def update(data):
# First run one gradient descent step for Q1 and Q2
q_optimizer.zero_grad()
loss_q, q_info = compute_loss_q(data)
loss_q.backward()
sync_grads(comm, q_params)
q_optimizer.step()
# 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 each pi.
[opt.zero_grad() for opt in pi_optimizers]
loss_pi = compute_losses_pi(data)
loss_pi.backward()
[sync_grads(comm, ac.pis[i].parameters()) for i in range(N)]
[opt.step() for opt in pi_optimizers]
# Unfreeze Q-networks so you can optimize it at next DDPG step.
for p in q_params:
p.requires_grad = True
# 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.mul_(polyak)
p_targ.data.add_((1 - polyak) * p.data)
# sync weights
sync_weights(comm, q_params)
[sync_weights(comm, pi.parameters()) for pi in ac.pis]
sync_weights(comm, ac_targ.parameters())
# save weights
if policy_record is not None:
torch.save(ac.state_dict(), data_dir + "ac.pt")
torch.save(ac_targ.state_dict(), data_dir + "ac_targ.pt")
def sac(env_fn,
comm,
data_dir,
policy_record=None,
eval_mode=False,
actor_critic=core.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=1):
"""
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.
"""
os.environ["OMP_NUM_THREADS"] = "1"
torch.set_num_threads(1)
torch.manual_seed(seed)
np.random.seed(seed)
env, test_env = env_fn(), env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape[0]
# 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)
# if records exist, load weights
if policy_record is not None:
if os.path.exists(data_dir + "ac.pt"):
ac.load_state_dict(torch.load(data_dir + "ac.pt"))
if os.path.exists(data_dir + "ac_targ.pt"):
ac_targ.load_state_dict(torch.load(data_dir + "ac_targ.pt"))
# initial weight sync
sync_weights(comm, ac.q1.parameters())
sync_weights(comm, ac.q2.parameters())
sync_weights(comm, ac.pi.parameters())
sync_weights(comm, ac_targ.parameters())
# 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)
q_params = itertools.chain(ac.q1.parameters(), ac.q2.parameters())
# Experience buffer
replay_buffer = ReplayBuffer(obs_dim=obs_dim,
act_dim=act_dim,
size=replay_size)
# Count variables (protip: try to get a feel for how different size networks behave!)
var_counts = tuple(
core.count_vars(module) for module in [ac.pi, ac.q1, ac.q2])
# 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']
q1 = ac.q1(o, a)
q2 = ac.q2(o, a)
# Bellman backup for Q functions
with torch.no_grad():
# Target actions come from *current* policy
a2, logp_a2 = ac.pi(o2)
# Target Q-values
q1_pi_targ = ac_targ.q1(o2, a2)
q2_pi_targ = ac_targ.q2(o2, a2)
q_pi_targ = torch.min(q1_pi_targ, q2_pi_targ)
backup = r + gamma * (1 - d) * (q_pi_targ - alpha * logp_a2)
# MSE loss against Bellman backup
loss_q1 = ((q1 - backup)**2).mean()
loss_q2 = ((q2 - backup)**2).mean()
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']
pi, logp_pi = ac.pi(o)
q1_pi = ac.q1(o, pi)
q2_pi = ac.q2(o, pi)
q_pi = torch.min(q1_pi, q2_pi)
# Entropy-regularized policy loss
loss_pi = (alpha * logp_pi - q_pi).mean()
# Useful info for logging
pi_info = dict(LogPi=logp_pi.detach().numpy())
return loss_pi, pi_info
# Set up optimizers for policy and q-function
pi_optimizer = Adam(ac.pi.parameters(), lr=lr)
q_optimizer = Adam(q_params, lr=lr)
def update(data):
# First run one gradient descent step for Q1 and Q2
q_optimizer.zero_grad()
loss_q, q_info = compute_loss_q(data)
loss_q.backward()
sync_grads(comm, q_params)
q_optimizer.step()
# 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()
sync_grads(comm, ac.pi.parameters())
pi_optimizer.step()
# Unfreeze Q-networks so you can optimize it at next DDPG step.
for p in q_params:
p.requires_grad = True
# 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.mul_(polyak)
p_targ.data.add_((1 - polyak) * p.data)
# sync weights
sync_weights(comm, q_params)
sync_weights(comm, ac.pi.parameters())
sync_weights(comm, ac_targ.parameters())
# save weights
if policy_record is not None:
torch.save(ac.state_dict(), data_dir + "ac.pt")
torch.save(ac_targ.state_dict(), data_dir + "ac_targ.pt")
def get_action(o, deterministic=False):
return ac.act(torch.as_tensor(o, dtype=torch.float32), deterministic)
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))
ep_ret += r
ep_len += 1
# Prepare for interaction with environment
total_steps = steps_per_epoch * epochs
start_time = time.time()
o, ep_ret, ep_len = env.reset(), 0, 0
# 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 eval_mode:
a = get_action(o, True)
else:
if t > start_steps:
a = get_action(o)
else:
a = env.action_space.sample()
# Step the env
o2, r, d, _ = env.step(a)
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 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:
o, ep_ret, ep_len = env.reset(), 0, 0
# Update handling (training only)
if not eval_mode:
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
def maddpg(env_fn, comm, data_dir, policy_record=None, eval_mode=False,
common_actor=False,
actor_critic=core.MLPActorCritic, ac_kwargs=dict(), seed=0,
steps_per_epoch=4000, epochs=100, replay_size=int(1e6), gamma=0.99,
polyak=0.995, pi_lr=1e-3, q_lr=1e-3, batch_size=100, start_steps=10000,
update_after=1000, update_every=50, act_noise=0.1, num_test_episodes=10,
max_ep_len=1000, logger_kwargs=dict(), save_freq=1):
"""
Deep Deterministic Policy Gradient (DDPG)
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, and a ``q`` module. The ``act`` method and
``pi`` module should accept batches of observations as inputs,
and ``q`` should accept a batch of observations and a batch of
actions as inputs. When called, these should return:
=========== ================ ======================================
Call Output Shape Description
=========== ================ ======================================
``act`` (batch, act_dim) | Numpy array of actions for each
| observation.
