def create_emnist_lines_support_files() -> None:
"""Create EMNIST Lines test images."""
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
# TODO: maybe have to add args to dataset.
dataset = EmnistLinesDataset(
init_token="<sos>",
pad_token="_",
eos_token="<eos>",
transform=[{"type": "ToTensor", "args": {}}],
target_transform=[
{
"type": "AddTokens",
"args": {"init_token": "<sos>", "pad_token": "_", "eos_token": "<eos>"},
}
],
) # nosec: S106
dataset.load_or_generate_data()
for index in [5, 7, 9]:
image, target = dataset[index]
if len(image.shape) == 3:
image = image.squeeze(0)
print(image.sum(), image.dtype)
label = "".join(dataset.mapper(label) for label in target[1:]).strip(
dataset.mapper.pad_token
)
print(label)
image = image.numpy()
util.write_image(image, str(SUPPORT_DIRNAME / f"{label}.png"))
def create_emnist_support_files():
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
dataset = EmnistDataset()
dataset.load_or_generate_data()
for ind in [5, 7, 9]:
image = dataset.x_test[ind]
label = dataset.mapping[np.argmax(dataset.y_test[ind])]
print(ind, label)
util.write_image(image, str(SUPPORT_DIRNAME / f'{label}.png'))
def create_iam_lines_support_files():
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
dataset = IamLinesDataset()
dataset.load_or_generate_data()
for ind in [0, 1, 3]:
image = dataset.x_test[ind]
label = ''.join(dataset.mapping[label] for label in np.argmax(dataset.y_test[ind], axis=-1).flatten()).strip(' _')
print(label)
util.write_image(image, str(SUPPORT_DIRNAME / f'{label}.png'))
def create_emnist_lines_support_files():
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
dataset = EmnistLinesDataset()
dataset.load_or_generate_data()
for ind in [0, 1, 3]:
image = dataset.x_test[ind]
print(image.sum(), image.dtype)
label = "".join(dataset.mapping[label] for label in np.argmax(
dataset.y_test[ind], axis=-1).flatten()).strip(" _")
print(label)
util.write_image(image, str(SUPPORT_DIRNAME / f"{label}.png"))
def create_emnist_support_files() -> None:
"""Create support images for test of CharacterPredictor class."""
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
dataset = EmnistDataset(train=False)
dataset.load_or_generate_data()
for index in [5, 7, 9]:
image, label = dataset[index]
if len(image.shape) == 3:
image = image.squeeze(0)
image = image.numpy()
label = dataset.mapper(int(label))
print(index, label)
write_image(image, str(SUPPORT_DIRNAME / f"{label}.png"))
def create_emnist_lines_support_files():
shutil.rmtree(SUPPORT_DIRNAME, ignore_errors=True)
SUPPORT_DIRNAME.mkdir()
dataset = EmnistLinesDataset()
dataset.load_or_generate_data()
for ind in [0, 1, 3]:
image = dataset.x_test[ind]
print(image.sum(), image.dtype)
label = ''.join(dataset.mapping[label] for label in np.argmax(
dataset.y_test[ind], axis=-1).flatten()).strip(' _')
print(label)
util.write_image(image, str(SUPPORT_DIRNAME / f'{label}.png'))
# Hide lines below until Lab 8
# Inverted version
image = -(image - 255)
util.write_image(image, str(SUPPORT_DIRNAME / f'{label}_i.png'))
def _crop_paragraph_image(filename: str, line_regions: Dict,
crop_dims: Tuple[int,
int], final_dims: Tuple) -> None:
image = util.read_image(filename, grayscale=True)
min_y1 = min(region["y1"] for region in line_regions) - PARAGRAPH_BUFFER
max_y2 = max(region["y2"] for region in line_regions) + PARAGRAPH_BUFFER
height = max_y2 - min_y1
crop_height = crop_dims[0]
buffer = (crop_height - height) // 2
# Generate image crop.
image_crop = 255 * np.ones(crop_dims, dtype=np.uint8)
try:
image_crop[buffer:buffer + height] = image[min_y1:max_y2]
except Exception as e: # pylint: disable=broad-except
logger.error(f"Rescued {filename}: {e}")
return
# Generate ground truth.
gt_image = np.zeros_like(image_crop, dtype=np.uint8)
for index, region in enumerate(line_regions):
gt_image[(region["y1"] - min_y1 + buffer):(region["y2"] - min_y1 +
buffer),
region["x1"]:region["x2"], ] = (index % 2 + 1)
