class MSAC(BasePolicy):
def __init__(
self,
model,
action_dim=1,
buffer_size=1000,
batch_size=100,
actor_learn_freq=1,
target_update_freq=5,
target_update_tau=0.1,
# learning_rate=1e-3,
actor_lr=1e-4,
critic_lr=1e-3,
discount_factor=0.99,
verbose=False,
update_iteration=10,
use_priority=False,
use_m=False,
n_step=1,
):
super().__init__()
# self.lr = learning_rate
self.eps = np.finfo(np.float32).eps.item()
self.tau = target_update_tau
self.actor_learn_freq = actor_learn_freq
self.target_update_freq = target_update_freq
self._gamma = discount_factor
self._target = target_update_freq > 0
self._update_iteration = update_iteration
self._sync_cnt = 0
# self._learn_cnt = 0
self._learn_critic_cnt = 0
self._learn_actor_cnt = 0
self._verbose = verbose
self._batch_size = batch_size
self.use_priority = use_priority
self.use_dist = model.value_net.use_dist
self.use_munchausen = use_m
self.n_step = n_step
if self.use_priority:
self.buffer = PriorityReplayBuffer(buffer_size, n_step=self.n_step)
else:
self.buffer = ReplayBuffer(buffer_size) # off-policy
if self.use_dist:
assert model.value_net.num_atoms > 1
# assert isinstance(model.value_net, CriticModelDist)
self.v_min = model.value_net.v_min
self.v_max = model.value_net.v_max
self.num_atoms = model.value_net.num_atoms
self.delta_z = (self.v_max - self.v_min) / (self.num_atoms - 1)
self.support = torch.linspace(self.v_min, self.v_max,
self.num_atoms)
self.actor_eval = model.policy_net.to(device).train()
self.critic_eval = model.value_net.to(device).train()
self.actor_target = self.copy_net(self.actor_eval)
self.critic_target = self.copy_net(self.critic_eval)
self.actor_eval_optim = optim.Adam(self.actor_eval.parameters(),
lr=actor_lr)
self.critic_eval_optim = optim.Adam(self.critic_eval.parameters(),
lr=critic_lr)
self.criterion = nn.SmoothL1Loss(reduction='none') # keep batch dim
self.target_entropy = -torch.tensor(action_dim).to(device)
self.log_alpha = torch.zeros(1, requires_grad=True, device=device)
self.alpha_optim = optim.Adam([self.log_alpha], lr=actor_lr)
self.alpha = self.log_alpha.exp()
def _tensor(self, data, use_cuda=False):
if np.array(data).ndim == 1:
data = torch.tensor(data, dtype=torch.float32).view(-1, 1)
else:
data = torch.tensor(data, dtype=torch.float32)
if use_cuda:
data = data.to(device)
return data
def learn_critic_dist(self, obs, act, rew, next_obs, mask):
with torch.no_grad():
next_act, next_log_pi = self.actor_target(next_obs)
# q(s, a) change to z(s, a) to discribe a distributional
p1_next, p2_next = self.critic_target.get_probs(
next_obs, next_act) # [batch_size, num_atoms]
p_next = torch.stack([
torch.where(z1.sum() < z2.sum(), z1, z2)
for z1, z2 in zip(p1_next, p2_next)
])
p_next -= (self.alpha * next_log_pi)
Tz = rew.unsqueeze(1) + mask * self.support.unsqueeze(0)
Tz = Tz.clamp(min=self.v_min, max=self.v_max)
b = (Tz - self.v_min) / self.delta_z
l, u = b.floor().to(torch.int64), b.ceil().to(torch.int64)
l[(u > 0) * (l == u)] -= 1
u[(l < (self.num_atoms - 1)) * (l == u)] += 1
m = obs.new_zeros(self._batch_size, self.num_atoms).cpu()
p_next = p_next.cpu()
# print (f'm device: {m.device}')
# print (f'p_next device: {p_next.device}')
offset = torch.linspace(0,
((self._batch_size - 1) * self.num_atoms),
self._batch_size).unsqueeze(1).expand(
