def sac_opp(
global_ac,
global_ac_targ,
global_cpc,
rank,
T,
E,
args,
scores,
wins,
buffer,
device=None,
tensorboard_dir=None,
):
torch.manual_seed(args.seed + rank)
np.random.seed(args.seed + rank)
# writer = GlobalSummaryWriter.getSummaryWriter()
tensorboard_dir = os.path.join(tensorboard_dir, str(rank))
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
if args.exp_name == "test":
env = gym.make("CartPole-v0")
elif args.non_station:
env = make_ftg_ram_nonstation(args.env,
p2_list=args.list,
total_episode=args.station_rounds,
stable=args.stable)
else:
env = make_ftg_ram(args.env, p2=args.p2)
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.n
print("set up child process env")
local_ac = MLPActorCritic(obs_dim + args.c_dim, act_dim,
**dict(hidden_sizes=[args.hid] *
args.l)).to(device)
local_ac.load_state_dict(global_ac.state_dict())
print("local ac load global ac")
# Freeze target networks with respect to optimizers (only update via polyak averaging)
# Async Version
for p in global_ac_targ.parameters():
p.requires_grad = False
# Experience buffer
if args.cpc:
replay_buffer = ReplayBufferOppo(obs_dim=obs_dim,
max_size=args.replay_size,
encoder=global_cpc)
else:
replay_buffer = ReplayBuffer(obs_dim=obs_dim, size=args.replay_size)
# Entropy Tuning
target_entropy = -torch.prod(
torch.Tensor(env.action_space.shape).to(
device)).item() # heuristic value from the paper
alpha = max(local_ac.log_alpha.exp().item(),
args.min_alpha) if not args.fix_alpha else args.min_alpha
# Set up optimizers for policy and q-function
# Async Version
pi_optimizer = Adam(global_ac.pi.parameters(), lr=args.lr, eps=1e-4)
q1_optimizer = Adam(global_ac.q1.parameters(), lr=args.lr, eps=1e-4)
q2_optimizer = Adam(global_ac.q2.parameters(), lr=args.lr, eps=1e-4)
cpc_optimizer = Adam(global_cpc.parameters(), lr=args.lr, eps=1e-4)
alpha_optim = Adam([global_ac.log_alpha], lr=args.lr, eps=1e-4)
# Prepare for interaction with environment
o, ep_ret, ep_len = env.reset(), 0, 0
if args.cpc:
c_hidden = global_cpc.init_hidden(1, args.c_dim, use_gpu=args.cuda)
c1, c_hidden = global_cpc.predict(o, c_hidden)
assert len(c1.shape) == 3
c1 = c1.flatten().cpu().numpy()
all_embeddings = []
meta = []
trajectory = list()
p2 = env.p2
p2_list = [str(p2)]
discard = False
uncertainties = []
local_t, local_e = 0, 0
t = T.value()
e = E.value()
glod_input = list()
glod_target = list()
# Main loop: collect experience in env and update/log each epoch
while e <= args.episode:
with torch.no_grad():
# Until start_steps have elapsed, randomly sample actions
# from a uniform distribution for better exploration. Afterwards,
# use the learned policy.
if t > args.start_steps:
if args.cpc:
a = local_ac.get_action(np.concatenate((o, c1), axis=0),
device=device)
a_prob = local_ac.act(
torch.as_tensor(np.expand_dims(np.concatenate((o, c1),
axis=0),
axis=0),
dtype=torch.float32,
device=device))
else:
a = local_ac.get_action(o, greedy=True, device=device)
a_prob = local_ac.act(
torch.as_tensor(np.expand_dims(o, axis=0),
dtype=torch.float32,
device=device))
else:
a = env.action_space.sample()
a_prob = local_ac.act(
torch.as_tensor(np.expand_dims(o, axis=0),
dtype=torch.float32,
device=device))
uncertainty = ood_scores(a_prob).item()
# Step the env
o2, r, d, info = env.step(a)
if info.get('no_data_receive', False):
discard = True
ep_ret += r
ep_len += 1
# Ignore the "done" signal if it comes from hitting the time
# horizon (that is, when it's an artificial terminal signal
# that isn't based on the agent's state)
d = False if (ep_len == args.max_ep_len) or discard else d
glod_input.append(o), glod_target.append(a)
if args.cpc:
# changed the trace structure for further analysis
c2, c_hidden = global_cpc.predict(o2, c_hidden)
assert len(c2.shape) == 3
c2 = c2.flatten().cpu().numpy()
replay_buffer.store(np.concatenate((o, c1), axis=0), a, r,
np.concatenate((o2, c2), axis=0), d)
trajectory.append([o, a, r, o2, d, c1, c2, ep_len])
all_embeddings.append(c1)
meta.append([env.p2, local_e, ep_len, r, a, uncertainty])
c1 = c2
trajectory.append([o, a, r, o2, d, c1, c2])
else:
replay_buffer.store(o, a, r, o2, d)
