def boxcorrupt()

in ebm_sandbox.py [0:0]

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

def boxcorrupt(test_dataloader, dataloader, weights, model, target_vars, logdir, sess):
    X, Y_GT, X_final = target_vars['X'], target_vars['Y_GT'], target_vars['X_final']
    eval_im = 10000

    data_diff = []
    for data_corrupt, data, label_gt in tqdm(dataloader):
        data, label_gt = data.numpy(), label_gt.numpy()
        data_uncorrupts =  []

        data_corrupt = data.copy()
        data_corrupt[:, 16:, :] = np.random.uniform(0, 1, (FLAGS.batch_size, 16, 32, 3))

        data_corrupt_init = data_corrupt

        for j in range(10):
            feed_dict = {X: data_corrupt, Y_GT: label_gt}
            data_corrupt = sess.run([X_final], feed_dict)[0]

        val = np.mean(np.square(data_corrupt - data), axis=(1, 2, 3))
        data_diff.extend(list(val))

        if len(data_diff) > eval_im:
            break

    print("Mean {} and std {} for train dataloader".format(np.mean(data_diff), np.std(data_diff)))

    np.save("data_diff_train_image.npy", data_diff)

    data_diff = []

    for data_corrupt, data, label_gt in tqdm(test_dataloader):
        data, label_gt = data.numpy(), label_gt.numpy()
        data_uncorrupts =  []

        data_corrupt = data.copy()
        data_corrupt[:, 16:, :] = np.random.uniform(0, 1, (FLAGS.batch_size, 16, 32, 3))

        data_corrupt_init = data_corrupt

        for j in range(10):
            feed_dict = {X: data_corrupt, Y_GT: label_gt}
            data_corrupt = sess.run([X_final], feed_dict)[0]

        data_diff.extend(list(np.mean(np.square(data_corrupt - data), axis=(1, 2, 3))))

        if len(data_diff) > eval_im:
            break

    print("Mean {} and std {} for test dataloader".format(np.mean(data_diff), np.std(data_diff)))

    np.save("data_diff_test_image.npy", data_diff)