def eval_performance(self, model):
print 'start evaluating...'
Agent = dqn_agent_nature.dqn_agent(
gpu_id=self.gpu_id,
state_dimention=1,
enable_controller=self.enable_controller)
Agent.agent_init()
Agent.DQN.model = model
Agent.DQN.model_to_gpu()
Agent.policyFrozen = True
profit_list = []
test_profit_list = []
files = os.listdir(self.target_folder)
#train term evaluation
print 'start evaluation train term...'
for f in files:
stock_agent = env_stockmarket.Stock_agent(Agent,
self.action_split_number)
try:
traindata, trainprice = self.market.get_trainData(
f, self.input_num)
except:
continue
profit_ratio = stock_agent.trading(self.input_num, trainprice,
traindata)
profit_list.append(profit_ratio)
train_ave = np.mean(np.array(profit_list))
train_ave_Q = Agent.get_average_Q()
train_ave_reward = Agent.get_average_reward()
#test term evaluation
print 'start evaluation test term...'
for f in files:
stock_agent = env_stockmarket.Stock_agent(Agent,
self.action_split_number)
try:
traindata, trainprice = self.market.get_testData(
f, self.input_num)
except:
continue
profit_ratio = stock_agent.trading(self.input_num, trainprice,
traindata)
test_profit_list.append(profit_ratio)
test_ave = np.mean(np.array(test_profit_list))
self.train_ave_profit_list.append(train_ave)
self.test_ave_profit_list.append(test_ave)
self.ave_Q_list.append(train_ave_Q)
self.ave_reward_list.append(train_ave_reward)
print 'finish evaluation'
return train_ave, test_ave, train_ave_Q, train_ave_reward
files = os.listdir("./nikkei100")
Agent.init_max_Q_list()
Agent.init_reward_list()
profit_list = []
for f in files:
print f
if args.online_update == 1:
#銘柄ごとに初期化
Agent.DQN.model = copy.deepcopy(org_model)
Agent.DQN.model_target = copy.deepcopy(org_model)
Agent.DQN.reset_optimizer()
stock_agent = env_stockmarket.Stock_agent(Agent, args.action_split_number)
try:
testdata, testprice = market.get_testData(f, args.input_num)
#testdata, testprice = market.get_trainData(f,END_TRAIN_DAY,args.input_num)
except:
print 'skip', f
continue
profit_ratio, proper, order, stocks, price, Q_list, ave_buyprice_list, reward_list = stock_agent.trading_test(
args.input_num, testprice, testdata)
profit_list.append(profit_ratio)
tools.listToCsv(folder + str(f).replace(".CSV", "") + '.csv', price,
proper, order, stocks, ave_buyprice_list, reward_list)
o.write('targetFlag:' + str(targetFlag) + '\n')
files = os.listdir(args.data_folder)
for epoch in tqdm(range(1,n_epoch + 1)):
Agent.init_max_Q_list()
Agent.init_reward_list()
#ファイルの順をシャッフル
random.shuffle(files)
#train_loop
Agent.policyFrozen = False
for f in tqdm(files):
stock_agent = env_stockmarket.Stock_agent(Agent)
try:
traindata,trainprice = market.get_trainData(f,args.input_num)
except:
continue
profit_ratio = stock_agent.trading(args.input_num,trainprice,traindata)
#model evaluation
eval_model = Agent.DQN.get_model_copy()
evaluater.eval_performance(eval_model)
evaluater.get_epsilon(Agent.epsilon)
evaluater.save_eval_result()
if epoch % 1 == 0: