Evaluation Toolkit

[Neptune-Trojans]

Hi,

I see that the format of the evaluation toolkit different then the training dataset.
Do you have a script that can take folder images structure and turn them to the evaluation format ?

Thanks

[mk-minchul]

Here is a sample function to create a verification evaluation dataset.

import pyrootutils
root = pyrootutils.setup_root(
    search_from=__file__,
    indicator=["__root__.txt"],
    pythonpath=True,
    dotenv=True,
)

import argparse
import numpy as np
np.bool = np.bool_  # fix bug for mxnet 1.9.1
import os
from datasets import Dataset
from functools import partial
import re


def natural_sort(l):
    convert = lambda text: int(text) if text.isdigit() else text.lower()
    alphanum_key = lambda key: [convert(c) for c in re.split('([0-9]+)', key)]
    return sorted(l, key=alphanum_key)


def entry_for_row(index, image, is_same):
    return {
        "image": image,
        "index": index,
        "is_same": is_same
    }

def generate_entries(image_list, is_same_list):
    for index, (image, is_same) in enumerate(zip(image_list, is_same_list)):
        yield entry_for_row(index, image, is_same)


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='Export bin to rec')
    parser.add_argument('--save_dir', type=str, default='/ssd2/data/faces/facerec_val',)
    args = parser.parse_args()

    pil_images, issame_list = some_function_to_get_list_of_images_and_issame_list()
    # pil_images = [Image.open(f) for f in some_path]
    # issame_list = [True, False] * (len(pil_images) // 2)

    # make issame_list repeat twice
    repeated_issame_list = np.stack([issame_list, issame_list], axis=0).transpose().flatten().tolist()
    assert repeated_issame_list[::2] == issame_list
    assert not repeated_issame_list[:len(issame_list)] == issame_list
    assert len(repeated_issame_list) == len(pil_images)


    generator = partial(generate_entries, pil_images, repeated_issame_list)
    ds = Dataset.from_generator(generator)
    dataset_name = 'some_name'
    print(dataset_name)
    ds.save_to_disk(os.path.join(args.save_dir, dataset_name), num_shards=1)

    # save some examples
    os.makedirs(os.path.join(args.save_dir, dataset_name, 'examples'))
    for i in range(5):
        ds[i]['image'].save(os.path.join(args.save_dir, dataset_name, 'examples', f'{i}.png'))

    # how to load
    # from datasets import load_dataset
    ds_file = Dataset.load_from_disk(os.path.join(args.save_dir, dataset_name))