon:
for i, trans in enumerate(sorted_lines):
polygon_img = extract_polygon_image(
img, polygon=trans.polygon, rect=trans.rect
)
self._save_line_image(page_id, i, polygon_img, manifest_fp, trans)
elif self.extraction_mode == Extraction.min_area_rect:
for i, trans in enumerate(sorted_lines):
min_rect_img = extract_min_area_rect_image(
img, polygon=trans.polygon, rect=trans.rect
)
self._save_line_image(page_id, i, min_rect_img, manifest_fp, trans)
elif self.extraction_mode == Extraction.deskew_polygon:
for i, trans in enumerate(sorted_lines):
# get angle from min area rect
rotate_angle = determine_rotate_angle(trans.polygon)
if abs(rotate_angle) > self.max_deskew_angle:
logger.warning(
f"Deskew angle ({rotate_angle}) over the limit ({self.max_deskew_angle}), won't rotate"
)
rotate_angle = 0
# get polygon image
polygon_img = extract_polygon_image(
img, polygon=trans.polygon, rect=trans.rect
)
trimmed_img = self.rotate_and_trim(polygon_img, rotate_angle)
self._save_line_image(page_id, i, trimmed_img, manifest_fp, trans)
elif self.extraction_mode == Extraction.deskew_min_area_rect:
for i, trans in enumerate(sorted_lines):
# get angle from min area rect
rotate_angle = determine_rotate_angle(trans.polygon)
if abs(rotate_angle) > self.max_deskew_angle:
logger.warning(
f"Deskew angle ({rotate_angle}) over the limit ({self.max_deskew_angle}), won't rotate"
)
rotate_angle = 0
min_rect_img = extract_min_area_rect_image(
img, polygon=trans.polygon, rect=trans.rect
)
trimmed_img = self.rotate_and_trim(min_rect_img, rotate_angle)
self._save_line_image(page_id, i, trimmed_img, manifest_fp, trans)
elif self.extraction_mode == Extraction.skew_polygon:
for i, trans in enumerate(sorted_lines):
rotate_angle = self.skew_angle
# get polygon image
polygon_img = extract_polygon_image(
img, polygon=trans.polygon, rect=trans.rect
)
trimmed_img = self.rotate_and_trim(polygon_img, rotate_angle)
self._save_line_image(page_id, i, trimmed_img, manifest_fp, trans)
elif self.extraction_mode == Extraction.skew_min_area_rect:
for i, trans in enumerate(sorted_lines):
rotate_angle = self.skew_angle
min_rect_img = extract_min_area_rect_image(
img, polygon=trans.polygon, rect=trans.rect
)
trimmed_img = self.rotate_and_trim(min_rect_img, rotate_angle)
self._save_line_image(page_id, i, trimmed_img, manifest_fp, trans)
else:
raise ValueError(f"Unsupported extraction mode: {self.extraction_mode}")
if self.format == "kraken":
manifest_fp.close()
for i, trans in enumerate(sorted_lines):
if self.format == "kraken":
write_file(f"{self.out_line_dir}/{page_id}_{i}.gt.txt", trans.text)
else:
write_file(f"{self.out_line_text_dir}/{page_id}_{i}.txt", trans.text)
def rotate_and_trim(self, img, rotate_angle):
"""
Rotate image by given an angle and trim extra whitespace left after rotating
"""
if self.grayscale:
background = WHITE
else:
background = (WHITE, WHITE, WHITE)
# rotate polygon image
deskewed_img = rotate(img, rotate_angle, background)
# trim extra whitespace left after rotating
trimmed_img = trim(deskewed_img, background)
trimmed_img = np.array(trimmed_img)
return trimmed_img
def run_pages(self, pages: list):
if all(isinstance(n, str) for n in pages):
for page in pages:
elt = self.api_client.request("RetrieveElement", id=page)
page_id = elt["id"]
image_data = elt["zone"]["image"]
logger.debug(f"Page {page_id}")
self.extract_lines(page_id, image_data)
else:
for page in tqdm.tqdm(pages):
