论文复现：Active Learning with the Furthest NearestNeighbor Criterion for Facial Age Estimation

论文复现：Active Learning with the Furthest NearestNeighbor Criterion for Facial Age Estimation

Furthest Nearest Neighbor 方法就是其他文章中的Descripency方法，是一种diversity samplig方法。 

由于特征空间是不断变化的，在特征空间上使用Descripency方法违背了该准则的初衷。

import os
import torch
import numpy as np
from copy import deepcopy
from collections import OrderedDict
from PIL import Image
del_selection import StratifiedKFoldclass FNN_2DLDA(object):def __init__(self, X_train, y_train, labeled, budget, X_test, y_test):self.X = X_trainself.y = y_trainself.X_test = X_testself.y_test = y_testself.nSample = X_train.shape[0]print("样本个数=",self.nSample)self.labeled = list(deepcopy(labeled))     # 已标记样本的索引self.unlabeled = self.init_unlabeled_index()self.labels = np.sort(np.unique(y_train))  # 标签列表self.nClass = len(self.labels)self.budget = deepcopy(budget)self.nRow, self.nCol = self.X[0].shape     # 图像样本的行数和列数self.K = 10self.class_mean, self.global_mean, self.class_count = _init_mean()self.S_bl, self.S_br, self.S_wl, self.S_wr = _init_Sbl_Sbr_Swl_Swr()self.Wl, self.Wr = _Wl_Wr()self.X_feature = _feature()self.batch_size = 5def init_unlabeled_index(self):# =============无标记样本索引===============unlabeled = [i for i in range(self.nSample)]for idx in self.ve(idx)return unlabeleddef get_init_mean(self):class_mean = s((self.nClass, self.nRow, self.nCol))class_count = s(self.nClass)global_mean = s((self.nRow, self.nCol))# ========计算各类样本中心========for i in range(self.nClass):# ==获取第i个类的样本的索引==ids = []for idx in self.labeled:if self.y[idx] == self.labels[i]:ids.append(idx)class_count[i] = len(ids)class_mean[i] = an(self.X[ids], dim=0)# ==========计算全局样本中心============for i in range(self.nClass):global_mean += (class_count[i] / len(self.labeled)) * class_mean[i]return class_mean, global_mean, class_countdef get_init_Sbl_Sbr_Swl_Swr(self):# =============计算Sbl和Sbr=================S_bl = s((self.nCol, self.nCol))S_br = s((self.nRow, self.nRow))for i in range(self.nClass):tmp = self.class_mean[i] - self.global_meanS_bl += self.class_count[i] * (tmp.T,tmp)S_br += self.class_count[i] * (tmp, tmp.T)# =============计算Swl和Swr=================S_wl = s((self.nCol, self.nCol))S_wr = s((self.nRow, self.nRow))for i in range(self.nClass):for idx in self.labeled:if self.y[idx] == self.labels[i]:tmp = self.X[idx] - self.class_mean[i]S_wl += (tmp.T, tmp)S_wr += (tmp, tmp.T)return S_bl, S_br, S_wl, S_wrdef get_Wl_Wr(self):Wl_eigen_val, Wl_eigen_vec = torch.linalg.(torch.linalg.pinv(self.S_wl), self.S_bl))Wr_eigen_val, Wr_eigen_vec = torch.linalg.(torch.linalg.pinv(self.S_wr), self.S_br))odx_Wl = np.flipud(np.argsort(Wl_eigen_val))odx_Wr = np.flipud(np.argsort(Wr_eigen_val))Wl = s((self.nCol, self.K))Wr = s((self.K,self.nRow))for i in range(self.K):Wr[i] = Wr_eigen_vec[odx_Wr[i]]Wl[:,i] = Wl_eigen_vec[odx_Wl[i]]return Wl, Wrdef get_feature(self):X_featrue = s((self.nSample, self.K, self.K))for idx in range(self.nSample):X_featrue[idx] = ((self.Wr,self.X[idx]),self.Wl)return X_featruedef incremental_update_X_feature(self, selected):# ========update self.class_mean============for i in range(self.nClass):tmp_count = 0tmp_mean = s((self.nRow, self.nCol))for idx in selected:if self.y[idx] == self.labels[i]:tmp_count += 1tmp_mean += self.X[idx]self.class_mean[i] = (self.class_count[i] * self.class_mean[i] + tmp_count * tmp_mean) / (self.class_count[i] + tmp_count)self.class_count[i] = self.class_count[i] + tmp_count# =========updata self.global_mean===========for i in range(self.nClass):self.global_mean = s((self.nRow, self.nCol))self.global_mean += (self.class_count[i] / len(self.labeled)) * self.class_mean[i]# =========updata S_bl & S_br ===========self.S_bl = s((self.nCol, self.nCol))self.S_br = s((self.nRow, self.nRow))for i in range(self.nClass):tmp = self.class_mean[i] - self.global_meanself.S_bl += self.class_count[i] * (tmp.T,tmp)self.S_br += self.class_count[i] * (tmp, tmp.T)# =============update Swl & Swr=================self.S_wl = s((self.nCol, self.nCol))self.S_wr = s((self.nRow, self.nRow))for i in range(self.nClass):for idx in self.labeled:if self.y[idx] == self.labels[i]:tmp = self.X[idx] - self.class_mean[i]self.S_wl += (tmp.T, tmp)self.S_wr += (tmp, tmp.T)Wl_eigen_val, Wl_eigen_vec = torch.linalg.(torch.linalg.pinv(self.S_wl), self.S_bl))Wr_eigen_val, Wr_eigen_vec = torch.linalg.(torch.linalg.pinv(self.S_wr), self.S_br))odx_Wl = np.flipud(np.argsort(Wl_eigen_val))odx_Wr = np.flipud(np.argsort(Wr_eigen_val))self.Wl = s((self.nCol, self.K))self.Wr = s((self.K,self.nRow))for i in range(self.K):self.Wr[i] = Wr_eigen_vec[odx_Wr[i]]self.Wl[:,i] = Wl_eigen_vec[odx_Wl[i]]# =============更新特征===============self.X_featrue = s((self.nSample, self.K, self.K))for idx in range(self.nSample):self.X_featrue[idx] = ((self.Wr,self.X[idx]),self.Wl)def image_select(self, batch_size):metric_dict = OrderedDict()for idx in self.labeled:min_dist = np.infmin_index = Nonefor jdx in self.unlabeled:dist_tmp = (self.X_feature[idx] - self.X_feature[jdx])if dist_tmp < min_dist:min_dist = dist_tmpmin_index = jdxmetric_dict[(idx,min_index)] = min_distselected = []for i in range(batch_size):tar_tuple = max(metric_dict, key=)selected.append(tar_tuple[1])self.labeled.append(tar_tuple[1])self.labeled.append(tar_tuple[1])ve(tar_tuple[1])del metric_dict[tar_tuple]for idx in [tar_tuple[0], tar_tuple[1]]:min_dist = np.infmin_index = Nonefor jdx in self.unlabeled:dist_tmp = (self.X_feature[idx] - self.X_feature[jdx])if dist_tmp < min_dist:min_dist = dist_tmpmin_index = jdxmetric_dict[(idx,min_index)] = min_distreturn selecteddef start(self):while self.budget > 0:if self.budget > self.batch_size:selected = self.image_select(batch_size=self.batch_size)self.budget -= self.batch_sizeelse:selected = self.image_select(batch_size=self.budget)self.budget = 0print("selected::",selected)# ==========如果标记预算还没用完，则还要更新模型============if self.budget > 0:self.incremental_update_X_feature(selected=selected)if __name__ == '__main__':path_dir = r"E:PycharmProjectsDataSetsFaceDatayalefaces"# ===============基础信息=================nSample = 165nClass = 11labels = [i for i in np.arange(1,nClass+1)]nRow = 243nCol = 320Budget = 30# ==============构造标签==================y = np.zeros(165)i = 0label = 1j = 1while i < 165:if i+1 <= label*11:y[i] = labeli += 1else:label +=1# ============读取图片数据=================X = s((nSample, nRow, nCol))index = 0for name in os.listdir(path_dir):if name.split(".")[0][:7] == "subject":img = np.array(Image.open(path_dir + "\" + name))X[index] = torch.from_numpy(img)index += 1SKF = StratifiedKFold(n_splits=5, shuffle=True)for train_idx, test_idx in SKF.split(X=X,y=y):X_train = X[train_idx]y_train = y[train_idx]X_test = X[test_idx]y_test = y[test_idx]labeled = []label_dict = OrderedDict()for lab in np.unique(y_train):label_dict[lab] = []for idx in range(len(y_train)):label_dict[y_train[idx]].append(idx)for idxlist in label_dict.values():for jdx in np.random.choice(idxlist,size=2, replace=False):labeled.append(jdx)model = FNN_2DLDA(X_train=X_train,y_train=y_train,labeled=labeled,budget=Budget,X_test=X_test,y_test=y_test)model.start()break