# Super critical, easy to overlook step: make sure to update
# most recent observation!
o = o2
T.increment()
t = T.value()
local_t += 1
# End of trajectory handling
if d or (ep_len == args.max_ep_len) or discard:
replay_buffer.store(trajectory)
E.increment()
e = E.value()
local_e += 1
# logger.store(EpRet=ep_ret, EpLen=ep_len)
if info.get('win', False):
wins.append(1)
else:
wins.append(0)
scores.append(ep_ret)
m_score = np.mean(scores[-100:])
win_rate = np.mean(wins[-100:])
print(
"Process {}, opponent:{}, # of global_episode :{}, # of global_steps :{}, round score: {}, mean score : {:.1f}, win_rate:{}, steps: {}, alpha: {}"
.format(rank, args.p2, e, t, ep_ret, m_score, win_rate, ep_len,
alpha))
writer.add_scalar("metrics/round_score", ep_ret, e)
writer.add_scalar("metrics/mean_score", m_score.item(), e)
writer.add_scalar("metrics/win_rate", win_rate.item(), e)
writer.add_scalar("metrics/round_step", ep_len, e)
writer.add_scalar("metrics/alpha", alpha, e)
# CPC update handing
if local_e > args.batch_size and local_e % args.update_every == 0 and args.cpc:
data, indexes, min_len = replay_buffer.sample_traj(
args.batch_size)
global_cpc.train()
cpc_optimizer.zero_grad()
c_hidden = global_cpc.init_hidden(len(data),
args.c_dim,
use_gpu=args.cuda)
acc, loss, latents = global_cpc(data, c_hidden)
replay_buffer.update_latent(indexes, min_len, latents.detach())
loss.backward()
# add gradient clipping
nn.utils.clip_grad_norm_(global_cpc.parameters(), 20)
cpc_optimizer.step()
writer.add_scalar("training/acc", acc, e)
writer.add_scalar("training/cpc_loss", loss.detach().item(), e)
all_embeddings = np.array(all_embeddings)
writer.add_embedding(mat=all_embeddings,
metadata=meta,
metadata_header=[
"opponent", "round", "step", "reward",
"action", "uncertainty"
])
c_hidden = global_cpc.init_hidden(1,
args.c_dim,
use_gpu=args.cuda)
o, ep_ret, ep_len = env.reset(), 0, 0
trajectory = list()
discard = False
# OOD update stage
if (t >= args.ood_update_step and local_t % args.ood_update_step == 0
or replay_buffer.is_full()) and args.ood:
# used all the data collected from the last args.ood_update_steps as the train data
print("Conduct OOD updating")
ood_train = (glod_input, glod_target)
glod_model = convert_to_glod(global_ac.pi,
train_loader=ood_train,
hidden_dim=args.hid,
act_dim=act_dim,
device=device)
glod_scores = retrieve_scores(
glod_model,
replay_buffer.obs_buf[:replay_buffer.size],
device=device,
k=args.ood_K)
glod_scores = glod_scores.detach().cpu().numpy()
print(len(glod_scores))
writer.add_histogram(values=glod_scores,
max_bins=300,
global_step=local_t,
tag="OOD")
drop_points = np.percentile(
a=glod_scores, q=[args.ood_drop_lower, args.ood_drop_upper])
lower, upper = drop_points[0], drop_points[1]
print(lower, upper)
mask = np.logical_and((glod_scores >= lower),
(glod_scores <= upper))
reserved_indexes = np.argwhere(mask).flatten()
print(len(reserved_indexes))
if len(reserved_indexes) > 0:
replay_buffer.ood_drop(reserved_indexes)
glod_input = list()
glod_target = list()
# SAC Update handling
if local_t >= args.update_after and local_t % args.update_every == 0:
for j in range(args.update_every):
batch = replay_buffer.sample_trans(batch_size=args.batch_size,
device=device)
# First run one gradient descent step for Q1 and Q2
q1_optimizer.zero_grad()
q2_optimizer.zero_grad()
loss_q = local_ac.compute_loss_q(batch, global_ac_targ,
args.gamma, alpha)
loss_q.backward()
