论文复现：Active Learning by Learning

DeniuHe
已于 2023-10-16 16:18:54 修改
阅读量206
点赞数
文章标签： python
于 2023-10-16 08:31:33 首次发布
本文链接：https://blog.youkuaiyun.com/DeniuHe/article/details/133850778
版权
这篇文章说实在的，写的差强人意。
实质性内容是两个现有方法的拼凑！
讲的模模糊糊！对于复现代码不太友好！
撸一点，通读一遍，再撸一点，通读一遍~~~
"""
注意：使用了训练集索引。
"""
import xlwt
import xlrd
import numpy as np
import pandas as pd
from pathlib import Path
from copy import deepcopy
from sklearn.preprocessing import StandardScaler
from time import time
from sklearn.metrics.pairwise import pairwise_distances
from numpy.linalg import inv
from sklearn.metrics import accuracy_score, mean_absolute_error, f1_score, mutual_info_score
from sklearn.neighbors import NearestNeighbors
np.seterr(divide='ignore',invalid='ignore')


class AL_ALBL():
    def __init__(self,X, y, labeled, budget, X_test, y_test):
        self.X = X
        self.y = y
        self.nSample, self.nDim = X.shape
        self.labels = sorted(np.unique(self.y))
        self.nClass = len(self.labels)
        self.M = np.array([[(i - j) ** 2 for i in range(self.nClass)] for j in range(self.nClass)])
        self.X_test = X_test
        self.y_test = y_test
        self.labeled = list(deepcopy(labeled))

        self.dist_matrix = pairwise_distances(X=self.X, metric='euclidean')
        self.K = -1.0 * self.dist_matrix  # 无标记样本池对应的核矩阵
        self.K_test_pool = -1.0 * pairwise_distances(X=self.X_test, Y=self.X, metric='euclidean')
        # -------------------------------------------------
        self.budgetLeft = deepcopy(budget)
        self.c = 0.01
        self.unlabeled = [i for i in range(self.nSample)]
        self.model_initial()
        # ------------------------------
        self._nstrategies = 2
        self._delta = 0.1
        self._w = np.ones(self._nstrategies)
        self._pmin = 1.0 / (self._nstrategies * 10.0)
        self._start = True
        self._aw = np.zeros(self.nSample)
        self._aw[self.labeled] = 1.0
        self._s_idx = None
        # -------------------------------
        self._pmin = 1.0 / (self._nstrategies * 10.0)
        self.PS = np.array([0.5,0.5])
        self.phi = np.zeros((2, self.nSample))
        self.Q = np.zeros(self.nSample)
        self.T = deepcopy(budget)
        self.hat_y_matrix = np.zeros((self.T, self.nSample))
        self.Wt = np.zeros((self.T, self.nSample))
        self.reward = 0.0
        self._nT = 1/ (self.nSample * self.T)



        # -------------------------------
        self.ACClist = []
        self.MZElist = []
        self.MAElist = []
        self.F1list = []
        self.MIlist = []
        self.ALC_ACC = 0.0
        self.ALC_MZE = 0.0
        self.ALC_MAE = 0.0
        self.ALC_F1 = 0.0
        self.ALC_MI = 0.0
        self.Redundancy = 0.0
        # -------------------------------

    def model_initial(self):
        self.T_labeled = self.M[self.y[self.labeled],:]
        self.K_labeled = self.K[np.ix_(self.labeled, self.labeled)]
        self.K_labeled_inv = inv(self.c * np.eye(len(self.labeled)) + self.K_labeled)
        self.Beta = self.K_labeled_inv @ self.T_labeled
        # -----------------------------
        for idx in self.labeled:
            self.unlabeled.remove(idx)
        return self

    def Block_Matrix_Inverse(self, A11_inv, A12, A21, A22):
        n = A11_inv.shape[0]
        m = A22.shape[0]
        M = np.zeros((m+n, m+n))
        B22 = inv(A22 - A21 @ A11_inv @ A12)
        B12 = -A11_inv @ A12 @ B22
        M[n:,n:] = B22
        M[:n,:n] = A11_inv - B12 @ (A21 @ A11_inv)
        M[:n,n:] = B12
        M[n:,:n] = -B22 @ A21 @ A11_inv
        return M

