class PPO:
def __init__(self,
env_id,
render=False,
num_process=4,
min_batch_size=2048,
lr_p=3e-4,
lr_v=3e-4,
gamma=0.99,
tau=0.95,
clip_epsilon=0.2,
ppo_epochs=10,
model_path=None,
seed=1):
self.env_id = env_id
self.gamma = gamma
self.tau = tau
self.ppo_epochs = ppo_epochs
self.clip_epsilon = clip_epsilon
self.render = render
self.num_process = num_process
self.lr_p = lr_p
self.lr_v = lr_v
self.min_batch_size = min_batch_size
self.model_path = model_path
self.seed = seed
self._init_model()
def _init_model(self):
"""init model from parameters"""
self.env, env_continuous, num_states, num_actions = get_env_info(
self.env_id)
# seeding
torch.manual_seed(self.seed)
self.env.seed(self.seed)
if env_continuous:
self.policy_net = Policy(num_states, num_actions).double().to(
device) # current policy
self.policy_net_old = Policy(num_states, num_actions).double().to(
device) # old policy
else:
self.policy_net = DiscretePolicy(num_states,
num_actions).double().to(device)
self.policy_net_old = DiscretePolicy(
num_states, num_actions).double().to(device)
self.value_net = Value(num_states).double().to(device)
self.running_state = ZFilter((num_states, ), clip=5)
if self.model_path:
print("Loading Saved Model {}_ppo.p".format(self.env_id))
self.policy_net, self.value_net, self.running_state = pickle.load(
open('{}/{}_ppo.p'.format(self.model_path, self.env_id), "rb"))
self.policy_net_old.load_state_dict(self.policy_net.state_dict())
self.collector = MemoryCollector(self.env,
self.policy_net_old,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
self.optimizer_p = optim.Adam(self.policy_net.parameters(),
lr=self.lr_p)
self.optimizer_v = optim.Adam(self.value_net.parameters(),
lr=self.lr_v)
def choose_action(self, state):
"""select action"""
state = DOUBLE(state).unsqueeze(0).to(device)
with torch.no_grad():
action, log_prob = self.policy_net.get_action_log_prob(state)
return action, log_prob
def eval(self, i_iter):
"""init model from parameters"""
state = self.env.reset()
test_reward = 0
while True:
self.env.render()
state = self.running_state(state)
action, _ = self.choose_action(state)
action = action.cpu().numpy()[0]
state, reward, done, _ = self.env.step(action)
test_reward += reward
if done:
break
print(f"Iter: {i_iter}, test Reward: {test_reward}")
self.env.close()
def learn(self, writer, i_iter):
"""learn model"""
memory, log = self.collector.collect_samples(self.min_batch_size)
print(
f"Iter: {i_iter}, num steps: {log['num_steps']}, total reward: {log['total_reward']: .4f}, "
f"min reward: {log['min_episode_reward']: .4f}, max reward: {log['max_episode_reward']: .4f}, "
f"average reward: {log['avg_reward']: .4f}, sample time: {log['sample_time']: .4f}"
)
# record reward information
writer.add_scalars(
"ppo", {
"total reward": log['total_reward'],
"average reward": log['avg_reward'],
"min reward": log['min_episode_reward'],
"max reward": log['max_episode_reward'],
"num steps": log['num_steps']
}, i_iter)
batch = memory.sample() # sample all items in memory
batch_state = DOUBLE(batch.state).to(device)
batch_action = DOUBLE(batch.action).to(device)
batch_reward = DOUBLE(batch.reward).to(device)
batch_mask = DOUBLE(batch.mask).to(device)
batch_log_prob = DOUBLE(batch.log_prob).to(device)
with torch.no_grad():
batch_value = self.value_net(batch_state)
batch_advantage, batch_return = estimate_advantages(
batch_reward, batch_mask, batch_value, self.gamma, self.tau)
v_loss, p_loss = torch.empty(1), torch.empty(1)
