def super_dqn_train(configs, time_data, sumoCmd):
from Agent.super_dqn import Trainer
if configs['model'] == 'base':
from Env.MultiEnv import GridEnv
# elif configs['model'] == 'frap':
# from Env.FRAP import TL3x3Env # will be added
side_list = ['u', 'r', 'd', 'l']
tl_rl_list = list()
for _, node in enumerate(configs['node_info']):
if node['id'][-1] not in side_list:
tl_rl_list.append(node['id'])
configs['tl_rl_list'] = tl_rl_list
NUM_EPOCHS = configs['num_epochs']
MAX_STEPS = configs['max_steps']
# init agent and tensorboard writer
agent = Trainer(configs)
writer = SummaryWriter(os.path.join(
configs['current_path'], 'training_data', time_data))
# save hyper parameters
agent.save_params(time_data)
# init training
epoch = 0
while epoch < NUM_EPOCHS:
traci.start(sumoCmd)
for tl_rl in tl_rl_list:
traci.trafficlight.setRedYellowGreenState(tl_rl, 'G{0}{3}rr{2}{3}rG{0}{3}rr{2}{3}r'.format(
'G'*configs['num_lanes'], 'G', 'r'*configs['num_lanes'], 'r'))
before_action= torch.ones((1,len(tl_rl_list),1),dtype=torch.int)
env = GridEnv(configs)
step = 0
done = False
# state initialization
# agent setting
total_reward = 0
reward = 0
arrived_vehicles = 0
state = env.get_state()
action_distribution = tuple()
a = time.time()
while step < MAX_STEPS:
action = agent.get_action(state)
action_distribution += tuple(action.unsqueeze(1))
# action 을 정하고
# action이 before_actio과 같으면 yellow없이 진행하고
for tl_rl in tl_rl_list:
idx=tl_rl_list.index(tl_rl)
if before_action[0][idx]==action[0][idx]:
traci.trafficlight.setRedYellowGreenState(
tl_rl, 'y'*(3+configs['num_lanes'])*4)
arrived_vehicles += simulation_step(env, 5)
step+=5
# environment에 적용
env.step(action) # action 적용함수
#적용 후 20초 진행
arrived_vehicles += simulation_step(env, 20)
step+=20
next_state = env.get_state() # 다음스테이트
reward = env.get_reward() # 20초 지연된 보상
agent.save_replay(state, action, reward, next_state) # dqn
agent.update(done)
state = next_state
total_reward += reward
before_action=action
# 20초 끝나고 yellow 4초
agent.update_hyperparams(epoch) # lr and epsilon upate
if epoch % 2 == 0:
agent.target_update() # dqn
b = time.time()
traci.close()
print("time:", b-a)
epoch += 1
# once in an epoch
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
print('======== {} epoch/ return: {} arrived number:{}'.format(epoch,
total_reward.sum(), arrived_vehicles))
if epoch % 50 == 0:
agent.save_weights(
configs['file_name']+'_{}_{}'.format(time_data, epoch))
writer.close()
def dqn_train(configs, time_data, sumoCmd):
# Environment Setting
from Agent.dqn import Trainer
if configs['model'] == 'base':
from Env.Env import TL3x3Env
elif configs['model'] == 'frap':
from Env.FRAP import TL3x3Env
# EXP_CONFIG Setting
NUM_EPOCHS = configs['num_epochs']
MAX_STEPS = configs['max_steps']
tl_rl_list = configs['tl_rl_list']
epoch = 0
# init agent and tensorboard writer
# agent setting
agent = Trainer(configs)
writer = SummaryWriter(os.path.join(
configs['current_path'], 'training_data', time_data))
# save hyper parameters
agent.save_params(time_data)
# init training
while epoch < NUM_EPOCHS:
# Epoch Start
traci.start(sumoCmd)
step = 0
action_distribution = tuple()
# Epoch Start setting
env = TL3x3Env(configs)
traci.trafficlight.setRedYellowGreenState(tl_rl_list[0], 'G{0}{3}rr{2}{3}rG{0}{3}rr{2}{3}r'.format(