    def model_incremental_train(self, new_ids):
        A12 = self.K[np.ix_(self.labeled, new_ids)]
        A22 = self.K[np.ix_(new_ids, new_ids)] + self.c * np.eye(len(new_ids))
        K_bar_inv = self.Block_Matrix_Inverse(A11_inv=self.K_labeled_inv, A12=A12, A21=A12.T, A22=A22)
        T_bar = np.vstack((self.T_labeled, self.M[self.y[new_ids],:]))
        Beta_bar = K_bar_inv @ T_bar
        # --------------------------
        self.K_labeled_inv = K_bar_inv
        self.T_labeled = T_bar
        self.Beta = Beta_bar
        return self

    def tmp_incremental_train(self, tmp_idx):
        A12 = self.K[np.ix_(self.labeled, [tmp_idx])]
        A22 = self.K[np.ix_([tmp_idx], [tmp_idx])] + self.c * np.eye(1)
        K_bar_inv = self.Block_Matrix_Inverse(A11_inv=self.K_labeled_inv, A12=A12, A21=A12.T, A22=A22)
        return K_bar_inv

    def predict_proba(self, ids):
        K_test_labeled = self.K[np.ix_(ids, self.labeled)]
        output = K_test_labeled @ self.Beta
        predictions = np.linalg.norm(output[:, None] - self.M, axis=2, ord=1)
        predictions = -predictions
        predictions = np.exp(predictions)
        predictions_sum = np.sum(predictions, axis=1, keepdims=True)
        proba_matrix = predictions / predictions_sum
        return proba_matrix

    def predict(self, ids):
        K_test_labeled = self.K[np.ix_(ids, self.labeled)]
        output = K_test_labeled @ self.Beta
        predictions = np.argmin(np.linalg.norm(output[:, None] - self.M, axis=2, ord=1), axis=1)
        return predictions

    def get_EC(self, ids):
        K_test_labeled = self.K[np.ix_(ids, self.labeled)]
        output = K_test_labeled @ self.Beta
        ids_norm1 = np.linalg.norm(output[:, None] - self.M, axis=2, ord=1)
        predictions = -1.0 * deepcopy(ids_norm1)
        predictions = np.exp(predictions)
        predictions_sum = np.sum(predictions, axis=1, keepdims=True)
        proba_matrix = predictions / predictions_sum
        return np.sum(ids_norm1 * proba_matrix, axis=1)



    def evaluation(self):
        output = self.K_test_pool[:,self.labeled] @ self.Beta
        y_hat = np.argmin(np.linalg.norm(output[:, None] - self.M, axis=2, ord=1), axis=1)
        self.ACClist.append(accuracy_score(self.y_test, y_hat))
        self.MZElist.append(1-accuracy_score(self.y_test, y_hat))
        self.MAElist.append(mean_absolute_error(self.y_test, y_hat))
        self.F1list.append(f1_score(self.y_test, y_hat, average='macro'))
        self.MIlist.append(mutual_info_score(labels_true=self.y_test, labels_pred=y_hat))
        self.ALC_ACC += self.ACClist[-1]
        self.ALC_MZE += self.MZElist[-1]
        self.ALC_MAE += self.MAElist[-1]
        self.ALC_F1 += self.F1list[-1]
        self.ALC_MI += self.MIlist[-1]

    def SoftMax(self,value_list):
        exp_values = np.exp(value_list)
        sum_exp_values = np.sum(exp_values)
        return exp_values / sum_exp_values

    def select(self):
        self.evaluation()
        t = 0  #迭代次数
        while self.budgetLeft > 0:
            if not self._start:
                self._w[self._s_idx] *= np.exp(self._pmin / 2.0 * (self.reward + 1.0 / self.last_p * np.sqrt(np.log(self._nstrategies / self._delta) / self._nstrategies)))
            self._start = False
            W = self._w.sum()
            p = (1.0 - self._nstrategies * self._pmin) * self._w / W + self._pmin

