def prepare_data(self, *args, **kwargs) -> None:
"""
Prepare IAM lines such that they can be used to generate synthetic paragraphs dataset in setup().
This method is IAMLines.prepare_data + resizing of line crops.
"""
if PROCESSED_DATA_DIRNAME.exists():
return
print(
"IAMSyntheticParagraphs.prepare_data: preparing IAM lines for synthetic IAM paragraph creation..."
)
print("Cropping IAM line regions and loading labels...")
iam = IAM()
iam.prepare_data()
crops_trainval, labels_trainval = line_crops_and_labels(
iam, "trainval")
crops_test, labels_test = line_crops_and_labels(iam, "test")
crops_trainval = [
resize_image(crop, IMAGE_SCALE_FACTOR) for crop in crops_trainval
]
crops_test = [
resize_image(crop, IMAGE_SCALE_FACTOR) for crop in crops_test
]
print(f"Saving images and labels at {PROCESSED_DATA_DIRNAME}...")
save_images_and_labels(crops_trainval, labels_trainval, "trainval",
PROCESSED_DATA_DIRNAME)
save_images_and_labels(crops_test, labels_test, "test",
PROCESSED_DATA_DIRNAME)
def predict(self, image_filename: Path) -> str:
"""Predict/infer text in input image filename."""
pil_img = util.read_image_pil(image_filename, grayscale=True)
pil_img = resize_image(
pil_img, IMAGE_SCALE_FACTOR
) # ideally resize should have been part of transform
img_tensor = self.transform(pil_img)
y_pred = self.lit_model(img_tensor.unsqueeze(axis=0))[0]
pred_str = convert_y_label_to_string(y=y_pred,
mapping=self.mapping,
ignore_tokens=self.ignore_tokens)
return pred_str
def predict(self, image: Union[str, Path, Image.Image]) -> str:
"""Predict/infer text in input image (which can be a file path)."""
image_pil = image
if not isinstance(image, Image.Image):
image_pil = util.read_image_pil(image, grayscale=True)
image_pil = resize_image(image_pil, IMAGE_SCALE_FACTOR)
image_tensor = self.transform(image_pil)
y_pred = self.scripted_model(image_tensor.unsqueeze(axis=0))[0]
pred_str = convert_y_label_to_string(y=y_pred,
mapping=self.mapping,
ignore_tokens=self.ignore_tokens)
return pred_str