page_id = page["id"]
image_data = page["zone"]["image"]
logger.debug(f"Page {page_id}")
self.extract_lines(page_id, image_data)
def run_volumes(self, volume_ids: list):
for volume_id in tqdm.tqdm(volume_ids):
logger.info(f"Volume {volume_id}")
pages = [
page
for page in self.api_client.cached_paginate(
"ListElementChildren", id=volume_id, recursive=True, type="page"
)
]
self.run_pages(pages)
def run_folders(self, element_ids: list, volume_type: str):
for elem_id in tqdm.tqdm(element_ids):
logger.info(f"Folder {elem_id}")
vol_ids = [
page["id"]
for page in self.api_client.cached_paginate(
"ListElementChildren", id=elem_id, recursive=True, type=volume_type
)
]
self.run_volumes(vol_ids)
def run_corpora(self, corpus_ids: list, volume_type: str):
for corpus_id in tqdm.tqdm(corpus_ids):
logger.info(f"Corpus {corpus_id}")
vol_ids = [
vol["id"]
for vol in self.api_client.cached_paginate(
"ListElements", corpus=corpus_id, type=volume_type
)
]
self.run_volumes(vol_ids)
class Split(Enum):
Train: int = 0
Test: int = 1
Validation: int = 2
@property
def short_name(self) -> str:
if self == self.Validation:
return "val"
return self.name.lower()
class KaldiPartitionSplitter:
def __init__(
self,
out_dir_base="/tmp/kaldi_data",
split_train_ratio=0.8,
split_test_ratio=0.1,
use_existing_split=False,
):
self.out_dir_base = out_dir_base
self.split_train_ratio = split_train_ratio
self.split_test_ratio = split_test_ratio
self.split_val_ratio = 1 - self.split_train_ratio - self.split_test_ratio
self.use_existing_split = use_existing_split
def page_level_split(self, line_ids: list) -> dict:
# need to sort again, because `set` will lose the order
page_ids = sorted({"_".join(line_id.split("_")[:-1]) for line_id in line_ids})
random.Random(SEED).shuffle(page_ids)
page_count = len(page_ids)
train_page_ids = page_ids[: round(page_count * self.split_train_ratio)]
page_ids = page_ids[round(page_count * self.split_train_ratio) :]
test_page_ids = page_ids[: round(page_count * self.split_test_ratio)]
page_ids = page_ids[round(page_count * self.split_test_ratio) :]
val_page_ids = page_ids
page_dict = {page_id: Split.Train.value for page_id in train_page_ids}
page_dict.update({page_id: Split.Test.value for page_id in test_page_ids})
page_dict.update({page_id: Split.Validation.value for page_id in val_page_ids})
return page_dict
def existing_split(self, line_ids: list) -> list:
split_dict = {split.short_name: [] for split in Split}
for line_id in line_ids:
split_prefix = line_id.split("/")[0].lower()
split_dict[split_prefix].append(line_id)
splits = [split_dict[split.short_name] for split in Split]
return splits
def create_partitions(self):
logger.info("Creating partitions")
lines_path = Path(f"{self.out_dir_base}/Lines")
line_ids = [
str(file.relative_to(lines_path).with_suffix(""))
for file in sorted(lines_path.glob("**/*.jpg"))
]
if self.use_existing_split:
logger.info("Using existing split")
datasets = self.existing_split(line_ids)
else:
page_dict = self.page_level_split(line_ids)
datasets = [[] for _ in range(3)]
for line_id in line_ids:
page_id = "_".join(line_id.split("_")[:-1])
split_id = page_dict[page_id]
datasets[split_id].append(line_id)
partitions_dir = os.path.join(self.out_dir_base, "Partitions")
os.makedirs(partitions_dir, exist_ok=True)
for i, dataset in enumerate(datasets):
if not dataset:
logger.info(f"Partition {Split(i).name} is empty! Skipping..")
