def test_raw_prediction(self):
args = PredictionAttrs()
predictor = Predictor.from_checkpoint(PredictorParams(
progress_bar=False, silent=True),
checkpoint=args.checkpoint[0])
images = [load_image(file) for file in args.files]
for result in predictor.predict_raw(images):
self.assertGreater(result.outputs.avg_char_probability, 0)
def test_raw_prediction_voted(self):
args = PredictionAttrs()
predictor = MultiPredictor.from_paths(checkpoints=args.checkpoint, predictor_params=PredictorParams(progress_bar=False, silent=True))
images = [load_image(file) for file in args.files]
for sample in predictor.predict_raw(images):
r, voted = sample.outputs
print([rn.sentence for rn in r])
print(voted.sentence)
def test_raw_prediction(self):
args = PredictionAttrs()
predictor = Predictor.from_checkpoint(PredictorParams(
progress_bar=False, silent=True),
checkpoint=args.checkpoint[0])
images = [load_image(file) for file in args.files]
for file, image in zip(args.files, images):
_, prediction, _ = list(predictor.predict_raw([image]))[0]
print(file, prediction.sentence)
def _load_line(self, image_path):
if image_path is None:
return None
if not os.path.exists(image_path):
if self._non_existing_as_empty:
return np.zeros((1, 1), dtype=np.uint8)
else:
raise Exception(
"Image file at '{}' does not exist".format(image_path))
try:
img = load_image(image_path)
except:
return None
return img
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 _load_sample(self, sample, text_only):
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 == DataSetMode.PREDICT or self.mode == DataSetMode.TRAIN or self.mode == DataSetMode.PRED_AND_EVAL:
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 == DataSetMode.PREDICT
or self.mode == DataSetMode.TRAIN
or self.mode == DataSetMode.PRED_AND_EVAL):
ly, lx = img.shape[:2]
line_img = PageXMLDataset.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.get('pad', None):
pad = self.args['pad']
img = np.pad(img,
pad,
mode='constant',
constant_values=img.max())
else:
line_img = None
yield line_img, text
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))
class DefaultDataAugmenter(DataAugmenterBase[DefaultDataAugmenterParams]):
def augment_single(self, data, gt_txt):
import calamari_ocr.thirdparty.ocrodeg as ocrodeg
original_dtype = data.dtype
data = data.astype(np.float)
m = data.max()
data = data / (1 if m == 0 else m)
data = ocrodeg.random_pad(data, (0, data.shape[1] * 2))
# data = ocrodeg.transform_image(data, **ocrodeg.random_transform(rotation=(-0.1, 0.1), translation=(-0.1, 0.1)))
for sigma in [2, 5]:
noise = ocrodeg.bounded_gaussian_noise(data.shape, sigma, 3.0)
data = ocrodeg.distort_with_noise(data, noise)
data = ocrodeg.printlike_multiscale(data, blur=1, inverted=True)
data = (data * 255 / data.max()).astype(original_dtype)
return data, gt_txt
if __name__ == '__main__':
import matplotlib.pyplot as plt
aug = DefaultDataAugmenterParams().create()
img = 255 - np.mean(load_image("../../test/data/uw3_50lines/train/010001.bin.png")[:, :, 0:2], axis=-1)
aug_img = [aug.augment_single(img.T, '')[0].T for _ in range(4)]
f, ax = plt.subplots(5, 1)
ax[0].imshow(255 - img, cmap='gray')
for i, x in enumerate(aug_img):
ax[i + 1].imshow(255 - x, cmap='gray')
plt.show()
data = data.astype(np.float)
m = data.max()
data = data / (1 if m == 0 else m)
data = ocrodeg.random_pad(data, (0, data.shape[1] * 2))
# data = ocrodeg.transform_image(data, **ocrodeg.random_transform(rotation=(-0.1, 0.1), translation=(-0.1, 0.1)))
for sigma in [2, 5]:
noise = ocrodeg.bounded_gaussian_noise(data.shape, sigma, 3.0)
data = ocrodeg.distort_with_noise(data, noise)
data = ocrodeg.printlike_multiscale(data, blur=1, inverted=True)
data = (data * 255 / data.max()).astype(original_dtype)
return data, gt_txt
NoopDataAugmenter.register()
SimpleDataAugmenter.register()
if __name__ == '__main__':
aug = SimpleDataAugmenter()
img = 255 - np.mean(
load_image("../../test/data/uw3_50lines/train/010001.bin.png")[:, :,
0:2],
axis=-1)
aug_img = [aug.augment_single(img.T, '')[0].T for _ in range(4)]
import matplotlib.pyplot as plt
f, ax = plt.subplots(5, 1)
ax[0].imshow(255 - img, cmap='gray')
for i, x in enumerate(aug_img):
ax[i + 1].imshow(255 - x, cmap='gray')
plt.show()