Function get_poisoned_loader

utils/generic.py:142–186

(
        dataset, batch_size, root='./data', train=True,
        noise_path=None, noise_rate=1.0, poisoned_indices_path=None, fitr=None)

def get_poisoned_loader(
        dataset, batch_size, root='./data', train=True,
        noise_path=None, noise_rate=1.0, poisoned_indices_path=None, fitr=None):

    if dataset == 'imagenet' or dataset == 'imagenet-mini':
        return imagenet_utils.get_poisoned_loader(
                dataset, batch_size, root, train, noise_path, noise_rate, poisoned_indices_path, fitr)

    target_set = get_dataset(dataset, root=root, train=train, fitr=fitr)

    if noise_path is not None:
        with open(noise_path, 'rb') as f:
            raw_noise = pickle.load(f)

        assert isinstance(raw_noise, np.ndarray)
        assert raw_noise.dtype == np.int8

        raw_noise = raw_noise.astype(np.int16)

        noise = np.zeros_like(raw_noise)

        if poisoned_indices_path is not None:
            with open(poisoned_indices_path, 'rb') as f:
                indices = pickle.load(f)
        else:
            indices = np.random.permutation(len(noise))[:int(len(noise)*noise_rate)]

        noise[indices] += raw_noise[indices]

        ''' restore noise (NCWH) for raw images (NHWC) '''
        noise = np.transpose(noise, [0,2,3,1])

        ''' add noise to images (uint8, 0~255) '''
        imgs = target_set.x.astype(np.int16) + noise
        imgs = imgs.clip(0,255).astype(np.uint8)
        target_set.x = imgs

    target_set = data.Dataset(x=target_set.x, y=target_set.y, transform=target_set.transform, fitr=target_set.fitr)

    if train:
        loader = data.Loader(target_set, batch_size=batch_size, shuffle=True, drop_last=True)
    else:
        loader = data.Loader(target_set, batch_size=batch_size, shuffle=False, drop_last=False)

    return loader


def get_clear_loader(