continue
file_name = f"{partitions_dir}/{Split(i).name}Lines.lst"
write_file(file_name, "\n".join(dataset) + "\n")
def create_parser():
user_name = getpass.getuser()
parser = argparse.ArgumentParser(
description="Script to generate Kaldi or kraken training data from annotations from Arkindex",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
"-f",
"--format",
type=str,
help="is the data generated going to be used for kaldi or kraken",
)
parser.add_argument(
"-n",
"--dataset_name",
type=str,
help="Name of the dataset being created for kaldi or kraken "
"(useful for distinguishing different datasets when in Lines or Transcriptions directory)",
)
parser.add_argument(
"-o", "--out_dir", type=str, required=True, help="output directory"
)
parser.add_argument(
"--train_ratio",
type=float,
default=0.8,
help="Ratio of pages to be used in train (between 0 and 1)",
)
parser.add_argument(
"--test_ratio",
type=float,
default=0.1,
help="Ratio of pages to be used in test (between 0 and 1 - train_ratio)",
)
parser.add_argument(
"--use_existing_split",
action="store_true",
default=False,
help="Use an existing split instead of random. "
"Expecting line_ids to be prefixed with (train, val and test)",
)
parser.add_argument(
"--split_only",
"--no_download",
action="store_true",
default=False,
help="Create the split from already downloaded lines, don't download the lines",
)
parser.add_argument(
"--no_split",
action="store_true",
default=False,
help="No splitting of the data to be done just download the line in the right format",
)
parser.add_argument(
"-e",
"--extraction_mode",
type=lambda x: Extraction[x],
default=Extraction.boundingRect,
help=f"Mode for extracting the line images: {[e.name for e in Extraction]}",
)
parser.add_argument(
"--max_deskew_angle",
type=int,
default=45,
help="Maximum angle by which deskewing is allowed to rotate the line image. "
"If the angle determined by deskew tool is bigger than max "
"then that line won't be deskewed/rotated.",
)
parser.add_argument(
"--skew_angle",
type=int,
default=0,
help="Angle by which the line image will be rotated. Useful for data augmnetation"
" - creating skewed text lines for a more robust model."
" Only used with skew_* extraction modes.",
)
parser.add_argument(
"--should_rotate",
action="store_true",
default=False,
help="Use text line rotation class to rotate lines if possible",
)
parser.add_argument(
"--transcription_type",
type=str,
default="text_line",
help="Which type of elements' transcriptions to use? (page, paragraph, text_line, etc)",
)
group = parser.add_mutually_exclusive_group(required=False)
group.add_argument(
"--grayscale",
action="store_true",
dest="grayscale",
help="Convert images to grayscale (By default grayscale)",
)
group.add_argument(
"--color", action="store_false", dest="grayscale", help="Use color images"
)
group.set_defaults(grayscale=True)
parser.add_argument(
"--corpora",
nargs="*",
help="List of corpus ids to be used, separated by spaces",
)
parser.add_argument(
"--folders",
type=str,
nargs="*",
help="List of folder ids to be used, separated by spaces. "
"Elements of `volume_type` will be searched recursively in these folders",
)
parser.add_argument(
"--volumes",
nargs="*",
help="List of volume ids to be used, separated by spaces",
)
parser.add_argument(
"--pages", nargs="*", help="List of page ids to be used, separated by spaces"
)
parser.add_argument(
"-v",
"--volume_type",
type=str,
default="volume",
help="Volumes (1 level above page) may have a different name on corpora",
)
parser.add_argument(
"--skip_vertical_lines",
action="store_true",
default=False,
help="skips vertical lines when downloading",
)
parser.add_argument(
"--ignored_classes",
nargs="*",
default=[],
help="List of ignored ml_class names. Filter lines by class",
)
parser.add_argument(
"--accepted_classes",
nargs="*",
default=[],
help="List of accepted ml_class names. Filter lines by class",
)
parser.add_argument(
"--accepted_worker_version_ids",
nargs="*",
default=[],
help="List of accepted worker version ids. Filter transcriptions by worker version ids."
"The order is important - only up to one transcription will be chosen per element (text_line)"
" and the worker version order defines the precedence. If there exists a transcription for"
" the first worker version then it will be chosen, otherwise will continue on to the next"
" worker version."