for _ in range(self.ppo_epochs):
v_loss, p_loss = ppo_step(self.policy_net, self.value_net,
self.optimizer_p, self.optimizer_v, 1,
batch_state, batch_action, batch_return,
batch_advantage, batch_log_prob,
self.clip_epsilon, 1e-3)
self.policy_net_old.load_state_dict(self.policy_net.state_dict())
return v_loss, p_loss
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.policy_net, self.value_net, self.running_state),
open('{}/{}_ppo.p'.format(save_path, self.env_id), 'wb'))
class GAIL:
def __init__(self,
render=False,
num_process=4,
config=None,
expert_data_path=None,
env_id=None):
self.render = render
self.env_id = env_id
self.num_process = num_process
self.expert_data_path = expert_data_path
self.config = config
self._load_expert_trajectory()
self._init_model()
def _load_expert_trajectory(self):
self.expert_dataset = ExpertDataset(
expert_data_path=self.expert_data_path,
train_fraction=self.config["expert_data"]["train_fraction"],
traj_limitation=self.config["expert_data"]["traj_limitation"],
shuffle=self.config["expert_data"]["shuffle"],
batch_size=self.config["expert_data"]["batch_size"])
def _init_model(self):
# seeding
seed = self.config["train"]["general"]["seed"]
torch.manual_seed(seed)
np.random.seed(seed)
self.env, env_continuous, num_states, num_actions = get_env_info(
self.env_id)
# check env
assert num_states == self.expert_dataset.num_states and num_actions == self.expert_dataset.num_actions, \
"Expected corresponding expert dataset and env"
dim_dict = {"dim_state": num_states, "dim_action": num_actions}
self.config["value"].update(dim_dict)
self.config["policy"].update(dim_dict)
self.config["discriminator"].update(dim_dict)
self.value = Value(dim_state=self.config["value"]["dim_state"],
dim_hidden=self.config["value"]["dim_hidden"],
activation=resolve_activate_function(
self.config["value"]["activation"]))
self.policy = Policy(config=self.config["policy"])
self.discriminator = Discriminator(
dim_state=self.config["discriminator"]["dim_state"],
dim_action=self.config["discriminator"]["dim_action"],
dim_hidden=self.config["discriminator"]["dim_hidden"],
activation=resolve_activate_function(
self.config["discriminator"]["activation"]))
self.discriminator_func = nn.BCELoss()
self.running_state = None
self.collector = MemoryCollector(self.env,
self.policy,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
print("Model Structure")
print(self.policy)
print(self.value)
print(self.discriminator)
print()
self.optimizer_policy = optim.Adam(
self.policy.parameters(),
lr=self.config["policy"]["learning_rate"])
self.optimizer_value = optim.Adam(
self.value.parameters(), lr=self.config["value"]["learning_rate"])
self.optimizer_discriminator = optim.Adam(
self.discriminator.parameters(),
lr=self.config["discriminator"]["learning_rate"])
to_device(self.value, self.policy, self.discriminator,
self.discriminator_func)
def choose_action(self, state):
"""select action"""
state = FLOAT(state).unsqueeze(0).to(device)
with torch.no_grad():
action, log_prob = self.policy.get_action_log_prob(state)
return action, log_prob
def learn(self, writer, i_iter):
memory, log = self.collector.collect_samples(
self.config["train"]["generator"]["sample_batch_size"])
self.policy.train()
self.value.train()
self.discriminator.train()
print(
f"Iter: {i_iter}, num steps: {log['num_steps']}, total reward: {log['total_reward']: .4f}, "
f"min reward: {log['min_episode_reward']: .4f}, max reward: {log['max_episode_reward']: .4f}, "
