def _generate_data(self, split: str) -> None:
print(f"EMNISTLines2 generating data for {split}...")
# pylint: disable=import-outside-toplevel
from text_recognizer.data.sentence_generator import SentenceGenerator
sentence_generator = SentenceGenerator(self.max_length - 2) # Subtract two because we will add start/end tokens
emnist = self.emnist
emnist.prepare_data()
emnist.setup()
if split == "train":
samples_by_char = get_samples_by_char(emnist.x_trainval, emnist.y_trainval, emnist.mapping)
num = self.num_train
elif split == "val":
samples_by_char = get_samples_by_char(emnist.x_trainval, emnist.y_trainval, emnist.mapping)
num = self.num_val
else:
samples_by_char = get_samples_by_char(emnist.x_test, emnist.y_test, emnist.mapping)
num = self.num_test
DATA_DIRNAME.mkdir(parents=True, exist_ok=True)
with h5py.File(self.data_filename, "a") as f:
x, y = create_dataset_of_images(
num, samples_by_char, sentence_generator, self.min_overlap, self.max_overlap, self.dims
)
y = convert_strings_to_labels(
y,
emnist.inverse_mapping,
length=MAX_OUTPUT_LENGTH,
)
f.create_dataset(f"x_{split}", data=x, dtype="u1", compression="lzf")
f.create_dataset(f"y_{split}", data=y, dtype="u1", compression="lzf")
def setup(self, stage: str = None) -> None:
print(f"IAMSyntheticParagraphs.setup({stage}): Loading trainval IAM paragraph regions and lines...")
if stage == "fit" or stage is None:
line_crops, line_labels = load_line_crops_and_labels("trainval", PROCESSED_DATA_DIRNAME)
X, para_labels = generate_synthetic_paragraphs(line_crops=line_crops, line_labels=line_labels)
Y = convert_strings_to_labels(strings=para_labels, mapping=self.inverse_mapping, length=self.output_dims[0])
transform = get_transform(image_shape=self.dims[1:], augment=self.augment) # type: ignore
self.data_train = BaseDataset(X, Y, transform=transform)
def _load_dataset(split: str, augment: bool) -> BaseDataset:
crops, labels = load_processed_crops_and_labels(split)
X = [resize_image(crop, IMAGE_SCALE_FACTOR) for crop in crops]
Y = convert_strings_to_labels(strings=labels,
mapping=self.inverse_mapping,
length=self.output_dims[0])
transform = get_transform(image_shape=self.dims[1:],
augment=augment) # type: ignore
return BaseDataset(X, Y, transform=transform)
def setup(self, stage: str = None) -> None:
with open(PROCESSED_DATA_DIRNAME / "_max_aspect_ratio.txt") as file:
max_aspect_ratio = float(file.read())
image_width = int(IMAGE_HEIGHT * max_aspect_ratio)
assert image_width <= IMAGE_WIDTH
if stage == "fit" or stage is None:
x_trainval, labels_trainval = load_line_crops_and_labels(
"trainval", PROCESSED_DATA_DIRNAME)
assert self.output_dims[0] >= max(
[len(_)
for _ in labels_trainval]) + 2 # Add 2 for start/end tokens.
y_trainval = convert_strings_to_labels(labels_trainval,
self.inverse_mapping,
length=self.output_dims[0])
data_trainval = BaseDataset(x_trainval,
y_trainval,
transform=get_transform(
IMAGE_WIDTH, self.augment))
self.data_train, self.data_val = split_dataset(
base_dataset=data_trainval, fraction=TRAIN_FRAC, seed=42)
# Note that test data does not go through augmentation transforms
if stage == "test" or stage is None:
x_test, labels_test = load_line_crops_and_labels(
"test", PROCESSED_DATA_DIRNAME)
assert self.output_dims[0] >= max([len(_)
for _ in labels_test]) + 2
y_test = convert_strings_to_labels(labels_test,
self.inverse_mapping,
length=self.output_dims[0])
self.data_test = BaseDataset(x_test,
y_test,
transform=get_transform(IMAGE_WIDTH))
if stage is None:
self._verify_output_dims(labels_trainval=labels_trainval,
labels_test=labels_test)
def setup(self, stage: str = None):
with open(PROCESSED_DATA_DIRNAME / '_max_aspect_ratio.txt') as file:
max_aspect_ratio = float(file.read())
image_width = int(IMAGE_HEIGHT * max_aspect_ratio)
assert image_width <= IMAGE_WIDTH
with open(PROCESSED_DATA_DIRNAME / 'trainval' / '_labels.json') as file:
labels_trainval = json.load(file)
with open(PROCESSED_DATA_DIRNAME / 'test' / '_labels.json') as file:
labels_test = json.load(file)
max_label_length = max([len(label) for label in labels_trainval + labels_test]) + 2 # Add 2 because of start and end tokens.
output_dims = (max_label_length, 1)
if output_dims != self.output_dims:
raise RuntimeError(dims, output_dims)
if stage == "fit" or stage is None:
filenames_trainval = sorted(
(PROCESSED_DATA_DIRNAME / 'trainval').glob('*.png'),
key=lambda filename: int(Path(filename).stem)
)
x_trainval = [Image.open(filename) for filename in filenames_trainval]
y_trainval = convert_strings_to_labels(labels_trainval, self.inverse_mapping, length=self.output_dims[0])
data_trainval = BaseDataset(x_trainval, y_trainval, transform=get_transform(IMAGE_WIDTH, self.augment))
train_size = int(TRAIN_FRAC * len(data_trainval))
val_size = len(data_trainval) - train_size
self.data_train, self.data_val = torch.utils.data.random_split(
data_trainval, [train_size, val_size], generator=torch.Generator().manual_seed(42)
)
# Note that test data does not go through augmentation transforms
if stage == "test" or stage is None:
filenames_test = sorted(
(PROCESSED_DATA_DIRNAME / 'test').glob('*.png'),
key=lambda filename: int(Path(filename).stem)
)
x_test = [Image.open(filename) for filename in filenames_test]
y_test = convert_strings_to_labels(labels_test, self.inverse_mapping, length=self.output_dims[0])
self.data_test = BaseDataset(x_test, y_test, transform=get_transform(IMAGE_WIDTH))