# Next run one gradient descent step for pi.
pi_optimizer.zero_grad()
loss_pi, entropy = local_ac.compute_loss_pi(batch, alpha)
loss_pi.backward()
alpha_optim.zero_grad()
alpha_loss = -(local_ac.log_alpha *
(entropy + target_entropy).detach()).mean()
alpha_loss.backward(retain_graph=False)
alpha = max(
local_ac.log_alpha.exp().item(),
args.min_alpha) if not args.fix_alpha else args.min_alpha
nn.utils.clip_grad_norm_(local_ac.parameters(), 20)
for global_param, local_param in zip(global_ac.parameters(),
local_ac.parameters()):
global_param._grad = local_param.grad
pi_optimizer.step()
q1_optimizer.step()
q2_optimizer.step()
alpha_optim.step()
state_dict = global_ac.state_dict()
local_ac.load_state_dict(state_dict)
# Finally, update target networks by polyak averaging.
with torch.no_grad():
for p, p_targ in zip(global_ac.parameters(),
global_ac_targ.parameters()):
p_targ.data.copy_((1 - args.polyak) * p.data +
args.polyak * p_targ.data)
writer.add_scalar("training/pi_loss",
loss_pi.detach().item(), t)
writer.add_scalar("training/q_loss", loss_q.detach().item(), t)
writer.add_scalar("training/alpha_loss",
alpha_loss.detach().item(), t)
writer.add_scalar("training/entropy",
entropy.detach().mean().item(), t)
if t % args.save_freq == 0 and t > 0:
torch.save(
global_ac.state_dict(),
os.path.join(args.save_dir, args.exp_name, args.model_para))
torch.save(
global_cpc.state_dict(),
os.path.join(args.save_dir, args.exp_name, args.cpc_para))
state_dict_trans(
global_ac.state_dict(),
os.path.join(args.save_dir, args.exp_name, args.numpy_para))
torch.save((e, t, list(scores), list(wins)),
os.path.join(args.save_dir, args.exp_name,
args.train_indicator))
print("Saving model at episode:{}".format(t))
def train(global_model, rank, T, scores):
env = make_ftg_ram(env_name, p2=p2)
state_shape = env.observation_space.shape[0]
action_shape = env.action_space.n
local_model = ActorCritic(state_shape, action_shape, hidden_size)
local_model.load_state_dict(global_model.state_dict())
# MODEL_STATE = "/home/byron/Repos/FTG4.50/OpenAI/ByronAI.numpy"
# test_model = ActorCriticNumpy(MODEL_STATE)
optimizer = optim.Adam(global_model.parameters(), lr=learning_rate)
while True:
discard = False
done = False
s = env.reset()
score = 0
sum_entropy = 0
step = 0
while not done:
s_lst, a_lst, r_lst = [], [], []
for t in range(update_interval):
prob = local_model.pi(torch.from_numpy(s).float())
# test_prob = test_model.pi(s)
# diff = prob.detach().numpy() - test_prob
# print(diff.sum())
m = Categorical(prob)
a = m.sample().item()
s_prime, r, done, info = env.step(a)
if info.get('no_data_receive', False):
discard = True
break
s_lst.append(s)
a_lst.append([a])
# r_lst.append(r/100.0)
r_lst.append(r)
s = s_prime
score += r
sum_entropy += Categorical(probs=prob.detach()).entropy()
step += 1
if done:
break
if discard:
break
s_final = torch.tensor(s_prime, dtype=torch.float)
R = 0.0 if done else local_model.v(s_final).item()
td_target_lst = []
for reward in r_lst[::-1]:
R = gamma * R + reward
td_target_lst.append([R])
td_target_lst.reverse()
s_batch, a_batch, td_target = torch.tensor(s_lst, dtype=torch.float), torch.tensor(a_lst), \
torch.tensor(td_target_lst)
advantage = td_target - local_model.v(s_batch)
pi = local_model.pi(s_batch, softmax_dim=1)
pi_a = pi.gather(1, a_batch)
loss = -torch.log(pi_a) * advantage.detach() + \
F.smooth_l1_loss(local_model.v(s_batch), td_target.detach()) - \
(entropy_weight * -(torch.log(pi) * pi).sum())
optimizer.zero_grad()
loss.mean().backward()
for global_param, local_param in zip(global_model.parameters(),
local_model.parameters()):
global_param._grad = local_param.grad
optimizer.step()
local_model.load_state_dict(global_model.state_dict())
if discard:
continue
T.increment()
t = T.value()
scores.append(score)
m_score = np.mean(scores[-100:])
print(
"Process {}, # of episode :{}, round score: {}, mean score : {:.1f}, entropy: {}, steps: {}"
.format(rank, t, score, m_score, sum_entropy / step, step))
if t % save_interval == 0 and t > 0:
torch.save(global_model.state_dict(),
os.path.join(save_dir, "model"))
state_dict_trans(global_model.state_dict(),
os.path.join(save_dir, numpy_para))
print("Saving model at episode:{}".format(t))
def sac(
global_ac,
global_ac_targ,
rank,
T,
E,
args,
scores,
wins,
buffer_q,
device=None,
tensorboard_dir=None,
):
torch.manual_seed(args.seed + rank)
np.random.seed(args.seed + rank)
# writer = GlobalSummaryWriter.getSummaryWriter()
tensorboard_dir = os.path.join(tensorboard_dir, str(rank))
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# if args.exp_name == "test":
# env = gym.make("CartPole-v0")
# elif args.non_station:
# env = make_ftg_ram_nonstation(args.env, p2_list=args.list, total_episode=args.station_rounds,stable=args.stable)
# else:
# env = make_ftg_ram(args.env, p2=args.p2)
# obs_dim = env.observation_space.shape[0]
# act_dim = env.action_space.n
env = Soccer()
# env = gym.make("CartPole-v0")
obs_dim = env.n_features
act_dim = env.n_actions
print("set up child process env")
local_ac = MLPActorCritic(obs_dim, act_dim,
**dict(hidden_sizes=[args.hid] *
args.l)).to(device)
state_dict = global_ac.state_dict()
local_ac.load_state_dict(state_dict)
print("local ac load global ac")
# Freeze target networks with respect to optimizers (only update via polyak averaging)
# Async Version
for p in global_ac_targ.parameters():
p.requires_grad = False
# Experience buffer
replay_buffer = ReplayBuffer(obs_dim=obs_dim, size=args.replay_size)
training_buffer = ReplayBuffer(obs_dim=obs_dim, size=args.replay_size)
# Entropy Tuning
target_entropy = -np.log((1.0 / act_dim)) * 0.5
alpha = max(local_ac.log_alpha.exp().item(),
args.min_alpha) if not args.fix_alpha else args.min_alpha
# Set up optimizers for policy and q-function
# Async Version
pi_optimizer = Adam(global_ac.pi.parameters(), lr=args.lr, eps=1e-4)
q1_optimizer = Adam(global_ac.q1.parameters(), lr=args.lr, eps=1e-4)
q2_optimizer = Adam(global_ac.q2.parameters(), lr=args.lr, eps=1e-4)
alpha_optim = Adam([global_ac.log_alpha], lr=args.lr, eps=1e-4)
# Prepare for interaction with environment
o, ep_ret, ep_len = env.reset(), 0, 0
discard = False
glod_model = None
glod_lower = None
glod_upper = None
last_updated = 0
saved_e = 0
t = T.value()
e = E.value()
local_t, local_e = 0, 0
# Main loop: collect experience in env and update/log each epoch
while e <= args.episode:
# Until start_steps have elapsed, randomly sample actions
# from a uniform distribution for better exploration. Afterwards,
# use the learned policy.
with torch.no_grad():
a = local_ac.get_action(o, device=device)
if hasattr(env, 'p2'):
p2 = env.p2
else:
p2 = None
# Step the env
o2, r, d, info = env.step(a, np.random.randint(act_dim))
env.render()
if info.get('no_data_receive', False):
discard = True
ep_ret += r
ep_len += 1
# Ignore the "done" signal if it comes from hitting the time
# horizon (that is, when it's an artificial terminal signal
# that isn't based on the agent's state)
# d = False if (ep_len == args.max_ep_len) or discard else d
# Store experience to replay buffer
if glod_model is None or not args.ood:
replay_buffer.store(o, a, r, o2, d, str(p2))
training_buffer.store(o, a, r, o2, d, str(p2))
else:
obs_glod_score = retrieve_scores(glod_model,
np.expand_dims(o, axis=0),
device=torch.device("cpu"),
k=args.ood_K)
if glod_lower <= obs_glod_score <= glod_upper:
training_buffer.store(o, a, r, o2, d, str(p2))
replay_buffer.store(o, a, r, o2, d, str(p2))
# Super critical, easy to overlook step: make sure to update
# most recent observation!
o = o2
T.increment()
t = T.value()
local_t += 1
# End of trajectory handling
if d or (ep_len == args.max_ep_len) or discard:
E.increment()
e = E.value()
local_e += 1
# logger.store(EpRet=ep_ret, EpLen=ep_len)
if info.get('win', False):
wins.append(1)
else:
wins.append(0)
scores.append(ep_ret)
m_score = np.mean(scores[-100:])
win_rate = np.mean(wins[-100:])
print(
"Process {}, opponent:{}, # of global_episode :{}, # of global_steps :{}, round score: {}, mean score : {:.1f}, win_rate:{}, steps: {}, alpha: {}"
.format(rank, args.p2, e, t, ep_ret, m_score, win_rate, ep_len,
alpha))
writer.add_scalar("metrics/round_score", ep_ret, e)
writer.add_scalar("metrics/mean_score", m_score.item(), e)
writer.add_scalar("metrics/win_rate", win_rate.item(), e)
writer.add_scalar("metrics/round_step", ep_len, e)
writer.add_scalar("metrics/alpha", alpha, e)
o, ep_ret, ep_len = env.reset(), 0, 0
discard = False
# OOD update stage, can only use CPU as the GPU memory can not hold so much data
if local_e >= args.ood_starts and local_e % args.ood_update_rounds == 0 and args.ood and local_e != last_updated:
print("OOD updating at rounds {}".format(e))
print("Replay Buffer Size: {}, Training Buffer Size: {}".format(
replay_buffer.size, training_buffer.size))
glod_idxs = np.random.randint(0,
training_buffer.size,
size=int(training_buffer.size *
args.ood_train_per))
glod_input = training_buffer.obs_buf[glod_idxs]
glod_target = training_buffer.act_buf[glod_idxs]
ood_train = (glod_input, glod_target)
glod_model = deepcopy(global_ac.pi).cpu()
glod_model = convert_to_glod(glod_model,
train_loader=ood_train,
hidden_dim=args.hid,
act_dim=act_dim,
device=torch.device("cpu"))
training_buffer = deepcopy(replay_buffer)
glod_scores = retrieve_scores(
glod_model,
replay_buffer.obs_buf[:training_buffer.size],
device=torch.device("cpu"),
k=args.ood_K)
glod_scores = glod_scores.detach().cpu().numpy()
glod_p2 = training_buffer.p2_buf[:training_buffer.size]
drop_points = np.percentile(
a=glod_scores, q=[args.ood_drop_lower, args.ood_drop_upper])
glod_lower, glod_upper = drop_points[0], drop_points[1]
mask = np.logical_and((glod_scores >= glod_lower),
(glod_scores <= glod_upper))
reserved_indexes = np.argwhere(mask).flatten()
if len(reserved_indexes) > 0:
training_buffer.ood_drop(reserved_indexes)
writer.add_histogram(values=glod_scores,
max_bins=300,
global_step=local_e,
tag="OOD")
print("Replay Buffer Size: {}, Training Buffer Size: {}".format(
replay_buffer.size, training_buffer.size))
torch.save(
(glod_scores, replay_buffer.p2_buf[:replay_buffer.size]),
os.path.join(args.save_dir, args.exp_name,
"glod_info_{}_{}".format(rank, local_e)))
last_updated = local_e
# SAC Update handling
if local_e >= args.update_after and local_t % args.update_every == 0:
for j in range(args.update_every):
batch = training_buffer.sample_trans(args.batch_size,
device=device)
# First run one gradient descent step for Q1 and Q2
q1_optimizer.zero_grad()
q2_optimizer.zero_grad()
loss_q = local_ac.compute_loss_q(batch, global_ac_targ,
args.gamma, alpha)
loss_q.backward()
nn.utils.clip_grad_norm_(global_ac.parameters(),
max_norm=20,
norm_type=2)
q1_optimizer.step()
q2_optimizer.step()
# Next run one gradient descent step for pi.
pi_optimizer.zero_grad()
loss_pi, entropy = local_ac.compute_loss_pi(batch, alpha)
loss_pi.backward()
nn.utils.clip_grad_norm_(global_ac.parameters(),
max_norm=20,
norm_type=2)
pi_optimizer.step()
alpha_optim.zero_grad()
alpha_loss = -(local_ac.log_alpha *
(entropy + target_entropy).detach()).mean()
alpha_loss.backward(retain_graph=False)
alpha = max(
local_ac.log_alpha.exp().item(),
args.min_alpha) if not args.fix_alpha else args.min_alpha
nn.utils.clip_grad_norm_(global_ac.parameters(),
max_norm=20,
norm_type=2)
alpha_optim.step()
for global_param, local_param in zip(global_ac.parameters(),
local_ac.parameters()):
global_param._grad = local_param.grad
state_dict = global_ac.state_dict()
local_ac.load_state_dict(state_dict)
# Finally, update target networks by polyak averaging.
with torch.no_grad():
for p, p_targ in zip(global_ac.parameters(),
global_ac_targ.parameters()):
p_targ.data.copy_((1 - args.polyak) * p.data +
args.polyak * p_targ.data)
writer.add_scalar("training/pi_loss",
loss_pi.detach().item(), t)
writer.add_scalar("training/q_loss", loss_q.detach().item(), t)
writer.add_scalar("training/alpha_loss",
alpha_loss.detach().item(), t)
writer.add_scalar("training/entropy",
entropy.detach().mean().item(), t)
if e % args.save_freq == 0 and e > 0 and e != saved_e:
torch.save(
global_ac.state_dict(),
os.path.join(args.save_dir, args.exp_name,
"model_torch_{}".format(e)))
state_dict_trans(
global_ac.state_dict(),
os.path.join(args.save_dir, args.exp_name,
"model_numpy_{}".format(e)))
torch.save((e, t, list(scores), list(wins)),
os.path.join(args.save_dir, args.exp_name,
"model_data_{}".format(e)))
print("Saving model at episode:{}".format(t))
saved_e = e
def sac(env_fn, actor_critic=MLPActorCritic, ac_kwargs=dict(), seed=0,
steps_per_epoch=4000, epochs=100, replay_size=int(1e6), gamma=0.99,
polyak=0.995, lr=1e-3, alpha=0.2, batch_size=100, start_steps=10000,
update_after=1000, update_every=50, num_test_episodes=10, max_ep_len=1000,
logger_kwargs=dict(), save_freq=1000, save_dir=None):
"""
Soft Actor-Critic (SAC)
Args:
env_fn : A function which creates a copy of the environment.
The environment must satisfy the OpenAI Gym API.
actor_critic: The constructor method for a PyTorch Module with an ``act``
method, a ``pi`` module, a ``q1`` module, and a ``q2`` module.
The ``act`` method and ``pi`` module should accept batches of
observations as inputs, and ``q1`` and ``q2`` should accept a batch
of observations and a batch of actions as inputs. When called,
``act``, ``q1``, and ``q2`` should return:
=========== ================ ======================================
Call Output Shape Description
=========== ================ ======================================
``act`` (batch, act_dim) | Numpy array of actions for each
| observation.
``q1`` (batch,) | Tensor containing one current estimate
| of Q* for the provided observations
| and actions. (Critical: make sure to
| flatten this!)
``q2`` (batch,) | Tensor containing the other current
| estimate of Q* for the provided observations
| and actions. (Critical: make sure to
| flatten this!)
=========== ================ ======================================
Calling ``pi`` should return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``a`` (batch, act_dim) | Tensor containing actions from policy
| given observations.
``logp_pi`` (batch,) | Tensor containing log probabilities of
| actions in ``a``. Importantly: gradients
| should be able to flow back into ``a``.
