def train():
best_reward, best_reward_timesteps = None, None
save_path = "model_save/"+MODEL_PATH+"/"
if save_path is not None:
os.makedirs(save_path, exist_ok=True)
# log_dir = f"model_save/"
log_dir = save_path
env, env_eval = ENV(util='train', par=PARAM, dt=DT), ENV(util='val', par=PARAM, dt=DT)
env, env_eval = Monitor(env, log_dir), Monitor(env_eval, log_dir)
env, env_eval = DummyVecEnv([lambda: env]), DummyVecEnv([lambda: env_eval])
# env = VecNormalize(env, norm_obs=True, norm_reward=True,
# clip_obs=10.)
if PARAM['algo']=='td3':
model = TD3('MlpPolicy', env, verbose=1, batch_size=PARAM['batch_size'], seed=PARAM['seed'],
learning_starts=PARAM['learning_starts'])
elif PARAM['algo']=='ddpg':
model = DDPG('MlpPolicy', env, verbose=1, batch_size=PARAM['batch_size'], seed=PARAM['seed'],
learning_starts=PARAM['learning_starts'])
elif PARAM['algo']=='ppo':
model = PPO('MlpPolicy', env, verbose=1, batch_size=PARAM['batch_size'], seed=PARAM['seed'])
eval_callback = EvalCallback(env_eval, best_model_save_path=save_path+MODEL_PATH+'_best_model',
log_path=log_dir, eval_freq=PARAM['eval_freq'], save_freq=PARAM['save_freq'],
deterministic=True, render=False)
model.learn(total_timesteps=int(PARAM['total_time_step']), callback=eval_callback, log_interval = 500)
print("best mean reward:", eval_callback.best_mean_reward_overall, "timesteps:", eval_callback.best_mean_reward_timestep)
model.save(save_path+MODEL_PATH+'_final_timesteps')
def test(MODEL_TEST):
log_dir = "model_save/" + MODEL_PATH + "/" + MODEL_PATH + MODEL_TEST
env = ENV(util='test', par=PARAM, dt=DT)
env.render = True
env = Monitor(env, log_dir)
if PARAM['algo']=='td3':
model = TD3.load(log_dir)
elif PARAM['algo']=='ddpg':
model = DDPG.load(log_dir)
elif PARAM['algo']=='ppo':
model = PPO.load(log_dir)
# plot_results(f"model_save/")
trade_dt = pd.DataFrame([]) # trade_dt: 所有股票的交易数据
result_dt = pd.DataFrame([]) # result_dt: 所有股票一年测试结果数据
for i in range(TEST_STOCK_NUM):
state = env.reset()
stock_bh_id = 'stock_bh_'+str(i) # 记录每个股票交易的buy_hold
stock_port_id = 'stock_port_'+str(i) # 记录每个股票交易的portfolio
stock_action_id = 'stock_action_' + str(i) # 记录每个股票交易的action
flow_L_id = 'stock_flow_' + str(i) # 记录每个股票的流水
stock_bh_dt, stock_port_dt, action_policy_dt, flow_L_dt = [], [], [], []
day = 0
while True:
action = model.predict(state)
next_state, reward, done, info = env.step(action[0])
state = next_state
# print("trying:",day,"reward:", reward,"now profit:",env.profit) # 测试每一步的交易policy
stock_bh_dt.append(env.buy_hold)
stock_port_dt.append(env.Portfolio_unit)
action_policy_dt.append(action[0][0]) # 用于记录policy
flow_L_dt.append(env.flow)
day+=1
if done:
print('stock: {}, total profit: {:.2f}%, buy hold: {:.2f}%, sp: {:.4f}, mdd: {:.2f}%, romad: {:.4f}'
.format(i, env.profit*100, env.buy_hold*100, env.sp, env.mdd*100, env.romad))
# 交易完后记录：股票ID，利润（单位100%），buy_hold（单位100%），夏普率，最大回撤率（单位100%），romad
result=pd.DataFrame([[i,env.profit*100,env.buy_hold*100,env.sp,env.mdd*100,env.romad]])
break
trade_dt_stock = pd.DataFrame({stock_port_id: stock_port_dt,
stock_bh_id: stock_bh_dt,
stock_action_id: action_policy_dt,
flow_L_id: flow_L_dt}) # 支股票的交易数据
trade_dt = pd.concat([trade_dt, trade_dt_stock], axis=1) # 所有股票交易数据合并（加行）
result_dt = pd.concat([result_dt,result],axis=0) # 所有股票结果数据合并（加列）
result_dt.columns = ['stock_id','prfit(100%)','buy_hold(100%)','sp','mdd(100%)','romad']
trade_dt.to_csv('out_dt/trade_'+MODEL_PATH+'.csv',index=False)
result_dt.to_csv('out_dt/result_'+MODEL_PATH+'.csv',index=False)
def train():
log_dir = f"model_save/"
env = ENV(istest=False)
env = Monitor(env, log_dir)
env = DummyVecEnv([lambda: env])
# env = VecNormalize(env, norm_obs=True, norm_reward=True,
# clip_obs=10.)
model = PPO('MlpPolicy',
env,
verbose=1,
batch_size=PARAM['batch_size'],
seed=PARAM['seed'])
callback = SaveOnBestTrainingRewardCallback(check_freq=480,
log_dir=log_dir)
model.learn(total_timesteps=int(PARAM['total_time_step']),
callback=callback,
log_interval=480)
model.save('model_save/' + MODEL_PATH)