def run(self):
ep = 0
while self.g_ep.value < 100:
# total_step = 1
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = [0. for i in range(self.agent_num)]
for step in range(1000):
# print(ep)
# if self.name == 'w00' and self.g_ep.value%10 == 0:
# path = "/Users/xue/Desktop/temp/temp%d"%self.g_ep.value
# if not os.path.exists(path):
# os.mkdir(path)
# self.env.render(path)
s0 = s[0]
a0, prob0 = self.lnet[0].choose_action(v_wrap(s0[None, :]), True)
a0 = [a0]
s = [np.concatenate((s[i],np.array(a0)),-1) for i in range(1, self.agent_num)]
s = [s0] + s
a = [self.lnet[i].choose_action(v_wrap(s[i][None, :]), True) for i in range(1, self.agent_num)]
prob = [elem[1] for elem in a]
a = a0 + [elem[0] for elem in a]
s_, r, done, _ = self.env.step(a,need_argmax=False)
# print(a)
# if done[0]: r = -1
ep_r = [ep_r[i] + r[i] for i in range(self.agent_num)]
x = self._influencer_reward(r[0], self.lnet[1:], prob0, a0, s[1:], prob)
r = [float(i) for i in r]
r[0] += x.numpy()
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if step % 5 == 0 and step != 0: # update global and assign to local net
_s0 = s_[0]
a0 = self.lnet[0].choose_action(v_wrap(_s0[None, :]), False)
a0 = [a0]
_s = [np.concatenate((s_[i], np.array(a0)), -1) for i in range(1, self.agent_num)]
_s = [_s0] + _s
# sync
done = [False for i in range(self.agent_num)]
[push_and_pull(self.opt[i], self.lnet[i], self.gnet[i], done[i],
_s[i], buffer_s, buffer_a, buffer_r, self.GAMMA, i)
for i in range(self.agent_num)]
[self.scheduler_lr[i].step() for i in range(self.agent_num)]
buffer_s, buffer_a, buffer_r = [], [], []
# if ep == 999: # done and print information
# record(self.g_ep, self.g_ep_r, sum(ep_r), self.res_queue, self.name)
# break
s = s_
# total_step += 1
print('ep%d'%ep, self.name, sum(ep_r))
ep+=1
if self.name == "w00":
self.sender.send([sum(ep_r),ep])
self.res_queue.put(None)
def expand_dispatch(self,dis = 1):
job = copy.deepcopy(self)
s = self.state
while(True):
a = v_wrap(np.repeat(dis,net.choose_action(v_wrap(s)).size))
s_,r,done = self.step(a.numpy())
s = s_
if done :
# print(self.total_time)
break
return job ,self.total_time
def expand(self, net):
job = copy.deepcopy(self)
s = self.state
while (True):
a = v_wrap(net.choose_action(v_wrap(s)))
s_, r, done = self.step(a.numpy())
s = s_
if done:
# print(self.total_time)
break
return job, self.total_time
def run(self):
total_step = 1
while self.g_ep.value < self.args.MAXEPS:
# print(self.g_ep.value)
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0
for t in range(self.args.MAXSTEP):
if self.name == 'worker0':
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
s_, r, done, _ = self.env.step(a.clip(-2, 2))
if t == self.args.MAXSTEP - 1:
done = True
ep_r += r
buffer_s.append(s)
buffer_a.append(a)
buffer_r.append((r + 8.1) / 8.1)
if total_step % self.args.updateperiod == 0 or done:
# print(total_step)
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_a, buffer_r,
self.args.gamma)
buffer_s, buffer_a, buffer_r = [], [], []
if done:
print('*')
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue,
self.name)
break
s = s_
total_step += 1
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
while True:
if self.name == 'w00':
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
s_, r, done, _ = self.env.step(a)
if done: r = -1
ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if total_step % UPDATE_GLOBAL_ITER == 0 or done:
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_a, buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done:
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue,
self.name)
break
s = s_
total_step += 1
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
for t in range(MAX_EP_STEP):
if self.name == 'w0':
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
s_, r, done, _ = self.env.step(a.clip(-2, 2))
if t == MAX_EP_STEP - 1:
done = True
ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append((r + 8.1) / 8.1) # normalize
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
# sync
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_a, buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done: # done and print information
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue,
self.name)
break
s = s_
total_step += 1
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
s = self.env.reset()
buffer_s, buffer_action, buffer_reward = [], [], []
ep_r = 0.
while True:
if self.name == 'w00':
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
s_, r, done, _ = self.env.step(a)
if done: r = -1
ep_r += r
buffer_action.append(a)
buffer_s.append(s)
buffer_reward.append(r)
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
# sync
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_action, buffer_reward,
GAMMA)
buffer_s, buffer_action, buffer_reward = [], [], []
if done: # done and print information
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue,
self.name)
break
s = s_
total_step += 1
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
try:
mapID = np.random.randint(
1,
6) #Choosing a map ID from 5 maps in Maps folder randomly
posID_x = np.random.randint(
MAP_MAX_X
) #Choosing a initial position of the DQN agent on X-axes randomly
posID_y = np.random.randint(
MAP_MAX_Y
) #Choosing a initial position of the DQN agent on Y-axes randomly
#Creating a request for initializing a map, initial position, the initial energy, and the maximum number of steps of the DQN agent
request = ("map" + str(mapID) + "," + str(posID_x) + "," +
str(posID_y) + ",50,100")
#Send the request to the game environment (GAME_SOCKET_DUMMY.py)
self.env.send_map_info(request)
self.env.reset()
s = self.env.get_state()
print(s.shape)
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
for t in range(MAX_EP_STEP):
a = self.lnet.choose_action(v_wrap(s[None, :]), .5)
self.env.step(str(a))
s_ = self.env.get_state()
r = self.env.get_reward()
done = self.env.check_terminate()
if t == MAX_EP_STEP - 1:
done = True
ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
# sync
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_a, buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done: # done and print information
record(self.g_ep, self.g_ep_r, ep_r,
self.res_queue, self.name)
break
s = s_
total_step += 1
except Exception as e:
import traceback
traceback.print_exc()
break
self.res_queue.put(None)
def _influencer_reward(self, e, nets, prob0, a0, s, p_a):
a_cf = []
for i in range(self.action_dim):
if i != a0[0]:
a_cf.append(i)
p_cf = []
s_cf = np.array([[np.concatenate((s[i][ :-1],np.array([a_cf[j]])),-1) for j in range(self.action_dim-1)] for i in range(self.agent_num-1)])
for i in range(len(nets)):
# _a = [nets[i].choose_action(v_wrap(s_cf[i][None, :]), True)[1] for i in range(self.agent_num-1)]
# temp = nets[i].choose_action(v_wrap(s_cf[i][None, :]), True)[1][0]
_a = [torch.mul(nets[i].choose_action(v_wrap(s_cf[i][None, :]), True)[1][0], prob0)]
# _a =
_a = torch.sum(_a[0],-2)
x = p_a[i][0]
y = _a.detach()
p_cf.append(torch.nn.functional.kl_div(torch.log(x),y,reduction="sum"))
# l = scipy.stats.entropy(x.numpy(), y.numpy())
return e + 50*self._sum(p_cf)/len(p_cf)
def run(self):
ep = 0
while self.g_ep.value < 100:
# total_step = 1
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = [0. for i in range(self.agent_num)]
for step in range(1000):
# print(ep)
# if self.name == 'w00' and self.g_ep.value%10 == 0:
# path = "/Users/xue/Desktop/temp/temp%d"%self.g_ep.value
# if not os.path.exists(path):
# os.mkdir(path)
# self.env.render(path)
a = [self.lnet[i].choose_action(v_wrap(s[i][None, :])) for i in range(self.agent_num)]
s_, r, done, _ = self.env.step(a, need_argmax=False)
# print(a)
# if done[0]: r = -1
ep_r = [ep_r[i] + r[i] for i in range(self.agent_num)]
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if step % 5 == 0: # update global and assign to local net
# sync
done = [False for i in range(self.agent_num)]
[push_and_pull(self.opt[i], self.lnet[i], self.gnet[i], done[i],
s_[i], buffer_s, buffer_a, buffer_r, self.GAMMA, i)
for i in range(self.agent_num)]
[self.scheduler_lr[i].step() for i in range(self.agent_num)]
buffer_s, buffer_a, buffer_r = [], [], []
# if ep == 999: # done and print information
# record(self.g_ep, self.g_ep_r, sum(ep_r), self.res_queue, self.name)
# break
s = s_
# total_step += 1
print('ep%d' % ep, self.name, sum(ep_r))
ep += 1
if self.name == "w00":
self.sender.send([sum(ep_r), ep])
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
s = np.transpose(self.env.reset(), (2, 0, 1)) / 255.0
lives = self.lives_sum
buffer_s, buffer_a, buffer_r = [], [], []
self.ep_r = 0.
actions = []
while True:
total_step += 1
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
# a = np.random.randint(low = 0, high = 8)
actions.append(str(a))
s_, r, done, info = self.env.step(a)
s_ = np.transpose(s_, (2, 0, 1)) / 255.0
livesLeft = info[
'ale.lives'] # punish everytime the agent loses life
if livesLeft != lives:
r = DIE_PENALTY
lives = livesLeft
self.ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
# sync
# if self.name == 'w0':
# self.env.render()
push_and_pull(self.opt, self.lnet, self.gnet, done, s_,
buffer_s, buffer_a, buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done: # done and print information
record(self.g_ep, self.ep_r, self.res_queue, self.name,
self.lives_sum, DIE_PENALTY)
break
s = s_
self.res_queue.put(None)
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
s = self.env.reset()
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
while True:
if self.name == 'w0':
self.env.render()
a = self.lnet.choose_action(v_wrap(s[None, :]))
action = np.zeros((N_A))
action[a] = 1
# print(self.name,action)
s_, r, done, _ = self.env.step(action)
#if done: r = -1
ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if done: # update global and assign to local net
break
s = s_
total_step += 1
def run(self):
total_step = 1
while self.g_ep.value < MAX_EP:
# get video -----------------------------
while(True):
video_random = random.random()
videoName = ""
for i in range(len(self.videoList)):
if video_random < self.videoList[i][1]:
videoName = self.videoList[i - 1][0]
break
if videoName == "":
videoName = self.videoList[-1][0]
else:
break
# get video -----------------------------
busyList = self.get_busyTrace()
bandwidth_fileName, rtt = self.getBandwidthFile()
reqBI = self.client.init(videoName, bandwidth_fileName, rtt, self.bwType)
# mask---------------------------
mask = [1] * A_DIM
randmCachedBICount = random.randint(1, 5)
BI = [0,1,2,3,4]
randomCachedBI = random.sample(BI, randmCachedBICount)
for bIndex in range(5):
if bIndex not in randomCachedBI:
mask[bIndex] = 0
# mask---------------------------
segNum = 0
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
busy = busyList[segNum%len(busyList)]
state_ = np.zeros(S_LEN)
state = state_.copy() #state =[reqBitrate, lastBitrate, buffer, hThroughput, mThroughput, busy, mask]
reqBitrate, lastBitrate, buffer, hThroughput, mThroughput, reward, reqBI, done, segNum, busy = [0] * 10
# start one epoch **********************************
while True:
if sum(mask) == 1:
a = mask.index(1)
break
# lnet.chose_action ****************************
a, logits = self.lnet.choose_action(mask, v_wrap(state[None, :]))
# lnet.choose_action ****************************
# print --------------------------------------------
if platform.system() == "Linux":
if random.randint(0,1000) == 1:
print("reqb=", reqBitrate, "lb=", lastBitrate, "buffer=", int(buffer), "hT=", int(hThroughput), "mT=", int(mThroughput), "busy=", round(busy, 2),
"mask=",mask, "action=", a, "reqBI=", reqBI, "reward=",round(reward,2), "logits=", logits)
else:
print("reqb=", reqBitrate, "lb=", round(lastBitrate,2), "buffer=", int(buffer), "hT=", int(hThroughput), "mT=", int(mThroughput), "busy=", round(busy, 2),
"mask=",mask, "action=", a, "reqBI=", reqBI, "reward=",round(reward,2), "logits=", logits)
# print --------------------------------------------
busy = busyList[segNum % len(busyList)]
# client.run ****************************
if a == 5:
hitFlag = False
else:
hitFlag = True
reqBitrate, lastBitrate, buffer, hThroughput, mThroughput, reward, reqBI, done, segNum = self.client.run(a, busy, hitFlag)
# client.run ****************************
state_[0] = reqBitrate / BITRATES[-1]
state_[1] = lastBitrate / BITRATES[-1]
state_[2] = (buffer/1000 - 30) / 10
state_[3] = (hThroughput - throughput_mean) / throughput_std
state_[4] = (mThroughput - throughput_mean) / throughput_std
print(state)
# state_[5] = (busy - busy_mean) / busy_std
reward = reward / 5
ep_r += reward
buffer_a.append(a)
buffer_s.append(state)
buffer_r.append(reward)
if total_step % UPDATE_GLOBAL_ITER == 0 or done: # update global and assign to local net
# sync
push_and_pull(self.opt, self.lnet, self.gnet, done, state_, buffer_s, buffer_a, buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done: # done and print information
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue, self.name)
break
state = state_.copy()
total_step += 1
self.res_queue.put(None)
print("end run")
c_loss = td.pow(2) # value_loss = c_loss
probs = F.softmax(logits, dim=1)
m = self.distribution(probs)
exp_v = m.log_prob(a) * td.detach().squeeze()
a_loss = -exp_v # policy_loss = a_loss
# entropy regularization ---
log_probs = F.log_softmax(logits, dim=1)
entropy = -(log_probs * probs).sum(1)
a_loss -= 0.5 * entropy
# entropy regularization ---
total_loss = (c_loss + a_loss).mean()
return total_loss
if __name__ == '__main__':
# An instance of your model.
lnet = Net()
lnet.load_state_dict(
torch.load('../../data/RL_model/2019-06-28_10-18-56/model/233293.pkl'))
# An example input you would normally provide to your model's forward() method.
data_list = [0.33333333, 0.33333333, -0.91149055, 0.00101571, -0.19378804]
data = np.array(data_list)
print(data_list)
# Use torch.jit.trace to generate a torch.jit.ScriptModule via tracing.
traced_script_module = torch.jit.trace(lnet, v_wrap(data[None, :]))
traced_script_module.save(
"../../data/RL_model/2019-06-28_10-18-56/model.pt")
return job ,self.total_time
timee = []
enet = Net(16,5)
tnet = Net(16,5)
#net.explore = False
job = job_shop_env("la11")
s = job.state
for i in range(50):
ot = 1222
j=0
while(True):
j += 1
job,pre = job.expand(enet)
a = v_wrap(enet.choose_action(v_wrap(s)))
a.numpy()
s_,r,done = job.step(a.numpy())
if abs(pre - ot ) == 0 :
r = 100 + r
else:
r = 100/abs(pre - ot )+r
opt = SharedAdam(enet.parameters(), lr=0.000001)
_,q_t = tnet.forward(v_wrap(s))
loss = enet.loss_func(v_wrap(s),a,v_wrap(r)+0.1*q_t)
if j % 2 ==0:
tnet.load_state_dict(enet.state_dict())
opt.zero_grad()
loss.backward()
opt.step()
s = s_
if done :
# print(self.total_time)
break
return job ,self.total_time
net = Net(16,5)
#net.explore = False
job = job_shop_env(part=0)
s = job.state
for i in range(50):
ot = optt(job)
while(True):
job,pre = job.expand(net)
a = v_wrap(net.choose_action(v_wrap(s)))
a.numpy()
s_,r,done = job.step(a.numpy())
if abs(pre - ot ) == 0 :
r = 10 + r
else:
r = 10/abs(pre - ot )+r
opt = SharedAdam(net.parameters(), lr=0.00001)
loss = net.loss_func(v_wrap(s),a,v_wrap(r))
opt.zero_grad()
loss.backward()
opt.step()
s = s_
if done :
print(job.total_time)
def run(self):
ptitle('Training Agent: {}'.format(self.rank))
config = self.config
check_point_episodes = config["check_point_episodes"]
check_point_folder = os.path.join(config["check_point_folder"],
config["env"])
setup_worker_logging(self.log_queue)
self.env = create_env(config["env"], self.seed)
observation_space = self.env.observation_space
action_space = IdToAct(self.env.action_space)
with open(os.path.join("data", f"{config['env']}_action_space.npz"),
'rb') as f:
archive = np.load(f)
action_space.init_converter(all_actions=archive[archive.files[0]])
self.action_space = action_space
all_actions = np.array(action_space.all_actions)
self.local_net = Net(self.state_size, self.action_mappings,
self.action_line_mappings) # local network
self.local_net = cuda(self.gpu_id, self.local_net)
total_step = 1
l_ep = 0
while self.g_ep.value < self.num_episodes:
self.print(
f"{self.env.name} - {self.env.chronics_handler.get_name()}")
if isinstance(self.env, MultiMixEnvironment):
obs = self.env.reset(random=True)
else:
obs = self.env.reset()
maintenance_list = obs.time_next_maintenance + obs.duration_next_maintenance
