kaldi_data_generator.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import argparse
import os
import random
from enum import Enum
from io import BytesIO
from pathlib import Path

import cv2
import numpy as np
import requests
from PIL import Image
from apistar.exceptions import ErrorResponse
from arkindex import ArkindexClient, options_from_env

api_client = ArkindexClient(**options_from_env())


def download_image(url):
    '''
    Download an image and open it with Pillow
    '''
    assert url.startswith('http'), 'Image URL must be HTTP(S)'
    # Download the image
    # Cannot use stream=True as urllib's responses do not support the seek(int) method,
    # which is explicitly required by Image.open on file-like objects
    resp = requests.get(url)
    resp.raise_for_status()

    # Preprocess the image and prepare it for classification
    image = Image.open(BytesIO(resp.content))
    print('Downloaded image {} - size={}x{}'.format(url,
                                                    image.size[0],
                                                    image.size[1]))

    return image


def write_file(file_name, content):
    with open(file_name, 'w') as f:
        f.write(content)


class KaldiDataGenerator:

    def __init__(self, dataset_name='foo', out_dir_base='/tmp/kaldi_data', grayscale=True):
        self.out_dir_base = out_dir_base
        self.dataset_name = dataset_name
        self.grayscale = grayscale

        self.out_line_text_dir = os.path.join(self.out_dir_base, 'Transcriptions', self.dataset_name)
        os.makedirs(self.out_line_text_dir, exist_ok=True)
        self.out_line_img_dir = os.path.join(self.out_dir_base, 'Lines', self.dataset_name)
        os.makedirs(self.out_line_img_dir, exist_ok=True)

    def get_image(self, image_url, page_id):
        out_full_img_dir = os.path.join(self.out_dir_base, 'full', page_id)
        os.makedirs(out_full_img_dir, exist_ok=True)
        out_full_img_path = os.path.join(out_full_img_dir, 'full.jpg')
        if self.grayscale:
            download_image(image_url).convert('L').save(
                out_full_img_path, format='jpeg')
            img = cv2.imread(out_full_img_path, cv2.IMREAD_GRAYSCALE)
        else:
            download_image(image_url).save(
                out_full_img_path, format='jpeg')
            img = cv2.imread(out_full_img_path)
        return img

    def extract_lines(self, page_id):
        count = 0
        line_bounding_rects = []
        line_polygons = []
        line_transcriptions = []
        try:
            for res in api_client.paginate('ListTranscriptions', id=page_id, type='line'):
                text = res['text']
                if not text or not text.strip():
                    continue
                line_transcriptions.append(text)
                polygon = res['zone']['polygon']
                line_polygons.append(polygon)
                [x, y, w, h] = cv2.boundingRect(np.asarray(polygon))
                line_bounding_rects.append([x, y, w, h])
                count += 1
        except ErrorResponse as e:
            print("ListTranscriptions failed", e.status_code, e.title, e.content, page_id)
            raise e
        print("Num of lines", count)
        full_image_url = res['zone']['image']['s3_url']

        img = self.get_image(full_image_url, page_id=page_id)

        for i, [x, y, w, h] in enumerate(line_bounding_rects):
            cropped = img[y:y + h, x:x + w].copy()
            cv2.imwrite(f'{self.out_line_img_dir}/{page_id}_{i}.jpg', cropped)

        for i, text in enumerate(line_transcriptions):
            write_file(f"{self.out_line_text_dir}/{page_id}_{i}.txt", text)

    def run_pages(self, page_ids):
        for page_id in page_ids:
            print("Page", page_id)
            self.extract_lines(page_id)

    def run_volumes(self, volume_ids):
        for volume_id in volume_ids:
            print("Vol", volume_id)
            page_ids = [page['id'] for page in api_client.paginate('ListElementChildren', id=volume_id)]
            self.run_pages(page_ids)


class Split(Enum):
    Train: int = 0
    Test: int = 1
    Validation: int = 2


class KaldiPartitionSplitter:

