data/split_data.py

import os, json, argparse
import numpy as np

datasets_type = ["Underwater", "Low_altitude", "Indoor", "High_altitude"]


def setup_parser():
    parser = argparse.ArgumentParser(description='Split the MSVQA datasets')
    parser.add_argument('--steps', type=int, default=10, help='Total number of steps (>=4)')
    parser.add_argument('--all', type=bool, default=True, help='Learn all scenarios in the first step')
    parser.add_argument('--shuffle', type=bool, default=True, help='Shuffle the order of learning scenarios.')
    parser.add_argument('--random_seed', type=int, default=1993)
    parser.add_argument('--train_anns_pth', type=str, default='./train_annfiles.json')
    parser.add_argument('--save_pth', type=str, default='./workspace')
    return parser


def load_json(p): return json.load(open(p, 'r', encoding='utf-8'))


def save_json(data_list, save_root, steps_num):
    if not os.path.exists(save_root): os.mkdir(save_root)
    for i, task_data in enumerate(data_list, 1):
        json.dump(task_data, open(os.path.join(save_root, f"{steps_num}-steps-{i}.json"), "w"), indent=4)


def split_datasets(img_data, ann_data, split_num, seed):
    np.random.seed(seed)
    np.random.shuffle(img_data)
    img_num, split_single = len(img_data), len(img_data) // split_num
    init_split = img_num - split_single * (split_num - 1)
    return [[ann for ann in ann_data if
             ann["image"] in img_data[init_split + split_single * (i - 1): init_split + split_single * i]] if i else
            [ann for ann in ann_data if ann["image"] in img_data[:init_split]] for i in range(split_num)]


if __name__ == "__main__":
    args = setup_parser().parse_args()
    if args.steps < 4: raise "Unsupported step size!"
    train_data = load_json(args.train_anns_pth)
    data, data_img = {}, {}
    for t in datasets_type: data[t], data_img[t] = [], []
    for ann in train_data:
        p = ann["perspective"]
        data[p].append(ann)
        if ann["image"] not in data_img[p]: data_img[p].append(ann["image"])

    omit_steps = args.steps - 4 * (args.steps // 4) - (1 if args.all else 0)
    for dataset in datasets_type:
        split_num = args.steps // 4 + (1 if omit_steps > 0 else 0) + (1 if args.all else 0)
        if omit_steps > 0: omit_steps -= 1
        data[dataset] = split_datasets(data_img[dataset], data[dataset], split_num, args.random_seed)

    if args.all:
        init_task = sum([data[t][0] for t in datasets_type], [])
        sub_tasks = sum([data[t][1:] for t in datasets_type], [])
        if args.shuffle: np.random.seed(args.random_seed); np.random.shuffle(sub_tasks)
        task_data = [init_task, *sub_tasks]
    else:
        sub_tasks = sum([data[t] for t in datasets_type], [])
        if args.shuffle: np.random.seed(args.random_seed); np.random.shuffle(sub_tasks)
        task_data = sub_tasks
    save_json(task_data, args.save_pth, args.steps)