'G'*configs['num_lanes'], 'G', 'r'*configs['num_lanes'], 'r'))
before_action=torch.ones((1,len(tl_rl_list)))
done = False
total_reward = 0
reward = 0
arrived_vehicles = 0
# state initialization
state = env.get_state()
# Time Check
a = time.time()
while step < MAX_STEPS:
action = agent.get_action(state)
action_distribution += tuple(action.unsqueeze(1))
# action 을 정하고
# action이 before_actio과 같으면 yellow없이 진행하고
if before_action!=action:
traci.trafficlight.setRedYellowGreenState(
tl_rl_list[0], 'y'*28)
arrived_vehicles += simulation_step(env, 5)
step+=5
# environment에 적용
env.step(action) # action 적용함수
#적용 후 20초 진행
arrived_vehicles += simulation_step(env, 20)
step+=20
next_state = env.get_state() # 다음스테이트
reward = env.get_reward() # 20초 지연된 보상
agent.save_replay(state, action, reward, next_state) # dqn
agent.update(done)
state = next_state
total_reward += reward
before_action=action
# 20초 끝나고 yellow 4초
b = time.time()
traci.close()
print("time:", b-a)
epoch += 1
# update hyper parameter
agent.update_hyperparams(epoch) # lr and epsilon upate
if epoch % agent.configs['target_update_period'] == 0:
agent.target_update() # dqn
# once in an epoch update tensorboard
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
print('======== {} epoch/ return: {} arrived number:{}'.format(epoch,
total_reward, arrived_vehicles))
if epoch % 50 == 0:
agent.save_weights(
configs['file_name']+'_{}_{}'.format(time_data, epoch))
writer.close()
val_psnr = 0
val_ssim = 0
for i in range(len(valid_hr_imgs)):
hr = valid_hr_imgs[i]
lr = valid_lr_imgs[i]
[lr, hr] = utils.datatype([lr, hr])
hr_expand = np.expand_dims(hr, axis=0)
lr_expand = np.expand_dims(lr, axis=0)
psnr, ssim, sr_expand = sess.run([PSNR, SSIM, tensor_sr], {
tensor_lr: lr_expand,
tensor_hr: hr_expand
})
sr = np.squeeze(sr_expand)
utils.update_tensorboard(epoch, writer, i, lr, sr, hr)
val_psnr += psnr
val_ssim += ssim
val_psnr = val_psnr / len(valid_hr_imgs)
val_ssim = val_ssim / len(valid_hr_imgs)
if val_psnr > best_psnr:
best_psnr = val_psnr
best_epoch = epoch
print('Saving new best model')
## save model
saver.save(sess, os.path.join(checkpoint_dir, 'model.ckpt'))
writer.add_scalar('Validate PSNR', val_psnr, epoch)
writer.add_scalar('Validate SSIM', val_ssim, epoch)
def dqn_train(configs, time_data, sumoCmd):
from Agent.dqn import Trainer
if configs['model'] == 'base':
from Env.Env import TL3x3Env
elif configs['model'] == 'frap':
from Env.FRAP import TL3x3Env
NUM_EPOCHS = configs['num_epochs']
MAX_STEPS = configs['max_steps']
tl_rl_list = configs['tl_rl_list']
# init agent and tensorboard writer
writer = SummaryWriter(
os.path.join(configs['current_path'], 'training_data', time_data))
agent = Trainer(configs)
# save hyper parameters
agent.save_params(time_data)
# init training
epoch = 0
while epoch < NUM_EPOCHS:
traci.start(sumoCmd)
env = TL3x3Env(configs)
traci.trafficlight.setRedYellowGreenState(
tl_rl_list[0], 'G{0}{3}rr{2}{3}rG{0}{3}rr{2}{3}r'.format(
'G' * configs['num_lanes'], 'G', 'r' * configs['num_lanes'],
'r'))
step = 0
done = False
# state initialization
# agent setting
total_reward = 0
reward = 0
arrived_vehicles = 0
state = env.get_state()
action_distribution = tuple()
a = time.time()
while step < MAX_STEPS:
'''
