import argparse import time import random from model import * import math import torch,gc import torch.nn as nn import torch.nn.functional as F from train_methods import * import logging import sys import numpy as np import matplotlib.pyplot as plt from Datasets import * from config import * from data_loader import * import mat73 from anchors import * from Cluster import * parser = argparse.ArgumentParser(description='CAPIMAC in PyTorch') parser.add_argument('--data', default='1', type=int, help='choice of dataset, 0-HW,1-3Sources,2BBC,3-Scene15, 4-Caltech101,5-ORL_mtv,6-Caltech_7,7-Reuters,' '8-20newsgroups,9-100leaves,10-BBC4,11-MSRCv1,12-BDGP,13-HandWritten,14-yale_mtv，15-Wikipedia-test,16-Movies,17-Prokaryotic,18-ALOI,19-flower17') parser.add_argument('-bs', '--batch-size', default='1024', type=int, help='number of batch size') parser.add_argument('-e', '--epochs', default='200', type=int, help='number of epochs to run') parser.add_argument('-lr', '--learn-rate', default='0.0001', type=float, help='learning rate of adam') parser.add_argument('-ap', '--aligned-prop', default='0.5', type=float, help='originally aligned proportions in the partially view-aligned data') parser.add_argument('--gpu', default=0, type=int, help='GPU device idx to use.') parser.add_argument('-cp', '--complete-prop', default='0.5', type=float, help='originally complete proportions in the partially sample-missing data') parser.add_argument('-m', '--margin', default='5', type=int, help='initial margin') parser.add_argument('-s', '--start-fine', default=True, type=bool, help='flag to start use robust loss or not') parser.add_argument('-np', '--neg-num', default='30', type=int, help='the ratio of negative to positive pairs') parser.add_argument('-noise', '--noisy-training', type=bool, default=True, help='training with real labels or noisy labels') parser.add_argument('-r', '--robust', default=1, type=int, help='use our robust loss or not') dim=0 class NoiseRobustLoss(nn.Module): def __init__(self): super(NoiseRobustLoss, self).__init__() def forward(self, pair_dist, P, margin, use_robust_loss, args): # print(max(pair_dist)) dist_sq = pair_dist * pair_dist P = P.to(torch.float32) N = len(P) if use_robust_loss == 1: if args.start_fine: loss = P * dist_sq + (1 - P) * (1 / margin) * torch.pow( torch.clamp(torch.pow(pair_dist, 0.5) * (0.5*margin - pair_dist), min=0.0), 2) else: loss = P * dist_sq + (1 - P) * torch.pow(torch.clamp(margin - pair_dist, min=0.0), 2) else: loss = P * dist_sq + (1 - P) * torch.pow(torch.clamp(margin - pair_dist, min=0.0), 2) loss = torch.sum(loss) / (2.0 * N) return loss def load_data(align_prop,complete_prop,neg_num,is_noise,dataset): global dim NetSeed = random.randint(1, 1000) # NetSeed=72 print(NetSeed) np.random.seed(NetSeed) torch.backends.cudnn.deterministic = True torch.manual_seed(NetSeed) # 为CPU设置随机种子 torch.cuda.manual_seed(NetSeed) # 为当前GPU设置随机种子 args = parser.parse_args() all_data = [] map_pairs = [] label = [] train_pairs = [] if dataset=='Caltech101_7': path = './datasets/' + dataset + '.mat' # 路径 mat = mat73.loadmat(path) # 加载mat文件 else: mat = sio.loadmat('./datasets/' + dataset + '.mat') if dataset == 'Scene15': data = mat['X'][0][0:2] # 20, 59 dimensions label = np.squeeze(mat['Y']) elif dataset == 'HandWritten': data = mat['X'][0][1:3] label = np.squeeze(mat['Y']) elif dataset == '3Sources': data = mat['X'][0][0:2] label = np.squeeze(mat['Y']) elif dataset == 'ALOI': data = mat['X'][0][0:2] label = np.squeeze(mat['gt']) elif dataset == 'BBCsports': data = mat['X'][0][0:2] label = np.squeeze(mat['Y']) elif dataset == 'Caltech101': data = mat['X'][0][0:2] label = np.squeeze(mat['Y']) elif dataset == 'Reuters_dim10': data = [] # 18758 samples data.append(normalize(np.vstack((mat['x_train'][0], mat['x_test'][0])))) data.append(normalize(np.vstack((mat['x_train'][1], mat['x_test'][1])))) label = np.squeeze(np.hstack((mat['y_train'], mat['y_test']))) elif dataset == 'ORL_mtv': data = mat['X'][0][0:2] label = np.squeeze(mat['gt']) elif dataset == 'Caltech101_7': data = mat['data'][3:5] data[0], data[1] = np.squeeze(data[0]), np.squeeze(data[1]) data[0], data[1] = np.array(data[0]), np.array(data[1]) label = np.squeeze(mat['labels']) elif dataset == 'Reuters': data = mat['X'][0][0:2] label = np.squeeze(mat['Y']) elif dataset == '20NewsGroups': data = mat['data'][0][1:3] label = np.squeeze(mat['truelabel'][0][0]) elif dataset == '100leaves': mat['data'][0][0], mat['data'][0][1] = mat['data'][0][0].T, mat['data'][0][1].T data = mat['data'][0][0:2] label = np.squeeze(mat['truelabel'][0][0]) elif dataset == 'BBC4': data = mat['data'][0][0:2] label = np.squeeze(mat['truelabel'][0][0]) # print(label) elif dataset == 'MSRCv1': data = mat['X'][0][1:3] label = np.squeeze(mat['Y']) elif dataset == 'BDGP': mat['X'][0][0], mat['X'][0][1] = mat['X'][0][0].T, mat['X'][0][1].T data = mat['X'][0][0:2] label = np.squeeze(mat['gt']) elif dataset == 'HandWritten': data = mat['X'][0][1:3] label = np.squeeze(mat['Y']) elif dataset == 'yale_mtv': mat['X'][0][0], mat['X'][0][1] = mat['X'][0][0].T, mat['X'][0][1].T data = mat['X'][0][0:2] # print((data)) label = np.squeeze(mat['gt']) elif dataset == 'Wikipedia-test': data = mat['X'][0:2][0:2] data = np.squeeze(data.T) # print(data) label = np.squeeze(mat['y']) elif dataset == 'Movies': data = mat['X'][0:2][0:2] data = np.squeeze(data.T) # print(data) label = np.squeeze(mat['y']) elif dataset == 'Prokaryotic': value1 = mat['X'][0][0] value2 = mat['X'][2][0] data = [value1, value2] # print(data) label = np.squeeze(mat['y']) elif dataset == 'flower17': data = mat['X'][0][0:2] label = np.squeeze(mat['Y']) divide_seed = random.randint(1, 1000) train_idx, test_idx = TT_split(len(label), 1 - align_prop, divide_seed) train_label, test_label = label[train_idx], label[test_idx] if dataset == 'Caltech101_7': data[0], data[1] = np.squeeze(data[0]), np.squeeze(data[1]) print(np.shape(data[0])) train_X, train_Y, test_X, test_Y = data[0][train_idx], data[1][train_idx], data[0][test_idx], data[1][test_idx] '''获取对齐部分的潜在表示''' map_pairs.append(train_X) map_pairs.append(train_Y) h0 , h1,epoch_time=pretrain(map_pairs, args) all_label = np.concatenate((train_label, test_label)) '''获取初始训练数据和测试数据''' if align_prop != 1: shuffle_idx = random.sample(range(len(test_Y)), len(test_Y)) test_Y = test_Y[shuffle_idx] test_label_X, test_label_Y = test_label, test_label[shuffle_idx] elif align_prop == 1: all_data.append(train_X.T) all_data.append(train_Y.T) '''不完整部分''' test_mask = get_sn(2, len(test_label), 1 - complete_prop) X_mask, Y_mask = test_mask[:, 0].astype(np.bool_), test_mask[:, 1].astype(np.bool_) # test_X[~X_mask] = 0 # test_Y[~Y_mask] = 0 test_X, test_Y = test_X[X_mask], test_Y[Y_mask] test_label_X, test_label_Y=test_label_X[X_mask], test_label_Y[Y_mask] if align_prop != 1: all_label_X = np.concatenate((train_label, test_label_X)) all_label_Y = np.concatenate((train_label, test_label_Y)) all_data.append(np.concatenate((train_X, test_X)).T) all_data.append(np.concatenate((train_Y, test_Y)).T) all_label = np.concatenate((train_label, test_label)) # all_label_X = test_label_X # all_label_Y = test_label_Y # all_data.append(test_X.T) # all_data.append(test_Y.T) # all_label = test_label elif align_prop == 1: all_label_X, all_label_Y = train_label, train_label all_label = train_label '''构建训练对''' view0, view1, noisy_labels, real_labels, _, _ = get_pairs(train_X, train_Y, neg_num, train_label) count = 0 for i in range(len(noisy_labels)): if noisy_labels[i] != real_labels[i]: count += 1 print('noise rate of the constructed neg. pairs is ', round(count / (len(noisy_labels) - len(train_X)), 2)) if is_noise == 0: # training with real_labels, v/t with real_labels print("----------------------Training with real_labels----------------------") train_pair_labels = real_labels else: # training with labels, v/t with real_labels print("----------------------Training with noisy_labels----------------------") train_pair_labels = noisy_labels '''初始化锚点''' num_unique_labels = np.unique(all_label).shape[0] anchors0,anchors1,len_indices=get_anchors(h0,h1,map_pairs,num_unique_labels)#h0是tensor '''数据重表示''' view0,view1,all_data[0],all_data[1]=torch.from_numpy(view0).float(),torch.from_numpy(view1).float(),torch.from_numpy(all_data[0]).float(),torch.from_numpy(all_data[1]).float() view0, view1, all_data[0],all_data[1]=find_nanchor(anchors0,view0),find_nanchor(anchors1,view1),find_nanchor(anchors0,all_data[0].T),find_nanchor(anchors1,all_data[1].T) #锚点数×样本数,增强锚点图 