def main():
args = argparser()
args.clip_rewards = False
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + 1122
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env)
model.load_state_dict(torch.load('model.pth', map_location='cpu'))
episode_reward, episode_length = 0, 0
state = env.reset()
while True:
if args.render:
env.render()
action, _ = model.act(torch.FloatTensor(np.array(state)), 0.)
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
episode_length += 1
if done:
state = env.reset()
print("Episode Length / Reward: {} / {}".format(
episode_length, episode_reward))
episode_reward = 0
episode_length = 0
def exploration(args, actor_id, param_queue):
writer = SummaryWriter(comment="-{}-eval".format(args.env))
args.clip_rewards = False
args.episode_life = False
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + actor_id
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env, args)
param = param_queue.get(block=True)
model.load_state_dict(param)
param = None
print("Received First Parameter!")
episode_reward, episode_length, episode_idx = 0, 0, 0
state = env.reset()
tb_dict = {k: [] for k in ['episode_reward', 'episode_length']}
while True:
action, _ = model.act(torch.FloatTensor(np.array(state)), 0.)
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
episode_length += 1
if done or episode_length == args.max_episode_length:
state = env.reset()
tb_dict["episode_reward"].append(episode_reward)
tb_dict["episode_length"].append(episode_length)
episode_reward = 0
episode_length = 0
episode_idx += 1
param = param_queue.get()
model.load_state_dict(param)
print(f"{datetime.now()} Updated Parameter..")
if (episode_idx *
args.num_envs_per_worker) % args.tb_interval == 0:
writer.add_scalar('evaluator/episode_reward_mean',
np.mean(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/episode_reward_max',
np.max(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/episode_reward_min',
np.min(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/episode_reward_std',
np.std(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/episode_length_mean',
np.mean(tb_dict['episode_length']),
episode_idx)
tb_dict['episode_reward'].clear()
tb_dict['episode_length'].clear()
def exploration(args, actor_id, n_actors, param_queue, send_queue,
req_param_queue):
writer = SummaryWriter(comment="-{}-actor{}".format(args.env, actor_id))
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + actor_id
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env)
epsilon = args.eps_base**(1 + actor_id / (n_actors - 1) * args.eps_alpha)
storage = BatchStorage(args.n_steps, args.gamma)
req_param_queue.put(True)
param = param_queue.get(block=True)
model.load_state_dict(param)
param = None
print("Received First Parameter!")
episode_reward, episode_length, episode_idx, actor_idx = 0, 0, 0, 0
state = env.reset()
while True:
action, q_values = model.act(torch.FloatTensor(np.array(state)),
epsilon)
next_state, reward, done, _ = env.step(action)
com_state = zlib.compress(np.array(state).tobytes())
storage.add(com_state, reward, action, done, q_values)
state = next_state
episode_reward += reward
episode_length += 1
actor_idx += 1
if done or episode_length == args.max_episode_length:
state = env.reset()
writer.add_scalar("actor/episode_reward", episode_reward,
episode_idx)
writer.add_scalar("actor/episode_length", episode_length,
episode_idx)
episode_reward = 0
episode_length = 0
episode_idx += 1
if actor_idx % args.update_interval == 0:
try:
req_param_queue.put(True)
param = param_queue.get(block=True)
model.load_state_dict(param)
print("Updated Parameter..")
except queue.Empty:
pass
if len(storage) == args.send_interval:
batch, prios = storage.make_batch()
send_queue.put((batch, prios))
batch, prios = None, None
storage.reset()
def main():
learner_ip = get_environ()
args = argparser()
writer = SummaryWriter(comment="-{}-eval".format(args.env))
ctx = zmq.Context()
param_socket = ctx.socket(zmq.SUB)
param_socket.setsockopt(zmq.SUBSCRIBE, b'')
param_socket.setsockopt(zmq.CONFLATE, 1)
param_socket.connect('tcp://{}:52001'.format(learner_ip))
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + 1122
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env)
data = param_socket.recv(copy=False)
param = pickle.loads(data)
model.load_state_dict(param)
print("Loaded first parameter from learner")
episode_reward, episode_length, episode_idx = 0, 0, 0
state = env.reset()
while True:
if args.render:
env.render()
action, _ = model.act(torch.FloatTensor(np.array(state)), 0.01)
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
episode_length += 1
if done:
state = env.reset()
writer.add_scalar("eval/episode_reward", episode_reward,
episode_idx)
writer.add_scalar("eval/episode_length", episode_length,
episode_idx)
episode_reward = 0
episode_length = 0
episode_idx += 1
if episode_idx % args.eval_update_interval == 0:
data = param_socket.recv(copy=False)
param = pickle.loads(data)
model.load_state_dict(param)
def exploration_eval(args, actor_id, param_queue):
writer = SummaryWriter(comment="-{}-eval".format(args.env))
args.clip_rewards = False
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + actor_id
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env)
param = param_queue.get(block=True)
model.load_state_dict(param)
param = None
print("Received First Parameter!")
episode_reward, episode_length, episode_idx = 0, 0, 0
state = env.reset()
while True:
action, _ = model.act(torch.FloatTensor(np.array(state)), 0.)
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
episode_length += 1
if done or episode_length == args.max_episode_length:
state = env.reset()
writer.add_scalar("evaluator/episode_reward", episode_reward,
episode_idx)
writer.add_scalar("evaluator/episode_length", episode_length,
episode_idx)
episode_reward = 0
episode_length = 0
episode_idx += 1
param = param_queue.get()
model.load_state_dict(param)
print("Updated Parameter..")
class DuelingAgent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = DuelingDQN(state_size, action_size,
seed).to(device)
self.qnetwork_target = DuelingDQN(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local.act(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
if random.uniform(0, 1) > 0.99:
print(loss)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
class train_DQN():
def __init__(self, env_id, max_step = 1e5, prior_alpha = 0.6, prior_beta_start = 0.4,
epsilon_start = 1.0, epsilon_final = 0.01, epsilon_decay = 500,
batch_size = 32, gamma = 0.99, target_update_interval=1000, save_interval = 1e4,
):
self.prior_beta_start = prior_beta_start
self.max_step = int(max_step)
self.batch_size = batch_size
self.gamma = gamma
self.target_update_interval = target_update_interval
self.save_interval = save_interval
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.env = gym.make(env_id)
self.model = DuelingDQN(self.env).to(self.device)
self.target_model = DuelingDQN(self.env).to(self.device)
self.target_model.load_state_dict(self.model.state_dict())
self.replay_buffer = PrioritizedReplayBuffer(100000,alpha=prior_alpha)
self.optimizer = optim.Adam(self.model.parameters())
self.writer = SummaryWriter(comment="-{}-learner".format(self.env.unwrapped.spec.id))
# decay function
self.scheduler = optim.lr_scheduler.StepLR(self.optimizer,step_size=1000,gamma=0.99)
self.beta_by_frame = lambda frame_idx: min(1.0, self.prior_beta_start + frame_idx * (1.0 - self.prior_beta_start) / 1000)
self.epsilon_by_frame = lambda frame_idx: epsilon_final + (epsilon_start - epsilon_final) * math.exp(-1. * frame_idx / epsilon_decay)
def update_target(self,current_model, target_model):
target_model.load_state_dict(current_model.state_dict())
def compute_td_loss(self,batch_size, beta):
state, action, reward, next_state, done, weights, indices = self.replay_buffer.sample(batch_size, beta)
state = torch.FloatTensor(state).to(self.device)
next_state = torch.FloatTensor(next_state).to(self.device)
action = torch.LongTensor(action).to(self.device)
reward = torch.FloatTensor(reward).to(self.device)
done = torch.FloatTensor(done).to(self.device)
weights = torch.FloatTensor(weights).to(self.device)
batch = (state, action, reward, next_state, done, weights)
# q_values = self.model(state)
# next_q_values = self.target_model(next_state)
# q_value = q_values.gather(1, action.unsqueeze(1)).squeeze(1)
# next_q_value = next_q_values.max(1)[0]
# expected_q_value = reward + self.gamma * next_q_value * (1 - done)
# td_error = torch.abs(expected_q_value.detach() - q_value)
# loss = (td_error).pow(2) * weights
# prios = loss+1e-5#0.9 * torch.max(td_error)+(1-0.9)*td_error
# loss = loss.mean()
loss, prios = utils.compute_loss(self.model,self.target_model, batch,1)
self.optimizer.zero_grad()
loss.backward()
self.scheduler.step()
self.replay_buffer.update_priorities(indices, prios)
self.optimizer.step()
return loss
def train(self):
losses = []
all_rewards = []
episode_reward = 0
episode_idx = 0
episode_length = 0
state = self.env.reset()
for frame_idx in range(self.max_step):
epsilon = self.epsilon_by_frame(frame_idx)
action,_ = self.model.act(torch.FloatTensor((state)).to(self.device), epsilon)
next_state, reward, done, _ = self.env.step(action)
self.replay_buffer.add(state, action, reward, next_state, done)
state = next_state
episode_reward += reward
episode_length += 1
if done:
state = self.env.reset()
all_rewards.append(episode_reward)
self.writer.add_scalar("actor/episode_reward", episode_reward, episode_idx)
self.writer.add_scalar("actor/episode_length", episode_length, episode_idx)
# print("episode: ",episode_idx, " reward: ", episode_reward)
episode_reward = 0
episode_length = 0
episode_idx += 1
if len(self.replay_buffer) > self.batch_size:
beta = self.beta_by_frame(frame_idx)
loss = self.compute_td_loss(self.batch_size, beta)
losses.append(loss.item())
self.writer.add_scalar("learner/loss", loss, frame_idx)
if frame_idx % self.target_update_interval == 0:
print("update target...")
self.update_target(self.model, self.target_model)
if frame_idx % self.save_interval == 0 or frame_idx == self.max_step-1:
print("save model...")
self.save_model(frame_idx)
def save_model(self, idx):
torch.save(self.model.state_dict(), "./model{}.pth".format(idx))
def load_model(self,idx):
with open("model{}.pth".format(idx), "rb") as f:
print("loading weights_{}".format(idx))
self.model.load_state_dict(torch.load(f,map_location="cpu"))
def evaluator(args):
comm = global_dict['comm_world']
writer = SummaryWriter(log_dir=os.path.join(args['log_dir'], 'eval'))
args['clip_rewards'] = False
args['episode_life'] = False
env = make_atari(args['env'])
env = wrap_atari_dqn(env, args)
seed = args['seed'] - 1
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
torch.set_num_threads(1)
model = DuelingDQN(env, args)
recv_param_buf = bytearray(100 * 1024 * 1024)
comm.Send(b'', dest=global_dict['rank_learner'])
comm.Recv(buf=recv_param_buf, source=global_dict['rank_learner'])
param = pickle.loads(recv_param_buf)
model.load_state_dict(param)
episode_reward, episode_length, episode_idx = 0, 0, 0
state = env.reset()
tb_dict = {k: [] for k in ['episode_reward', 'episode_length']}
while True:
action, _ = model.act(torch.FloatTensor(np.array(state)), 0.)
next_state, reward, done, _ = env.step(action)
state = next_state
episode_reward += reward
episode_length += 1
if done or episode_length == args['max_episode_length']:
state = env.reset()
tb_dict["episode_reward"].append(episode_reward)
tb_dict["episode_length"].append(episode_length)
episode_reward = 0
episode_length = 0
episode_idx += 1
comm.Send(b'', dest=global_dict['rank_learner'])
comm.Recv(buf=recv_param_buf, source=global_dict['rank_learner'])
param = pickle.loads(recv_param_buf)
model.load_state_dict(param)
if (episode_idx *
args['num_envs_per_worker']) % args['tb_interval'] == 0:
writer.add_scalar('evaluator/1_episode_reward_mean',
np.mean(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/2_episode_reward_max',
np.max(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/3_episode_reward_min',
np.min(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/4_episode_reward_std',
np.std(tb_dict['episode_reward']),
episode_idx)
writer.add_scalar('evaluator/5_episode_length_mean',
np.mean(tb_dict['episode_length']),
episode_idx)
tb_dict['episode_reward'].clear()
tb_dict['episode_length'].clear()
class Worker(mp.Process):
def __init__(self, worker_id, env_id, seed, epsilon, size, lock,
n_steps, gamma, send_interval, task_queue, buffer, max_episode_length):
mp.Process.__init__(self)
self.worker_id = worker_id
self.env = gym.make(env_id)
self.seed = seed
self.task_queue = task_queue
self.buffer = buffer
self.max_episode_length = max_episode_length
self.send_interval = send_interval
self.storage = BatchStorage(n_steps, gamma)
self.size = size
self.memory = []
self.model = DuelingDQN(self.env)
# self.writer = SummaryWriter(comment="-{}-actor{}".format(env_id, worker_id))
self.lock = lock
self.set_all_seeds()
self.epsilon = epsilon
def set_all_seeds(self):
self.env.seed(self.seed)
np.random.seed(self.seed)
random.seed(self.seed)
torch.manual_seed(self.seed)
torch.cuda.manual_seed_all(self.seed)
def update_weights(self, DQN_state_dict):
self.model.load_state_dict(DQN_state_dict)
def record_batch(self):
episode_reward, episode_length, episode_idx, actor_idx = 0, 0, 0, 0
state = self.env.reset()
self.storage.reset()
self.memory = []
