#
# # # Just checking if 2 class fashion mnist works ###########
# # idx_to_flip = np.random.choice(num_train_examples, size=num_flips,
# # replace=False) # Generates random indices from num_train_examples
# # Y_train_flipped[idx_to_flip] = 1 - Y_train[idx_to_flip]
# # # Just checking if 2 class fashion mnist works ###########
#
# # 1: Randomly change labels to another class ---------------------------------------------------------------
# for i in range(num_flips):
# # https://stackoverflow.com/questions/42999093/generate-random-number-in-range-excluding-some-numbers/42999212
# # Y_train_flipped[i] = random.choice([j for j in range(10) if j != Y_train[i]])
# Y_train_flipped[i] = random.choice([j for j in range(num_classes) if j != Y_train[i]])
# 1: Randomly change labels to another class ---------------------------------------------------------------
tf_model.update_train_x_y(X_train, Y_train_flipped)
# 2: Include parameters in model.train() to be for ALL_CNN_C for now ---------------------------------------
tf_model.train()
# 2: Include parameters in model.train() to be for ALL_CNN_C for now --------------------------------------
flipped_results[flips_idx, random_seed_idx, 1:] = tf_model.sess.run(
[tf_model.loss_no_reg, tf_model.accuracy_op],
feed_dict=tf_model.all_test_feed_dict)
print('Flipped loss: %.5f. Accuracy: %.3f' %
(flipped_results[flips_idx, random_seed_idx, 1],
flipped_results[flips_idx, random_seed_idx, 2]))
train_losses = tf_model.sess.run(
tf_model.indiv_loss_no_reg, feed_dict=tf_model.all_train_feed_dict)
train_loo_influences = tf_model.get_loo_influences()
# 3 -------------------------------------------------------------------------------------------------------------------#
new_better_X_train = X_train[good_influence_mask] new_better_Y_train = np.array((Y_train[good_influence_mask] + 1) / 2, dtype=int) print "Number of bad points removed: ", len(X_train) - len(new_better_X_train) influences_sevens = influences[Y_train == -1] influences_hifen_sevens = influences_sevens[np.array(peaks) == 2] new_train_7s = train_7s[np.array(peaks) != 2] new_X_train = np.concatenate((train_1s, new_train_7s), axis=0) new_Y_train = np.concatenate((np.ones(len(train_1s)), np.zeros(len(new_train_7s))), axis=0) total_population_influence = np.sum(influences_hifen_sevens) print "Influence of the population", total_population_influence print "Test class", Y_test[test_idx] print "Probability score before removing the population", 1 - expit(np.array(tf_model.return_params()).dot(X_test[test_idx])) tf_model.update_train_x_y(new_X_train, new_Y_train) tf_model.train() print "Probability score after removing the population", 1 - expit(np.array(tf_model.return_params()).dot(X_test[test_idx])) plt.imshow(X_test[test_idx].reshape(28,-1)) plt.show() tf_model.update_train_x_y(new_better_X_train, new_better_Y_train) tf_model.train() print "Probability score after removing the population", 1 - expit(np.array(tf_model.return_params()).dot(X_test[test_idx])) plt.figure(1) plt.imshow(removed_example_1) plt.figure(2) plt.imshow(removed_example_2) plt.show()