self._batch_size, self.num_atoms).to(l)
m.view(-1).index_add_(
0, (l + offset).view(-1),
(p_next *
(u.float() - b)).view(-1)) # m_l = m_l + p(s_t+n, a*)(u - b)
m.view(-1).index_add_(
0, (u + offset).view(-1),
(p_next *
(b - l.float())).view(-1)) # m_u = m_u + p(s_t+n, a*)(b - l)
m = m.to(device)
log_z1, log_z2 = self.critic_eval.get_probs(obs, act, log=True)
loss1 = -(m * log_z1).sum(dim=1)
loss2 = -(m * log_z2).sum(dim=1)
return 0.5 * (loss1 + loss2)
def learn_critic(self, obs, act, rew, next_obs, mask):
with torch.no_grad():
next_A, next_log = self.actor_target.evaluate(next_obs)
# print (f'nextA shape is {next_A.shape}')
q1_next, q2_next = self.critic_target.twinQ(next_obs, next_A)
# print (f'shape q1 {q1_next.shape}, q2 {q2_next.shape}, next_obs {next_obs.shape}, next_A {next_A.shape}')
# print (f'q1_next shape is {q1_next.shape}')
# q_next = torch.stack([torch.where(q1.sum() < q2.sum(), q1, q2) for q1, q2 in zip(q1_next, q2_next)])
# print (f'shape stack q_next {q_next.shape} ')
q_next = torch.min(q1_next, q2_next) - self.alpha * next_log
# print (f'q_next shape is {q_next.shape}')
# print(f'shpae rew {rew.shape}, mask {mask.shape}, q_next {q_next.shape}')
q_target = rew + mask * self._gamma * q_next.cpu()
if self.use_priority:
q_target = rew + mask * (self._gamma**
self.n_step) * q_next.cpu()
# print (f'q_target shape is {q_target.shape}')
q_target = q_target.to(device)
# q_loss
q1_eval, q2_eval = self.critic_eval.twinQ(obs, act)
criterion = nn.SmoothL1Loss(reduction='none')
# print (f'q1_eval shape is {q1_eval.shape}')
loss1 = criterion(q1_eval, q_target)
loss2 = criterion(q2_eval, q_target)
return 0.5 * (loss1 + loss2)
def learn_actor_dist(self, obs):
curr_act, curr_log = self.actor_eval.evaluate(obs)
p1_next, p2_next = self.critic_eval.get_probs(obs, curr_act)
p_next = torch.stack([
torch.where(p1.sum() < p2.sum(), p1, p2)
for p1, p2 in zip(p1_next, p2_next)
])
num_atoms = torch.tensor(self.num_atoms,
dtype=torch.float32,
device=device)
# actor_loss = p_next * num_atoms
# actor_loss = torch.sum(actor_loss, dim=1)
# actor_loss = -(actor_loss + self.alpha * curr_log).mean()
actor_loss = (self.alpha * curr_log - p_next)
actor_loss = torch.sum(actor_loss, dim=1)
actor_loss = actor_loss.mean()
return actor_loss, curr_log
def learn_actor(self, obs):
curr_act, curr_log = self.actor_eval.evaluate(obs)
q1_next, q2_next = self.critic_eval.twinQ(obs, curr_act)
q_next = torch.min(q1_next, q2_next)
actor_loss = (self.alpha * curr_log - q_next).mean()
return actor_loss, curr_log
def get_munchausen_rew(self, obs, act, rew):
self.m_alpha = 0.9
self.m_tau = 0.03
self.lo = -1
mu, log_std = self.actor_eval(obs)
std = log_std.exp()
dist = Normal(mu, std)
log_pi_a = self.m_tau * dist.log_prob(act).mean(1).unsqueeze(1).cpu()
m_rew = rew + self.m_alpha * torch.clamp(log_pi_a, min=self.lo, max=0)
return m_rew
def learn(self):
pg_loss, q_loss, a_loss = 0, 0, 0
for _ in range(self._update_iteration):
if self.use_priority:
S, A, R, S_, M, indices, weights = self.buffer.sample(
self._batch_size)
W = torch.tensor(weights, dtype=torch.float32,
device=device).view(-1, 1)
else:
batch_split = self.buffer.split_batch(self._batch_size)
S, A, M, R, S_ = batch_split['s'], batch_split[
'a'], batch_split['m'], batch_split['r'], batch_split['s_']
# print ('after sampling from buffer!')
R = torch.tensor(R, dtype=torch.float32).view(-1, 1)
S = torch.tensor(S, dtype=torch.float32, device=device)
# A = torch.tensor(A, dtype=torch.float32, device=device).view(-1, 1)
A = torch.tensor(A, dtype=torch.float32, device=device).squeeze(1)
# print (f'A shape {A.shape}')
M = torch.tensor(M, dtype=torch.float32).view(-1, 1)
S_ = torch.tensor(S_, dtype=torch.float32, device=device)
# self.use_munchausen = True
if self.use_munchausen:
R = self.get_munchausen_rew(S, A, R)
# print (f'shape S:{S.shape}, A:{A.shape}, M:{M.shape}, R:{R.shape}, S_:{S_.shape}')
if self.use_dist:
# D = torch.from_numpy(np.array([1^int(mask.item()) for mask in M])).view(-1, 1)
# print (f'size S:{S.shape}, A:{A.shape}, M:{M.shape}, R:{R.shape}, S_:{S_.shape}, D:{D.shape}')
# assert 0
batch_loss = self.learn_critic_dist(S, A, R, S_, M)
else:
batch_loss = self.learn_critic(S, A, R, S_, M)
if self.use_priority:
critic_loss = (W * batch_loss).mean()
td_errors = batch_loss.detach().cpu().numpy().sum(1)
# print(batch_loss)
# print(td_errors)
self.buffer.update_priorities(indices,
np.abs(td_errors) + 1e-6)
else:
critic_loss = batch_loss.mean()
self.critic_eval_optim.zero_grad()
critic_loss.backward()
self.critic_eval_optim.step()
self._learn_critic_cnt += 1
actor_loss = torch.tensor(0)
alpha_loss = torch.tensor(0)
if self._learn_critic_cnt % self.actor_learn_freq == 0:
if self.use_dist:
actor_loss, curr_log = self.learn_actor_dist(S)
else:
actor_loss, curr_log = self.learn_actor(S)
self.actor_eval_optim.zero_grad()
actor_loss.backward()
self.actor_eval_optim.step()
# alpha loss
alpha_loss = -(
self.log_alpha *
(curr_log + self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = float(self.log_alpha.exp().detach().cpu().numpy())
q_loss += critic_loss.item()
pg_loss += actor_loss.item()
a_loss += alpha_loss.item()
if self._learn_critic_cnt % self.target_update_freq:
self.soft_sync_weight(self.critic_target, self.critic_eval,
self.tau)
self.soft_sync_weight(self.actor_target, self.actor_eval,
self.tau)
return pg_loss, q_loss, a_loss
class SACV(BasePolicy):
def __init__(
self,
model,
buffer_size=1000,
batch_size=100,
actor_learn_freq=1,
target_update_freq=5,
target_update_tau=1e-2,
learning_rate=1e-3,
discount_factor=0.99,
update_iteration=10,
verbose=False,
use_priority=False,
act_dim=None,
):
super().__init__()
self.lr = learning_rate
self.eps = np.finfo(np.float32).eps.item()
self.tau = target_update_tau
self.actor_learn_freq = actor_learn_freq
self.target_update_freq = target_update_freq
self._gamma = discount_factor
self._target = target_update_freq > 0
self._update_iteration = update_iteration
self._sync_cnt = 0
# self._learn_cnt = 0
self._learn_critic_cnt = 0
self._learn_actor_cnt = 0
self._verbose = verbose
self._batch_size = batch_size
self.use_priority = use_priority
self.use_dist = model.value_net.use_dist
if self.use_priority:
self.buffer = PriorityReplayBuffer(buffer_size)
else:
self.buffer = ReplayBuffer(buffer_size) # off-policy
if self.use_dist:
assert model.value_net.num_atoms > 1
