def NAPFD_rank(self):
# 计算NAPFD的值
NAPFD_list = self.data_grouped.apply(
lambda NAPFD_stage: result_analysis.tc_NAPFD(NAPFD_stage)
if len(NAPFD_stage) >= self.mini_batch else -1).reset_index()
NAPFD_list.columns = ['Stage', 'NAPFD']
x_y = NAPFD_list[NAPFD_list.NAPFD != -1]
return x_y
def NAPFD_inf(self):
# 计算NAPFD曲线的下界
NAPFD_list_min = self.data_grouped.apply(
lambda NAPFD_stage: result_analysis.tc_NAPFD(NAPFD_stage, sort_by=['Verdict', 'Duration'],
ascending=[True, False]) if len(
NAPFD_stage) >= self.mini_batch else -1
).reset_index()
NAPFD_list_min.columns = ['Stage', 'NAPFD']
x_y_min = NAPFD_list_min[NAPFD_list_min.NAPFD != -1]
return x_y_min
def TcData2Result(tc_Stage=None,
rank_net=None,
PATH=None,
PATH_ori=None,
N_input=None,
N_output=None,
N_hidden=None,
N_layers=None,
dataset_name=None,
data_save_path=None,
save_name='tc_result.csv',
memory_setting=None,
train_setting=None,
train_flag=None,
param_read=None,
save_model_ori=None,
save_model=None,
history_length=None,
alpha=0.5,
gamma=0.5,
learn_rate=0.2,
epoch=1,
NAPFD_view=False):
"""
根据可选参数实现深度强化学习的测试用例排序
:param PATH: 模型训练结果参数保存路径
:param PATH_ori: 模型初始参数保存路径
:param rank_net: 模型输入
:param epoch: 记忆回放次数
:param learn_rate: 排序网络学习率
:param alpha: 奖励作用程度
:param gamma: 预期奖励折扣
:param tc_Stage: 数据的阶段标签
:param N_input: 智能体决策网络输入向量维度
:param N_output: 智能体决策网络输出向量维度
:param N_hidden: 智能体决策网络隐藏层节点个数
:param N_layers: 智能体决策网络层数
:param dataset_name: 数据集名称
:param data_save_path: 排序结果存储路径
:param save_name: 结果保存
:param stage_flag: 根据阶段的区分标签生成阶段标签
:param memory_setting: 记忆存储类型
:param train_setting: 智能体的学习类型
:param train_flag: 是否进行训练
:param param_read: 参数读取
:param save_model_ori: 是否保存网络初始参数
:param save_model: 是否保存网络的最终参数
:param history_length: 使用的历史信息长度
:return: 生成测试用例排序结果文件
"""
if dataset_name is None:
print("没有给定数据集,将采用默认数据集iofrol.csv")
dataset_name = "iofrol"
file_name = r"./data/" + dataset_name + ".csv"
if data_save_path is None:
data_save_path = './result/result_data/result_of_' + dataset_name + '/' + save_name
# 网络参数设定
if history_length is None:
history_length = 4 # 使用历史信息长度
if N_input is None:
N_input = 6 # 网络输入向量维度
if N_output is None:
N_output = 1 # 网络输出向量维度
if N_hidden is None:
N_hidden = 12
if N_layers is None:
N_layers = 2
if train_flag is None:
train_flag = True # 是否训练
if param_read is None:
param_read = False # 是否使用现有参数
if save_model_ori is None:
save_model_ori = False # 是否保存网络初始权重
if save_model is None:
save_model = False # 是否保存网络权重
if memory_setting is None:
memory_setting = "current"
if train_setting is None:
train_setting = "DNN"
if PATH is None:
PATH = './params/' + train_setting + '_' + memory_setting + '_' + str(N_layers) + '_' + \
str(N_hidden) + '_' + dataset_name + '_' + str(learn_rate) + '_' + str(alpha) + \
'_' + str(gamma) + '_' + str(epoch) + '.pth' # 模型保存路径
# PATH_ori = './params/' + train_setting + '_' + memory_setting + '_' + str(N_layers) + '_' + \
# str(N_hidden) + '_' + dataset_name + '_ori.pth' # 模型保存路径
if PATH_ori is None:
PATH_ori = './params/' + dataset_name + '.pth' # 模型保存路径
# 排序网络所使用的数据
data_list = ['Duration', 'RunTime', 'FreeTime'] + \
