class_means_dict = dict()

preds_dict = dict()

k=0

for K in range(2,6):

    pred = np.zeros(shape = (N,K))

    class_means = np.zeros(shape=(K,alpha))

    # Initialize Class Means from different values from the dataset.

    #First get the random indices.

    initial_indices = np.random.choice(N,K)

    #Pick the random values from the dataset based on random initialize indices.

    for k in range(K):

        class_means[k,:] = Z[initial_indices[k],:]

    old_class_means = np.zeros(shape=(K,alpha))

   

    while np.linalg.norm(old_class_means-class_means)/np.linalg.norm(class_means) > 0.001:

        old_class_means = np.array(class_means)

        for i in range(N): # Iterate over the dataset.

            distances_to_mean = np.zeros(K)

            # For each point in dataset compute the distance from the means.

            for k in range(K):

                distances_to_mean[k] = np.linalg.norm(class_means[k]-Z[i])

            nearest_mean = np.argmin(distances_to_mean)

            #Re-intialize the means

            pred[i,:]=np.zeros(K)

            # Update the means.

            pred[i,nearest_mean]=1

        #compute the new class means.

        for k in range(K):

            class_means[k] = np.mean(Z[np.where(pred[:,k]==1)],axis=0)

 

    class_means_dict[K] = class_means

    preds_dict[K] = pred

    np.set_printoptions(precision=2)

    probabilities = np.ndarray((3,K))

    print K

    for predicted in range(K):

        for true_label in range(3):

            probabilities[true_label,predicted]=np.intersect1d(np.where(labels[:,true_label]==1),np.where(pred[:,predicted]==1)).size/float(np.where(labels[:,true_label]==1)[0].size)

            print "%.2f" %probabilities[true_label,predicted],

        print '\n'