s = self.convert_obs(observation_space, obs)
s = v_wrap(s[None, :])
s = cuda(self.gpu_id, s)
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
ep_step = 0
ep_agent_num_dmd = 0
ep_agent_num_acts = 0
while True:
rho = obs.rho.copy()
rho[rho == 0.0] = 1.0
lines_overload = rho > config["danger_threshold"]
expert_act = expert_rules(self.name, maintenance_list, ep_step,
action_space, obs)
if expert_act is not None:
a = np.where(all_actions == expert_act)[0][0]
choosen_actions = np.array([a])
#print(f"Expert act: {a}")
elif not np.any(lines_overload):
choosen_actions = np.array([0])
else:
lines_overload = cuda(
self.gpu_id,
torch.tensor(lines_overload.astype(int)).float())
attention = torch.matmul(lines_overload.reshape(1, -1),
self.action_line_mappings)
attention[attention > 1] = 1
choosen_actions = self.local_net.choose_action(
s, attention, self.g_num_candidate_acts.value)
ep_agent_num_dmd += 1
obs_previous = obs
a, obs_forecasted, obs_do_nothing = forecast_actions(
choosen_actions,
self.action_space,
obs,
min_threshold=0.95)
logging.info(f"{self.name}_act|||{a}")
act = self.action_space.convert_act(a)
obs, r, done, info = self.env.step(act)
r = lreward(a,
self.env,
obs_previous,
obs_do_nothing,
obs_forecasted,
obs,
done,
info,
threshold_safe=0.85)
if a > 0:
if r > 0:
print("+", end="")
elif r < 0:
print("-", end="")
elif len(choosen_actions) > 0:
print("*", end="")
else:
print("x", end="")
else:
if len(choosen_actions) > 0:
print("o", end="")
else:
print("0", end="")
if r > 0:
ep_agent_num_acts += 1
s_ = self.convert_obs(observation_space, obs)
s_ = v_wrap(s_[None, :])
s_ = cuda(self.gpu_id, s_)
ep_r += r
buffer_a.append(a)
buffer_s.append(s)
buffer_r.append(r)
if total_step % self.update_global_iter == 0 or done: # update global and assign to local net
# sync
# if len(buffer_r) > 0 and np.mean(np.abs(buffer_r)) > 0:
buffer_a = cuda(self.gpu_id,
torch.tensor(buffer_a, dtype=torch.long))
buffer_s = cuda(self.gpu_id, torch.cat(buffer_s))
push_and_pull(self.opt, self.local_net,
check_point_episodes, check_point_folder,
self.g_ep, l_ep, self.name, self.rank,
self.global_net, done, s_, buffer_s,
buffer_a, buffer_r, self.gamma, self.gpu_id)
buffer_s, buffer_a, buffer_r = [], [], []
if done: # done and print information
print("")
record(config["starting_num_candidate_acts"],
config["num_candidate_acts_decay_iter"],
self.g_ep, self.g_step,
self.g_num_candidate_acts, self.g_ep_r, ep_r,
self.res_queue, self.name, ep_step,
ep_agent_num_dmd, ep_agent_num_acts)
break
s = s_
total_step += 1
ep_step += 1
l_ep += 1
self.res_queue.put(None)
opt = SharedAdam(gnet.parameters(), lr=5e-3) # global optimizer
global_ep, global_ep_r, res_queue = mp.Value('i', 0), mp.Value('d', 0.), mp.Queue()
# parallel training
workers = [Worker(gnet, opt, global_ep, global_ep_r, res_queue, i) for i in range(mp.cpu_count())]
[w.start() for w in workers]
res = [] # record episode reward to plot
while True:
r = res_queue.get()
if r is not None:
res.append(r)
else:
break
[w.join() for w in workers]
import matplotlib.pyplot as plt
plt.plot(res)
plt.ylabel('Moving average ep reward')
plt.xlabel('Step')
plt.show()
for _ in range(3):
s = env.reset()
while True:
env.render()
a = gnet.choose_action(v_wrap(s[None, :]))
s_, r, done, _ = env.step(a)
s = s_
if done:
break
def run(self):
self.getBandwidthFileList()
# print("bandwidth file count=", len(self.bandwidth_fileList))
r_avg_sum = 0
resDir = self.modelDir + "/test_res_" + str(self.traceIndex)
if os.path.exists(resDir) == False:
os.makedirs(resDir)
resFile = open(resDir + "/" + self.policy + "_" + self.bwType + ".txt",'w')
for bandwidth_file in self.bandwidth_fileList:
bandwidth_fileName = bandwidth_file[0]
rtt = bandwidth_file[1]
bwType = bandwidth_file[2]
busyList = self.get_busyTrace()
# get video -----------------------------
videoName = ""
