def train(args, n_actors, batch_queue, prios_queue, param_queue):
env = wrapper.make_atari(args.env)
env = wrapper.wrap_atari_dqn(env, args)
utils.set_global_seeds(args.seed, use_torch=True)
model = DuelingDQN(env, args).to(args.device)
# model.load_state_dict(torch.load('model_30h.pth'))
tgt_model = DuelingDQN(env, args).to(args.device)
tgt_model.load_state_dict(model.state_dict())
writer = SummaryWriter(comment="-{}-learner".format(args.env))
optimizer = torch.optim.Adam(model.parameters(), args.lr)
# optimizer = torch.optim.RMSprop(model.parameters(), args.lr, alpha=0.95, eps=1.5e-7, centered=True)
check_connection(n_actors)
param_queue.put(model.state_dict())
learn_idx = 0
ts = time.time()
tb_dict = {
k: []
for k in ['loss', 'grad_norm', 'max_q', 'mean_q', 'min_q']
}
while True:
*batch, idxes = batch_queue.get()
loss, prios, q_values = utils.compute_loss(model, tgt_model, batch,
args.n_steps, args.gamma)
grad_norm = utils.update_parameters(loss, model, optimizer,
args.max_norm)
prios_queue.put((idxes, prios))
batch, idxes, prios = None, None, None
learn_idx += 1
tb_dict["loss"].append(float(loss))
tb_dict["grad_norm"].append(float(grad_norm))
tb_dict["max_q"].append(float(torch.max(q_values)))
tb_dict["mean_q"].append(float(torch.mean(q_values)))
tb_dict["min_q"].append(float(torch.min(q_values)))
if args.soft_target_update:
tau = args.tau
for p_tgt, p in zip(tgt_model.parameters(), model.parameters()):
p_tgt.data *= 1 - tau
p_tgt.data += tau * p
elif learn_idx % args.target_update_interval == 0:
print("Updating Target Network..")
tgt_model.load_state_dict(model.state_dict())
if learn_idx % args.save_interval == 0:
print("Saving Model..")
torch.save(model.state_dict(), "model.pth")
if learn_idx % args.publish_param_interval == 0:
param_queue.put(model.state_dict())
if learn_idx % args.tb_interval == 0:
bps = args.tb_interval / (time.time() - ts)
print("Step: {:8} / BPS: {:.2f}".format(learn_idx, bps))
writer.add_scalar("learner/BPS", bps, learn_idx)
for k, v in tb_dict.items():
writer.add_scalar(f'learner/{k}', np.mean(v), learn_idx)
v.clear()
ts = time.time()
def train(args, n_actors, batch_queue, prios_queue, param_queue):
env = RunTagEnv(width=5,
height=5,
number_of_subordinates=1,
max_steps=1000)
#env = wrapper.make_atari(args.env)
#env = wrapper.wrap_atari_dqn(env, args)
utils.set_global_seeds(args.seed, use_torch=True)
model = DuelingDQN(env).to(args.device)
tgt_model = DuelingDQN(env).to(args.device)
tgt_model.load_state_dict(model.state_dict())
writer = SummaryWriter(comment="-{}-learner".format(args.env))
# optimizer = torch.optim.Adam(model.parameters(), args.lr)
optimizer = torch.optim.RMSprop(model.parameters(),
args.lr,
alpha=0.95,
eps=1.5e-7,
centered=True)
check_connection(n_actors)
param_queue.put(model.state_dict())
learn_idx = 0
ts = time.time()
while True:
*batch, idxes = batch_queue.get()
loss, prios = utils.compute_loss(model, tgt_model, batch, args.n_steps,
args.gamma)
grad_norm = utils.update_parameters(loss, model, optimizer,
args.max_norm)
print('Updated parameters!')
prios_queue.put((idxes, prios))
batch, idxes, prios = None, None, None
learn_idx += 1
writer.add_scalar("learner/loss", loss, learn_idx)
writer.add_scalar("learner/grad_norm", grad_norm, learn_idx)
if learn_idx % args.target_update_interval == 0:
print("Updating Target Network..")
tgt_model.load_state_dict(model.state_dict())
if learn_idx % args.save_interval == 0:
print("Saving Model..")
torch.save(model.state_dict(), "model.pth")
if learn_idx % args.publish_param_interval == 0:
param_queue.put(model.state_dict())
if learn_idx % args.bps_interval == 0:
bps = args.bps_interval / (time.time() - ts)
print("Step: {:8} / BPS: {:.2f}".format(learn_idx, bps))
writer.add_scalar("learner/BPS", bps, learn_idx)
ts = time.time()
class DoubleDuelingDQNAgent(DoubleDQNAgent):
"""
Interacts with and learns from the environment.
Double Dueling DQN.
"""
def __init__(self, state_size, action_size, seed):
"""
Initialize an Agent object.
:param state_size: dimension of each state;
:param action_size: dimension of each action;
:param seed: random seed.
"""
super().__init__(state_size, action_size, seed)
# Q-Network
self.network_local = DuelingDQN(state_size, action_size,
seed).to(DEVICE)
self.network_target = DuelingDQN(state_size, action_size,
seed).to(DEVICE)
self.optimizer = optim.Adam(self.network_local.parameters(), lr=LR)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed, model="QNetwork"):
"""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
if model == "QNetwork":
self.qnetwork_local = QNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size,
seed).to(device)
if model == "QNetworkConvolutional":
self.qnetwork_local = QNetworkConvolutional(
state_size, action_size, seed).to(device)
self.qnetwork_target = QNetworkConvolutional(
state_size, action_size, seed).to(device)
if model == "DuelingDQN":
self.qnetwork_local = DuelingDQN(state_size, action_size,
seed).to(device)
self.qnetwork_target = DuelingDQN(state_size, action_size,
seed).to(device)
print("Model: " + model)
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(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.Variable]): 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)
# 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)
def learner(args):
comm_cross = global_dict['comm_cross']
hvd.init(comm=comm_cross)
torch.cuda.set_device(hvd.local_rank())
env = wrap_atari_dqn(make_atari(args['env']), args)
# utils.set_global_seeds(args['seed'], use_torch=True)
device = args['device']
model = DuelingDQN(env, args).to(device)
if os.path.exists('model.pth'):
# model.load_state_dict(torch.load('model.pth'))
pass
tgt_model = DuelingDQN(env, args).to(device)
del env
writer = SummaryWriter(log_dir=os.path.join(
args['log_dir'], f'{global_dict["unit_idx"]}-learner'))
# optimizer = torch.optim.SGD(model.parameters(), 1e-5 * args['num_units'], momentum=0.8)
# optimizer = torch.optim.RMSprop(model.parameters(), args['lr'], alpha=0.95, eps=1.5e-7, centered=True)
optimizer = torch.optim.Adam(model.parameters(),
args['lr'] * args['num_units'])
optimizer = hvd.DistributedOptimizer(
optimizer, named_parameters=model.named_parameters())
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
tgt_model.load_state_dict(model.state_dict())
if args['dynamic_gradient_clip']:
grad_norm_running_mean = args['gradient_norm_running_mean']
grad_norm_lambda = args['gradient_norm_lambda']
batch_queue = queue.Queue(maxsize=3)
prios_queue = queue.Queue(maxsize=4)
param_queue = queue.Queue(maxsize=3)
threading.Thread(target=recv_batch, args=(batch_queue, )).start()
threading.Thread(target=send_prios, args=(prios_queue, )).start()
threading.Thread(target=send_param, args=(param_queue, )).start()
if global_dict['unit_idx'] == 0:
threading.Thread(target=send_param_evaluator,
args=(param_queue, )).start()
prefetcher = data_prefetcher(batch_queue, args['cuda'])
learn_idx = 0
ts = time.time()
tb_dict = {
k: []
for k in [
'loss', 'grad_norm', 'max_q', 'mean_q', 'min_q',
'batch_queue_size', 'prios_queue_size'
]
}
first_rount = True
while True:
(*batch, idxes) = prefetcher.next()
if first_rount:
print("start training")
sys.stdout.flush()
first_rount = False
loss, prios, q_values = utils.compute_loss(model, tgt_model, batch,
args['n_steps'],
args['gamma'])
optimizer.zero_grad()
loss.backward()
if args['dynamic_gradient_clip']:
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(),
grad_norm_running_mean * args['clipping_threshold'])
grad_norm_running_mean = grad_norm_running_mean * grad_norm_lambda + \
min(grad_norm, grad_norm_running_mean * args['clipping_threshold']) * (1-grad_norm_lambda)
else:
grad_norm = torch.norm(
torch.stack([
torch.norm(p.grad.detach(), 2) for p in model.parameters()
]), 2)
# global_prios_sum = np.array(prios_sum)
# comm_cross.Allreduce(MPI.IN_PLACE, global_prios_sum.data)
# global_prios_sum = float(global_prios_sum)
# scale = prios_sum / global_prios_sum
if args['dynamic_gradient_clip'] and args[
'dropping_threshold'] and grad_norm > grad_norm_running_mean * args[
'dropping_threshold']:
pass
else:
optimizer.step()
prios_queue.put((idxes, prios))
learn_idx += 1
tb_dict["loss"].append(float(loss))
tb_dict["grad_norm"].append(float(grad_norm))
tb_dict["max_q"].append(float(torch.max(q_values)))
tb_dict["mean_q"].append(float(torch.mean(q_values)))
tb_dict["min_q"].append(float(torch.min(q_values)))
tb_dict["batch_queue_size"].append(batch_queue.qsize())
tb_dict["prios_queue_size"].append(prios_queue.qsize())
if learn_idx % args['target_update_interval'] == 0:
tgt_model.load_state_dict(model.state_dict())
if learn_idx % args['save_interval'] == 0 and global_dict[
'unit_idx'] == 0:
torch.save(model.state_dict(), "model.pth")
if learn_idx % args['publish_param_interval'] == 0:
param_queue.put(model.state_dict())
if learn_idx % args['tb_interval'] == 0:
bps = args['tb_interval'] / (time.time() - ts)
for i, (k, v) in enumerate(tb_dict.items()):
writer.add_scalar(f'learner/{i+1}_{k}', np.mean(v), learn_idx)
v.clear()
writer.add_scalar(f"learner/{i+2}_BPS", bps, learn_idx)
ts = time.time()
class PERDoubleDuelingDQNAgent(DoubleDuelingDQNAgent):
"""
Interacts with and learns from the environment.
Double Dueling DQN with prioritized experience replay.
"""
def __init__(self, state_size, action_size, seed):
"""
Initialize an Agent object.
:param state_size: dimension of each state;
:param action_size: dimension of each action;
:param seed: random seed.
"""
super().__init__(state_size, action_size, seed)
# Replay memory
self.memory = PrioritizedReplayBuffer(BUFFER_SIZE, BATCH_SIZE,
state_size, seed)
# Q-Network
self.network_local = DuelingDQN(state_size, action_size,
seed).to(DEVICE)
self.network_target = DuelingDQN(state_size, action_size,
seed).to(DEVICE)
self.optimizer = optim.Adam(self.network_local.parameters(), lr=LR)
def learn(self, experiences, gamma):
"""
Update value parameters using given batch of experience tuples.
:param experiences: (Tuple[torch.Tensor]) tuple of (s, a, r, s', done) tuples;
:param gamma: discount factor.
"""
tree_idx, states, actions, rewards, next_states, dones, ISWeights = experiences
# Get expected Q values from local model
Q_expected = self.network_local(states).gather(1, actions)
# Get next actions based on local network
next_actions = self.network_local(next_states).detach().max(
1)[1].unsqueeze(1)
# Get max predicted Q values (for next states) from target model based on local model next actions
Q_targets_next = self.network_target(next_states).detach().gather(
1, next_actions)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Update transition priorities
self.memory.batch_update(tree_idx, np.ravel(np.abs(Q_targets.numpy())))
# Compute loss
loss = (torch.Tensor(ISWeights).float().to(DEVICE) *
F.mse_loss(Q_expected, Q_targets)).mean()
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.network_local, self.network_target, TAU)
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"))
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed, max_t=1000):
"""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
self.prio_b = PRIO_B
self.b_step = 0
self.max_b_step = 2000
self.learnFirst = True
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
#self.memory.add(state, action, reward, next_state, done)
# Hassan : Save the experience in prioritized replay memory
self.memory.prio_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)
# Hassan : prioritized replay memory
self.b_step = self.b_step + 1
experiences, indices = self.memory.prio_sample()
self.learn(experiences, GAMMA, indices)
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(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 get_beta(self, t):
'''
Return the current exponent β based on its schedul. Linearly anneal β
from its initial value β0 to 1, at the end of learning.
:param t: integer. Current time step in the episode
:return current_beta: float. Current exponent beta
'''
#f_frac = min(float(t) / self.max_b_step, 1.0)
#current_beta = self.prio_b + f_frac * (1. - self.prio_b)
#current_beta = min(1,current_beta)
self.prio_b = min(1, self.prio_b + PRIO_B_INC)
return self.prio_b
def learn(self, experiences, gamma, indices):
"""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, probabilities = experiences
## TODO: compute and minimize the loss
"*** YOUR CODE HERE ***"
# Get max predicted Q values (for next states) from target model
# Hassan : Action is selected using greedy policy
#Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
# Hassan : Double DQN
# Selecting actions which maximizes while taking w (qnetwork_local)
next_actions = self.qnetwork_local(next_states).detach().argmax(
dim=1).unsqueeze(1)
#next_actions_test = self.qnetwork_local(next_states).detach().max(1)[1].unsqueeze(1) # Hassan : from the example
#print(torch.sum(next_actions-next_actions_test)) # Hassan : no difference found
# Selecting q values of these actions using w' (qnetwork_target)
Q_targets_next = self.qnetwork_target(next_states).gather(
1, next_actions)
# Compute Q targets for current states
# Hassan : This is TD Target
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
# Hassan : This is current value
Q_expected = self.qnetwork_local(states).gather(1, actions)
#Hassan : Compute the td_error
td_error = Q_targets - Q_expected
#print(td_error.detach().numpy())
#self.prio_b = min(1, PRIO_B_INC+self.prio_b)
f_currbeta = self.get_beta(0)
#print(f_currbeta)
#f_currbeta = self.get_beta(self.b_step)
#print(self.b_step)
#print(t)
#print(self.prio_b)
weights_importance = probabilities.mul_(
self.memory.__len__()).pow_(-f_currbeta)
# Hassan : calculate max_weights_importance
#probabilities_min = min(self.memory.priorities)/self.memory.cum_priorities
probabilities_min = self.memory.min_priority / self.memory.cum_priorities
max_weights_importance = (probabilities_min *
self.memory.__len__())**(-f_currbeta)
# Hassan : divide the weights importance with the max_weights_importance
# Hassan : Improvement why not calculating the max_weights_importance = max(weights_importance)??
# Hassan : this will only calculating on the current list not the complete one
#print(weights_importance)
#print(weights_importance.max(0)[0])
#print(max_weights_importance)
#if self.learnFirst:
# self.learnFirst = False
#else :
# max_weights_importance = max_weights_importance[0]
weights_final = weights_importance.div_(max_weights_importance)
square_weighted_error = td_error.pow_(2).mul_(weights_final)
loss = square_weighted_error.mean()
# Hassan : after the observations observation from example, update was done after the weights calculation
if self.prio_b > 0.5:
self.memory.prio_update(indices,
td_error.detach().numpy(), PRIO_E, PRIO_A)
# Compute loss
#loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
# Hassan : Here not after C steps w is changed though cahnged slightly after every learn step
# Hassan : We can modify to change this after ever C steps
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 Agent():
"""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)
self.priority_alpha = 0.0 #current best: 03
self.priority_beta_start = 0.4
self.priority_beta_frames = BUFFER_SIZE
# Replay memory
self.memory = PrioritizedReplayMemory(BUFFER_SIZE, self.priority_alpha,
self.priority_beta_start,
self.priority_beta_frames)
# 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.push((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 self.memory.storage_size() > BATCH_SIZE:
#print("storage == ", self.memory.storage_size())
experiences, idxes, weights = self.memory.sample(BATCH_SIZE)
self.learn(experiences, idxes, weights, 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(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, idxes, weights, 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 = zip(*experiences)
states = torch.from_numpy(
np.vstack([state for state in states
if state is not None])).float().to(device)
actions = torch.from_numpy(
np.vstack([action for action in actions
if action is not None])).long().to(device)
rewards = torch.from_numpy(
np.vstack([reward for reward in rewards
if reward is not None])).float().to(device)
next_states = torch.from_numpy(
np.vstack([
next_state for next_state in next_states
if next_state is not None
])).float().to(device)
dones = torch.from_numpy(
np.vstack([done for done in dones if done is not None
]).astype(np.uint8)).float().to(device)
# Get max predicted Q values (for next states) from target model
#print("state-action values:")
#print(self.qnetwork_target(next_states).detach())
#print(next_states)
next_target_Q = self.qnetwork_target.forward(next_states)
#print("next_target_Q == ", next_target_Q)
_, next_local_Q_index = torch.max(
self.qnetwork_local.forward(next_states), axis=1)
#Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
Q_targets_next = next_target_Q[range(next_target_Q.shape[0]),
next_local_Q_index]
Q_targets_next1 = Q_targets_next.reshape((len(Q_targets_next), 1))
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next1 * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
#print(Q_expected)
#print(Q_targets)
diff = Q_expected - Q_targets
#print(diff)
#diff = diff.mean()
#print(idxes)
#print(diff.detach().squeeze().abs().cpu().numpy().tolist())
#update the priority of the replay buffer
self.memory.update_priorities(
idxes,
diff.detach().squeeze().abs().cpu().numpy().tolist())
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets) * weights
loss = loss.mean()
# 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 Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed,max_t=1000):
"""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
self.prio_b = PRIO_B
self.b_step = 0
self.max_b_step = 2000
self.learnFirst = True
def step(self, state, action, reward, next_state, done):
# Hassan : Save the experience in prioritized replay memory
self.memory.prio_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:
self.b_step = self.b_step + 1
experiences, indices = self.memory.prio_sample()
self.learn(experiences, GAMMA, indices)
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(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 get_beta(self, t):
'''
Return the current exponent β based on its schedul. Linearly anneal β
from its initial value β0 to 1, at the end of learning.
:param t: integer. Current time step in the episode
:return current_beta: float. Current exponent beta
'''
#f_frac = min(float(t) / self.max_b_step, 1.0)
#current_beta = self.prio_b + f_frac * (1. - self.prio_b)
#current_beta = min(1,current_beta)
self.prio_b = min(1,self.prio_b + PRIO_B_INC)
return self.prio_b
def learn(self, experiences, gamma, indices):
"""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, probabilities = experiences
"*** YOUR CODE HERE ***"
# Double DQN implementation
# Selecting actions which maximizes while taking w (qnetwork_local)
next_actions = self.qnetwork_local(next_states).detach().argmax(dim=1).unsqueeze(1)
# evluate best actions using w' (qnetwork_target)
Q_targets_next = self.qnetwork_target(next_states).gather(1, next_actions)
# Compute Q targets for current states (TD Target)
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 the td_error
td_error = Q_targets - Q_expected
f_currbeta = self.get_beta(0)
# Prioritized experience replay : calculating the final weights for calculating loss function
weights_importance = probabilities.mul_(self.memory.__len__()).pow_(-f_currbeta)
probabilities_min = self.memory.min_priority/self.memory.cum_priorities
max_weights_importance = (probabilities_min * self.memory.__len__())**(-f_currbeta)
weights_final = weights_importance.div_(max_weights_importance)
# Compute mean squared weighted error
square_weighted_error = td_error.pow_(2).mul_(weights_final)
loss = square_weighted_error.mean()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Prioritized experience replay : updating the priority of experience tuple in replay buffer
self.memory.prio_update(indices,td_error.detach().numpy(),PRIO_E,PRIO_A)
# ------------------- 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 Agent:
"""Interacts with and learns from the environment."""
def __init__(self, config):
"""Initialize an Agent object"""
self.seed = random.seed(config["general"]["seed"])
self.config = config
# Q-Network
self.q = DuelingDQN(config).to(DEVICE)
self.q_target = DuelingDQN(config).to(DEVICE)
self.optimizer = optim.RMSprop(self.q.parameters(),
lr=config["agent"]["learning_rate"])
self.criterion = F.mse_loss
self.memory = ReplayBuffer(config)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def save_experiences(self, state, action, reward, next_state, done):
"""Prepare and save experience in replay memory"""
reward = np.clip(reward, -1.0, 1.0)
self.memory.add(state, action, reward, next_state, done)
def _current_step_is_a_learning_step(self):
"""Check if the current step is an update step"""
self.t_step = (self.t_step + 1) % self.config["agent"]["update_rate"]
return self.t_step == 0
def _enough_samples_in_memory(self):
"""Check if minimum amount of samples are in memory"""
return len(self.memory) > self.config["train"]["batch_size"]
def epsilon_greedy_action_selection(self, action_values, eps):
"""Epsilon-greedy action selection"""
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(
np.arange(self.config["general"]["action_size"]))
def act(self, state, eps=0.0):
"""Returns actions for given state as per current policy"""
state = torch.from_numpy(state).float().unsqueeze(0).to(DEVICE)
self.q.eval()
with torch.no_grad():
action_values = self.q(state)
self.q.train()
return self.epsilon_greedy_action_selection(action_values, eps)
def _calc_loss(self, states, actions, rewards, next_states, dones):
"""Calculates loss for a given experience batch"""
q_eval = self.q(states).gather(1, actions)
q_eval_next = self.q(next_states)
_, q_argmax = q_eval_next.detach().max(1)
q_next = self.q_target(next_states)
q_next = q_next.gather(1, q_argmax.unsqueeze(1))
q_target = rewards + (self.config["agent"]["gamma"] * q_next *
(1 - dones))
loss = self.criterion(q_eval, q_target)
return loss
def _update_weights(self, loss):
"""update the q network weights"""
torch.nn.utils.clip_grad.clip_grad_value_(self.q.parameters(), 1.0)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
def learn(self):
"""Update network using one sample of experience from memory"""
if self._current_step_is_a_learning_step(
) and self._enough_samples_in_memory():
states, actions, rewards, next_states, dones = self.memory.sample(
self.config["train"]["batch_size"])
loss = self._calc_loss(states, actions, rewards, next_states,
dones)
self._update_weights(loss)
self._soft_update(self.q, self.q_target)
def _soft_update(self, local_model, target_model):
"""Soft update target network parameters: θ_target = τ*θ_local + (1 - τ)*θ_target"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(
self.config["agent"]["tau"] * local_param.data +
(1.0 - self.config["agent"]["tau"]) * target_param.data)
def save(self):
"""Save the network weights"""
helper.mkdir(
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"]))
current_date_time = helper.get_current_date_time()
current_date_time = current_date_time.replace(" ", "__").replace(
"/", "_").replace(":", "_")
torch.save(
self.q.state_dict(),
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"],
"ckpt_" + current_date_time))
def load(self):
"""Load latest available network weights"""
list_of_files = glob.glob(
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"], "*"))
latest_file = max(list_of_files, key=os.path.getctime)
self.q.load_state_dict(torch.load(latest_file))
self.q_target.load_state_dict(torch.load(latest_file))
class train_DQN():
def __init__(self,
env_id,
seed=0,
lr=1e-5,
n_step=3,
gamma=0.99,
n_workers=20,
max_norm=40,
target_update_interval=2500,
save_interval=5000,
batch_size=64,
buffer_size=1e6,
prior_alpha=0.6,
prior_beta=0.4,
publish_param_interval=32,
max_step=1e5):
self.env = gym.make(env_id)
self.seed = seed
self.lr = lr
self.n_step = n_step
self.gamma = gamma
self.max_norm = max_norm
self.target_update_interval = target_update_interval
self.save_interval = save_interval
self.publish_param_interval = publish_param_interval
self.batch_size = batch_size
self.prior_beta = prior_beta
self.max_step = max_step
self.buffer = CustomPrioritizedReplayBuffer(size=buffer_size,
alpha=prior_alpha)
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
self.model = DuelingDQN(self.env).to(self.device)
self.tgt_model = DuelingDQN(self.env).to(self.device)
self.tgt_model.load_state_dict(self.model.state_dict())
self.optimizer = torch.optim.RMSprop(self.model.parameters(),
self.lr,
alpha=0.95,
eps=1.5e-7,
centered=True)
self.scheduler = torch.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 + frame_idx * (1.0 - self.prior_beta) / 1000)
self.batch_recorder = BatchRecorder(env_id=env_id,
env_seed=seed,
n_workers=n_workers,
buffer=self.buffer,
n_steps=n_step,
gamma=gamma,
max_episode_length=50000)
self.writer = SummaryWriter(
comment="-{}-learner".format(self.env.unwrapped.spec.id))
def train(self):
utils.set_global_seeds(self.seed, use_torch=True)
learn_idx = 0
while True:
beta = self.beta_by_frame(learn_idx)
states, actions, rewards, next_states, dones, weights, idxes = self.buffer.sample(
self.batch_size, beta)
states = torch.FloatTensor(states).to(self.device)
actions = torch.LongTensor(actions).to(self.device)
rewards = torch.FloatTensor(rewards).to(self.device)
next_states = torch.FloatTensor(next_states).to(self.device)
dones = torch.FloatTensor(dones).to(self.device)
weights = torch.FloatTensor(weights).to(self.device)
batch = (states, actions, rewards, next_states, dones, weights)
loss, prios = utils.compute_loss(self.model, self.tgt_model, batch,
self.n_step, self.gamma)
self.scheduler.step()
grad_norm = utils.update_parameters(loss, self.model,
self.optimizer, self.max_norm)
self.buffer.update_priorities(idxes, prios)
batch, idxes, prios = None, None, None
learn_idx += 1
self.writer.add_scalar("learner/loss", loss, learn_idx)
self.writer.add_scalar("learner/grad_norm", grad_norm, learn_idx)
if learn_idx % self.target_update_interval == 0:
print("Updating Target Network..")
self.tgt_model.load_state_dict(self.model.state_dict())
if learn_idx % self.save_interval == 0:
print("Saving Model..")
torch.save(self.model.state_dict(),
"model{}.pth".format(learn_idx))
if learn_idx % self.publish_param_interval == 0:
self.batch_recorder.set_worker_weights(
copy.deepcopy(self.model))
if learn_idx >= self.max_step:
torch.save(self.model.state_dict(),
"model{}.pth".format(learn_idx))
self.batch_recorder.cleanup()
break
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 sampling_data(self):
self.batch_recorder.record_batch()
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
seed,
gamma=GAMMA,
buffer_size=BUFFER_SIZE,
batch_size=BATCH_SIZE,
update_every=UPDATE_EVERY,
lr=LR,
tau=TAU
):
"""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)
self.gamma = gamma
self.batch_size = batch_size
# Q-Network
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.model_local = DuelingDQN(state_size, action_size, seed).to(self.device)
self.model_target = DuelingDQN(state_size, action_size, seed).to(self.device)
self.optimizer = optim.Adam(self.model_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(
action_size=action_size,
buffer_size=BUFFER_SIZE,
batch_size=BATCH_SIZE,
seed=seed,
device=self.device
)
# 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) > self.batch_size:
experiences = self.memory.sample()
self.update(experiences)
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.FloatTensor(state).float().unsqueeze(0).to(self.device)
self.model_local.eval()
with torch.no_grad():
qvals = self.model_local.forward(state)
self.model_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
action = np.argmax(qvals.cpu().detach().numpy())
return action
else:
return random.choice(np.arange(self.action_size))
def update(self, batch):
"""Update value parameters using given batch of experience tuples.
Params
======
batch (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = batch
# Get expected Q values from local model
curr_Q = self.model_local.forward(states).gather(1, actions)
# curr_Q = curr_Q.squeeze(1)
# Get max predicted Q values (for next states) from target model
max_next_Q = self.model_target.forward(next_states).detach().max(1)[0].unsqueeze(1)
expected_Q = rewards + (self.gamma * max_next_Q * (1 - dones))
loss = F.mse_loss(curr_Q, expected_Q)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# update target model
self.update_target(self.model_local, self.model_target, TAU)
def update_target(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)
def train(args, n_actors, batch_queue, prios_queue, param_queue):
"""
thread to fill parameter queue
"""
def _fill_param():
while True:
model_dict = {}
state_dict = model.state_dict()
for k, v in state_dict.items():
model_dict[k] = v.cpu().numpy()
param_queue.put(model_dict)
env = wrapper.make_atari(args.env)
env = wrapper.wrap_atari_dqn(env, args)
utils.set_global_seeds(args.seed, use_torch=True)
model = DuelingDQN(env).to(args.device)
tgt_model = DuelingDQN(env).to(args.device)
tgt_model.load_state_dict(model.state_dict())
writer = SummaryWriter(comment="-{}-learner".format(args.env))
# optimizer = torch.optim.Adam(model.parameters(), args.lr)
optimizer = torch.optim.RMSprop(model.parameters(),
args.lr,
alpha=0.95,
eps=1.5e-7,
centered=True)
model_dict = {}
state_dict = model.state_dict()
for k, v in state_dict.items():
model_dict[k] = v.cpu().numpy()
param_queue.put(model_dict)
threading.Thread(target=_fill_param).start()
learn_idx = 0
ts = time.time()
while True:
#if batch_queue.empty():
# print("batch queue size:{}".format(batch_queue.qsize()))
*batch, idxes = batch_queue.get()
loss, prios = utils.compute_loss(model, tgt_model, batch, args.n_steps,
args.gamma)
grad_norm = utils.update_parameters(loss, model, optimizer,
args.max_norm)
prios_queue.put((idxes, prios))
batch, idxes, prios = None, None, None
learn_idx += 1
if learn_idx % args.tensorboard_update_interval == 0:
writer.add_scalar("learner/loss", loss, learn_idx)
writer.add_scalar("learner/grad_norm", grad_norm, learn_idx)
if learn_idx % args.target_update_interval == 0:
print("Updating Target Network..")
tgt_model.load_state_dict(model.state_dict())
if learn_idx % args.save_interval == 0:
print("Saving Model..")
torch.save(model.state_dict(), "model.pth")
if learn_idx % args.publish_param_interval == 0:
param_queue.get()
if learn_idx % args.bps_interval == 0:
bps = args.bps_interval / (time.time() - ts)
print("Step: {:8} / BPS: {:.2f}".format(learn_idx, bps))
writer.add_scalar("learner/BPS", bps, learn_idx)
ts = time.time()