def init_transform()

in src/data_manager.py [0:0]

• 25 lines of code
• 7 McCabe index (conditional complexity)

    def init_transform(root, samples, class_to_idx, seed,
                       keep_file=keep_file,
                       training=training):
        """ Transforms applied to dataset at the start of training """

        new_targets, new_samples = [], []
        if training and (keep_file is not None) and os.path.exists(keep_file):
            logger.info(f'Using {keep_file}')
            with open(keep_file, 'r') as rfile:
                for line in rfile:
                    class_name = line.split('_')[0]
                    target = class_to_idx[class_name]
                    img = line.split('\n')[0]
                    new_samples.append(
                        (os.path.join(root, class_name, img),
                         target))
                    new_targets.append(target)
        else:
            logger.info('flipping coin to keep labels')
            g = torch.Generator()
            g.manual_seed(seed)
            for sample in samples:
                if torch.bernoulli(torch.tensor(unlabel_prob), generator=g) == 0:
                    target = sample[1]
                    new_samples.append((sample[0], target))
                    new_targets.append(target)

        return np.array(new_targets), np.array(new_samples)