# assert isinstance(model.value_net, CriticModelDist)
self.v_min = model.value_net.v_min
self.v_max = model.value_net.v_max
self.num_atoms = model.value_net.num_atoms
self.delta_z = (self.v_max - self.v_min) / (self.num_atoms - 1)
self.support = torch.linspace(self.v_min, self.v_max,
self.num_atoms)
self.actor_eval = model.policy_net.to(device).train()
self.critic_eval = model.value_net.to(device).train()
self.actor_target = self.copy_net(self.actor_eval)
self.critic_target = self.copy_net(self.critic_eval)
self.actor_eval_optim = optim.Adam(self.actor_eval.parameters(),
lr=self.lr)
self.critic_eval_optim = optim.Adam(self.critic_eval.parameters(),
lr=self.lr)
self.criterion = nn.SmoothL1Loss(reduction='none') # keep batch dim
self.act_dim = act_dim
self.target_entropy = -torch.tensor(1).to(device)
self.log_alpha = torch.zeros(1, requires_grad=True, device=device)
self.alpha_optim = optim.Adam([self.log_alpha], lr=self.lr)
self.alpha = self.log_alpha.exp()
def _tensor(self, data, use_cuda=False):
if np.array(data).ndim == 1:
data = torch.tensor(data, dtype=torch.float32).view(-1, 1)
else:
data = torch.tensor(data, dtype=torch.float32)
if use_cuda:
data = data.to(device)
return data
def learn_dist(self, obs, act, rew, next_obs, mask):
with torch.no_grad():
next_act, next_log_pi = self.actor_target(next_obs)
# q(s, a) change to z(s, a) to discribe a distributional
z1_next, z2_next = self.critic_target.get_probs(
next_obs, next_act) # [batch_size, num_atoms]
p_next = torch.stack([
torch.where(z1.sum() < z2.sum(), z1, z2)
for z1, z2 in zip(z1_next, z2_next)
])
p_next -= (self.alpha * next_log_pi)
Tz = rew.unsqueeze(1) + mask * self.support.unsqueeze(0)
Tz = Tz.clamp(min=self.v_min, max=self.v_max)
b = (Tz - self.v_min) / self.delta_z
l, u = b.floor().to(torch.int64), b.ceil().to(torch.int64)
l[(u > 0) * (l == u)] -= 1
u[(l < (self.num_atoms - 1)) * (l == u)] += 1
m = obs.new_zeros(self._batch_size, self.num_atoms).cpu()
p_next = p_next.cpu()
# print (f'm device: {m.device}')
# print (f'p_next device: {p_next.device}')
offset = torch.linspace(0,
((self._batch_size - 1) * self.num_atoms),
self._batch_size).unsqueeze(1).expand(
self._batch_size, self.num_atoms).to(l)
m.view(-1).index_add_(
0, (l + offset).view(-1),
(p_next *
(u.float() - b)).view(-1)) # m_l = m_l + p(s_t+n, a*)(u - b)
m.view(-1).index_add_(
0, (u + offset).view(-1),
(p_next *
(b - l.float())).view(-1)) # m_u = m_u + p(s_t+n, a*)(b - l)
m = m.to(device)
log_z1, log_z2 = self.critic_eval.get_probs(obs, act, log=True)
loss1 = -(m * log_z1).sum(dim=1)
loss2 = -(m * log_z2).sum(dim=1)
batch_loss = 0.5 * (loss1 + loss2)
return batch_loss
def learn(self):
pg_loss, q_loss, a_loss = 0, 0, 0
for _ in range(self._update_iteration):
if self.use_priority:
# s_{t}, n-step_rewards, s_{t+n}
tree_idxs, S, A, R, S_, M, weights = self.buffer.sample(
self._batch_size)
W = torch.tensor(weights, dtype=torch.float32,
device=device).view(-1, 1)
else:
batch_split = self.buffer.split_batch(self._batch_size)
S, A, M, R, S_ = batch_split['s'], batch_split[
'a'], batch_split['m'], batch_split['r'], batch_split['s_']