f"average reward: {log['avg_reward']: .4f}, sample time: {log['sample_time']: .4f}"
)
# record reward information
writer.add_scalar("gail/average reward", log['avg_reward'], i_iter)
writer.add_scalar("gail/num steps", log['num_steps'], i_iter)
# collect generated batch
# gen_batch = self.collect_samples(self.config["ppo"]["sample_batch_size"])
gen_batch = memory.sample()
gen_batch_state = FLOAT(gen_batch.state).to(
device) # [batch size, state size]
gen_batch_action = FLOAT(gen_batch.action).to(
device) # [batch size, action size]
gen_batch_old_log_prob = FLOAT(gen_batch.log_prob).to(
device) # [batch size, 1]
gen_batch_mask = FLOAT(gen_batch.mask).to(device) # [batch, 1]
####################################################
# update discriminator
####################################################
d_optim_i_iters = self.config["train"]["discriminator"]["optim_step"]
if i_iter % d_optim_i_iters == 0:
for step, (expert_batch_state, expert_batch_action) in enumerate(
self.expert_dataset.train_loader):
if step >= d_optim_i_iters:
break
# calculate probs and logits
gen_prob, gen_logits = self.discriminator(
gen_batch_state, gen_batch_action)
expert_prob, expert_logits = self.discriminator(
expert_batch_state.to(device),
expert_batch_action.to(device))
# calculate accuracy
gen_acc = torch.mean((gen_prob < 0.5).float())
expert_acc = torch.mean((expert_prob > 0.5).float())
# calculate regression loss
expert_labels = torch.ones_like(expert_prob)
gen_labels = torch.zeros_like(gen_prob)
e_loss = self.discriminator_func(expert_prob,
target=expert_labels)
g_loss = self.discriminator_func(gen_prob, target=gen_labels)
d_loss = e_loss + g_loss
# calculate entropy loss
logits = torch.cat([gen_logits, expert_logits], 0)
entropy = ((1. - torch.sigmoid(logits)) * logits -
torch.nn.functional.logsigmoid(logits)).mean()
entropy_loss = - \
self.config["train"]["discriminator"]["ent_coeff"] * entropy
total_loss = d_loss + entropy_loss
self.optimizer_discriminator.zero_grad()
total_loss.backward()
self.optimizer_discriminator.step()
writer.add_scalar('discriminator/d_loss', d_loss.item(), i_iter)
writer.add_scalar("discriminator/e_loss", e_loss.item(), i_iter)
writer.add_scalar("discriminator/g_loss", g_loss.item(), i_iter)
writer.add_scalar("discriminator/ent", entropy.item(), i_iter)
writer.add_scalar('discriminator/expert_acc', gen_acc.item(), i_iter)
writer.add_scalar('discriminator/gen_acc', expert_acc.item(), i_iter)
####################################################
# update policy by ppo [mini_batch]
####################################################
with torch.no_grad():
gen_batch_value = self.value(gen_batch_state)
d_out, _ = self.discriminator(gen_batch_state, gen_batch_action)
gen_batch_reward = -torch.log(1 - d_out + 1e-6)
gen_batch_advantage, gen_batch_return = estimate_advantages(
gen_batch_reward, gen_batch_mask, gen_batch_value,
self.config["train"]["generator"]["gamma"],
self.config["train"]["generator"]["tau"])
ppo_optim_i_iters = self.config["train"]["generator"]["optim_step"]
ppo_mini_batch_size = self.config["train"]["generator"][
"mini_batch_size"]
for _ in range(ppo_optim_i_iters):
if ppo_mini_batch_size > 0:
gen_batch_size = gen_batch_state.shape[0]
optim_iter_num = int(
math.ceil(gen_batch_size / ppo_mini_batch_size))
perm = torch.randperm(gen_batch_size)
for i in range(optim_iter_num):
ind = perm[slice(
i * ppo_mini_batch_size,
min((i + 1) * ppo_mini_batch_size, gen_batch_size))]