" Use `--accepted_worker_version_ids manual` to get only manual transcriptions",
)
parser.add_argument(
"--style",
type=lambda x: Style[x.lower()],
default=None,
help=f"Filter line images by style class. 'other' corresponds to line elements that "
f"have neither handwritten or typewritten class : {[s.name for s in Style]}",
)
parser.add_argument(
"--scale_x",
type=float,
default=None,
help="Ratio of how much to scale the polygon horizontally (1.0 means no rescaling)",
)
parser.add_argument(
"--scale_y_top",
type=float,
default=None,
help="Ratio of how much to scale the polygon vertically on the top (1.0 means no rescaling)",
)
parser.add_argument(
"--scale_y_bottom",
type=float,
default=None,
help="Ratio of how much to scale the polygon vertically on the bottom (1.0 means no rescaling)",
)
parser.add_argument(
"--cache_dir",
type=Path,
default=Path(f"/tmp/kaldi_data_generator_{user_name}/cache/"),
help="Cache dir where to save the full size downloaded images. Change it to force redownload.",
)
return parser
def main():
parser = create_parser()
args = parser.parse_args()
if not args.dataset_name and not args.split_only and not args.format == "kraken":
parser.error("--dataset_name must be specified (unless --split-only)")
if args.accepted_classes and args.ignored_classes:
if set(args.accepted_classes) & set(args.ignored_classes):
parser.error(
f"--accepted_classes and --ignored_classes values must not overlap ({args.accepted_classes} - {args.ignored_classes})"
)
if args.style and (args.accepted_classes or args.ignored_classes):
if set(STYLE_CLASSES) & (
set(args.accepted_classes) | set(args.ignored_classes)
):
parser.error(
f"--style class values ({STYLE_CLASSES}) shouldn't be in the accepted_classes list "
f"(or ignored_classes list) "
"if both --style and --accepted_classes (or --ignored_classes) are used together."
)
logger.info(f"ARGS {args} \n")
api_client = create_api_client(args.cache_dir)
if not args.split_only:
data_generator = HTRDataGenerator(
format=args.format,
dataset_name=args.dataset_name,
out_dir_base=args.out_dir,
grayscale=args.grayscale,
extraction=args.extraction_mode,
accepted_classes=args.accepted_classes,
ignored_classes=args.ignored_classes,
style=args.style,
skip_vertical_lines=args.skip_vertical_lines,
transcription_type=args.transcription_type,
accepted_worker_version_ids=args.accepted_worker_version_ids,
max_deskew_angle=args.max_deskew_angle,
skew_angle=args.skew_angle,
should_rotate=args.should_rotate,
scale_x=args.scale_x,
scale_y_top=args.scale_y_top,
scale_y_bottom=args.scale_y_bottom,
cache_dir=args.cache_dir,
api_client=api_client,
)
# extract all the lines and transcriptions
if args.pages:
data_generator.run_pages(args.pages)
if args.volumes:
data_generator.run_volumes(args.volumes)
if args.folders:
data_generator.run_folders(args.folders, args.volume_type)
if args.corpora:
data_generator.run_corpora(args.corpora, args.volume_type)
if data_generator.skipped_vertical_lines_count > 0:
logger.info(
f"Number of skipped pages: {data_generator.skipped_pages_count}"
)
_skipped_vertical_count = data_generator.skipped_vertical_lines_count
_total_count = _skipped_vertical_count + data_generator.accepted_lines_count
skipped_ratio = _skipped_vertical_count / _total_count * 100
logger.info(
f"Skipped {data_generator.skipped_vertical_lines_count} vertical lines ({round(skipped_ratio, 2)}%)"
)
else:
logger.info("Creating a split from already downloaded files")
if not args.no_split:
kaldi_partitioner = KaldiPartitionSplitter(
out_dir_base=args.out_dir,
split_train_ratio=args.train_ratio,
split_test_ratio=args.test_ratio,
use_existing_split=args.use_existing_split,
)
# create partitions from all the extracted data
kaldi_partitioner.create_partitions()
else:
logger.info("No split to be done")
logger.info("DONE")
if __name__ == "__main__":
main()