# while actor_idx < self.size:
while True:
action, q_values = self.model.act(torch.FloatTensor(np.array(state)), self.epsilon)
next_state, reward, done, _ = self.env.step(action)
self.storage.add(state, reward, action, done, q_values)
state = next_state
episode_reward += reward
episode_length += 1
actor_idx += 1
if done or episode_length >= self.max_episode_length:
state = self.env.reset()
# self.writer.add_scalar("actor/episode_reward", episode_reward, episode_idx)
# self.writer.add_scalar("actor/episode_length", episode_length, episode_idx)
episode_reward = 0
episode_length = 0
episode_idx += 1
if done or len(self.storage) == self.send_interval:
batch, prios = self.storage.make_batch()
self.memory.append((*batch, prios))
# for i in range(len(prios)):
# self.buffer.add(batch[0][i],batch[1][i],batch[2][i],batch[3][i],batch[4][i],prios[i])
batch, prios = None, None
self.storage.reset()
if done:
break
def run(self):
while True:
########## run loop
task = self.task_queue.get(block=True)
if task["desc"] == "record_batch":
# print("start record batch")
self.record_batch()
self.buffer.put(self.memory)
self.task_queue.task_done()
# print("record batch done")
elif task["desc"] == "set_pi_weights":
# print("set weight")
self.update_weights(task["pi_state_dict"])
self.task_queue.task_done()
# print("set weight done")
elif task["desc"] == "cleanup":
# print("clean up")
self.env.close()
self.task_queue.task_done()
def exploration(args, actor_id, n_actors, replay_ip, param_queue, sample_enque, sample_deque):
ctx = zmq.Context()
batch_socket = ctx.socket(zmq.DEALER)
batch_socket.setsockopt(zmq.IDENTITY, pickle.dumps('actor-{}'.format(actor_id)))
batch_socket.connect('tcp://{}:51001'.format(replay_ip))
writer = SummaryWriter(comment="-{}-actor{}".format(args.env, actor_id))
env = make_atari(args.env)
env = wrap_atari_dqn(env, args)
seed = args.seed + actor_id
utils.set_global_seeds(seed, use_torch=True)
env.seed(seed)
model = DuelingDQN(env)
epsilon = args.eps_base ** (1 + actor_id / (n_actors - 1) * args.eps_alpha)
storage = BatchStorage(args.n_steps, args.gamma)
param = param_queue.get(block=True)
model.load_state_dict(param)
param = None
print("Received First Parameter!")
episode_reward, episode_length, episode_idx, actor_idx = 0, 0, 0, 0
state = env.reset()
while True:
action, q_values = model.act(torch.FloatTensor(np.array(state)), epsilon)
next_state, reward, done, _ = env.step(action)
storage.add(state, reward, action, done, q_values)
state = next_state
episode_reward += reward
episode_length += 1
actor_idx += 1
if done or episode_length == args.max_episode_length:
state = env.reset()
writer.add_scalar("actor/episode_reward", episode_reward, episode_idx)
writer.add_scalar("actor/episode_length", episode_length, episode_idx)
episode_reward = 0
episode_length = 0
episode_idx += 1
if actor_idx % args.update_interval == 0:
try:
param = param_queue.get(block=False)
model.load_state_dict(param)
print("Updated Parameter..")
except queue.Empty:
pass
# get sample batch after each step
while sample_enque:
idxes = sample_enque.get()
sample_deque.put(storage.get_sample_batch(idxes))
# only pass the prios and get indexes from ReplayBuffer
if len(storage) == args.send_interval:
batch, prios = storage.make_batch()
data = pickle.dumps(prios)
batch, prios = None, None
storage.reset()
batch_socket.send(data, copy=False)
_, idxes = batch_socket.recv_multipart(copy=False)
storage.add_batch(batch, idxes)