def _generate_epoch(self, text_only) -> Generator[InputSample, None, None]:
filenames = list(self.filenames)
if self.mode == PipelineMode.Training:
shuffle(filenames)
for filename in filenames:
with h5py.File(filename, 'r') as f:
codec = list(map(chr, f['codec']))
if text_only:
for i, (text, idx) in enumerate(
zip(f['transcripts'],
range(len(f['transcripts'])))):
text = "".join([codec[c] for c in text])
fold_id = idx % self.n_folds if self.n_folds > 0 else -1
yield InputSample(
None, text,
SampleMeta(id=f"{filename}/{i}", fold_id=fold_id))
else:
gen = zip(f['images'], f['images_dims'], f['transcripts'],
range(len(f['images'])))
if self.mode == PipelineMode.Training:
gen = list(gen)
shuffle(gen)
for i, (image, shape, text, idx) in enumerate(gen):
image = np.reshape(image, shape)
text = "".join([codec[c] for c in text])
fold_id = idx % self.n_folds if self.n_folds > 0 else -1
yield InputSample(
image, text,
SampleMeta(id=f"{filename}/{i}", fold_id=fold_id))
def _load_sample(self, sample, text_only):
if text_only:
yield InputSample(
None,
self._load_gt_txt(sample["text_path"]),
SampleMeta(sample['id'], fold_id=sample['fold_id']),
)
else:
yield InputSample(
self._load_line(sample["image_path"]),
self._load_gt_txt(sample["text_path"]),
SampleMeta(sample['id'], fold_id=sample['fold_id']),
)
def _generate_epoch(self, text_only) -> Generator[InputSample, None, None]:
fold_id = -1
for p, page in enumerate(self.book.pages):
if self.mode in INPUT_PROCESSOR:
img = load_image(page.imgFile)
if self.binary:
img = img > 0.9
else:
img = None
for l, line in enumerate(page.getLines()):
for f, fo in enumerate(line.formats):
fold_id += 1
sample_id = "{}_{}_{}_{}".format(
os.path.splitext(
page.xmlFile if page.xmlFile else page.imgFile)[0],
p, l, f)
text = None
if self.mode in TARGETS_PROCESSOR:
text = fo.text
if text_only:
yield InputSample(
None, text,
SampleMeta(id=sample_id, fold_id=fold_id))
else:
cut_img = None
if self.mode in INPUT_PROCESSOR:
ly, lx = img.shape
# Cut the Image
cut_img = img[line.rect.top:-ly + line.rect.bottom,
line.rect.left:-lx + line.rect.right]
# add padding as required from normal files
cut_img = np.pad(cut_img, ((3, 3), (0, 0)),
mode='constant',
constant_values=cut_img.max())
yield InputSample(
cut_img, text,
SampleMeta(id=sample_id, fold_id=fold_id))
def _generate_epoch(self, text_only) -> Generator[InputSample, None, None]:
for filename in self.filenames:
f = h5py.File(filename, 'r')
codec = list(map(chr, f['codec']))
if text_only:
for i, text in enumerate(f['transcripts']):
text = "".join([codec[c] for c in text])
yield InputSample(None, text,
SampleMeta(id=f"{filename}/{i}"))
else:
gen = zip(f['images'], f['images_dims'], f['transcripts'])
if self.mode == PipelineMode.Training:
gen = list(gen)
shuffle(gen)
for i, (image, shape, text) in enumerate(gen):
image = np.reshape(image, shape)
text = "".join([codec[c] for c in text])
yield InputSample(image, text,
SampleMeta(id=f"{filename}/{i}"))
def __init__(self, mode: PipelineMode, images=None, texts=None, meta=None):
""" Create a dataset from memory
Since this dataset already contains all data in the memory, this dataset may not be loaded
Parameters
----------
images : list of images
the images of the dataset
texts : list of str
the texts of this dataset
"""
super().__init__(mode, skip_invalid=False, remove_invalid=False)
if images is None and texts is None:
raise Exception(
"Empty data set is not allowed. Both images and text files are None"
)
if images is not None and texts is not None and len(
images) == 0 and len(texts) == 0:
raise Exception("Empty data set provided.")
if texts is None or len(texts) == 0:
if images is None:
raise Exception("Empty data set.")
# No gt provided, probably prediction
texts = [None] * len(images)
if images is None or len(images) == 0:
if len(texts) is None:
raise Exception("Empty data set.")
# No images provided, probably evaluation
images = [None] * len(texts)
if not meta:
meta = [SampleMeta(str(i), None) for i in range(len(images))]
for image, text, meta in zip(images, texts, meta):
self.add_sample({
"image": image,
"text": text,
"id": meta.id,
"meta": meta,
})
self.loaded = True
def _load_sample(self, sample,
text_only) -> Generator[InputSample, None, None]:
loader = PageXMLDatasetLoader(self.mode, self._non_existing_as_empty,
self.text_index, self.skip_invalid)
image_path, xml_path = sample
img = None
if self.mode in INPUT_PROCESSOR:
img = load_image(image_path)
for sample in loader.load(image_path, xml_path):
text = sample["text"]
orientation = sample["orientation"]
if not text_only and self.mode in INPUT_PROCESSOR:
ly, lx = img.shape[:2]
line_img = PageXMLReader.cutout(img, sample['coords'],
lx / sample['img_width'])
# rotate by orientation angle in clockwise direction to correct present skew
# (skimage rotates in counter-clockwise direction)
if orientation and orientation % 360 != 0:
line_img = rotate(line_img,
orientation * -1,
resize=True,
mode='constant',
cval=line_img.max(),
preserve_range=True).astype(np.uint8)
# add padding as required from normal files
if self.args.pad:
pad = self.args.pad
img = np.pad(img,
pad,
mode='constant',
constant_values=img.max())
else:
line_img = None
yield InputSample(line_img, text, SampleMeta(id=sample['id']))
def _load_sample(self, sample,
text_only) -> Generator[InputSample, None, None]:
loader = PageXMLDatasetLoader(self.mode, self._non_existing_as_empty,
self.text_index, self.skip_invalid)
image_path, xml_path, idx = sample
img = None
if self.mode in INPUT_PROCESSOR:
img = load_image(image_path)
for i, sample in enumerate(loader.load(image_path, xml_path)):
fold_id = (idx + i) % self.n_folds if self.n_folds > 0 else -1
text = sample["text"]
orientation = sample["orientation"]
if not text_only and self.mode in INPUT_PROCESSOR:
ly, lx = img.shape[:2]
# rotate by orientation angle in clockwise direction to correct present skew
angle = orientation if orientation and orientation % 360 != 0 else 0
line_img = PageXMLReader.cutout(img,
sample['coords'],
mode=CutMode.POLYGON,
angle=angle,
cval=None,
scale=lx / sample['img_width'])
# add padding as required from normal files
if self.args.pad:
pad = self.args.pad
img = np.pad(img,
pad,
mode='constant',
constant_values=img.max())
else:
line_img = None
yield InputSample(line_img, text,
SampleMeta(id=sample['id'], fold_id=fold_id))
def _load_sample(self, sample, text_only):
image, text = self.data_queue.get()
fold_id = -1 if self.n_folds <= 0 else np.random.randint(self.n_folds)
yield InputSample(image, text,
SampleMeta(id=sample['id'], fold_id=fold_id))
def _load_sample(self, sample, text_only):
image, text = self.data_queue.get()
yield InputSample(image, text, SampleMeta(id=sample['id']))