1 files changed, 224 insertions, 0 deletions
diff --git a/text_recognizer/data/iam_synthetic_paragraphs.py b/text_recognizer/data/iam_synthetic_paragraphs.py
new file mode 100644
index 0000000..9f1bd12
--- /dev/null
+++ b/text_recognizer/data/iam_synthetic_paragraphs.py
@@ -0,0 +1,224 @@
+"""IAM Synthetic Paragraphs Dataset class."""
+import itertools
+from pathlib import Path
+import random
+import time
+from typing import Any, List, Sequence, Tuple
+
+from loguru import logger
+import numpy as np
+from PIL import Image
+
+from text_recognizer.data.base_dataset import (
+    BaseDataset,
+    convert_strings_to_labels,
+    split_dataset,
+)
+from text_recognizer.data.base_data_module import BaseDataModule, load_and_print_info
+from text_recognizer.data.iam_paragraphs import (
+    get_dataset_properties,
+    get_transform,
+    NEW_LINE_TOKEN,
+    IAMParagraphs,
+    IMAGE_SCALE_FACTOR,
+    resize_image,
+)
+from text_recognizer.data.iam import IAM
+from text_recognizer.data.iam_lines import (
+    line_crops_and_labels,
+    load_line_crops_and_labels,
+    save_images_and_labels,
+)
+
+
+PROCESSED_DATA_DIRNAME = (
+    BaseDataModule.data_dirname() / "processed" / "iam_synthetic_paragraphs"
+)
+
+
+class IAMSyntheticParagraphs(IAMParagraphs):
+    """IAM Handwriting database of synthetic paragraphs."""
+
+    def __init__(
+        self,
+        batch_size: int = 128,
+        num_workers: int = 0,
+        train_fraction: float = 0.8,
+        augment: bool = True,
+    ) -> None:
+        super().__init__(batch_size, num_workers, train_fraction, augment)
+
+    def prepare_data(self) -> None:
+        """Prepare IAM lines to be used to generate paragraphs."""
+        if PROCESSED_DATA_DIRNAME.exists():
+            return
+
+        logger.info("Preparing IAM lines for synthetic paragraphs dataset.")
+        logger.info("Cropping IAM line regions and loading labels.")
+
+        iam = IAM()
+        iam.prepare_data()
+
+        crops_train, labels_train = line_crops_and_labels(iam, "train")
+        crops_test, labels_test = line_crops_and_labels(iam, "test")
+
+        crops_train = [resize_image(crop, IMAGE_SCALE_FACTOR) for crop in crops_train]
+        crops_test = [resize_image(crop, IMAGE_SCALE_FACTOR) for crop in crops_test]
+
+        logger.info(f"Saving images and labels at {PROCESSED_DATA_DIRNAME}")
+        save_images_and_labels(
+            crops_train, labels_train, "train", PROCESSED_DATA_DIRNAME
+        )
+        save_images_and_labels(crops_test, labels_test, "test", PROCESSED_DATA_DIRNAME)
+
+    def setup(self, stage: str = None) -> None:
+        """Loading synthetic dataset."""
+
+        logger.info(f"IAM Synthetic dataset steup for stage {stage}")
+
+        if stage == "fit" or stage is None:
+            line_crops, line_labels = load_line_crops_and_labels(
+                "train", PROCESSED_DATA_DIRNAME
+            )
+            data, paragraphs_labels = generate_synthetic_paragraphs(
+                line_crops=line_crops, line_labels=line_labels
+            )
+
+            targets = convert_strings_to_labels(
+                strings=paragraphs_labels,
+                mapping=self.inverse_mapping,
+                length=self.output_dims[0],
+            )
+            self.data_train = BaseDataset(
+                data,
+                targets,
+                transform=get_transform(
+                    image_shape=self.dims[1:], augment=self.augment
+                ),
+            )
+
+    def __repr__(self) -> str:
+        """Return information about the dataset."""