            s_idx = np.random.choice(np.arange(self._nstrategies), p=p)
            tar_idx = None
            if s_idx == 0:
                print("Div")
                # ----------------Diversity sampling criterion
                dist_D_L = self.dist_matrix[np.ix_(np.arange(self.nSample), self.labeled)]
                Div = np.min(dist_D_L, axis=1)
                tar_idx = np.argmax(Div)
            elif s_idx == 1:
                print("Expected misclassification cost")
                # ----------------Least Confidence criterion
                EC = self.get_EC(ids=np.arange(self.nSample))
                tar_idx = np.argmax(EC)
                # proba_matrix = self.predict_proba(ids=np.arange(self.nSample))
                # proba_max = np.max(proba_matrix, axis=1)
                # tar_idx = np.argmin(proba_max)
            # ==========================================
            self.last_p = p[s_idx]
            # ==========================================
            if tar_idx in self.labeled:
                """不用更新模型"""
                """计算奖励"""
                hat_y = self.predict(ids=[tar_idx])
                if hat_y == self.y[tar_idx]:
                    self.reward += self._nT / p[s_idx]
            elif tar_idx not in self.labeled:
                """更新模型"""
                self.model_incremental_train(new_ids=[tar_idx])
                self.unlabeled.remove(tar_idx)
                self.labeled.append(tar_idx)
                self.budgetLeft -= 1
                self.evaluation()
                """计算奖励"""
                hat_y = self.predict(ids=[tar_idx])
                if hat_y == self.y[tar_idx]:
                    self.reward += self._nT / p[s_idx] # the IW-ACC

            # -----------------------------------------------
            t += 1  #迭代次数加一
        neigh = NearestNeighbors(n_neighbors=1)
        neigh.fit(X=self.X[self.labeled])
        self.Redundancy = (1/np.mean(neigh.kneighbors()[0].flatten()))

if __name__ == '__main__':

    # name_list = ["Balance-scale","Toy","Cleveland","Knowledge","Glass",
    #              "Melanoma","Housing-5bin","Housing-10bin","Car"]
    name_list = ["Student","Balance-scale","Newthyroid","CTGs","Knowledge","Car","Nursery",
                 "Toy","Melanoma","Eucalyptus","Glass","Obesity1","stock-10bin","Computer-10bin"]

    class results():
        def __init__(self):
            self.ACCList = []
            self.MZEList = []
            self.MAEList = []
            self.F1List = []
            self.MIList = []
            self.ALC_ACC = []
            self.ALC_MZE = []
            self.ALC_MAE = []
            self.ALC_F1 = []
            self.ALC_MI = []
            self.Redun = []

    class stores():
        def __init__(self):
            self.num_labeled_mean = []
            self.num_labeled_std = []
            #-----------------------
            self.ACCList_mean = []
            self.ACCList_std = []
            #-----------------------
            self.MZEList_mean = []
            self.MZEList_std = []
            # -----------------
            self.MAEList_mean = []
            self.MAEList_std = []
            # -----------------
            self.F1List_mean = []
            self.F1List_std = []
            # -----------------
            self.MIList_mean = []
            self.MIList_std = []
            # -----------------
            self.ALC_ACC_mean = []
            self.ALC_ACC_std = []
            # -----------------
            self.ALC_MZE_mean = []
            self.ALC_MZE_std = []
            # -----------------
            self.ALC_MAE_mean = []
            self.ALC_MAE_std = []
            # -----------------
            self.ALC_F1_mean = []
            self.ALC_F1_std = []
            # -----------------
            self.ALC_MI_mean = []
            self.ALC_MI_std = []
            # -----------------
            self.ALC_ACC_list = []
            self.ALC_MZE_list = []
            self.ALC_MAE_list = []
            self.ALC_F1_list = []
            self.ALC_MI_list = []
            # -----------------
            self.Redun_list = []#TODO
            self.Redun_mean = []#TODO
            self.Redun_std = []#TODO