[('LastResult' + str(R_name)) for R_name in range(1, history_length + 1)]
data_list_act = ['RunTime', 'FreeTime'] + \
[('LastResult' + str(R_name)) for R_name in range(1, history_length + 1)]
data_list_learn = ['Duration', 'FreeTime', 'Verdict'] + \
[('LastResult' + str(R_name)) for R_name in range(1, history_length)]
NAPFD_frame = pd.DataFrame(columns=['Stage', 'NAPFD'])
if memory_setting == "current":
TS_minimal = 5 # 只有当次经验记忆时,当本次测试套件规模大于TS_minimal时智能体才进行学习
elif memory_setting == "replay":
memory_capacity = 10000 # 具有记忆能力时的记忆容量
memory_replay = 1000 # 每次回放次数
else:
pass
memory_storage = pd.DataFrame(columns=data_list)
# 重新定义持续集成阶段
if tc_Stage is None:
tc_Stage = pd.read_csv(file_name,
sep=';',
usecols=["Id", "Stage"],
index_col="Id")
tc_Name = pd.read_csv(file_name,
sep=';',
usecols=["Id", "Name"],
index_col="Id")
# 定义排序网络
if rank_net is None:
rank_net = tcNet.Rank_net(N_input, N_output, N_hidden)
# 是否使用已经训练好的模型
if not param_read:
# 网络参数初始化
tcNet.initNetParams(rank_net) # 读取训练好的模型
else:
rank_net.load_state_dict(torch.load(PATH_ori))
criterion = nn.MSELoss()
optimizer = optim.Adam(rank_net.parameters(), lr=learn_rate)
# 保存初始参数
if save_model_ori:
print("网络初始参数为:\n", list(rank_net.parameters()))
torch.save(rank_net.state_dict(), PATH_ori)
if train_setting == "DQN":
# 构建目标网络
rank_net_target = copy.deepcopy(rank_net)
print("此数据包含", str(tc_Stage.Stage.max()), "个阶段")
time_start = time.clock()
print("程序起始运行时间为",
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
for stage in range(1, max(tc_Stage.Stage) + 1):
"""读入本阶段数据"""
# 读取指定阶段的测试数据
tc_data_stage = tcDataRecon.tc_data_stage_read(tc_Stage, stage,
file_name)
tc_data_stage['Stage'] = stage
start_index = tc_data_stage.index.min()
# 产生网络输入数据
run_time_average = 0
run_time_std = 0
if not memory_storage.empty:
run_time_average = memory_storage.Duration.mean()
run_time_std = memory_storage.Duration.std()
tc_data_stage = tcDataRecon.tc_data_generalize_FreeTime(
tc_data_stage, tc_Name, file_name)
tc_data_stage = tcDataRecon.tc_data_generalize_RunTime(
tc_data_stage, tc_Name, file_name, run_time_average, run_time_std)
tc_data_stage = tcDataRecon.tc_LastResults_split(
tc_data_stage, history_length=history_length)
# 加入结果保存列
tc_data_stage = tc_data_stage.reindex(
columns=list(tc_data_stage.columns) + ['Rank'], fill_value=0)
# 更新记忆存储库
if memory_setting == 'current':
memory_storage = tc_data_stage
elif memory_setting == 'replay':
memory_storage = pd.concat([memory_storage, tc_data_stage],
ignore_index=True,
sort=True)
# 将记忆存储库进行删减
index = 0
while memory_storage.shape[0] > memory_capacity:
memory_storage = memory_storage.drop(index=index)
index += 1
memory_storage.reset_index(drop=True)
else:
pass
# 对更新后的记忆存储库进行max-min标准化
memory_storage_std = memory_storage.copy(deep=True)
memory_storage_std.FreeTime = TcData_std(memory_storage_std.FreeTime,
std_setting='max-min')
memory_storage_std.RunTime = TcData_std(memory_storage_std.RunTime,
std_setting='max-min')
memory_storage_std.RunTime = TcData_filled(memory_storage_std.RunTime)