mini_batch_state, mini_batch_action, mini_batch_advantage, mini_batch_return, \
mini_batch_old_log_prob = gen_batch_state[ind], gen_batch_action[ind], \
gen_batch_advantage[ind], gen_batch_return[ind], gen_batch_old_log_prob[
ind]
v_loss, p_loss, ent_loss = ppo_step(
policy_net=self.policy,
value_net=self.value,
optimizer_policy=self.optimizer_policy,
optimizer_value=self.optimizer_value,
optim_value_iternum=self.config["value"]
["optim_value_iter"],
states=mini_batch_state,
actions=mini_batch_action,
returns=mini_batch_return,
old_log_probs=mini_batch_old_log_prob,
advantages=mini_batch_advantage,
clip_epsilon=self.config["train"]["generator"]
["clip_ratio"],
l2_reg=self.config["value"]["l2_reg"])
else:
v_loss, p_loss, ent_loss = ppo_step(
policy_net=self.policy,
value_net=self.value,
optimizer_policy=self.optimizer_policy,
optimizer_value=self.optimizer_value,
optim_value_iternum=self.config["value"]
["optim_value_iter"],
states=gen_batch_state,
actions=gen_batch_action,
returns=gen_batch_return,
old_log_probs=gen_batch_old_log_prob,
advantages=gen_batch_advantage,
clip_epsilon=self.config["train"]["generator"]
["clip_ratio"],
l2_reg=self.config["value"]["l2_reg"])
writer.add_scalar('generator/p_loss', p_loss, i_iter)
writer.add_scalar('generator/v_loss', v_loss, i_iter)
writer.add_scalar('generator/ent_loss', ent_loss, i_iter)
print(f" Training episode:{i_iter} ".center(80, "#"))
print('d_gen_prob:', gen_prob.mean().item())
print('d_expert_prob:', expert_prob.mean().item())
print('d_loss:', d_loss.item())
print('e_loss:', e_loss.item())
print("d/bernoulli_entropy:", entropy.item())
def eval(self, i_iter, render=False):
self.policy.eval()
self.value.eval()
self.discriminator.eval()
state = self.env.reset()
test_reward = 0
while True:
if render:
self.env.render()
if self.running_state:
state = self.running_state(state)
action, _ = self.choose_action(state)
action = action.cpu().numpy()[0]
state, reward, done, _ = self.env.step(action)
test_reward += reward
if done:
break
print(f"Iter: {i_iter}, test Reward: {test_reward}")
self.env.close()
def save_model(self, save_path):
check_path(save_path)
# torch.save((self.discriminator, self.policy, self.value), f"{save_path}/{self.exp_name}.pt")
torch.save(self.discriminator,
f"{save_path}/{self.env_id}_Discriminator.pt")
torch.save(self.policy, f"{save_path}/{self.env_id}_Policy.pt")
torch.save(self.value, f"{save_path}/{self.env_id}_Value.pt")
def load_model(self, model_path):
# load entire model
# self.discriminator, self.policy, self.value = torch.load(model_path, map_location=device)
self.discriminator = torch.load(f"{model_path}_Discriminator.pt",
map_location=device)
self.policy = torch.load(f"{model_path}_Policy.pt",
map_location=device)
self.value = torch.load(f"{model_path}_Value.pt", map_location=device)
class A2C():
def __init__(self,
env_id,
render=False,
num_process=4,
min_batch_size=2048,
lr_ac=3e-4,
value_net_coeff=0.5,
entropy_coeff=1e-2,
gamma=0.99,
tau=0.95,
seed=1,
model_path=None):
self.env_id = env_id
self.gamma = gamma
self.tau = tau
self.render = render
self.num_process = num_process
self.lr_ac = lr_ac
self.value_net_coeff = value_net_coeff
self.entropy_coeff = entropy_coeff
self.min_batch_size = min_batch_size
self.model_path = model_path
self.seed = seed
self._init_model()
def _init_model(self):
"""init model from parameters"""
self.env, env_continuous, num_states, num_actions = get_env_info(
self.env_id)
# seeding
torch.manual_seed(self.seed)
self.env.seed(self.seed)
