def __init__(self,
alpha,
beta,
input_dims,
tau,
env,
env_id,
gamma=0.99,
n_actions=2,
max_size=1000000,
layer_1_size=256,
layer_2_size=256,
batch_size=100,
reward_scale=2):
self.gamma = gamma
self.tau = tau
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.batch_size = batch_size
self.n_actions = n_actions
self.scale = reward_scale
self.actor = ActorNetwork(alpha,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_actor',
max_action=env.action_space.high)
self.critic_1 = CriticNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_critic_1')
self.critic_2 = CriticNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_critic_2')
self.value = ValueNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
name=env_id + '_value')
self.target_value = ValueNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
name=env_id + '_target_value')
self.update_network_parameters(tau=1)
def __init__(self, load_from=None, will_train=True):
self.env = TorcsEnv(
path='/usr/local/share/games/torcs/config/raceman/quickrace.xml')
self.args = SAC_args()
self.buffer = ReplayBuffer(self.args.buffer_size)
action_dim = self.env.action_space.shape[0]
state_dim = self.env.observation_space.shape[0]
hidden_dim = 256
self.action_size = action_dim
self.state_size = state_dim
self.value_net = ValueNetwork(state_dim,
hidden_dim).to(self.args.device)
self.target_value_net = ValueNetwork(state_dim,
hidden_dim).to(self.args.device)
self.soft_q_net1 = SoftQNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.soft_q_net2 = SoftQNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.policy_net = PolicyNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.target_value_net.load_state_dict(self.value_net.state_dict())
self.value_criterion = nn.MSELoss()
self.soft_q_loss1 = nn.MSELoss()
self.soft_q_loss2 = nn.MSELoss()
self.value_opt = optim.Adam(self.value_net.parameters(),
lr=self.args.lr)
self.soft_q_opt1 = optim.Adam(self.soft_q_net1.parameters(),
lr=self.args.lr)
self.soft_q_opt2 = optim.Adam(self.soft_q_net2.parameters(),
lr=self.args.lr)
self.policy_opt = optim.Adam(self.policy_net.parameters(),
lr=self.args.lr)
if will_train:
current_time = time.strftime('%d-%b-%y-%H.%M.%S', time.localtime())
self.plot_folder = f'plots/{current_time}'
self.model_save_folder = f'model/{current_time}'
make_sure_dir_exists(self.plot_folder)
make_sure_dir_exists(self.model_save_folder)
self.cp = Checkpoint(self.model_save_folder)
if load_from is not None:
try:
self.load_checkpoint(load_from)
except FileNotFoundError:
print(f'{load_from} not found. Running default.')
else:
print('Starting from scratch.')
def __init__(self, config):
self.config = config
# Create session
self.session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True)))
# Create networks
self.prior_network = PolicyNetwork(
scope=config.prior_network,
temperature=config.prior_temperature,
use_symmetry=config.use_symmetry)
self.rollout_network = PolicyNetwork(
scope=config.rollout_network,
temperature=config.rollout_temperature,
reuse=config.prior_network == config.rollout_network,
use_symmetry=config.use_symmetry)
self.value_network = ValueNetwork(
scope=config.value_network, use_symmetry=config.use_symmetry)
# Load networks from checkpoints
run_dir = util.run_directory(config)
util.restore_network_or_fail(self.session, run_dir, self.prior_network)
util.restore_network_or_fail(self.session, run_dir, self.rollout_network)
util.restore_network_or_fail(self.session, run_dir, self.value_network)
# Create queues
self.prior_queue = AllQueue()
self.rollout_queue = AllQueue(maxsize=16)
