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
else:
self.policy_net = 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 {}_trpo.p".format(self.env_id))
self.policy_net, self.value_net, self.running_state = pickle.load(
open('{}/{}_trpo.p'.format(self.model_path, self.env_id),
"rb"))
self.collector = MemoryCollector(self.env,
self.policy_net,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
self.optimizer_v = optim.Adam(self.value_net.parameters(),
lr=self.lr_v)
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,
ppo_mini_batch_size=64,
seed=1,
model_path=None):
self.env_id = env_id
self.gamma = gamma
self.tau = tau
self.ppo_epochs = ppo_epochs
self.ppo_mini_batch_size = ppo_mini_batch_size
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).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)
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.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"""
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, render=False):
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_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
# ('state', 'action', 'reward', 'next_state', 'mask', 'log_prob')
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)
batch_log_prob = FLOAT(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):
if self.ppo_mini_batch_size:
batch_size = batch_state.shape[0]
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_return[
ind], batch_advantage[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)
else:
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)
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 TRPO:
def __init__(self,
env_id,
render=False,
num_process=1,
min_batch_size=2048,
lr_v=3e-4,
gamma=0.99,
tau=0.95,
max_kl=1e-2,
damping=1e-2,
seed=1,
model_path=None):
self.env_id = env_id
self.gamma = gamma
self.tau = tau
self.max_kl = max_kl
self.damping = damping
self.render = render
self.num_process = num_process
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
else:
self.policy_net = 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 {}_trpo.p".format(self.env_id))
self.policy_net, self.value_net, self.running_state = pickle.load(
open('{}/{}_trpo.p'.format(self.model_path, self.env_id),
"rb"))
self.collector = MemoryCollector(self.env,
self.policy_net,
render=self.render,
running_state=self.running_state,
num_process=self.num_process)
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, render=False):
"""evaluate model"""
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_scalars(
"trpo", {
"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)
# update by TRPO
trpo_step(self.policy_net, self.value_net, batch_state, batch_action,
batch_return, batch_advantage, batch_log_prob, self.max_kl,
self.damping, 1e-3, None)
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.policy_net, self.value_net, self.running_state),
open('{}/{}_trpo.p'.format(save_path, self.env_id), 'wb'))
class REINFORCE:
def __init__(self,
env_id,
render=False,
num_process=1,
min_batch_size=2048,
lr_p=3e-4,
gamma=0.99,
reinforce_epochs=5,
seed=1,
model_path=None):
self.env_id = env_id
self.render = render
self.num_process = num_process
self.min_batch_size = min_batch_size
self.lr_p = lr_p
self.gamma = gamma
self.reinforce_epochs = reinforce_epochs
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) # current policy
else:
self.policy_net = DiscretePolicy(num_states,
num_actions).to(device)
self.running_state = ZFilter((num_states, ), clip=5)
if self.model_path:
print("Loading Saved Model {}_reinforce.p".format(self.env_id))
self.policy_net, self.running_state = pickle.load(
open('{}/{}_reinforce.p'.format(self.model_path, self.env_id),
"rb"))
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)
def choose_action(self, state):
"""select action"""
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, 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_scalars(
"reinforce", {
"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_reward = FLOAT(batch.reward).to(device)
batch_mask = FLOAT(batch.mask).to(device)
p_loss = torch.empty(1)
for _ in range(self.reinforce_epochs):
p_loss = reinforce_step(self.policy_net, self.optimizer_p,
batch_state, batch_action, batch_reward,
batch_mask, self.gamma)
return p_loss
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.policy_net, self.running_state),
open('{}/{}_reinforce.p'.format(save_path, self.env_id),
'wb'))