    def __init__(self, out_dir_base='/tmp/kaldi_data', split_train_ratio=0.8, split_test_ratio=0.1):
        self.out_dir_base = out_dir_base
        self.split_train_ratio = split_train_ratio
        self.split_test_ratio = split_test_ratio
        self.split_val_ratio = 1 - self.split_train_ratio - self.split_test_ratio

    def page_level_split(self, line_ids):
        page_ids = list({'_'.join(line_id.split('_')[:-1]) for line_id in line_ids})
        random.shuffle(page_ids)
        page_count = len(page_ids)

        train_page_ids = page_ids[:round(page_count * self.split_train_ratio)]
        page_ids = page_ids[round(page_count * self.split_train_ratio):]

        test_page_ids = page_ids[:round(page_count * self.split_test_ratio)]
        page_ids = page_ids[round(page_count * self.split_test_ratio):]

        val_page_ids = page_ids

        page_dict = {page_id: Split.Train.value for page_id in train_page_ids}
        page_dict.update({page_id: Split.Test.value for page_id in test_page_ids})
        page_dict.update({page_id: Split.Validation.value for page_id in val_page_ids})
        return page_dict

    def create_partitions(self):
        print("Creating partitions")
        lines_path = Path(f'{self.out_dir_base}/Lines')
        line_ids = [str(file.relative_to(lines_path).with_suffix('')) for file in lines_path.glob('**/*.jpg')]

        page_dict = self.page_level_split(line_ids)
        datasets = [[] for _ in range(3)]
        for line_id in line_ids:
            page_id = '_'.join(line_id.split('_')[:-1])
            split_id = page_dict[page_id]
            datasets[split_id].append(line_id)

        partitions_dir = os.path.join(self.out_dir_base, 'Partitions')
        os.makedirs(partitions_dir, exist_ok=True)
        for i, dataset in enumerate(datasets):
            if not dataset:
                print(f"Partition {Split(i).name} is empty! Skipping..")
                continue
            file_name = f"{partitions_dir}/{Split(i).name}Lines.lst"
            write_file(file_name, '\n'.join(dataset) + '\n')


def create_parser():
    parser = argparse.ArgumentParser(
        description="Script to generate Kaldi training data from annotations from Arkindex")
    parser.add_argument('-n', '--dataset_name', type=str, required=True,
                        help='Name of the dataset being created for kaldi '
                             '(useful for distinguishing different datasets when in Lines or Transcriptions directory)')
    parser.add_argument('-o', '--out_dir', type=str, required=True,
                        help='output directory')
    parser.add_argument('--train_ratio', type=float, default=0.8,
                        help='Ratio of pages to be used in train (between 0 and 1)')
    parser.add_argument('--test_ratio', type=float, default=0.1,
                        help='Ratio of pages to be used in train (between 0 and 1 - train_ratio)')

    group = parser.add_mutually_exclusive_group(required=False)
    group.add_argument('--grayscale', action='store_true',
                       help='Convert images to grayscale')
    group.add_argument('--color', action='store_false',
                       help='Use color images')
    parser.set_defaults(grayscale=True)

    parser.add_argument('--volumes', nargs='*',
                        help='List of volume ids to be used, separated by spaces')
    parser.add_argument('--pages', nargs='*',
                        help='List of page ids to be used, separated by spaces')

    return parser


def main():
    args = create_parser().parse_args()

    print("ARGS", args, '\n')

    kaldi_data_generator = KaldiDataGenerator(dataset_name=args.dataset_name,
                                              out_dir_base=args.out_dir,
                                              grayscale=args.grayscale)

    kaldi_partitioner = KaldiPartitionSplitter(out_dir_base=args.out_dir,
                                               split_train_ratio=args.train_ratio,
                                               split_test_ratio=args.test_ratio)
    # extract all the lines and transcriptions
    if args.pages:
        kaldi_data_generator.run_pages(args.pages)
    if args.volumes:
        kaldi_data_generator.run_volumes(args.volumes)

    print()
    # create partitions from all the extracted data
    kaldi_partitioner.create_partitions()

    print("DONE")


if __name__ == '__main__':
    main()