memory_storage_std.Duration = TcData_std(memory_storage_std.Duration,
std_setting='max-min')
tc_data_stage.FreeTime = TcData_std(tc_data_stage.FreeTime,
std_setting='max-min')
tc_data_stage.RunTime = TcData_std(tc_data_stage.RunTime,
std_setting='max-min')
tc_data_stage.RunTime = TcData_filled(tc_data_stage.RunTime)
"""智能体决策"""
# 将DataFrame转换为torch.tensor
data_vec = torch.tensor(tc_data_stage.loc[:, data_list_act].apply(
pd.to_numeric).values,
dtype=torch.float,
requires_grad=True) # 网络输入值
# 对输入的测试用例输出rank
with torch.no_grad():
for index in range(len(data_vec)):
# 每个样本数据的处理
data_input = data_vec[index].clone().detach()
data_output = rank_net(data_input)
tc_data_stage.loc[start_index + index,
'Rank'] = data_output.item()
if stage == 1:
tc_data_stage.to_csv(data_save_path,
columns=[
'Name', 'Duration', 'Verdict', 'Rank',
'Stage', 'Cycle'
],
float_format="%.3f",
sep=';')
else:
tc_data_stage.to_csv(data_save_path,
mode='a',
columns=[
'Name', 'Duration', 'Verdict', 'Rank',
'Stage', 'Cycle'
],
header=False,
float_format="%.3f",
sep=';')
# 计算排序效果
if NAPFD_view and stage % 25 == 0 and stage >= 25:
NAPFD_frame.loc[stage, 'Stage'] = stage
NAPFD_frame.loc[stage,
'NAPFD'] = result_analysis.tc_NAPFD(tc_data_stage)
result_analysis.result_analysis(N_hidden=N_hidden,
N_layers=N_layers,
save_name=save_name,
train_setting=train_setting,
memory_setting=memory_setting,
dataset_name=dataset_name,
save_flag=False,
show_flag=True,
plot_NAPFD=False,
plot_NAPFD_adj=True,
mini_batch=10)
"""智能体学习"""
if train_flag:
# 记忆读取
for epoch_curr in range(epoch):
if memory_setting == 'current': # 当前轮次经验学习
data_replay = memory_storage_std
elif memory_setting == 'replay': # 记忆回放
if memory_storage_std.shape[0] >= memory_replay:
data_replay = memory_storage_std.sample(memory_replay)
else:
data_replay = memory_storage_std
else:
pass
# 智能体开始学习
for index in list(data_replay.index):
rank_net.zero_grad() # 梯度清零
# 将测试用例数据转化为tensor
data_input = torch.tensor(data_replay.loc[index,
data_list_learn],
dtype=torch.float,
requires_grad=True)
# 计算测试用例的奖励值
reward = reward_F.reward_function(
data_replay.loc[index, 'Verdict'],
function_name='tc_reward')
reward = torch.tensor(reward,
dtype=torch.float,
requires_grad=False)
data_output = rank_net(data_input)
if train_setting == 'DNN':
# DNN:Q = Q + alpha * r
data_target = data_output.clone().detach(
) + alpha * reward
elif train_setting == 'DQN':
# DQN:Q = Q + alpha * （r + gamma * Q' - Q）
data_target = data_output.clone().detach() + \
alpha * (reward + gamma * rank_net_target(data_input) - data_output)
else:
pass
loss = criterion(data_output, data_target)
loss.backward()
optimizer.step()
# 智能体学习结束
# 如果训练类型为DQN那么更新目标网络
if train_setting == 'DNN':
break # 仅训练一次
elif train_setting == 'DQN':
rank_net_target = copy.deepcopy(rank_net)
else:
pass
print("第" + str(stage) + "阶段学习完成,awa~")
# 如果需要保存训练参数,那么保存训练后
if save_model:
torch.save(rank_net.state_dict(), PATH)
time_end = time.clock()
print("程序终止运行时间为",
time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
print('程序耗时', time_end - time_start)
return time_end - time_start
