Multioutput Classification in scikit-learn: clean noise from pictures

from sklearn.datasets import fetch_openml mnist = fetch_openml('mnist_784', version=1) mnist.keys()

images = mnist['data'] print("Images array has the following shape: {}".format(images.shape)) print("One image has the following shape: {}".format(images.iloc[0].shape))

import numpy as np import matplotlib.pyplot as plt np.random.seed(0) def plot_imgs(images): # pick 4 random indexes of images indexes = np.random.choice(range(0, len(images)), 4) fig, axs = plt.subplots(1, 4, sharey=True, figsize=(12, 4), dpi=200) axs[0].imshow(np.array(images.iloc[indexes[0]]).reshape(28, 28), cmap='binary') axs[1].imshow(np.array(images.iloc[indexes[1]]).reshape(28, 28), cmap='binary') axs[2].imshow(np.array(images.iloc[indexes[2]]).reshape(28, 28), cmap='binary') axs[3].imshow(np.array(images.iloc[indexes[3]]).reshape(28, 28), cmap='binary') for ax in axs: ax.axis('off')

plot_imgs(images)

print(noise.shape) print(y_train.shape) print(y_test.shape)

y = images y_train = y[:60000] # take images up to the 60000th image for training y_test = y[60000:] # take images from the 60000th image to the last image noise = np.random.randint(0, 500, (len(y_train), 784)) X_train = y_train + noise # use different (unseen) noise for testing noise = np.random.randint(0, 500, (len(y_test), 784)) X_test = y_test + noise

def plot_imgs_with_labels(images, labels): # pick 4 random indexes of images indexes = np.random.choice(range(0, len(images)), 4) fig, axs = plt.subplots(2, 4, figsize=(12, 8), sharey=True, dpi=200) axs[0, 0].imshow(np.array(images.iloc[indexes[0]]).reshape(28, 28), cmap='binary') axs[0, 1].imshow(np.array(labels.iloc[indexes[0]]).reshape(28, 28), cmap='binary') axs[0, 2].imshow(np.array(images.iloc[indexes[1]]).reshape(28, 28), cmap='binary') axs[0, 3].imshow(np.array(labels.iloc[indexes[1]]).reshape(28, 28), cmap='binary') axs[1, 0].imshow(np.array(images.iloc[indexes[2]]).reshape(28, 28), cmap='binary') axs[1, 1].imshow(np.array(labels.iloc[indexes[2]]).reshape(28, 28), cmap='binary') axs[1, 2].imshow(np.array(images.iloc[indexes[3]]).reshape(28, 28), cmap='binary') axs[1, 3].imshow(np.array(labels.iloc[indexes[3]]).reshape(28, 28), cmap='binary') for ax in axs: for idx in ax: idx.axis('off') axs[0, 0].set_title('Dirty image') axs[0, 1].set_title('Clean image') axs[0, 2].set_title('Dirty image') axs[0, 3].set_title('Clean image') plot_imgs_with_labels(X_train, y_train)

from sklearn.neighbors import KNeighborsClassifier knn_clf = KNeighborsClassifier(n_jobs=-1) knn_clf.fit(X_train, y_train)

def plot_imgs_with_predictions(images, labels, model): # pick 4 random indexes of images indexes = np.random.choice(range(0, len(images)), 12) fig, axs = plt.subplots(4, 6, figsize=(12, 8), sharey=True) axs[0, 0].imshow(np.array(images.iloc[indexes[0]]).reshape(28, 28), cmap='binary') axs[0, 1].imshow(model.predict([np.array(images.iloc[indexes[0]])]).reshape(28, 28), cmap='binary') axs[0, 2].imshow(np.array(labels.iloc[indexes[0]]).reshape(28, 28), cmap='binary') axs[1, 0].imshow(np.array(images.iloc[indexes[1]]).reshape(28, 28), cmap='binary') axs[1, 1].imshow(model.predict([np.array(images.iloc[indexes[1]])]).reshape(28, 28), cmap='binary') axs[1, 2].imshow(np.array(labels.iloc[indexes[1]]).reshape(28, 28), cmap='binary') axs[2, 0].imshow(np.array(images.iloc[indexes[2]]).reshape(28, 28), cmap='binary') axs[2, 1].imshow(model.predict([np.array(images.iloc[indexes[2]])]).reshape(28, 28), cmap='binary') axs[2, 2].imshow(np.array(labels.iloc[indexes[2]]).reshape(28, 28), cmap='binary') axs[3, 0].imshow(np.array(images.iloc[indexes[3]]).reshape(28, 28), cmap='binary') axs[3, 1].imshow(model.predict([np.array(images.iloc[indexes[3]])]).reshape(28, 28), cmap='binary') axs[3, 2].imshow(np.array(labels.iloc[indexes[3]]).reshape(28, 28), cmap='binary') axs[0, 3].imshow(np.array(images.iloc[indexes[4]]).reshape(28, 28), cmap='binary') axs[0, 4].imshow(model.predict([np.array(images.iloc[indexes[4]])]).reshape(28, 28), cmap='binary') axs[0, 5].imshow(np.array(labels.iloc[indexes[4]]).reshape(28, 28), cmap='binary') axs[1, 3].imshow(np.array(images.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[1, 4].imshow(model.predict([np.array(images.iloc[indexes[5]])]).reshape(28, 28), cmap='binary') axs[1, 5].imshow(np.array(labels.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[2, 3].imshow(np.array(images.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[2, 4].imshow(model.predict([np.array(images.iloc[indexes[6]])]).reshape(28, 28), cmap='binary') axs[2, 5].imshow(np.array(labels.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[3, 3].imshow(np.array(images.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[3, 4].imshow(model.predict([np.array(images.iloc[indexes[7]])]).reshape(28, 28), cmap='binary') axs[3, 5].imshow(np.array(labels.iloc[indexes[5]]).reshape(28, 28), cmap='binary') axs[0, 0].set_title('input image') axs[0, 1].set_title('predicted image') axs[0, 2].set_title('true image') axs[0, 3].set_title('input image') axs[0, 4].set_title('predicted image') axs[0, 5].set_title('true image') for ax in axs: for idx in ax: idx.axis('off') plot_imgs_with_predictions(X_test, y_test, knn_clf)

y_test_np = np.array(y_test) perfect_prediction = y_test_np[0] perfect_diff = y_test_np[0] - perfect_prediction plt.imshow(perfect_diff.reshape(28, 28), cmap='binary') plt.title('Perfect image prediction') plt.show()

img = np.random.randint(0, len(y_test_np)) # pick a random image - everytime you run it changes - fig, axs = plt.subplots(1, 4, figsize=(12, 3), sharey=True) axs[0].imshow(y_test_np[img].reshape(28, 28), cmap='binary') axs[0].set_title('Test image') axs[1].imshow(np.array(X_test)[img].reshape(28, 28), cmap='binary') axs[1].set_title('model input') predicted = knn_clf.predict([np.array(X_test)[img]]) axs[2].imshow(predicted.reshape(28, 28), cmap='binary') axs[2].set_title('model output') diff = y_test_np[img] - predicted axs[3].imshow(np.maximum(diff, 0).reshape(28, 28), cmap='binary') axs[3].set_title('Difference') plt.show()

from sklearn.metrics import mean_squared_error mse = mean_squared_error(y_test_np[0], predicted.flatten()) # flattening the array to 1D print("Root mean squared error is: {:.2f}".format(np.sqrt(mse)))