else:
svm = SVC(gamma=0.1, C=10)
# 特征标准化
X_train_std = std.fit_transform(X_train)
X_test_std = std.transform(X_test)
# 训练模型
svm.fit(X_train_std, Y_train)
# 保存模型
file_path = r'.\model\svm_model_30features\\' + str(
signal_len) + 's-train_' + str(k + 1) + '-fold.pickle'
with open(file_path, 'wb') as fw:
pickle.dump(svm, fw)
# 预测
pred_testY = svm.predict(X_test_std)
Se_test, Sp_test, ber = calc_ber(Y_test, pred_testY)
Se_list.append(Se_test)
Sp_list.append(Sp_test)
ber_list.append(ber)
#保存实验结果
result = np.vstack([Se_list, Sp_list, ber_list])
np.save(r'important_data\svm_30features_result\svm_' + str(signal_len) + 's',
result)
# 读取实验结果
result = np.load(r'important_data\svm_30features_result\svm_' +
str(signal_len) + 's.npy')
print('30features+svm,信号%ds' % (signal_len))
print('mean_Se:%.2f%%' % (np.mean(result[0, :])))
print('mean_Sp:%.2f%%' % (np.mean(result[1, :])))
Y_train_oh,
batch_size=256,
epochs=40,
validation_data=(X_test, Y_test_oh))
# 保存模型
model.save(r'.\model\\' + path + '\\' + str(signal_len) + 's-train_' +
str(k + 1) + '-fold.h5py')
# 预测
prob_trainY = model.predict(X_train)
prob_testY = model.predict(X_test)
pred_trainY = np.argmax(prob_trainY, axis=1)
pred_testY = np.argmax(prob_testY, axis=1)
# 评价指标
Se_tr, Sp_tr, Ber_tr = calc_ber(Y_train, pred_trainY)
Se_te, Sp_te, Ber_te = calc_ber(Y_test, pred_testY)
Auc_tr = calc_auc(Y_train, np.argmax(prob_trainY, axis=1), 1)
Auc_te = calc_auc(Y_test, np.argmax(prob_testY, axis=1), 1)
# 绘制代价函数曲线
train_loss = history.history['loss']
test_loss = history.history['val_loss']
train_auc = history.history['auc']
test_auc = history.history['val_auc']
# 保存重要数据
important_data = {
'train_loss': train_loss,
name='final_output',
use_bias='False')(tensor)
model = Model(inputs=input0, outputs=prediction)
# 加载训练好的模型权重
model.load_weights('.\model\\' + path + '\\' + str(signal_len) +
's-train_' + str(k + 1) + '-fold.h5py')
# 预测
prob_testY = model.predict(X_test)
# 加载重要参数
data_path = r'.\important_data\\' + path + '\\' + str(
signal_len) + 's-train_' + str(k + 1) + '-fold.pickle'
with open(data_path, 'rb') as f:
parameter = pickle.load(f)
thr = parameter['threshold']
# 概率转换成0-1标签
pred_testY = (prob_testY >= thr).astype('int8')
pred_testY = pred_testY.reshape(-1)
Ber.append(calc_ber(Y_test, pred_testY)[2])
# 找出错误样本的索引
error_index = np.argwhere((pred_testY - Y_test) != 0).reshape(-1)
Error.append(error_index)
print(k + 1)
print('%.2f' % np.mean(Ber))
error = np.hstack([Error[0], Error[1], Error[2], Error[3], Error[4]])
for i in annotation:
one = np.argwhere(original_label == i).reshape(-1) # 423 AF index
# 错误样本索引与当前心律的索引的交集
c = list(set(error).intersection(set(one)))
acc = (len(one) - len(c)) / len(one)
print('%.2f' % acc)