# Generate image for debugging.
import matplotlib.pyplot as plt
cmap = plt.get_cmap("Set1")
image_crop_for_debug = np.dstack([image_crop, image_crop, image_crop])
for index, region in enumerate(line_regions):
color = [255 * _ for _ in cmap(index)[:-1]]
cv2.rectangle(
image_crop_for_debug,
(region["x1"], region["y1"] - min_y1 + buffer),
(region["x2"], region["y2"] - min_y1 + buffer),
color,
3,
)
image_crop_for_debug = cv2.resize(image_crop_for_debug,
final_dims,
interpolation=cv2.INTER_AREA)
util.write_image(image_crop_for_debug,
DEBUG_CROPS_DIRNAME / f"{filename.stem}.jpg")
image_crop = cv2.resize(image_crop,
final_dims,
interpolation=cv2.INTER_AREA)
util.write_image(image_crop, CROPS_DIRNAME / f"{filename.stem}.jpg")
gt_image = cv2.resize(gt_image,
final_dims,
interpolation=cv2.INTER_NEAREST)
util.write_image(gt_image, GT_DIRNAME / f"{filename.stem}.png")
def _crop_paragraph_image(filename, line_regions, crop_dims, final_dims): # pylint: disable=too-many-locals
image = util.read_image(filename, grayscale=True)
min_y1 = min(r["y1"] for r in line_regions) - PARAGRAPH_BUFFER
max_y2 = max(r["y2"] for r in line_regions) + PARAGRAPH_BUFFER
height = max_y2 - min_y1
crop_height = crop_dims[0]
buffer = (crop_height - height) // 2
# Generate image crop
image_crop = 255 * np.ones(crop_dims, dtype=np.uint8)
try:
image_crop[buffer:buffer + height] = image[min_y1:max_y2]
except Exception as e: # pylint: disable=broad-except
print(f"Rescued {filename}: {e}")
return
# Generate ground truth
gt_image = np.zeros_like(image_crop, dtype=np.uint8)
for ind, region in enumerate(line_regions):
gt_image[(region["y1"] - min_y1 + buffer):(region["y2"] - min_y1 +
buffer),
region["x1"]:region["x2"]] = (ind % 2 + 1)
# Generate image for debugging
import matplotlib.pyplot as plt # pylint: disable=import-outside-toplevel
cmap = plt.get_cmap("Set1")
image_crop_for_debug = np.dstack([image_crop, image_crop, image_crop])
for ind, region in enumerate(line_regions):
color = [255 * _ for _ in cmap(ind)[:-1]]
cv2.rectangle(
image_crop_for_debug,
(region["x1"], region["y1"] - min_y1 + buffer),
(region["x2"], region["y2"] - min_y1 + buffer),
color,
3,
)
image_crop_for_debug = cv2.resize(image_crop_for_debug,
final_dims,
interpolation=cv2.INTER_AREA)
util.write_image(image_crop_for_debug,
DEBUG_CROPS_DIRNAME / f"{filename.stem}.jpg")
image_crop = cv2.resize(
image_crop, final_dims,
interpolation=cv2.INTER_AREA) # Quality interpolation for input
util.write_image(image_crop, CROPS_DIRNAME / f"{filename.stem}.jpg")
gt_image = cv2.resize(
gt_image, final_dims,
interpolation=cv2.INTER_NEAREST) # No interpolation for labels
util.write_image(gt_image, GT_DIRNAME / f"{filename.stem}.png")
def _crop_paragraph_image(filename, line_regions, crop_dims, final_dims):
image = util.read_image(filename, grayscale=True)
min_y1 = min(r['y1'] for r in line_regions) - PARAGRAPH_BUFFER
max_y2 = max(r['y2'] for r in line_regions) + PARAGRAPH_BUFFER
height = max_y2 - min_y1
crop_height = crop_dims[0]
buffer = (crop_height - height) // 2
# Generate image crop
image_crop = 255 * np.ones(crop_dims, dtype=np.uint8)
try:
image_crop[buffer:buffer + height] = image[min_y1:max_y2]
except Exception as e:
print(f'Rescued {filename}: {e}')
return
# Generate ground truth
gt_image = np.zeros_like(image_crop, dtype=np.uint8)
for ind, region in enumerate(line_regions):
gt_image[
(region['y1'] - min_y1 + buffer):(region['y2'] - min_y1 + buffer),
region['x1']:region['x2']
] = ind % 2 + 1
# Generate image for debugging
import matplotlib.pyplot as plt
cmap = plt.get_cmap('Set1')
image_crop_for_debug = np.dstack([image_crop, image_crop, image_crop])
for ind, region in enumerate(line_regions):
color = [255 * _ for _ in cmap(ind)[:-1]]
cv2.rectangle(
image_crop_for_debug,
(region['x1'], region['y1'] - min_y1 + buffer),
(region['x2'], region['y2'] - min_y1 + buffer),
color,
3
)
image_crop_for_debug = cv2.resize(image_crop_for_debug, final_dims, interpolation=cv2.INTER_AREA)
util.write_image(image_crop_for_debug, DEBUG_CROPS_DIRNAME / f'{filename.stem}.jpg')
image_crop = cv2.resize(image_crop, final_dims, interpolation=cv2.INTER_AREA) # Quality interpolation for input
util.write_image(image_crop, CROPS_DIRNAME / f'{filename.stem}.jpg')
gt_image = cv2.resize(gt_image, final_dims, interpolation=cv2.INTER_NEAREST) # No interpolation for labels
util.write_image(gt_image, GT_DIRNAME / f'{filename.stem}.png')