self.value_queue = AllQueue(maxsize=16)
self.new_game()
def train():
tf.reset_default_graph()
# Create a global step variable
global_step = tf.Variable(0, name="global_step", trainable=False)
policy_estimator = PolicyNetwork(input_size=len(env.state), output_size=env.action_space.n,
summaries_dir=experiment_dir)
value_estimator = ValueNetwork(input_size=len(env.state), output_size=1)
# Object-aware Reward Network
reward_estimator = ObjectAwareRewardNetwork(input_size=len(env.state), output_size=1, action_num=env.action_space.n)
# # Reward Network
# reward_estimator = RewardNetwork(input_size=len(env.state), output_size=1, action_num=env.action_space.n)
saver = tf.train.Saver()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(tf.initialize_all_variables())
reinforce(env,
policy_estimator,
value_estimator,
reward_estimator,
max_num_episodes,
sess,
discount_factor=0.99,
uniform_sample=False,
saver=saver,
model_dir=model_dir,
figure_dir=figure_dir)
def __init__(self, alpha = 0.0003, beta = 0.0003, input_dims = [8],
env = None, gamma = 0.99, tau = 0.005, n_actions = 2, max_size = 1000000,
layer1_size = 256, layer2_size = 256, batch_size = 256, reward_scale = 2):
self.gamma = gamma
self.tau = tau
self.batch_size = batch_size
self.n_actions = n_actions
self.scale = reward_scale
self.memory = ReplayBuffer(max_size, input_dims, n_actions = n_actions)
self.actor = ActorNetwork(alpha, input_dims, n_actions = n_actions, max_action = env.action_space.high)
self.critic1 = CriticNetwork(beta, input_dims, n_actions = n_actions, name = 'critic1')
self.critic2 = CriticNetwork(beta, input_dims, n_actions = n_actions, name = 'critic2')
self.value = ValueNetwork(beta, input_dims, name = 'value')
self.target_value = ValueNetwork(beta, input_dims, name = 'target')
self.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
self.update_network_params(tau = 1)
def __init__(self,
alpha=0.0003,
beta=.0003,
input_dims=[8],
env=None,
gamma=.99,
n_actions=2,
max_size=1000000,
layer1_size=256,
layer2_size=256,
tau=.005,
batch_size=256,
reward_scale=2):
# reward scales depends on action convention for the environment
self.gamma = gamma
self.tau = tau
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.batch_size = batch_size
self.n_actions = n_actions
# set up classes
self.actor = ActorNetwork(alpha,
input_dims,
max_action=env.action_space.high,
n_actions=n_actions,
name='actor')
self.critic1 = CriticNetwork(beta,
input_dims,
n_actions=n_actions,
name='critic_1')
self.critic2 = CriticNetwork(beta,
input_dims,
n_actions=n_actions,
name='critic_2')
self.value = ValueNetwork(beta, input_dims, name='value')
# target value
self.target_value = ValueNetwork(beta, input_dims, name='target_value')
self.scale = reward_scale
self.update_network_parameters(tau=1)
def __init__(self, config):
self.config = config
self.run_dir = util.run_directory(config)
self.position_targets = PositionTargets(config, self.run_dir)
self.session = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(allow_growth=True)))
self.value_network = ValueNetwork('value')
util.restore_or_initialize_network(self.session, self.run_dir,
self.value_network)
# Train ops
self.create_train_op(self.value_network)
self.writer = tf.summary.FileWriter(self.run_dir)
util.restore_or_initialize_scope(self.session, self.run_dir,
self.training_scope.name)
class Agent():
def __init__(self, alpha = 0.0003, beta = 0.0003, input_dims = [8],
env = None, gamma = 0.99, tau = 0.005, n_actions = 2, max_size = 1000000,
layer1_size = 256, layer2_size = 256, batch_size = 256, reward_scale = 2):