# state=env.get_state(action) #partial하게는 env에서 조정
action=agent.get_action(state)
env.step(action)
reward=env.get_reward()
next_state=env.get_state()
# if traci.inductionloop.getLastStepVehicleNumber("0") > 0:
store transition in D (experience replay)
Sample random minibatch from D
step += 1
state=next_state
set yi
'''
action = agent.get_action(state)
action_distribution += tuple(action.unsqueeze(1))
env.step(action) # action 적용함수
for _ in range(20): # 10초마다 행동 갱신
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
next_state = env.get_state() # 다음스테이트
traci.trafficlight.setRedYellowGreenState(tl_rl_list[0], 'y' * 28)
for _ in range(5): # 4번더
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
reward = env.get_reward() # 25초 지연된 보상
agent.save_replay(state, action, reward, next_state) # dqn
agent.update(done)
state = next_state
total_reward += reward
# 20초 끝나고 yellow 4초
agent.update_hyperparams(epoch) # lr and epsilon upate
if epoch % 2 == 0:
agent.target_update() # dqn
b = time.time()
traci.close()
print("time:", b - a)
epoch += 1
# once in an epoch
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
print('======== {} epoch/ return: {} arrived number:{}'.format(
epoch, total_reward, arrived_vehicles))
if epoch % 50 == 0:
agent.save_weights(configs['file_name'] +
'_{}_{}'.format(time_data, epoch))
writer.close()
def ppo_train(configs, time_data, sumoCmd):
from Agent.ppo import Trainer
if configs['model'] == 'base':
from Env.Env import TL3x3Env
elif configs['model'] == 'frap':
from Env.FRAP import TL3x3Env
tl_rl_list = configs['tl_rl_list']
NUM_EPOCHS = configs['num_epochs']
MAX_STEPS = configs['max_steps']
# init agent and tensorboard writer
agent = Trainer(configs)
writer = SummaryWriter(
os.path.join(configs['current_path'], 'training_data', time_data))
# save hyper parameters
agent.save_params(time_data)
# init training
epoch = 0
ppo_update_step = 0
while epoch < NUM_EPOCHS:
traci.start(sumoCmd)
traci.trafficlight.setRedYellowGreenState(
tl_rl_list[0], 'G{0}{3}rr{2}{3}rG{0}{3}rr{2}{3}r'.format(
'G' * configs['num_lanes'], 'G', 'r' * configs['num_lanes'],
'r'))
env = TL3x3Env(configs)
# env = GridEnv( configs)
step = 0
done = False
# state initialization
# agent setting
total_reward = 0
reward = 0
arrived_vehicles = 0
state = env.get_state()
action_distribution = tuple()
a = time.time()
while step < MAX_STEPS:
action = agent.get_action(state)
action_distribution += tuple(action.unsqueeze(1))
env.step(action) # action 적용함수
ppo_update_step += 1
for _ in range(20): # 10초마다 행동 갱신
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
next_state = env.get_state() # 다음스테이트
traci.trafficlight.setRedYellowGreenState(tl_rl_list[0], 'y' * 28)
for _ in range(5): # 4번더
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
reward = env.get_reward() # 25초 지연된 보상
agent.memory.rewards.append(reward)
if step >= MAX_STEPS:
done = True
agent.memory.dones.append(done)
state = next_state
total_reward += reward
if ppo_update_step % 400 == 0:
agent.update()
agent.update_hyperparams(epoch) # lr update
ppo_update_step = 0
b = time.time()
traci.close()
print("time:", b - a)
epoch += 1
# once in an epoch
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
print('======== {} epoch/ return: {} arrived number:{}'.format(
