def output_autoencoder_performance(autoencoder_state, autoencoder_connections, check_data):
data = check_data[0]
labels = check_data[1]
neuron_states = autoencoder_state[0]
total_error = 0.0
for i in np.arange(data.shape[0]):
neuron_states[0] = data[i]
update_autoencoder(autoencoder_state, autoencoder_connections)
total_error += np.linalg.norm((neuron_states[-1] - labels[i]))
output = open(output_file, 'a')
output.write('{} mean squared error.\n\n'.format(total_error / float(data.shape[0])))
output.close()
# Show some pictures!
if display_autoencoder_images:
random_indices = np.random.randint(0, data.shape[0], 10)
inputs = np.copy(data[random_indices])
outputs = np.ndarray((10, data.shape[1]))
for i in np.arange(10):
neuron_states[0] = data[random_indices[i]]
update_autoencoder(autoencoder_state, autoencoder_connections)
outputs[i] = np.copy(neuron_states[-1])
input_viewable = denormalize(inputs)
output_viewable = denormalize(outputs)
mnist.visualize(np.concatenate((input_viewable, output_viewable)))
None
def visualize_digits(assignment, data_matrix):
groups = [[] for i in range(10)]
for i in range(len(data_matrix)):
digit = assignment[i]
groups[digit].append(data_matrix[i])
for digit in range(len(groups)):
print "Printing for digit", digit
mnist.visualize(np.array(groups[digit]))
def main():
"""
DO NOT TOUCH THIS FUNCTION. IT IS USED FOR COMPUTER EVALUATION OF YOUR CODE
"""
results = my_info() + "\t\t"
print results + "\t\t"
X, Y = mnist.read_mnist_training_data(500)
centriods = X[:10]
cm, c = kmeans(X, centriods)
mnist.visualize(cm)
# for mean, cluster in zip(cm, c):
# mnist.visualize(np.insert(cluster, 0, mean, axis=0))
centriods_unique = np.array([X[np.where(Y == i)[0][0]] for i in range(10)])
cm, c = kmeans(X, centriods_unique)
mnist.visualize(cm)
# for mean, cluster in zip(cm, c):
# mnist.visualize(np.insert(cluster, 0, mean, axis=0))
distances = distance.cdist(X, X, "euclidean")
medoids_idx, clusters = kmedoids(distances, list(range(10)))
medoids = np.array([X[int(i)] for i in medoids_idx])
c = np.array([X[clusters == i] for i in range(10)])
mnist.visualize(medoids)
# for mean, cluster in zip(cm, c):
# mnist.visualize(np.insert(cluster, 0, mean, axis=0))
mediod_idx = [np.where(Y == i)[0][0] for i in range(10)]
medoids_idx, clusters = kmedoids(distances, mediod_idx)
medoids = np.array([X[int(i)] for i in medoids_idx])
c = np.array([X[clusters == i] for i in range(10)])
mnist.visualize(medoids)
def main():
Xin = X[0:500]
print "=== k-means ==="
assignment1, cluster_means1 = kmeans(Xin, X[0:10])
print "= First iteration"
print "Cluster means"
mnist.visualize(cluster_means1)
print "Clusters"
visualize_digits(assignment1, Xin)
print "= Second iteration"
distinct_means = X[0:10].copy()
digit_set = set()
for i in range(len(Xin)):
digit = y[i]
if digit not in digit_set:
digit_set.add(digit)
distinct_means[digit] = Xin[i]
if len(digit_set) == 10:
break
print digit_set
assignment1, cluster_means1 = kmeans(Xin, distinct_means)
print "Cluster means"
mnist.visualize(cluster_means1)
print "Clusters"
visualize_digits(assignment1, Xin)
print "=== k-medoids ==="
dissimilarity_matrix = compute_dissimilarity_matrix(Xin)
print "= First iteration"
assignment2, cluster_medoids1 = kmedoids(dissimilarity_matrix, X[0:10])
print "Cluster medoids"
mnist.visualize(cluster_medoids1)
print "Clusters"
visualize_digits(assignment2, Xin)
print "= Second iteration"
assignment2, cluster_medoids1 = kmedoids(dissimilarity_matrix, distinct_means)
print "Cluster medoids"
mnist.visualize(cluster_medoids1)
print "Clusters"
visualize_digits(assignment2, Xin)
def sanity_check():
indices = np.random.choice(5000, 100, replace=False)
print bmatrix(y[indices].reshape(10,10))
mnist.visualize(X[indices])
new_mediod_indices[i] = new_mediod_i
changed = (new_mediod_indices != mediod_indices).any()
mediod_indices = np.copy(new_mediod_indices)
return new_mediod_indices, cluster_indices
X, Y = mnist_load_show.read_mnist_training_data(SAMPLE_SIZE)
first_ten = X[:10]
# select first instance of each label
first_label_instance = np.array([X[np.where(Y == i)[0][0]] for i in range(10)])
cluster_means, clusters = k_means(X, first_ten)
mnist_load_show.visualize(cluster_means)
for mean, cluster in zip(cluster_means, clusters):
mnist_load_show.visualize(np.insert(cluster, 0, mean, axis=0))
cluster_means, clusters = k_means(X, first_label_instance)
mnist_load_show.visualize(cluster_means)
for mean, cluster in zip(cluster_means, clusters):
mnist_load_show.visualize(np.insert(cluster, 0, mean, axis=0))
distances = distance.cdist(X, X, 'euclidean')
cluster_medoids_indices, clusters_indices = k_medoids(distances,
list(range(10)))
cluster_medoids = np.array([X[int(i)] for i in cluster_medoids_indices])
clusters = np.array([X[clusters_indices == i] for i in range(10)])
mnist_load_show.visualize(cluster_medoids)
for mediod, cluster in zip(cluster_medoids, clusters):
new_mediod_indices[i] = new_mediod_i
changed = (new_mediod_indices != mediod_indices).any()
mediod_indices = np.copy(new_mediod_indices)
return new_mediod_indices, cluster_indices
X, Y = mnist_load_show.read_mnist_training_data(SAMPLE_SIZE)
first_ten = X[:10]
# select first instance of each label
first_label_instance = np.array([ X[np.where(Y == i)[0][0]] for i in range(10) ])
cluster_means, clusters = k_means(X, first_ten)
mnist_load_show.visualize(cluster_means)
for mean, cluster in zip(cluster_means, clusters):
mnist_load_show.visualize(np.insert(cluster, 0, mean, axis=0))
cluster_means, clusters = k_means(X, first_label_instance)
mnist_load_show.visualize(cluster_means)
for mean, cluster in zip(cluster_means, clusters):
mnist_load_show.visualize(np.insert(cluster, 0, mean, axis=0))
distances = distance.cdist(X, X, 'euclidean')
cluster_medoids_indices, clusters_indices = k_medoids(distances, list(range(10)))
cluster_medoids = np.array([X[int(i)] for i in cluster_medoids_indices])
clusters = np.array([ X[clusters_indices == i] for i in range(10) ])
mnist_load_show.visualize(cluster_medoids)
for mediod, cluster in zip(cluster_medoids, clusters):
mnist_load_show.visualize(np.insert(cluster, 0, mediod, axis=0))
def verify(n, xs, ys): for i in random.sample(range(0,len(xs)), n): print ys[i] mnist.visualize(xs[i])
def verify(n, xs, ys):
for i in random.sample(range(0, len(xs)), n):
print ys[i]
mnist.visualize(xs[i])