self.gamma = gamma
self.tau = tau
self.batch_size = batch_size
self.n_actions = n_actions
self.scale = reward_scale
self.memory = ReplayBuffer(max_size, input_dims, n_actions = n_actions)
self.actor = ActorNetwork(alpha, input_dims, n_actions = n_actions, max_action = env.action_space.high)
self.critic1 = CriticNetwork(beta, input_dims, n_actions = n_actions, name = 'critic1')
self.critic2 = CriticNetwork(beta, input_dims, n_actions = n_actions, name = 'critic2')
self.value = ValueNetwork(beta, input_dims, name = 'value')
self.target_value = ValueNetwork(beta, input_dims, name = 'target')
self.device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
self.update_network_params(tau = 1)
def choose_action(self, obs):
state = torch.tensor([obs],dtype=torch.float32).to(self.device)
actions, _ = self.actor.sample_normal(state, reparam = False)
return actions.cpu().detach().numpy()[0]
def store_trans(self, state, action, reward, new_state, done):
self.memory.store_trans(state, action, reward, new_state, done)
def update_network_params(self, tau = None):
if tau is None:
tau = self.tau
target_value_params = self.target_value.named_parameters()
value_params = self.value.named_parameters()
target_value_state_dict = dict(target_value_params)
value_state_dict = dict(value_params)
for name in value_state_dict.keys():
value_state_dict[name] = tau * value_state_dict[name].clone() + \
(1 - tau) * target_value_state_dict[name].clone()
self.target_value.load_state_dict(value_state_dict)
def save_models(self):
self.actor.save_checkpoint()
self.value.save_checkpoint()
self.target_value.save_checkpoint()
self.critic1.save_checkpoint()
self.critic2.save_checkpoint()
print('saving models')
def load_models(self):
self.actor.load_checkpoint()
self.value.load_checkpoint()
self.target_value.load_checkpoint()
self.critic1.load_checkpoint()
self.critic2.load_checkpoint()
print('loading models')
def get_critic_val_log_prob(self, state, reparam):
actions, log_probs = self.actor.sample_normal(state, reparam = False)
log_probs = log_probs.view(-1)
q1_new = self.critic1(state, actions)
q2_new = self.critic2(state, actions)
critic_value = torch.min(q1_new, q2_new)
critic_value = critic_value.view(-1)
return log_probs, critic_value
def learn(self):
if self.memory.mem_counter < self.batch_size:
return
state, action, reward, new_state, done = \
self.memory.sample_buffer(self.batch_size)
reward = torch.tensor(reward, dtype=torch.float).to(self.actor.device)
done = torch.tensor(done).to(self.actor.device)
state_ = torch.tensor(new_state, dtype=torch.float).to(self.actor.device)
state = torch.tensor(state, dtype=torch.float).to(self.actor.device)
action = torch.tensor(action, dtype=torch.float).to(self.actor.device)
value = self.value(state).view(-1)
value_ = self.target_value(state_).view(-1)
value_[done] = 0.0
actions, log_probs = self.actor.sample_normal(state, reparam=False)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic1.forward(state, actions)
q2_new_policy = self.critic2.forward(state, actions)
critic_value = torch.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
self.value.optimizer.zero_grad()
value_target = critic_value - log_probs
value_loss = 0.5 * F.mse_loss(value, value_target)
value_loss.backward(retain_graph=True)
self.value.optimizer.step()
actions, log_probs = self.actor.sample_normal(state, reparam=True)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic1.forward(state, actions)
q2_new_policy = self.critic2.forward(state, actions)
critic_value = torch.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