view0, view1, all_data[0], all_data[1]=np.array(view0),np.array(view1),np.array(all_data[0]),np.array(all_data[1]) print(np.shape(view0),'view0') train_pairs.append(view0) train_pairs.append(view1) train_pair_real_labels = real_labels dim=view0.shape[0] return train_pairs, train_pair_labels, train_pair_real_labels, all_data, all_label, all_label_X, all_label_Y, dim,num_unique_labels,divide_seed def normalize(x): x = (x - np.tile(np.min(x, axis=0), (x.shape[0], 1))) / np.tile((np.max(x, axis=0) - np.min(x, axis=0)), (x.shape[0], 1)) return x def loader(train_bs, align_prop, complete_prop,neg_num, is_noise, dataset): """ :param train_bs: batch size for training, default is 1024 :param neg_prop: negative / positive pairs' ratio :param test_prop: known aligned proportions for training MvCLN :param is_noise: training with noisy labels or not, 0 --- not, 1 --- yes :param data_idx: choice of dataset :return: train_pair_loader including the constructed pos. and neg. pairs used for training MvCLN, all_loader including originally aligned and unaligned data used for testing MvCLN """ train_pairs, train_pair_labels, train_pair_real_labels, all_data, all_label, all_label_X, all_label_Y, dim,num_unique_labels,divide_seed\ = load_data(align_prop,complete_prop,neg_num,is_noise, dataset) train_pair_dataset = GetDataset(train_pairs, train_pair_labels, train_pair_real_labels) train_pair_loader = DataLoader( train_pair_dataset, batch_size=train_bs, shuffle=True, drop_last=True ) return train_pair_loader, all_data, all_label, all_label_X, all_label_Y, dim,num_unique_labels,divide_seed if __name__ == '__main__': for i in range(1): args = parser.parse_args() data_name = ['HandWritten', '3Sources', 'BBCsports', 'Scene15', 'Caltech101', 'ORL_mtv', 'Caltech101_7', 'Reuters', '20NewsGroups','100leaves','BBC4','MSRCv1','BDGP','HandWritten','yale_mtv','Wikipedia-test','Movies','Prokaryotic','ALOI','flower17'] train_pair_loader, all_data, all_label, all_label_X, all_label_Y, dim, outfeature ,divide_seed=loader(args.batch_size, args.aligned_prop,args.complete_prop,args.neg_num,args.noisy_training,data_name[args.data]) model = Anchormodel(dim,outfeature).to(args.gpu) criterion = NoiseRobustLoss().to(args.gpu) # criterion_mse = nn.MSELoss().to(args.gpu) optimizer = torch.optim.Adam(model.parameters(), lr=args.learn_rate) CAR_list = [] acc_list, nmi_list, ari_list,f_list,f1_list,pre_list,pre2_list,rec_list,pur_list = [], [], [],[], [], [],[], [], [] train_time = 0 all_data[0], all_data[1]=torch.from_numpy(all_data[0]), torch.from_numpy(all_data[1]) for i in range(0, args.epochs + 1): if i == 0: with torch.no_grad(): epoch_time = train2(train_pair_loader, model, criterion, optimizer, i, args) else: epoch_time = train2(train_pair_loader, model, criterion, optimizer, i, args) # test v0, v1, pred_label, alignment_rate = tiny_infer(model, args.gpu, all_data, all_label_X, all_label_Y) CAR_list.append(alignment_rate) data = [] data.append(v0) data.append(v1) y_pred, ret, accuracy, nmi, ari, f_score, f_score2, precision, precision2, recall, purity = Clustering(data, pred_label) if i % 10 == 0: print(accuracy, nmi, ari, f_score, f_score2, precision, precision2, recall, purity) # logging.info("******** testing ********") # logging.info( # "CAR={} kmeans: acc={} nmi={} ari={}".format(round(alignment_rate, 4), ret['kmeans']['accuracy'], # ret['kmeans']['NMI'], ret['kmeans']['ARI'])) acc_list.append(ret['kmeans']['ACC']) nmi_list.append(ret['kmeans']['NMI']) ari_list.append(ret['kmeans']['ARI']) f_list.append(ret['kmeans']['F1']) f1_list.append(ret['kmeans']['F2']) pre_list.append(ret['kmeans']['PRE']) pre2_list.append(ret['kmeans']['PRE2']) rec_list.append(ret['kmeans']['REC']) pur_list.append(ret['kmeans']['PUR']) print('ACC:', max(acc_list)) print("NMI:", max(nmi_list)) print("ARI:", max(ari_list)) print("F1:", max(f_list)) print("F2:", max(f1_list)) print("PRE:", max(pre_list)) print("PRE2:", max(pre2_list)) print("REC:", max(rec_list)) print("PUR:", max(pur_list)) logging.info('******** End, training time = {} s ********'.format(round(train_time, 2)))