if env_continuous:
self.policy_net = Policy(num_states, num_actions).to(device)
else:
self.policy_net = DiscretePolicy(num_states,
num_actions).to(device)
self.value_net = Value(num_states).to(device)
self.ac_net = Actor_Critic(self.policy_net, self.value_net).to(device)
self.running_state = ZFilter((num_states, ), clip=5)
if self.model_path:
print("Loading Saved Model {}_a2c.p".format(self.env_id))
self.ac_net, self.running_state = pickle.load(
open('{}/{}_a2c.p'.format(self.model_path, self.env_id), "rb"))
self.collector = MemoryCollector(self.env,
self.ac_net,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
self.optimizer_ac = optim.Adam(self.ac_net.parameters(), lr=self.lr_ac)
def choose_action(self, state):
"""select action"""
state = FLOAT(state).unsqueeze(0).to(device)
with torch.no_grad():
action, log_prob = self.ac_net.get_action_log_prob(state)
return action, log_prob
def eval(self, i_iter, render=False):
"""init model from parameters"""
state = self.env.reset()
test_reward = 0
while True:
if render:
self.env.render()
state = self.running_state(state)
action, _ = self.choose_action(state)
action = action.cpu().numpy()[0]
state, reward, done, _ = self.env.step(action)
test_reward += reward
if done:
break
print(f"Iter: {i_iter}, test Reward: {test_reward}")
self.env.close()
def learn(self, writer, i_iter):
"""learn model"""
memory, log = self.collector.collect_samples(self.min_batch_size)
print(
f"Iter: {i_iter}, num steps: {log['num_steps']}, total reward: {log['total_reward']: .4f}, "
f"min reward: {log['min_episode_reward']: .4f}, max reward: {log['max_episode_reward']: .4f}, "
f"average reward: {log['avg_reward']: .4f}, sample time: {log['sample_time']: .4f}"
)
# record reward information
writer.add_scalar("total reward", log['total_reward'], i_iter)
writer.add_scalar("average reward", log['avg_reward'], i_iter)
writer.add_scalar("min reward", log['min_episode_reward'], i_iter)
writer.add_scalar("max reward", log['max_episode_reward'], i_iter)
writer.add_scalar("num steps", log['num_steps'], i_iter)
batch = memory.sample() # sample all items in memory
batch_state = FLOAT(batch.state).to(device)
batch_action = FLOAT(batch.action).to(device)
batch_reward = FLOAT(batch.reward).to(device)
batch_mask = FLOAT(batch.mask).to(device)
with torch.no_grad():
batch_value = self.ac_net.get_value(batch_state)
batch_advantage, batch_return = estimate_advantages(
batch_reward, batch_mask, batch_value, self.gamma, self.tau)
alg_step_stats = a2c_step(self.ac_net, self.optimizer_ac, batch_state,
batch_action, batch_return, batch_advantage,
self.value_net_coeff, self.entropy_coeff)
return alg_step_stats
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.ac_net, self.running_state),
open('{}/{}_a2c.p'.format(save_path, self.env_id), 'wb'))
class PPO_Minibatch:
def __init__(self,
env_id,
render=False,
num_process=4,
min_batch_size=2048,
lr_p=3e-4,
lr_v=3e-4,
gamma=0.99,
tau=0.95,
clip_epsilon=0.2,
ppo_mini_batch_size=64,
ppo_epochs=10,
model_path=None,
seed=1):
self.env_id = env_id
self.gamma = gamma
self.tau = tau
self.clip_epsilon = clip_epsilon
self.ppo_mini_batch_size = ppo_mini_batch_size
self.ppo_epochs = ppo_epochs
self.render = render
self.num_process = num_process
self.min_batch_size = min_batch_size
self.lr_p = lr_p
self.lr_v = lr_v
self.model_path = model_path
self.seed = seed
self._init_model()
def _init_model(self):
"""init model from parameters"""
self.env, env_continuous, num_states, num_actions = get_env_info(