``pi`` (batch, act_dim) | Tensor containing actions from policy
| given observations.
``q`` (batch,) | Tensor containing the current estimate
| of Q* for the provided observations
| and actions. (Critical: make sure to
| flatten this!)
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the ActorCritic object
you provided to DDPG.
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.)
pi_lr (float): Learning rate for policy.
q_lr (float): Learning rate for Q-networks.
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.
act_noise (float): Stddev for Gaussian exploration noise added to
policy at training time. (At test time, no noise is added.)
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.
"""
os.environ["OMP_NUM_THREADS"] = "1"
torch.set_num_threads(1)
torch.manual_seed(seed)
np.random.seed(seed)
env, test_env = env_fn(), env_fn()
N = len(env.observation_spaces)
obs_dim = env.observation_spaces[0].shape
act_dim = env.action_spaces[0].shape[0]
# Action limit for clamping: critically, assumes all dimensions share the same bound!
act_limit = env.action_spaces[0].high[0]
# Create actor-critic module and target networks
ac = actor_critic(env.observation_spaces, env.action_spaces, common_actor, **ac_kwargs)
ac_targ = deepcopy(ac)
# if records exist, load weights
if policy_record is not None:
if os.path.exists(data_dir+"ac.pt"):
ac.load_state_dict(torch.load(data_dir+"ac.pt"))
if os.path.exists(data_dir+"ac_targ.pt"):
ac_targ.load_state_dict(torch.load(data_dir+"ac_targ.pt"))
# initial weight sync
sync_weights(comm, ac.parameters())
sync_weights(comm, ac_targ.parameters())
# for param in ac.named_parameters():
# print(param)
# Freeze target networks with respect to optimizers (only update via polyak averaging)
for p in ac_targ.parameters():
p.requires_grad = False
# Experience buffer
replay_buffer = ReplayBuffer(N, obs_dim=obs_dim, act_dim=act_dim, size=replay_size)
# Set up function for computing DDPG Q-loss
def compute_loss_q(data):
o, a, r, o2, d = data['obs'], data['act'], data['rew'], data['obs2'], data['done']
q = ac.q(o,a)
# Bellman backup for Q function
with torch.no_grad():
q_pi_targ = ac_targ.q(o2, [ac_targ.pis[i](o2[i]) for i in range(N)])
backup = r + gamma * (1 - d) * q_pi_targ
# MSE loss against Bellman backup
loss_q = ((q - backup)**2).mean()
# Useful info for logging
loss_info = dict(QVals=q.detach().numpy())
return loss_q, loss_info
# Set up function for computing DDPG pi loss
def compute_loss_pi(data):
o = data['obs']
q_pi = ac.q(o, [ac.pis[i](o[i]) for i in range(N)])
return -q_pi.mean()
# Set up optimizers for policy and q-function
pi_optimizers = [Adam(p.parameters(), lr=pi_lr) for p in ac.unique_pis]
q_optimizer = Adam(ac.q.parameters(), lr=q_lr)
def update(data):
# First run one gradient descent step for Q.
q_optimizer.zero_grad()
loss_q, loss_info = compute_loss_q(data)
loss_q.backward()
sync_grads(comm, ac.q.parameters())
q_optimizer.step()
# Freeze Q-network so you don't waste computational effort
# computing gradients for it during the policy learning step.
for p in ac.q.parameters():
p.requires_grad = False
# Next run one gradient descent step for pi.
[opt.zero_grad() for opt in pi_optimizers]
loss_pi = compute_loss_pi(data)
loss_pi.backward()
[sync_grads(comm, p.parameters()) for p in ac.unique_pis]
[opt.step() for opt in pi_optimizers]
# Unfreeze Q-network so you can optimize it at next DDPG step.
for p in ac.q.parameters():
p.requires_grad = True
# 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.mul_(polyak)
p_targ.data.add_((1 - polyak) * p.data)
# sync weights
sync_weights(comm, ac.q.parameters())
[sync_weights(comm, pi.parameters()) for pi in ac.unique_pis]
sync_weights(comm, ac_targ.parameters())
# save weights
if policy_record is not None:
torch.save(ac.state_dict(), data_dir+"ac.pt")
torch.save(ac_targ.state_dict(), data_dir+"ac_targ.pt")
def get_action(o, noise_scale):
a = ac.act(torch.as_tensor(o, dtype=torch.float32))
a += noise_scale * np.random.randn(act_dim)
return np.clip(a, -act_limit, act_limit)
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 (noise_scale=0)
o, r, d, _ = test_env.step(get_action(o, 0))
ep_ret += r
ep_len += 1
logger.store(TestEpRet=ep_ret, TestEpLen=ep_len)
# Prepare for interaction with environment
total_steps = steps_per_epoch * epochs
start_time = time.time()
o, ep_ret, ep_len = env.reset(), 0, 0
# 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 (with some noise, via act_noise).
if eval_mode:
a = get_action(o, 0.0)
else:
if t > start_steps:
a = get_action(o, act_noise)
else:
a = [asp.sample() for asp in env.action_spaces]
# Step the env
o2, rs, d, _ = env.step(a)
r = sum(rs)
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 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):
o, ep_ret, ep_len = env.reset(), 0, 0
# Update handling
if not eval_mode:
if t >= update_after and t % update_every == 0:
for _ 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