def show_result(filename,
q=50,
lr=0.1,
thresh=0.01,
epoch=10000,
test_size=0.3):
dataSet_name = filename.split('/')[1].split('.')[0]
print("\033[31m------------------------" + dataSet_name +
"------------------------\033[0m")
dataSet = BP_revalue(filename)
train_data, test_data = utils.splitDataSet1(dataSet, test_size=test_size)
v2, gamma2, w2, out2, errHistory2, accHistory2 = ABP(train_data,
q=q,
lr=lr,
thresh=thresh,
epoch=epoch)
v3, gamma3, w3, out3, errHistory3, accHistory3 = ABP(train_data,
q=q,
lr=lr,
thresh=thresh,
epoch=epoch,
pro=True)
plt.plot(np.arange(len(errHistory2)) * 10, errHistory2, 'r', label='ABP')
plt.plot(np.arange(len(errHistory3)) * 10, errHistory3, 'b', label='改进ABP')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.title(dataSet_name + "-ABP/改进ABP训练损失变化图")
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.legend()
plt.show()
plt.plot(np.arange(len(accHistory2)) * 10, accHistory2, 'r', label='ABP')
plt.plot(np.arange(len(accHistory3)) * 10, accHistory3, 'b', label='改进ABP')
# plt.ylim(0, 1)
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.title(dataSet_name + "-ABP/改进ABP训练准确率变化图")
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.legend()
plt.show()
#ABP算法测试
err, acc, real_list, predict_list = calErr(test_data, v2, gamma2, w2, out2)
print("------------ABP算法------------")
acc, p, r, f1 = utils.calAccuracy(predict_list, real_list)
print("正确率:{:.2%}\t查准率:{:.4f}\t查全率:{:.4f}\tF1:{:.4f}".format(
acc, p, r, f1))
#改进BP算法测试
err, acc, real_list, predict_list = calErr(test_data, v3, gamma3, w3, out3)
print("------------改进ABP算法------------")
acc, p, r, f1 = utils.calAccuracy(predict_list, real_list)
print("正确率:{:.2%}\t查准率:{:.4f}\t查全率:{:.4f}\tF1:{:.4f}".format(
acc, p, r, f1))
def evaluate_train_epoch(fixednet, trainDataLoader, criterion, device,
optimizer, bAuxiliary, auxiliary_weight):
fixednet.train()
objs = utils.AvgrageMeter()
top1 = utils.AvgrageMeter()
top5 = utils.AvgrageMeter()
for batch_idx, (traininputs, traintargets) in enumerate(trainDataLoader):
traininputs, traintargets = traininputs.to(device), traintargets.to(
device)
optimizer.zero_grad()
logits, logits_aux = fixednet(traininputs)
loss = criterion(logits, traintargets)
if bAuxiliary:
loss_aux = criterion(logits_aux, traintargets)
loss += auxiliary_weight * loss_aux
loss.backward()
nn.utils.clip_grad_norm_(fixednet.parameters(), 5)
optimizer.step()
#
prec1, prec5 = utils.calAccuracy(logits, traintargets, topk=(1, 5))
tmpBatchSize = traininputs.size(0)
objs.update(loss.data, tmpBatchSize)
top1.update(prec1.data, tmpBatchSize)
top5.update(prec5.data, tmpBatchSize)
return objs.avg, top1.avg, top5.avg
def test(filename):
X, Y_, _ = utils.getData(filename)
Y = XGboost_revalue(Y_)
dataSet_name = filename.split('/')[1].split('.')[0]
print("------------------------" + dataSet_name +
"------------------------")
train_data, train_label, test_data, test_label = utils.splitDataSet(
X, Y, test_size=0.3)
# 转换为DMatrix数据格式
dtrain = xgb.DMatrix(train_data, label=train_label)
dtest = xgb.DMatrix(test_data, label=test_label)
# 设置参数
parameters = {
'eta': 0.01,
'subsample': 0.75,
'objective': 'multi:softmax', # error evaluation for multiclass tasks
'num_class': 2, # number of classes to predic
'max_depth': 8 # depth of the trees in the boosting process
}
num_round = 500 # the number of training iterations
bst = xgb.train(parameters, dtrain, num_round)
preds = bst.predict(dtest) #输出的是概率
acc, p, r, f1 = utils.calAccuracy(preds, test_label)
print("正确率:{:.2%}\t查准率:{:.4f}\t查全率:{:.4f}\tF1:{:.4f}".format(
acc, p, r, f1))
def SBC(train_data, train_label, valid_data, valid_label):
"""
SBC: 改进朴素贝叶斯：基于分类精度和贪婪算法的属性选择方法，
参考Langley P, Sage S. Induction of selective Bayesian classifiers[M]//Uncertainty Proceedings 1994. Morgan Kaufmann, 1994: 399-406.