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the ActorCritic object
you provided to SAC.
seed (int): Seed for random number generators.
steps_per_epoch (int): Number of steps of interaction (state-action pairs)
for the agent and the environment in each epoch.
epochs (int): Number of epochs to run and train agent.
replay_size (int): Maximum length of replay buffer.
gamma (float): Discount factor. (Always between 0 and 1.)
polyak (float): Interpolation factor in polyak averaging for target
networks. Target networks are updated towards main networks
according to:
.. math:: \\theta_{\\text{targ}} \\leftarrow
\\rho \\theta_{\\text{targ}} + (1-\\rho) \\theta
where :math:`\\rho` is polyak. (Always between 0 and 1, usually
close to 1.)
lr (float): Learning rate (used for both policy and value learning).
alpha (float): Entropy regularization coefficient. (Equivalent to
inverse of reward scale in the original SAC paper.)
batch_size (int): Minibatch size for SGD.
start_steps (int): Number of steps for uniform-random action selection,
before running real policy. Helps exploration.
update_after (int): Number of env interactions to collect before
starting to do gradient descent updates. Ensures replay buffer
is full enough for useful updates.
update_every (int): Number of env interactions that should elapse
between gradient descent updates. Note: Regardless of how long
you wait between updates, the ratio of env steps to gradient steps
is locked to 1.
num_test_episodes (int): Number of episodes to test the deterministic
policy at the end of each epoch.
max_ep_len (int): Maximum length of trajectory / episode / rollout.
logger_kwargs (dict): Keyword args for EpochLogger.
save_freq (int): How often (in terms of gap between epochs) to save
the current policy and value function.
"""
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
torch.manual_seed(seed)
np.random.seed(seed)
env, test_env = env_fn(), env_fn()
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.n
# Action limit for clamping: critically, assumes all dimensions share the same bound!
# act_limit = env.action_space.high[0]
# Create actor-critic module and target networks
ac = actor_critic(env.observation_space, env.action_space, **ac_kwargs)
ac_targ = deepcopy(ac)
# Freeze target networks with respect to optimizers (only update via polyak averaging)
for p in ac_targ.parameters():
p.requires_grad = False
# List of parameters for both Q-networks (save this for convenience)
# Experience buffer
replay_buffer = ReplayBuffer(obs_dim=obs_dim, size=replay_size)
# Count variables (protip: try to get a feel for how different size networks behave!)
var_counts = tuple(count_vars(module) for module in [ac.pi, ac.q1, ac.q2])
logger.log('\nNumber of parameters: \t pi: %d, \t q1: %d, \t q2: %d\n' % var_counts)
# Set up optimizers for policy and q-function
pi_optimizer = Adam(ac.pi.parameters(), lr=lr)
q1_optimizer = Adam(ac.q1.parameters(), lr=lr)
q2_optimizer = Adam(ac.q2.parameters(), lr=lr)
# Set up model saving
logger.setup_pytorch_saver(ac)
# product action
def get_actions_info(a_prob):
a_dis = Categorical(a_prob)
max_a = torch.argmax(a_prob)
sample_a = a_dis.sample().cpu()
z = a_prob == 0.0
z = z.float() * 1e-8
log_a_prob = torch.log(a_prob + z)
return a_prob, log_a_prob, sample_a, max_a
# Set up function for computing SAC Q-losses
def compute_loss_q(data):
o, a, r, o2, d = data['obs'], data['act'], data['rew'], data['obs2'], data['done']
# Bellman backup for Q functions
with torch.no_grad():
# Target actions come from *current* policy
a_prob, log_a_prob, sample_a, max_a = get_actions_info(ac.pi(o2))
# Target Q-values
q1_pi_targ = ac_targ.q1(o2)
q2_pi_targ = ac_targ.q2(o2)
q_pi_targ = torch.min(q1_pi_targ, q2_pi_targ)
backup = r + gamma * (1 - d) * (a_prob * (q_pi_targ - alpha * log_a_prob)).sum(dim=1)
# MSE loss against Bellman backup
q1 = ac.q1(o).gather(1, a.unsqueeze(-1).long())
q2 = ac.q2(o).gather(1, a.unsqueeze(-1).long())
loss_q1 = F.mse_loss(q1, backup.unsqueeze(-1))
loss_q2 = F.mse_loss(q2, backup.unsqueeze(-1))
loss_q = loss_q1 + loss_q2
# Useful info for logging
q_info = dict(Q1Vals=q1.detach().numpy(),
Q2Vals=q2.detach().numpy())
return loss_q, q_info
# Set up function for computing SAC pi loss
def compute_loss_pi(data):
o = data['obs']
a_prob, log_a_prob, sample_a, max_a = get_actions_info(ac.pi(o))
q1_pi = ac.q1(o)
q2_pi = ac.q2(o)
q_pi = torch.min(q1_pi, q2_pi)
# Entropy-regularized policy loss
loss_pi = (a_prob * (alpha * log_a_prob - q_pi)).mean()
entropy = torch.sum(log_a_prob * a_prob, dim=1).detach()
# Useful info for logging
pi_info = dict(LogPi=entropy.numpy())
return loss_pi, pi_info
def update(data):
# First run one gradient descent step for Q1 and Q2
q1_optimizer.zero_grad()
q2_optimizer.zero_grad()
loss_q, q_info = compute_loss_q(data)
loss_q.backward()
q1_optimizer.step()
q2_optimizer.step()
# Record things
logger.store(LossQ=loss_q.detach().item(), **q_info)