epoch, total_reward, arrived_vehicles))
if epoch % 50 == 0:
agent.save_weights(configs['file_name'] +
'_{}_{}'.format(time_data, epoch))
writer.close()
def REINFORCE_train(configs, time_data, sumoCmd):
from Agent.REINFORCE import Trainer
from Agent.REINFORCE import DEFAULT_CONFIG
from Env.Env import TL3x3Env
tl_rl_list = configs['tl_rl_list']
NUM_EPOCHS = configs['num_epochs']
MAX_STEPS = configs['max_steps']
# init agent and tensorboard writer
agent = Trainer(configs)
writer = SummaryWriter(
os.path.join(configs['current_path'], 'training_data', time_data))
# save hyper parameters
agent.save_params(time_data)
# init training
epoch = 0
while epoch < NUM_EPOCHS:
traci.start(sumoCmd)
traci.trafficlight.setRedYellowGreenState(
tl_rl_list[0], 'G{0}{3}rr{2}{3}rG{0}{3}rr{2}{3}r'.format(
'G' * configs['num_lanes'], 'G', 'r' * configs['num_lanes'],
'r'))
env = TL3x3Env(configs)
# env = GridEnv( configs)
step = 0
done = False
# state initialization
# agent setting
total_reward = 0
reward = 0
arrived_vehicles = 0
state = env.get_state()
while step < MAX_STEPS:
action = agent.get_action(state)
env.step(action) # action 적용함수
next_state = env.get_state()
for _ in range(20): # 10초마다 행동 갱신
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
# 20초 끝나고 yellow 4초
traci.trafficlight.setRedYellowGreenState(tl_rl_list[0], 'y' * 28)
for _ in range(5): # 4번더
traci.simulationStep()
env.collect_state()
step += 1
arrived_vehicles += traci.simulation.getArrivedNumber(
) # throughput
reward = env.get_reward()
prob = agent.get_prob()
agent.put_data((reward, prob[action]))
state = next_state
total_reward += reward
if step > MAX_STEPS:
done = True
agent.update(done)
agent.update_hyperparams(epoch) # lr and epsilon upate
traci.close()
epoch += 1
# once in an epoch
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
print('======== {} epoch/ return: {} arrived number:{}'.format(
epoch, total_reward, arrived_vehicles))
writer.close()
def city_dqn_train(configs, time_data, sumoCmd):
from Agent.super_dqn import Trainer
if configs['model'] == 'city':
from Env.CityEnv import CityEnv
phase_num_matrix = torch.tensor( # 각 tl이 갖는 최대 phase갯수
[len(configs['traffic_node_info'][index]['phase_duration']) for _, index in enumerate(configs['traffic_node_info'])])
# init agent and tensorboard writer
writer = SummaryWriter(os.path.join(
configs['current_path'], 'training_data', time_data))
agent = Trainer(configs)
# save hyper parameters
agent.save_params(time_data)
# init training
NUM_AGENT = configs['num_agent']
DEVICE = configs['device']
TL_RL_LIST = configs['tl_rl_list']
MAX_PHASES = configs['max_phase_num']
MAX_STEPS = configs['max_steps']
OFFSET = torch.tensor(configs['offset'], # i*10
device=DEVICE, dtype=torch.int)
TL_PERIOD = torch.tensor(
configs['tl_period'], device=DEVICE, dtype=torch.int)
epoch = 0
while epoch < configs['num_epochs']:
step = 0
if configs['randomness'] == True:
tmp_sumoCmd = sumoCmd+['--scale', str(1.5+random())] # 1.5~2.5
else:
if configs['network'] == 'dunsan' or 'grid' in configs['network']:
tmp_sumoCmd = sumoCmd+['--scale', str(configs['scale'])]
else:
tmp_sumoCmd = sumoCmd
traci.start(tmp_sumoCmd)
env = CityEnv(configs)
# Total Initialization
actions = torch.zeros(