    for name in name_list:
        print("########################{}".format(name))
        data_path = Path("D:\Chapter1\DataSet")
        partition_path = Path(r"D:\Chapter1\Partition")
        """--------------read the whole data--------------------"""
        read_data_path = data_path.joinpath(name + ".csv")
        data = np.array(pd.read_csv(read_data_path, header=None))
        X = np.asarray(data[:, :-1], np.float64)
        scaler = StandardScaler()
        X = scaler.fit_transform(X)
        y = data[:, -1]
        y -= y.min()
        dist_matrix = pairwise_distances(X=X, metric="euclidean")
        nClass = len(np.unique(y))
        nSample = len(y)
        Budget = 25 * nClass
        """--------read the partitions--------"""
        read_partition_path = str(partition_path.joinpath(name + ".xls"))
        book_partition = xlrd.open_workbook(read_partition_path)
        workbook = xlwt.Workbook()
        count = 0
        # --------------------------------------
        RESULT = results()
        STORE = stores()
        # --------------------------------------
        for SN in book_partition.sheet_names():
            print("================{}".format(SN))
            S_Time = time()
            train_ids = []
            test_ids = []
            labeled = []
            table_partition = book_partition.sheet_by_name(SN)
            for idx in table_partition.col_values(0):
                if isinstance(idx,float):
                    train_ids.append(int(idx))
            for idx in table_partition.col_values(1):
                if isinstance(idx,float):
                    test_ids.append(int(idx))
            for idx in table_partition.col_values(2):
                if isinstance(idx,float):
                    labeled.append(int(idx))
            X_train = X[train_ids]
            y_train = y[train_ids].astype(np.int32)
            X_test = X[test_ids]
            y_test = y[test_ids]

            model = AL_ALBL(X=X_train, y=y_train, labeled=labeled, budget=Budget, X_test=X_test, y_test=y_test)
            model.select()

            RESULT.ACCList.append(model.ACClist)
            RESULT.MZEList.append(model.MZElist)
            RESULT.MAEList.append(model.MAElist)
            RESULT.F1List.append(model.F1list)
            RESULT.MIList.append(model.MIlist)
            RESULT.ALC_ACC.append(model.ALC_ACC)
            RESULT.ALC_MZE.append(model.ALC_MZE)
            RESULT.ALC_MAE.append(model.ALC_MAE)
            RESULT.ALC_F1.append(model.ALC_F1)
            RESULT.ALC_MI.append(model.ALC_MI)
            RESULT.Redun.append(model.Redundancy) # TODO
            print("SN===",SN, "time:",time()-S_Time)

        STORE.ACCList_mean = np.mean(RESULT.ACCList, axis=0)
        STORE.ACCList_std = np.std(RESULT.ACCList, axis=0)
        STORE.MZEList_mean = np.mean(RESULT.MZEList, axis=0)
        STORE.MZEList_std = np.std(RESULT.MZEList, axis=0)
        STORE.MAEList_mean = np.mean(RESULT.MAEList, axis=0)
        STORE.MAEList_std = np.std(RESULT.MAEList, axis=0)
        STORE.F1List_mean = np.mean(RESULT.F1List, axis=0)
        STORE.F1List_std = np.std(RESULT.F1List, axis=0)
        STORE.MIList_mean = np.mean(RESULT.MIList, axis=0)
        STORE.MIList_std = np.std(RESULT.MIList, axis=0)
        STORE.ALC_ACC_mean = np.mean(RESULT.ALC_ACC)
        STORE.ALC_ACC_std = np.std(RESULT.ALC_ACC)
        STORE.ALC_MZE_mean = np.mean(RESULT.ALC_MZE)
        STORE.ALC_MZE_std = np.std(RESULT.ALC_MZE)
        STORE.ALC_MAE_mean = np.mean(RESULT.ALC_MAE)
        STORE.ALC_MAE_std = np.std(RESULT.ALC_MAE)
        STORE.ALC_F1_mean = np.mean(RESULT.ALC_F1)
        STORE.ALC_F1_std = np.std(RESULT.ALC_F1)
        STORE.ALC_MI_mean = np.mean(RESULT.ALC_MI)
        STORE.ALC_MI_std = np.std(RESULT.ALC_MI)
        STORE.ALC_ACC_list = RESULT.ALC_ACC
        STORE.ALC_MZE_list = RESULT.ALC_MZE
        STORE.ALC_MAE_list = RESULT.ALC_MAE
        STORE.ALC_F1_list = RESULT.ALC_F1
        STORE.ALC_MI_list = RESULT.ALC_MI
        STORE.Redun_list = RESULT.Redun # TODO
        STORE.Redun_mean = np.mean(RESULT.Redun)# TODO
        STORE.Redun_std = np.std(RESULT.Redun)# TODO