# Freeze Q-networks so you don't waste computational effort
# computing gradients for them during the policy learning step.
# for p in q_params:
# p.requires_grad = False
# Next run one gradient descent step for pi.
pi_optimizer.zero_grad()
loss_pi, pi_info = compute_loss_pi(data)
loss_pi.backward()
pi_optimizer.step()
# Unfreeze Q-networks so you can optimize it at next DDPG step.
# for p in q_params:
# p.requires_grad = True
# Record things
logger.store(LossPi=loss_pi.item(), **pi_info)
# Finally, update target networks by polyak averaging.
with torch.no_grad():
for p, p_targ in zip(ac.parameters(), ac_targ.parameters()):
# NB: We use an in-place operations "mul_", "add_" to update target
# params, as opposed to "mul" and "add", which would make new tensors.
p_targ.data.copy_((1 - polyak) * p.data + polyak * p_targ.data)
# p_targ.data.mul_(polyak)
# p_targ.data.add_((1 - polyak) * p.data)
def get_action(o, greedy=False):
if len(o.shape) == 1:
o = np.expand_dims(o, axis=0)
a_prob = ac.act(torch.as_tensor(o, dtype=torch.float32), greedy)
a_prob, log_a_prob, sample_a, max_a = get_actions_info(a_prob)
action = sample_a if not greedy else max_a
return action.item()
def test_agent():
for j in range(num_test_episodes):
o, d, ep_ret, ep_len = test_env.reset(), False, 0, 0
while not (d or (ep_len == max_ep_len)):
# Take deterministic actions at test time
o, r, d, _ = test_env.step(get_action(o, True))
# test_env.render()
ep_ret += r
ep_len += 1
logger.store(TestEpRet=ep_ret, TestEpLen=ep_len)
print(ep_ret)
# Prepare for interaction with environment
total_steps = steps_per_epoch * epochs
start_time = time.time()
scores = []
o, ep_ret, ep_len = env.reset(), 0, 0
discard = False
# Main loop: collect experience in env and update/log each epoch
for t in range(total_steps):
# Until start_steps have elapsed, randomly sample actions
# from a uniform distribution for better exploration. Afterwards,
# use the learned policy.
if t > start_steps:
a = get_action(o)
else:
a = env.action_space.sample()
# Step the env
o2, r, d, info = env.step(a)
if info.get('no_data_receive', False):
discard = True
env.render()
ep_ret += r
ep_len += 1
# Ignore the "done" signal if it comes from hitting the time
# horizon (that is, when it's an artificial terminal signal
# that isn't based on the agent's state)
d = False if ep_len == max_ep_len or discard else d
# Store experience to replay buffer
replay_buffer.store(o, a, r, o2, d)
# Super critical, easy to overlook step: make sure to update
# most recent observation!
o = o2
# End of trajectory handling
if d or (ep_len == max_ep_len) or discard:
logger.store(EpRet=ep_ret, EpLen=ep_len)
scores.append(ep_ret)
print("round len:{}, round score: {}, mean score: {}".format(ep_len, ep_ret, np.mean(scores[-100:])))
o, ep_ret, ep_len = env.reset(), 0, 0
discard = False
# Update handling
if t >= update_after and t % update_every == 0:
for j in range(update_every):
batch = replay_buffer.sample_batch(batch_size)
update(data=batch)
# End of epoch handling
if (t + 1) % steps_per_epoch == 0:
epoch = (t + 1) // steps_per_epoch
# Save model
if t % save_freq == 0 and t > 0:
torch.save(ac.state_dict(), os.path.join(save_dir, "model"))
state_dict_trans(ac.state_dict(), os.path.join(save_dir, "SAC_Toothless.numpy"))
print("Saving model at episode:{}".format(t))
if (epoch % save_freq == 0) or (epoch == epochs):
logger.save_state({'env': env}, None)