actor_loss = log_probs - critic_value
actor_loss = torch.mean(actor_loss)
self.actor.optimizer.zero_grad()
actor_loss.backward(retain_graph=True)
self.actor.optimizer.step()
self.critic1.optimizer.zero_grad()
self.critic2.optimizer.zero_grad()
q_hat = self.scale*reward + self.gamma*value_
q1_old_policy = self.critic1.forward(state, action).view(-1)
q2_old_policy = self.critic2.forward(state, action).view(-1)
critic1_loss = 0.5 * F.mse_loss(q1_old_policy, q_hat)
critic2_loss = 0.5 * F.mse_loss(q2_old_policy, q_hat)
critic_loss = critic1_loss + critic2_loss
critic_loss.backward()
self.critic1.optimizer.step()
self.critic2.optimizer.step()
self.update_network_params()
def main():
# ENVIROMENT
env_name = "CartPole-v1"
# env_name = "LunarLander-v2"
env = gym.make(env_name)
n_actions = env.action_space.n
feature_dim = env.observation_space.shape[0]
# PARAMETERS
learning_rate = 1e-3
state_scale = 1.0
reward_scale = 1.0
clip = 0.2
n_epoch = 4
max_episodes = 10
max_timesteps = 200
batch_size = 32
max_iterations = 200
gamma = 0.99
gae_lambda = 0.95
entropy_coefficient = 0.01
# NETWORK
value_model = ValueNetwork(in_dim=feature_dim).to(device)
value_optimizer = optim.Adam(value_model.parameters(), lr=learning_rate)
policy_model = PolicyNetwork(in_dim=feature_dim, n=n_actions).to(device)
policy_optimizer = optim.Adam(policy_model.parameters(), lr=learning_rate)
# INIT
history = History()
observation = env.reset()
epoch_ite = 0
episode_ite = 0
train_ite = 0
running_reward = -500
# TENSORBOARD
timestr = time.strftime("%d%m%Y-%H%M%S-")
log_dir = "./runs/" + timestr + env_name + "-BS" + str(batch_size) + "-E" + \
str(max_episodes) + "-MT" + str(max_timesteps) + "-NE" + str(n_epoch) + \
"-LR" + str(learning_rate) + "-G" + str(gamma) + "-L" + str(gae_lambda)
writer = SummaryWriter(log_dir=log_dir)
# LOAD MODEL
# Create folder models
if not Path("./models").exists():
print("Creating Models folder")
Path("./models").mkdir()
model_path = Path("./models/" + env_name + ".tar")
if model_path.exists():
print("Loading model!")
#Load model
checkpoint = torch.load(model_path)
policy_model.load_state_dict(checkpoint['policy_model'])
policy_optimizer.load_state_dict(checkpoint['policy_optimizer'])
value_model.load_state_dict(checkpoint['value_model'])
value_optimizer.load_state_dict(checkpoint['value_optimizer'])
running_reward = checkpoint['running_reward']
for ite in tqdm(range(max_iterations), ascii=True):
if ite % 5 == 0:
torch.save({
'policy_model': policy_model.state_dict(),
'policy_optimizer': policy_optimizer.state_dict(),
'value_model': value_model.state_dict(),
'value_optimizer': value_optimizer.state_dict(),
'running_reward': running_reward}, model_path)
episode_ite, running_reward = collect(episode_ite, running_reward, env,
max_episodes, max_timesteps, state_scale,
reward_scale, writer, history, policy_model,
value_model, gamma, gae_lambda, device)
# Here we have collected N trajectories.
history.build_dataset()
data_loader = DataLoader(history, batch_size=batch_size, shuffle=True, drop_last=True)
policy_loss, value_loss, train_ite = train_network(data_loader, policy_model, value_model,
policy_optimizer, value_optimizer ,n_epoch, clip,
train_ite, writer, entropy_coefficient)
for p_l, v_l in zip(policy_loss, value_loss):
epoch_ite += 1
writer.add_scalar("Policy Loss", p_l, epoch_ite)
writer.add_scalar("Value Loss", v_l, epoch_ite)
history.free_memory()
# print("\n", running_reward)
writer.add_scalar("Running Reward", running_reward, epoch_ite)
if (running_reward > env.spec.reward_threshold):
print("\nSolved!")
