path: root/text_recognizer/data/iam_lines.py

"""Class for IAM Lines dataset.

If not created, will generate a handwritten lines dataset from the IAM paragraphs
dataset.
"""
import json
from pathlib import Path
import random
from typing import List, Sequence, Tuple

import attr
from loguru import logger as log
import numpy as np
from PIL import Image, ImageFile, ImageOps
from torch import Tensor
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

from text_recognizer.data import image_utils
from text_recognizer.data.base_data_module import BaseDataModule, load_and_print_info
from text_recognizer.data.base_dataset import (
    BaseDataset,
    convert_strings_to_labels,
    split_dataset,
)
from text_recognizer.data.emnist_mapping import EmnistMapping
from text_recognizer.data.iam_paragraphs import get_target_transform
from text_recognizer.data.iam import IAM


ImageFile.LOAD_TRUNCATED_IMAGES = True

SEED = 4711
PROCESSED_DATA_DIRNAME = BaseDataModule.data_dirname() / "processed" / "iam_lines"
IMAGE_HEIGHT = 56
IMAGE_WIDTH = 1024
MAX_LABEL_LENGTH = 89


@attr.s(auto_attribs=True, repr=False)
class IAMLines(BaseDataModule):
    """IAM handwritten lines dataset."""

    word_pieces: bool = attr.ib(default=False)
    augment: bool = attr.ib(default=True)
    train_fraction: float = attr.ib(default=0.8)
    dims: Tuple[int, int, int] = attr.ib(
        init=False, default=(1, IMAGE_HEIGHT, IMAGE_WIDTH)
    )
    output_dims: Tuple[int, int] = attr.ib(init=False, default=(MAX_LABEL_LENGTH, 1))

    def prepare_data(self) -> None:
        """Creates the IAM lines dataset if not existing."""
        if PROCESSED_DATA_DIRNAME.exists():
            return

        log.info("Cropping IAM lines regions...")
        iam = IAM(mapping=EmnistMapping())
        iam.prepare_data()
        crops_train, labels_train = line_crops_and_labels(iam, "train")
        crops_test, labels_test = line_crops_and_labels(iam, "test")

        shapes = np.array([crop.size for crop in crops_train + crops_test])
        aspect_ratios = shapes[:, 0] / shapes[:, 1]

        log.info("Saving images, labels, and statistics...")
        save_images_and_labels(
            crops_train, labels_train, "train", PROCESSED_DATA_DIRNAME
        )
        save_images_and_labels(crops_test, labels_test, "test", PROCESSED_DATA_DIRNAME)

        with (PROCESSED_DATA_DIRNAME / "_max_aspect_ratio.txt").open(mode="w") as f:
            f.write(str(aspect_ratios.max()))

    def setup(self, stage: str = None) -> None:
        """Load data for training/testing."""
        with (PROCESSED_DATA_DIRNAME / "_max_aspect_ratio.txt").open(mode="r") as f:
            max_aspect_ratio = float(f.read())
            image_width = int(IMAGE_HEIGHT * max_aspect_ratio)
            if image_width >= IMAGE_WIDTH:
                raise ValueError("image_width equal or greater than IMAGE_WIDTH")

        if stage == "fit" or stage is None:
            x_train, labels_train = load_line_crops_and_labels(
                "train", PROCESSED_DATA_DIRNAME
            )
            if self.output_dims[0] < max([len(labels) for labels in labels_train]) + 2:
                raise ValueError("Target length longer than max output length.")

            y_train = convert_strings_to_labels(
                labels_train, self.mapping.inverse_mapping, length=self.output_dims[0]
            )
            data_train = BaseDataset(
                x_train,
                y_train,
                transform=get_transform(IMAGE_WIDTH, self.augment),
                target_transform=get_target_transform(self.word_pieces),
            )

            self.data_train, self.data_val = split_dataset(
                dataset=data_train, fraction=self.train_fraction, seed=SEED
            )

        if stage == "test" or stage is None:
            x_test, labels_test = load_line_crops_and_labels(
                "test", PROCESSED_DATA_DIRNAME
            )

            if self.output_dims[0] < max([len(labels) for labels in labels_test]) + 2:
                raise ValueError("Taget length longer than max output length.")

            y_test = convert_strings_to_labels(
                labels_test, self.mapping.inverse_mapping, length=self.output_dims[0]
            )
            self.data_test = BaseDataset(
                x_test,
                y_test,
                transform=get_transform(IMAGE_WIDTH),
                target_transform=get_target_transform(self.word_pieces),
            )

        if stage is None:
            self._verify_output_dims(labels_train, labels_test)

    def _verify_output_dims(self, labels_train: Tensor, labels_test: Tensor) -> None:
        max_label_length = max([len(label) for label in labels_train + labels_test]) + 2
        output_dims = (max_label_length, 1)
        if output_dims != self.output_dims:
            raise ValueError("Output dim does not match expected output dims.")