if PUREFLAF:
video_random = random.random()
for i in range(len(self.videoList)):
if video_random < self.videoList[i][1]:
videoName = self.videoList[i - 1][0]
break
if videoName == "":
videoName = self.videoList[-1][0]
else:
if self.videoTraceIndex == len(self.videoTraceList):
self.videoTraceIndex = 0
videoName = self.videoTraceList[self.videoTraceIndex]
# print(videoName)
# get video ----------------------------
reqBI = self.client.init(videoName=videoName, bandwidthFileName=bandwidth_fileName, rtt=rtt, bwType=bwType)
# mask---------------------------
mask = [1] * A_DIM
for bIndex in range(5):
if PUREFLAF:
mask[bIndex] = 0
else:
if videoName + "_" + str(bIndex + 1) not in self.cachedList:
mask[bIndex] = 0.
# mask---------------------------
state_ = np.zeros(S_LEN)
state = state_.copy() #state =[lastBitrate buffer hThroughput mThroughput busy mask]
# 开始测试------------------------------
total_step = 0
segNum = 0
r_sum = 0
reqBitrate, lastBitrate, buffer, hThroughput, mThroughput, reward, reqBI, done, segNum, busy = [0] * 10
while True:
if sum(mask) == 1:
a = mask.index(1)
else:
if self.policy == "no_policy":
a = 5
elif self.policy == "RL":
a, logits = self.lnet.choose_action(mask, v_wrap(state[None, :]))
self.saveLogits(videoName, logits)
elif self.policy == "lower":
a = self.choose_action_lower(mask, reqBI)
elif self.policy == "closest":
a = self.choose_action_closest(mask, reqBI)
elif self.policy == "highest":
a = self.choose_action_highest(mask, reqBI)
elif self.policy == "prefetch":
a = self.choose_action_prefetch(mask, reqBI)
else:
print("想啥呢")
return
# if random.randint(0, 1000) == 1:
# print("reqb=", reqBitrate, "lb=", lastBitrate, "buffer=", int(buffer), "hT=", int(hThroughput),
# "mT=", int(mThroughput), "busy=", round(busy, 2),
# "mask=", mask, "action=", a, "reqBI=", reqBI, "reward=", round(reward, 2), "logits=", logits)
busy = busyList[segNum % len(busyList)]
if a == 5:
hitFlag = False
else:
hitFlag = True
reqBitrate, lastBitrate, buffer, hThroughput, mThroughput, reward, reqBI, done, segNum = self.client.run(a, busy, hitFlag)
state_[0] = reqBitrate / BITRATES[-1]
state_[1] = lastBitrate / BITRATES[-1]
state_[2] = (buffer/1000 - 30) / 10
state_[3] = (hThroughput - throughput_mean) / throughput_std
state_[4] = (mThroughput - throughput_mean) / throughput_std
reward = reward / 5
r_sum += reward
total_step += 1
if done:
break
state = state_.copy()
# 结束测试------------------------------
r_avg = r_sum / total_step
r_avg_sum += r_avg
resFile.write(str(r_avg) + "\n")
resFile.flush()
print(self.bwType, self.policy, videoName, self.bandwidth_fileList.index(bandwidth_file),"/",len(self.bandwidth_fileList), r_avg)
self.videoTraceIndex += 1
def generate_saliency(observation,
game_step,
net,
game_tracker,
record_json_dir,
actual_action,
actual_probs,
opponent_action,
opponent_id,
new_val=True):
"""Generate saliency data for a given observation
observation -- the observation to be used
game_step -- the current timestep of the game, needed because of an off-by-one in the observation's counter
net -- the network to generate saliency from
game_tracker -- Needed for generate_NN_input
record_json_dir -- output directory for saliency data files. Should be kept with the data from env.render
actual_action -- action decided from actor distribution by the unmodified observation
actual_probs -- action distribution produced by the unmodified observation
opponent_action -- the action the opponent took at this time step
opponent_id -- the id of the opponent
new_val -- if false uses passage/wall modifications, otherwise uses previously unknown value modifications
"""
data = {}
data['step'] = game_step
data[
'actual_action'] = actual_action #have to pass actual action in from outside the func because the agent selects an action from the prob distribution in the main code