break
def main(env_name, lr, state_scale, reward_scale, clip, train_epoch,
max_episodes, max_timesteps, batch_size, max_iterations, gamma,
gae_lambda, entropy_coefficient, start_running_reward, update_rate):
# ENVIROMENT
env_name = env_name
env = ChessEnv()
# PARAMETERS
learning_rate = lr
state_scale = state_scale
reward_scale = reward_scale
clip = clip
n_epoch = train_epoch
max_episodes = max_episodes
max_timesteps = max_timesteps
batch_size = batch_size
max_iterations = max_iterations
gamma = gamma
gae_lambda = gae_lambda
entropy_coefficient = entropy_coefficient
# NETWORK
value_model = ValueNetwork().to(device)
value_optimizer = optim.Adam(value_model.parameters(), lr=learning_rate)
policy_model = PolicyNetwork().to(device)
policy_optimizer = optim.Adam(policy_model.parameters(), lr=learning_rate)
# INIT
history = History()
epoch_ite = 0
episode_ite = 0
train_ite = 0
running_reward = start_running_reward
# TENSORBOARD
timestr = time.strftime("%d%m%Y-%H%M%S-")
log_dir = "./runs/" + timestr + env_name + "-BS" + str(batch_size) + "-E" + \
str(max_episodes) + "-MT" + str(max_timesteps) + "-NE" + str(n_epoch) + \
"-LR" + str(learning_rate) + "-G" + str(gamma) + "-L" + str(gae_lambda)
writer = SummaryWriter(log_dir=log_dir)
# LOAD MODEL
# Create folder models
if not Path("./models").exists():
print("Creating Models folder")
Path("./models").mkdir()
model_path = Path("./models/" + env_name + ".tar")
if model_path.exists():
print("Loading model!")
#Load model
checkpoint = torch.load(model_path)
policy_model.load_state_dict(checkpoint['policy_model'])
policy_optimizer.load_state_dict(checkpoint['policy_optimizer'])
value_model.load_state_dict(checkpoint['value_model'])
value_optimizer.load_state_dict(checkpoint['value_optimizer'])
running_reward = checkpoint['running_reward']
# Create SavedEnvs queue
SavedEnv = queue.SimpleQueue()
for _ in range(max_episodes):
env = ChessEnv()
SavedEnv.put((env, env.reset(), 0))
# START ITERATING
for ite in tqdm(range(max_iterations), ascii=True):
# Load model to rival each update_rate epochs
if ite % update_rate == 0:
print("\nUpdating")
rival_policy = PolicyNetwork().to(device)
rival_policy.load_state_dict(policy_model.state_dict())
if ite % 5 == 0:
torch.save(
{
'policy_model': policy_model.state_dict(),
'policy_optimizer': policy_optimizer.state_dict(),
'value_model': value_model.state_dict(),
'value_optimizer': value_optimizer.state_dict(),
'running_reward': running_reward
}, model_path)
print("\nSimulating")
start_simulation = time.perf_counter()
q = queue.SimpleQueue()
env_list = []
while not SavedEnv.empty():
env_list.append(SavedEnv.get())
threads = []
for saved_env in env_list:
t = threading.Thread(target=collect,
args=[
q, env_name, saved_env, SavedEnv,
max_timesteps, state_scale, reward_scale,
policy_model, value_model, gamma,
gae_lambda, device, rival_policy
])
t.start()
threads.append(t)
for t in threads:
t.join()
# for saved_env in env_list:
# if ite % 20 == 0:
# update_policy = True
# else:
# update_policy = False
# collect(q, env_name, saved_env,
# SavedEnv, max_timesteps, state_scale, reward_scale,
# policy_model, value_model, gamma,
# gae_lambda, device, update_policy)
avg_episode_reward = []
# Write all episodes from queue to history buffer
while not q.empty():
episode, done = q.get()
history.episodes.append(episode)
avg_episode_reward.append((episode.reward, done))
end_simulation = time.perf_counter()
print(f"Simulation time: {end_simulation-start_simulation:.2f} ")
for ep_reward, done in avg_episode_reward:
if done:
running_reward = 0.05 * ep_reward + (1 - 0.05) * running_reward
writer.add_scalar("Average Episode Reward", ep_reward,
episode_ite)
episode_ite += 1
# avg_ep_reward = sum(avg_episode_reward) / len(avg_episode_reward)
# Here we have collected N trajectories and prepare dataset
history.build_dataset()
data_loader = DataLoader(history,
batch_size=batch_size,
shuffle=True,
drop_last=True)
print("Training")
policy_loss, value_loss, train_ite = train_network(
data_loader, policy_model, value_model, policy_optimizer,
value_optimizer, n_epoch, clip, train_ite, writer,
entropy_coefficient)
end_training = time.perf_counter()
print(f"Training time: {end_training-end_simulation:.2f}")
for p_l, v_l in zip(policy_loss, value_loss):
epoch_ite += 1
writer.add_scalar("Policy Loss", p_l, epoch_ite)
writer.add_scalar("Value Loss", v_l, epoch_ite)
history.free_memory()
# print("\n", running_reward)
writer.add_scalar("Running Reward", running_reward, epoch_ite)
if (running_reward > 0):
print("\nSolved!")