        sheet_names = ["ACC","MZE","MAE","F1","MI",
                       "ALC_ACC_list","ALC_MZE_list","ALC_MAE_list","ALC_F1_list","ALC_MI_list",
                       "ALC_ACC", "ALC_MZE","ALC_MAE", "ALC_F1", "ALC_MI",
                       "Redun_list","Redun"]
        workbook = xlwt.Workbook()

        for sn in sheet_names:
            print("sn::",sn)
            sheet = workbook.add_sheet(sn)
            n_col = len(STORE.MZEList_mean)
            if sn == "ACC":
                sheet.write(0, 0, sn)
                for j in range(1,n_col + 1):
                    sheet.write(j,0,STORE.ACCList_mean[j - 1])
                    sheet.write(j,1,STORE.ACCList_std[j - 1])
            elif sn == "MZE":
                sheet.write(0, 0, sn)
                for j in range(1,n_col + 1):
                    sheet.write(j,0,STORE.MZEList_mean[j - 1])
                    sheet.write(j,1,STORE.MZEList_std[j - 1])
            elif sn == "MAE":
                sheet.write(0, 0, sn)
                for j in range(1,n_col + 1):
                    sheet.write(j,0,STORE.MAEList_mean[j - 1])
                    sheet.write(j,1,STORE.MAEList_std[j - 1])
            elif sn == "F1":
                sheet.write(0, 0, sn)
                for j in range(1,n_col + 1):
                    sheet.write(j,0,STORE.F1List_mean[j - 1])
                    sheet.write(j,1,STORE.F1List_std[j - 1])
            elif sn == "MI":
                sheet.write(0, 0, sn)
                for j in range(1,n_col + 1):
                    sheet.write(j,0,STORE.MIList_mean[j - 1])
                    sheet.write(j,1,STORE.MIList_std[j - 1])

            # ---------------------------------------------------
            elif sn == "ALC_ACC_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.ALC_ACC_list) + 1):
                    sheet.write(j,0,STORE.ALC_ACC_list[j - 1])
            elif sn == "ALC_MZE_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.ALC_MZE_list) + 1):
                    sheet.write(j,0,STORE.ALC_MZE_list[j - 1])
            elif sn == "ALC_MAE_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.ALC_MAE_list) + 1):
                    sheet.write(j,0,STORE.ALC_MAE_list[j - 1])
            elif sn == "ALC_F1_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.ALC_F1_list) + 1):
                    sheet.write(j,0,STORE.ALC_F1_list[j - 1])
            elif sn == "ALC_MI_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.ALC_MI_list) + 1):
                    sheet.write(j,0,STORE.ALC_MI_list[j - 1])

            # -----------------
            elif sn == "ALC_ACC":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.ALC_ACC_mean)
                sheet.write(2, 0, STORE.ALC_ACC_std)
            elif sn == "ALC_MZE":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.ALC_MZE_mean)
                sheet.write(2, 0, STORE.ALC_MZE_std)
            elif sn == "ALC_MAE":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.ALC_MAE_mean)
                sheet.write(2, 0, STORE.ALC_MAE_std)
            elif sn == "ALC_F1":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.ALC_F1_mean)
                sheet.write(2, 0, STORE.ALC_F1_std)
            elif sn == "ALC_MI":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.ALC_MI_mean)
                sheet.write(2, 0, STORE.ALC_MI_std)
            elif sn == "Redun_list":
                sheet.write(0, 0, sn)
                for j in range(1,len(STORE.Redun_list) + 1):
                    sheet.write(j,0,STORE.Redun_list[j - 1])
            elif sn == "Redun":
                sheet.write(0, 0, sn)
                sheet.write(1, 0, STORE.Redun_mean)
                sheet.write(2, 0, STORE.Redun_std)

        save_path = Path(r"D:\Chapter1\ALresult\ALBL")
        save_path = str(save_path.joinpath(name + ".xls"))
        workbook.save(save_path)