data['actual_probs'] = actual_probs
# data['actual_terminal_prediction'] = actual_terminal_predicton #ignoring critic for now. TODO?
#below gets populated by the algorithm
data['mods'] = []
data['actions'] = []
data['opponent_action'] = opponent_action
data['opponent_id'] = opponent_id
# data['predictions'] = [] #again, ignoring critic so ignoring collection of critic value changes.
board_size = len(observation['board'][0]) #assuming board is square
for i in range(board_size):
mod_list = []
action_list = []
# prediction_list = []
for j in range(board_size):
mod_observation = deepcopy(observation)
#remove the player from the 'alive' channel as well just to truly remove the data.
if mod_observation['board'][i][j] >= 10:
mod_observation['alive'].remove(mod_observation['board'][i][j])
#replace this tile's value with a new value.
if new_val:
mod_observation['board'][i][
j] = 14 #first number that doesn't have a definition already
else:
#if the tile here is a passage, make it a wall. Otherwise, make this tile a passage.
#prevents no change from occuring in passages
mod_observation['board'][i][
j] = 1 if mod_observation['board'][i][j] == 0 else 0
#won't ever be a bomb in this tile. possible values are 0, 1, and 14
mod_observation['bomb_blast_strength'][i][j] = 0
mod_observation['bomb_life'][i][j] = 0
this_state = generate_NN_input(10, mod_observation, game_step,
game_tracker)
m_this_state = v_wrap(this_state).unsqueeze(0)
this_action, _, _, this_probs, this_terminal_prediction = net.choose_action(
m_this_state)
diffs = [
this_probs[k] - actual_probs[k] for k in range(len(this_probs))
]
mod_list.append(
deepcopy(diffs)) # Have to deppcopy to avoid pointer hell.
action_list.append(this_action)
# prediction_list.append(this_terminal_prediction)
data['mods'].append(deepcopy(mod_list))
data['actions'].append(deepcopy(action_list))
# data['predictions'].append(deepcopy(prediction_list))
with open(f"{record_json_dir}/d{game_step:03d}.json", 'w') as f:
#Uses MyEncoder to properly clean numpy data to types which json will serialize.
#json.dumps won't serialize numpy data by default
f.write(json.dumps(data, cls=MyEncoder))
env = gym.make(env_id) #Pendulum VibrationEnv input_size = env.observation_space.shape[0] output_size = env.action_space.shape[0] env.reset() s_, r, done, _ = env.step([1]) # s_ = torch.from_numpy(s_.reshape([batch, input_size,seq_len])).float() # s_ = s_.reshape([1, input_size,-1]).float() # s_ = s_.float() print(s_) sys = TCN(input_size, output_size, num_channels = [25, 25, 30], kernel_size = 8, dropout = 0) # lstm_input = torch.randn(1, input_size, 1) # mu, sigma, values = sys(lstm_input) # print(lstm_input) # print(mu, sigma, values) s_ = s_.reshape([batch, input_size, seq_len]) v_wrap(s_) mu, sigma, values = sys(v_wrap(s_)) print(mu, sigma, values) # a = sys.choose_action(v_wrap(s_[None, :])) a = sys.choose_action(v_wrap(s_)) print(a)