:param train_data:训练数据
:param train_label: 训练标签
:param valid_data:验证数据
:param valid_label: 验证标签
:returns:p1：好瓜先验概率
:returns:px1_list:好瓜中每个属性的条件概率，是一个二维列表，离散值直接返回而连续值返回的方差和均值
:returns:px0_list:坏瓜中每个属性的条件概率，是一个二维列表，离散值直接返回而连续值返回的方差和均值
:returns:col_del:删除的列名
"""
current_data = train_data #当前样本，后面会依次去掉样本
col_name = train_data.columns.tolist()
p1_best, px1_list_best, px0_list_best = train(train_data,
train_label,
is_Laplacian=True) #考虑所有样本时
pred = predict(valid_data, p1_best, px1_list_best, px0_list_best)
max_acc, p, r, f1 = utils.calAccuracy(pred, valid_label)
print("当保留所有列的时候，准确率为：", max_acc)
col_del = [] #记录被删除的列
for col in col_name: #尝试依次去掉每一列
p1, px1_list, px0_list = train(current_data.drop(columns=[col]),
train_label)
pred = predict(valid_data.drop(columns=[col]), p1, px1_list, px0_list)
acc, p, r, f1 = utils.calAccuracy(pred, valid_label)
if acc >= max_acc: #当前准确率更大
current_data = current_data.drop(columns=[col]) #去掉当前列,同时更新参数
valid_data = valid_data.drop(columns=[col])
col_del.append(col)
print("由于删除【{}】列后准确率由{}大于等于此前最大准确率{}因此删除该列！".format(
col, max_acc, acc))
max_acc = acc
p1_best = p1
px1_list_best = px1_list
px0_list_best = px0_list
if len(current_data.columns == 1):
break
else:
continue
return p1_best, px1_list_best, px0_list_best, col_del
def evaluate_test_epoch(fixednet, testDataLoader, criterion, device):
fixednet.eval()
objs = utils.AvgrageMeter()
top1 = utils.AvgrageMeter()
top5 = utils.AvgrageMeter()
#
for batch_idx, (testinputs, testtargets) in enumerate(testDataLoader):
testinputs, testtargets = testinputs.to(device), testtargets.to(device)
logits = fixednet(testinputs)
loss = criterion(logits, testtargets)
prec1, prec5 = utils.calAccuracy(logits, testtargets, topk=(1, 5))
n = testinputs.size(0)
objs.update(loss.data, n)
top1.update(prec1.data, n)
top5.update(prec5.data, n)
#break
return objs.avg, top1.avg, top5.avg
test_size = 0.3
seed = 1111
print(
"\033[31m------------------------haberman------------------------\033[0m"
)
print("\033[4;32m*************朴素贝叶斯*************\033[0m")
X, Y_, col_name = utils.getData("data/haberman.data")
Y = Bayes_revalue(Y_)
col_name = col_name.tolist()[:-1]
train_data, train_label, test_data, test_label = utils.splitDataSet(
X, Y, test_size=test_size, seed=seed)
train_data = pd.DataFrame(train_data, columns=col_name)
test_data = pd.DataFrame(test_data, columns=col_name)
p1, px1_list, px0_list = train(train_data, train_label, is_Laplacian=True)
pred = predict(test_data, p1, px1_list, px0_list)
acc, p, r, f1 = utils.calAccuracy(pred, test_label)
print("正确率:{:.2%}\t查准率:{:.4f}\t查全率:{:.4f}\tF1:{:.4f}".format(
acc, p, r, f1))
print("\033[4;32m*************朴素贝叶斯结构扩展*************\033[0m")
filename = "data/haberman.data" # haberman.data【】\heart.dat【】
dataSet = utils.getDataSet(filename)
train_data, test_data = utils.splitDataSet1(dataSet,
test_size=test_size,
seed=seed)
train_data, test_data = pd.DataFrame(train_data), pd.DataFrame(test_data)
train_label = train_data.iloc[:, -1].astype(int)
train_data = train_data.iloc[:, :-1].astype(int)
test_label = test_data.iloc[:, -1].astype(int)
test_data = test_data.iloc[:, :-1].astype(int)
pred = AODE_Predict(train_data, train_label, test_data)
filename="data/heart.dat"#haberman.data【3,0.3】\heart.dat【】
dataSet = utils.getDataSet(filename)
train_data, test_data = utils.splitDataSet1(dataSet, test_size=0.3)
test_data_data=test_data[:,:-1]
real_label=list(test_data[:,-1])
for i in range(len(real_label)):
real_label[i]=int(real_label[i])
# KNN算法测试[1,2,3,4,5,6,7,8,9,10,15,17,23][3,7,11][3,6,11]
for K in [3,5,7,11]:
pred=KNN_Predict(train_data=train_data,test_data=test_data_data,K=K)
print("------------"+filename+"---KNN算法---K="+str(K)+"------------")
# print("true:", real_label)
# print("pred:", pred_label)
acc, p, r, f1 = utils.calAccuracy(pred, real_label)
print("正确率:{:.2%}\t查准率:{:.4f}\t查全率:{:.4f}\tF1:{:.4f}".format(acc, p, r, f1))
# 读取数据并划分训练集以及测试集
filename="data/haberman.data"#haberman.data【3,0.3】\heart.dat【】
dataSet = utils.getDataSet(filename)
train_data, test_data = utils.splitDataSet1(dataSet, test_size=0.3)
print()
test_data_data=test_data[:,:-1]
real_label=list(test_data[:,-1])
for i in range(len(real_label)):
real_label[i]=int(real_label[i])
for K in [3,5,7,11]:
pred=KNN_Predict(train_data=train_data,test_data=test_data_data,K=K)
print("------------"+filename+"---KNN算法---K="+str(K)+"------------")
# print("true:", real_label)