{ "cells": [ { "cell_type": "code", "execution_count": null, "id": "1b149ac1", "metadata": {}, "outputs": [], "source": [ "import numpy as np\n", "import math\n", "import scipy\n", "import warnings\n", "import os\n", "warnings.filterwarnings(\"ignore\")\n", "import numpy as np\n", "from numpy import linalg as npla\n", "import scipy.spatial.distance\n", "from scipy.sparse import csgraph\n", "import sklearn\n", "from sklearn.cluster import KMeans\n", "from sklearn.metrics import silhouette_score\n", "from itertools import permutations\n", "from sklearn.metrics.cluster import adjusted_mutual_info_score\n", "import time\n", "def construct_L(X):\n", " n = len(X[:, 0])\n", " Dist_temp = scipy.spatial.distance.pdist(X, 'euclidean')\n", " W = np.zeros([n,n])\n", " Dist = scipy.spatial.distance.squareform(Dist_temp)\n", " sigma = np.amax(Dist)\n", " for i in range(0,n):\n", " for j in range(0,n):\n", " W[i,j] = np.exp((-Dist[i,j]**2)/(sigma*sigma))\n", " L = csgraph.laplacian(W,normed = True)\n", " L_shift = (2*np.eye(n))-L\n", " return (L_shift)\n", "def eigen(A):\n", " eigenValues, eigenVectors = npla.eigh(A)\n", " idx = np.argsort(eigenValues)[::-1]\n", " eigenValues = eigenValues[idx]\n", " eigenVectors = eigenVectors[:,idx]\n", " return (eigenValues, eigenVectors)\n", "def orthogonalize(U, eps=1e-15):\n", " n = len(U[0])\n", " V = U.T\n", " for i in range(n):\n", " prev_basis = V[0:i]\n", " coeff_vec = np.dot(prev_basis, V[i].T)\n", " V[i] -= np.dot(coeff_vec, prev_basis).T\n", " if npla.norm(V[i]) < eps:\n", " V[i][V[i] < eps] = 0\n", " else:\n", " V[i] /= npla.norm(V[i])\n", " return V.T\n", "def acc_clusters(pred_labels, true_labels):\n", " pred_labels, true_labels = np.array(pred_labels), np.array(true_labels)\n", " assert pred_labels.ndim == 1 == true_labels.ndim\n", " assert len(pred_labels) == len(true_labels)\n", " cluster_names = np.unique(pred_labels)\n", " accuracy = 0\n", " perms = np.array(list(permutations(np.unique(true_labels))))\n", " remapped_labels = true_labels\n", " for perm in perms:\n", " flipped_labels = np.zeros(len(true_labels))\n", " for label_index, label in enumerate(cluster_names):\n", " flipped_labels[pred_labels == label] = perm[label_index]\n", " testAcc = np.sum(flipped_labels == true_labels) / len(true_labels)\n", " if testAcc > accuracy:\n", " accuracy = testAcc\n", " remapped_labels = flipped_labels\n", " return accuracy\n", "a = next(os.walk('./'))[1]\n", "dataset_list = []\n", "for diter in range(len(a)):\n", " start = time.time()\n", " dataset_list.append(a[diter])\n", " data_name = a[diter]\n", " modality_list = next(os.walk('./'+dataset_list[diter]))[1]\n", " data = []\n", " for miter in range(len(modality_list)):\n", " data_modality = dataset_list[diter]+'/'+modality_list[miter]+'/'\n", " cluster_data = data_modality+\"clustering.txt\"\n", " cluster_file = open(cluster_data,\"r\")\n", " clus_data = np.asarray([[float(num) for num in line.split()] for line in cluster_file])\n", " data.append(clus_data)\n", " ground_truth_file = open(dataset_list[diter]+'/cluster_ground_truth.txt',\"r\")\n", " ground_truth = np.asarray([[float(num) for num in line.split()] for line in ground_truth_file])\n", " groundtruth = np.empty(len(ground_truth))\n", " groundtruth = np.empty(len(ground_truth))\n", " for i in range(len(ground_truth)):\n", " groundtruth[i] =int(ground_truth[i,0])\n", " v = len(data)\n", " rank_lower = int(np.ceil(len(np.unique(ground_truth))/len(data)))\n", " rank_upper = int(np.ceil(len(ground_truth)/len(data)))\n", " eigval = []\n", " eigvec = []\n", " eigenVector = []\n", " laplacian = []\n", " view_acc_order = []\n", " weight_init = []\n", " single_view_rank_table = np.empty([len(data),(rank_upper-rank_lower)+1])\n", " for view in range(len(data)):\n", " L = construct_L(data[view])\n", " sveigval,sveigvec = eigen(L)\n", " eigval.append(sveigval[1])\n", " eigvec.append(sveigvec)\n", " laplacian.append(L)\n", " for rank in range(rank_lower,rank_upper+1):\n", " eigenvec = sveigvec[:,0:rank+1]\n", " kmeans = KMeans(n_clusters=len(np.unique(groundtruth))).fit(eigenvec)\n", " cluslabel=kmeans.labels_\n", " sil_view = silhouette_score(data[view],cluslabel)+1\n", " sil_view = 0.25*sveigval[1]*sil_view\n", " single_view_rank_table[view,rank-rank_lower] = sil_view\n", " weight_view = np.argmax(single_view_rank_table[view])+1+rank_lower\n", " temp = eigvec[view]\n", " eigenVector.append(temp[:,0:(weight_view)])\n", " weight_init.append(np.max(single_view_rank_table[view])/(weight_view))\n", " view_order = weight_init\n", " view_descend = np.multiply(view_order,-1)\n", " indices = (np.argsort(view_descend))\n", " u = (eigenVector[indices[0]])\n", " for i in range(len(data)-1):\n", " utemp = eigenVector[indices[i+1]]\n", " s = u.T@utemp\n", " p = u@s\n", " q = utemp-p\n", " u = np.concatenate((u,q),axis=1)\n", " u,R = np.linalg.qr(u)\n", " H = np.empty_like(u.T@laplacian[indices[0]]@u)\n", " w = weight_init\n", " index = indices.astype(float)\n", " damp = np.ones_like(w)\n", " damp = np.ones_like(w)\n", " for i in range(len(indices)):\n", " temp = 1.25**(i+1)\n", " damp[indices[i]] = 1.0/temp \n", " W = np.multiply(w,damp)\n", " w_prime = W/np.sum(W)\n", " for i in range(len(data)):\n", " htemp = u.T@laplacian[i]@u\n", " H = H+(w_prime[i]*htemp)\n", " P,R = eigen(H)\n", " V = u@R\n", " w_new = w_prime\n", " for witer in range(1,2):\n", " H_learned = np.zeros_like(H)\n", " t = np.zeros(len(data))\n", " for viter in range(len(data)):\n", " tr = V.T@laplacian[indices[viter]]@V\n", " T = W[indices[viter]]\n", " t[indices[viter]] = T\n", " for viter in range(len(data)):\n", " w_new[viter]=math.exp(t[viter])\n", " W_new = w_new/np.sum(w_new)\n", " for viter in range(len(data)):\n", " hl_temp = u.T@laplacian[viter]@u\n", " H_learned = H_learned+(W_new[viter]*hl_temp)\n", " P1,R1 = eigen(H_learned)\n", " V = u@R1\n", " silhoutte_list = []\n", " acc_list = []\n", " ami_list =[]\n", " ari_list =[]\n", " sil_check = -1\n", " labels = []\n", " for k in range(len(np.unique(groundtruth)),u.shape[1]):\n", " V_new = V[:,0:k]\n", " kmeans = KMeans(n_clusters=len(np.unique(groundtruth))).fit(V_new)\n", " label1=kmeans.labels_\n", " silhoutte = silhouette_score(V_new,label1)\n", " acc = acc_clusters(label1,groundtruth)\n", " nmi = adjusted_mutual_info_score(groundtruth,label1)\n", " ari = sklearn.metrics.adjusted_rand_score(label1, groundtruth)\n", " silhoutte_list.append(silhoutte)\n", " acc_list.append(acc)\n", " ami_list.append(nmi)\n", " ari_list.append(ari)\n", " if sil_check