    def __repr__(self) -> str:
        """Return information about the dataset."""
        basic = (
            "IAM Lines dataset\n"
            f"Num classes: {len(self.mapping)}\n"
            f"Input dims: {self.dims}\n"
            f"Output dims: {self.output_dims}\n"
        )

        if not any([self.data_train, self.data_val, self.data_test]):
            return basic

        x, y = next(iter(self.train_dataloader()))
        xt, yt = next(iter(self.test_dataloader()))
        data = (
            "Train/val/test sizes: "
            f"{len(self.data_train)}, "
            f"{len(self.data_val)}, "
            f"{len(self.data_test)}\n"
            "Train Batch x stats: "
            f"{(x.shape, x.dtype, x.min(), x.mean(), x.std(), x.max())}\n"
            "Train Batch y stats: "
            f"{(y.shape, y.dtype, y.min(), y.max())}\n"
            "Test Batch x stats: "
            f"{(xt.shape, xt.dtype, xt.min(), xt.mean(), xt.std(), xt.max())}\n"
            "Test Batch y stats: "
            f"{(yt.shape, yt.dtype, yt.min(), yt.max())}\n"
        )
        return basic + data


def line_crops_and_labels(iam: IAM, split: str) -> Tuple[List, List]:
    """Load IAM line labels and regions, and load image crops."""
    crops = []
    labels = []
    for filename in iam.form_filenames:
        if not iam.split_by_id[filename.stem] == split:
            continue
        image = image_utils.read_image_pil(filename)
        image = ImageOps.grayscale(image)
        image = ImageOps.invert(image)
        labels += iam.line_strings_by_id[filename.stem]
        crops += [
            image.crop([region[box] for box in ["x1", "y1", "x2", "y2"]])
            for region in iam.line_regions_by_id[filename.stem]
        ]
    if len(crops) != len(labels):
        raise ValueError("Length of crops does not match length of labels")
    return crops, labels


def save_images_and_labels(
    crops: Sequence[Image.Image], labels: Sequence[str], split: str, data_dirname: Path
) -> None:
    """Saves generated images and labels to disk."""
    (data_dirname / split).mkdir(parents=True, exist_ok=True)

    with (data_dirname / split / "_labels.json").open(mode="w") as f:
        json.dump(labels, f)

    for index, crop in enumerate(crops):
        crop.save(data_dirname / split / f"{index}.png")


def load_line_crops_and_labels(split: str, data_dirname: Path) -> Tuple[List, List]:
    """Load line crops and labels for given split from processed directoru."""
    with (data_dirname / split / "_labels.json").open(mode="r") as f:
        labels = json.load(f)

    crop_filenames = sorted(
        (data_dirname / split).glob("*.png"),
        key=lambda filename: int(Path(filename).stem),
    )
    crops = [
        image_utils.read_image_pil(filename, grayscale=True)
        for filename in crop_filenames
    ]

    if len(crops) != len(labels):
        raise ValueError("Length of crops does not match length of labels")

    return crops, labels


def get_transform(image_width: int, augment: bool = False) -> T.Compose:
    """Augment with brigthness, rotation, slant, translation, scale, and noise."""

    def embed_crop(
        crop: Image, augment: bool = augment, image_width: int = image_width
    ) -> Image:
        # Crop is PIL.Image of dtype="L" (so value range is [0, 255])
        image = Image.new("L", (image_width, IMAGE_HEIGHT))

        # Resize crop.
        crop_width, crop_height = crop.size
        new_crop_height = IMAGE_HEIGHT
        new_crop_width = int(new_crop_height * (crop_width / crop_height))

        if augment:
            # Add random stretching
            new_crop_width = int(new_crop_width * random.uniform(0.9, 1.1))
            new_crop_width = min(new_crop_width, image_width)
        crop_resized = crop.resize(
            (new_crop_width, new_crop_height), resample=Image.BILINEAR
        )

        # Embed in image
        x = min(28, image_width - new_crop_width)
        y = IMAGE_HEIGHT - new_crop_height
        image.paste(crop_resized, (x, y))

        return image

    transfroms_list = [T.Lambda(embed_crop)]

    if augment:
        transfroms_list += [
            T.ColorJitter(brightness=(0.8, 1.6)),
            T.RandomAffine(
                degrees=1,
                shear=(-30, 20),
                interpolation=InterpolationMode.BILINEAR,
                fill=0,
            ),
        ]
    transfroms_list.append(T.ToTensor())
    return T.Compose(transfroms_list)


def generate_iam_lines() -> None:
    """Displays Iam Lines dataset statistics."""
    load_and_print_info(IAMLines)