(NUM_AGENT, configs['action_size']), dtype=torch.int, device=DEVICE)
# Mask Matrix : TL_Period가 끝나면 True
mask_matrix = torch.zeros(
(NUM_AGENT), dtype=torch.bool, device=DEVICE)
# MAX Period까지만 증가하는 t
t_agent = torch.zeros(
(NUM_AGENT), dtype=torch.int, device=DEVICE)
t_agent -= OFFSET
# Action configs['offset']on Matrix : 비교해서 동일할 때 collect_state, 없는 state는 zero padding
action_matrix = torch.zeros(
(NUM_AGENT, MAX_PHASES), dtype=torch.int, device=DEVICE) # 노란불 3초 해줘야됨
action_index_matrix = torch.zeros(
(NUM_AGENT), dtype=torch.long, device=DEVICE) # 현재 몇번째 phase인지
action_update_mask = torch.eq( # action이 지금 update해야되는지 확인
t_agent, action_matrix[0, action_index_matrix]).view(NUM_AGENT) # 0,인 이유는 인덱싱
# 최대에 도달하면 0으로 초기화 (offset과 비교)
clear_matrix = torch.eq(t_agent % TL_PERIOD, 0)
t_agent[clear_matrix] = 0
# action 넘어가야된다면 action index증가 (by tensor slicing)
action_index_matrix[action_update_mask] += 1
action_index_matrix[clear_matrix] = 0
# mask update, matrix True로 전환
mask_matrix[clear_matrix] = True
mask_matrix[~clear_matrix] = False
# state initialization
state = env.collect_state(
action_update_mask, action_index_matrix, mask_matrix)
total_reward = 0
# agent setting
arrived_vehicles = 0
a = time.time()
while step < MAX_STEPS:
# action 을 정하고
actions = agent.get_action(state, mask_matrix)
if mask_matrix.sum()>0:
print(actions.transpose(1,2))
# action형태로 변환 # 다음으로 넘어가야할 시점에 대한 matrix
action_matrix = env.calc_action(
action_matrix, actions, mask_matrix)
# 누적값으로 나타남
# environment에 적용
# action 적용함수, traci.simulationStep 있음
env.step(
actions, mask_matrix, action_index_matrix, action_update_mask)
# 전체 1초증가 # traci는 env.step에
step += 1
t_agent += 1
# 최대에 도달하면 0으로 초기화 (offset과 비교)
clear_matrix = torch.eq(t_agent % TL_PERIOD, 0)
# action 넘어가야된다면 action index증가 (by tensor slicing)
for idx,_ in enumerate(TL_RL_LIST):
action_update_mask[idx] = torch.eq( # update는 단순히 진짜 현시만 받아서 결정해야됨
t_agent[idx], action_matrix[idx, action_index_matrix[idx]].view(-1)) # 0,인 이유는 인덱싱
action_index_matrix[action_update_mask] += 1
# agent의 최대 phase를 넘어가면 해당 agent의 action index 0으로 초기화
action_index_matrix[clear_matrix] = 0
# mask update, matrix True로 전환
t_agent[clear_matrix] = 0
# print(t_agent,action_index_matrix,step,action_update_mask)
mask_matrix[clear_matrix] = True
mask_matrix[~clear_matrix] = False
next_state = env.collect_state(
action_update_mask, action_index_matrix, mask_matrix)
# env속에 agent별 state를 꺼내옴, max_offset+period 이상일 때 시작
if step >= int(torch.max(OFFSET)+torch.max(TL_PERIOD)) and mask_matrix.sum() > 0:
rep_state, rep_action, rep_reward, rep_next_state = env.get_state(
mask_matrix)
agent.save_replay(rep_state, rep_action, rep_reward,
rep_next_state, mask_matrix) # dqn
# update
agent.update(mask_matrix)
state = next_state
# info
arrived_vehicles += traci.simulation.getArrivedNumber()
agent.target_update(epoch)
agent.update_hyperparams(epoch) # lr and epsilon upate
b = time.time()
traci.close()
print("time:", b-a)
epoch += 1
# once in an epoch
print('======== {} epoch/ return: {:.5f} arrived number:{}'.format(epoch,
env.cum_reward.sum(), arrived_vehicles))
update_tensorboard(writer, epoch, env, agent, arrived_vehicles)
env.test_val=0
if epoch % 50 == 0:
agent.save_weights(
configs['file_name']+'_{}'.format(epoch))
writer.close()