# print ('after sampling from buffer!')
if self.act_dim is None:
self.act_dim = A.shape[-1]
self.target_entropy = -torch.tensor(self.act_dim).to(device)
print(self.target_entropy)
assert 0
R = torch.tensor(R, dtype=torch.float32).view(-1, 1)
S = torch.tensor(S, dtype=torch.float32, device=device)
# A = torch.tensor(A, dtype=torch.float32, device=device).view(-1, 1)
A = torch.tensor(A, dtype=torch.float32,
device=device).view(-1, self.act_dim)
# print (f'A shape {A.shape}')
M = torch.tensor(M, dtype=torch.float32).view(-1, 1)
S_ = torch.tensor(S_, dtype=torch.float32, device=device)
# print (f'size S:{S.size()}, A:{A.size()}, M:{M.size()}, R:{R.size()}, S_:{S_.size()}')
if self.use_dist:
# print (M[0].size())
# print (M[0])
# print (M[0].item())
# assert 0
# D = torch.from_numpy(np.array([1^int(mask.item()) for mask in M])).view(-1, 1)
# print (f'size S:{S.shape}, A:{A.size()}, M:{M.size()}, R:{R.size()}, S_:{S_.size()}, D:{D.size()}')
# assert 0
batch_loss = self.learn_dist(S, A, R, S_, M)
else:
with torch.no_grad():
next_A, next_log = self.actor_target.evaluate(S_)
q1_next, q2_next = self.critic_target(S_, next_A)
q_next = torch.min(q1_next,
q2_next) - self.alpha * next_log
q_target = R + M * self._gamma * q_next.cpu()
q_target = q_target.to(device)
# q_loss
q1_eval, q2_eval = self.critic_eval(S, A)
loss1 = self.criterion(q1_eval, q_target)
loss2 = self.criterion(q2_eval, q_target)
# print(f'q_eval {q1_eval.shape}, q_target {q_target.shape}')
batch_loss = 0.5 * (loss1 + loss2)
if self.use_priority:
critic_loss = (W * batch_loss).mean()
self.buffer.update_priorities(
tree_idxs,
np.abs(batch_loss.detach().cpu().numpy()) + 1e-6)
else:
critic_loss = batch_loss.mean()
self.critic_eval_optim.zero_grad()
critic_loss.backward()
self.critic_eval_optim.step()
self._learn_critic_cnt += 1
actor_loss = torch.tensor(0)
alpha_loss = torch.tensor(0)
if self._learn_critic_cnt % self.actor_learn_freq == 0:
curr_A, curr_log = self.actor_eval.evaluate(S)
if self.use_dist:
z1_next, z2_next = self.critic_eval.get_probs(S, curr_A)
p_next = torch.stack([
torch.where(z1.sum() < z2.sum(), z1, z2)
for z1, z2 in zip(z1_next, z2_next)
])
num_atoms = torch.tensor(self.num_atoms,
dtype=torch.float32,
device=device)
# actor_loss = p_next * num_atoms
# actor_loss = torch.sum(actor_loss, dim=1)
# actor_loss = -(actor_loss + self.alpha * curr_log).mean()
actor_loss = (self.alpha * curr_log - p_next)
actor_loss = torch.sum(actor_loss, dim=1)
actor_loss = actor_loss.mean()
else:
q1_next, q2_next = self.critic_eval(S, curr_A)
q_next = torch.min(q1_next, q2_next)
# pg_loss
actor_loss = (self.alpha * curr_log - q_next).mean()
self.actor_eval_optim.zero_grad()
actor_loss.backward()
self.actor_eval_optim.step()
# alpha loss
alpha_loss = -(
self.log_alpha *
(curr_log + self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = float(self.log_alpha.exp().detach().cpu().numpy())
q_loss += critic_loss.item()
pg_loss += actor_loss.item()
a_loss += alpha_loss.item()
if self._learn_critic_cnt % self.target_update_freq:
self.soft_sync_weight(self.critic_target, self.critic_eval,
self.tau)
self.soft_sync_weight(self.actor_target, self.actor_eval,
self.tau)
return pg_loss, q_loss, a_loss