+        basic = (
+            "IAM Synthetic Paragraphs Dataset\n"  # pylint: disable=no-member
+            f"Num classes: {len(self.mapping)}\n"
+            f"Input dims : {self.dims}\n"
+            f"Output dims: {self.output_dims}\n"
+        )
+        if self.data_train is None:
+            return basic
+
+        x, y = next(iter(self.train_dataloader()))
+        data = (
+            f"Train/val/test sizes: {len(self.data_train)}, 0, 0\n"
+            f"Train Batch x stats: {(x.shape, x.dtype, x.min(), x.mean(), x.std(), x.max())}\n"
+            f"Train Batch y stats: {(y.shape, y.dtype, y.min(), y.max())}\n"
+        )
+        return basic + data
+
+
+def generate_synthetic_paragraphs(
+    line_crops: List[Image.Image], line_labels: List[str], max_batch_size: int = 9
+) -> Tuple[List[Image.Image], List[str]]:
+    """Generate synthetic paragraphs from randomly joining different subsets."""
+    paragraphs_properties = get_dataset_properties()
+
+    indices = list(range(len(line_labels)))
+
+    if max_batch_size >= paragraphs_properties["num_lines"]["max"]:
+        raise ValueError("max_batch_size greater or equalt to max num lines.")
+
+    batched_indices_list = [[index] for index in indices]
+    batched_indices_list.extend(
+        generate_random_batches(
+            values=indices, min_batch_size=2, max_batch_size=max_batch_size // 2
+        )
+    )
+    batched_indices_list.extend(
+        generate_random_batches(
+            values=indices, min_batch_size=2, max_batch_size=max_batch_size
+        )
+    )
+    batched_indices_list.extend(
+        generate_random_batches(
+            values=indices,
+            min_batch_size=max_batch_size // 2 + 1,
+            max_batch_size=max_batch_size,
+        )
+    )
+
+    paragraphs_crops, paragraphs_labels = [], []
+    for paragraph_indices in batched_indices_list:
+        paragraph_label = NEW_LINE_TOKEN.join(
+            [line_labels[i] for i in paragraph_indices]
+        )
+        if len(paragraph_label) > paragraphs_properties["label_length"]["max"]:
+            logger.info(
+                "Label longer than longest label in original IAM paragraph dataset - hence dropping."
+            )
+            continue
+
+        paragraph_crop = join_line_crops_to_form_paragraph(
+            [line_crops[i] for i in paragraph_indices]
+        )
+        max_paragraph_shape = paragraphs_properties["crop_shape"]["max"]
+
+        if (
+            paragraph_crop.height > max_paragraph_shape[0]
+            or paragraph_crop.width > max_paragraph_shape[1]
+        ):
+            logger.info(
+                "Crop larger than largest crop in original IAM paragraphs dataset - hence dropping"
+            )
+            continue
+
+        paragraphs_crops.append(paragraph_crop)
+        paragraphs_labels.append(paragraph_label)
+
+    if len(paragraphs_crops) != len(paragraphs_labels):
+        raise ValueError("Number of crops does not match number of labels.")
+
+    return paragraphs_crops, paragraphs_labels
+
+
+def join_line_crops_to_form_paragraph(line_crops: Sequence[Image.Image]) -> Image.Image:
+    """Horizontally stack line crops and return a single image forming a paragraph."""
+    crop_shapes = np.array([line.size[::-1] for line in line_crops])
+    paragraph_height = crop_shapes[:, 0].sum()
+    paragraph_width = crop_shapes[:, 1].max()
+
+    paragraph_image = Image.new(
+        mode="L", size=(paragraph_width, paragraph_height), color=0
+    )
+    current_height = 0
+    for line_crop in line_crops:
+        paragraph_image.paste(line_crop, box=(0, current_height))
+        current_height += line_crop.height
+
+    return paragraph_image
+
+
+def generate_random_batches(
+    values: List[Any], min_batch_size: int, max_batch_size: int
+) -> List[List[Any]]:
+    """Generate random batches of elements in values without replacement."""
+    shuffled_values = values.copy()
+    random.shuffle(shuffled_values)
+
+    start_index = 0
+    grouped_values_list = []
+    while start_index < len(shuffled_values):
+        num_values = random.randint(min_batch_size, max_batch_size)
+        grouped_values_list.append(
+            shuffled_values[start_index : start_index + num_values]
+        )
+        start_index += num_values
+
+    if sum([len(grp) for grp in grouped_values_list]) != len(values):
+        raise ValueError("Length of groups does not match length of values.")
+
+    return grouped_values_list
+
+
+def create_synthetic_iam_paragraphs() -> None:
+    load_and_print_info(IAMSyntheticParagraphs)