self.env_id)
# seeding
torch.manual_seed(self.seed)
self.env.seed(self.seed)
if self.model_path:
print("Loading Saved Model {}_ppo_mini.p".format(self.env_id))
self.policy_net, self.value_net, self.running_state = pickle.load(
open('{}/{}_ppo_mini.p'.format(self.model_path, self.env_id),
"rb"))
else:
if env_continuous:
self.policy_net = Policy(num_states, num_actions).to(device)
else:
self.policy_net = DiscretePolicy(num_states,
num_actions).to(device)
self.value_net = Value(num_states).to(device)
self.running_state = ZFilter((num_states, ), clip=5)
self.collector = MemoryCollector(self.env,
self.policy_net,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
self.optimizer_p = optim.Adam(self.policy_net.parameters(),
lr=self.lr_p)
self.optimizer_v = optim.Adam(self.value_net.parameters(),
lr=self.lr_v)
def choose_action(self, state):
"""select action according to policy"""
state = FLOAT(state).unsqueeze(0).to(device)
with torch.no_grad():
action, log_prob = self.policy_net.get_action_log_prob(state)
return action, log_prob
def eval(self, i_iter):
"""evaluate current model"""
state = self.env.reset()
test_reward = 0
while True:
self.env.render()
state = self.running_state(state)
action, _ = self.choose_action(state)
action = action.cpu().numpy()[0]
state, reward, done, _ = self.env.step(action)
test_reward += reward
if done:
break
print(f"Iter: {i_iter}, test Reward: {test_reward}")
self.env.close()
def learn(self, writer, i_iter):
"""learn model"""
memory, log = self.collector.collect_samples(self.min_batch_size)
print(
f"Iter: {i_iter}, num steps: {log['num_steps']}, total reward: {log['total_reward']: .4f}, "
f"min reward: {log['min_episode_reward']: .4f}, max reward: {log['max_episode_reward']: .4f}, "
f"average reward: {log['avg_reward']: .4f}, sample time: {log['sample_time']: .4f}"
)
# record reward information
writer.add_scalars(
"PPO_mini_batch", {
"total reward": log['total_reward'],
"average reward": log['avg_reward'],
"min reward": log['min_episode_reward'],
"max reward": log['max_episode_reward'],
"num steps": log['num_steps']
}, i_iter)
batch = memory.sample() # sample all items in memory
batch_state = FLOAT(batch.state).to(device)
batch_action = FLOAT(batch.action).to(device)
batch_log_prob = FLOAT(batch.log_prob).to(device)
batch_reward = FLOAT(batch.reward).to(device)
batch_mask = FLOAT(batch.mask).to(device)
batch_size = batch_state.shape[0]
with torch.no_grad():
batch_values = self.value_net(batch_state)
batch_advantages, batch_returns = estimate_advantages(
batch_reward, batch_mask, batch_values, self.gamma, self.tau)
v_loss, p_loss = torch.empty(1), torch.empty(1)
mini_batch_num = int(math.ceil(batch_size / self.ppo_mini_batch_size))
# update with mini-batch
for _ in range(self.ppo_epochs):
index = torch.randperm(batch_size)
for i in range(mini_batch_num):
ind = index[slice(
i * self.ppo_mini_batch_size,
min(batch_size, (i + 1) * self.ppo_mini_batch_size))]
state, action, returns, advantages, old_log_pis = batch_state[
ind], batch_action[ind], batch_returns[
ind], batch_advantages[ind], batch_log_prob[ind]
v_loss, p_loss = ppo_step(self.policy_net, self.value_net,
self.optimizer_p, self.optimizer_v,
1, state, action, returns,
advantages, old_log_pis,
self.clip_epsilon, 1e-3)
return v_loss, p_loss
def save(self, save_path):
"""save model"""
pickle.dump((self.policy_net, self.value_net, self.running_state),
open('{}/{}_ppo_mini.p'.format(save_path, self.env_id),
'wb'))