break
def main():
# ENVIROMENT
# env_name = "CartPole-v1"
# env_name = "LunarLander-v2"
# env_name = "Acrobot-v1"
env_name = "MountainCar-v0"
env = gym.make(env_name)
n_actions = env.action_space.n
feature_dim = env.observation_space.shape[0]
# PARAMETERS
learning_rate = 1e-3
state_scale = 1.0
reward_scale = 1.0
clip = 0.2
n_epoch = 4
max_episodes = 10
max_timesteps = 100
batch_size = 32
max_iterations = 1000
gamma = 0.99
gae_lambda = 0.95
entropy_coefficient = 0.01
env_threshold = env.spec.reward_threshold
# NETWORK
value_model = ValueNetwork(in_dim=feature_dim).to(device)
value_optimizer = optim.Adam(value_model.parameters(), lr=learning_rate)
policy_model = PolicyNetwork(in_dim=feature_dim, n=n_actions).to(device)
policy_optimizer = optim.Adam(policy_model.parameters(), lr=learning_rate)
# INIT
history = History()
observation = env.reset()
epoch_ite = 0
episode_ite = 0
train_ite = 0
running_reward = -500
# TENSORBOARD
timestr = time.strftime("%d%m%Y-%H%M%S-")
log_dir = "./runs/" + timestr + env_name + "-BS" + str(batch_size) + "-E" + \
str(max_episodes) + "-MT" + str(max_timesteps) + "-NE" + str(n_epoch) + \
"-LR" + str(learning_rate) + "-G" + str(gamma) + "-L" + str(gae_lambda)
writer = SummaryWriter(log_dir=log_dir)
# LOAD MODEL
# Create folder models
if not Path("./models").exists():
print("Creating Models folder")
Path("./models").mkdir()
model_path = Path("./models/" + env_name + ".tar")
if model_path.exists():
print("Loading model!")
#Load model
checkpoint = torch.load(model_path)
policy_model.load_state_dict(checkpoint['policy_model'])
policy_optimizer.load_state_dict(checkpoint['policy_optimizer'])
value_model.load_state_dict(checkpoint['value_model'])
value_optimizer.load_state_dict(checkpoint['value_optimizer'])
running_reward = checkpoint['running_reward']
EnvQueue = queue.SimpleQueue()
for _ in range(max_episodes):
env = gym.make(env_name)
observation = env.reset()
EnvQueue.put((env, observation, 0))
for ite in tqdm(range(max_iterations), ascii=True):
if ite % 5 == 0:
torch.save(
{
'policy_model': policy_model.state_dict(),
'policy_optimizer': policy_optimizer.state_dict(),
'value_model': value_model.state_dict(),
'value_optimizer': value_optimizer.state_dict(),
'running_reward': running_reward
}, model_path)
q = queue.SimpleQueue()
env_list = []
while not EnvQueue.empty():
env_list.append(EnvQueue.get())
threads = []
for env in env_list:
t = threading.Thread(target=collect,
args=[
q, env_name, env, EnvQueue, max_timesteps,
state_scale, reward_scale, policy_model,
value_model, gamma, gae_lambda, device
])
t.start()
threads.append(t)
for t in threads:
t.join()
avg_episode_reward = []
# Write all episodes from queue to history buffer
while not q.empty():
episode, done = q.get()
history.episodes.append(episode)
avg_episode_reward.append((episode.reward, done))
for ep_reward, done in avg_episode_reward:
if done:
running_reward = 0.05 * ep_reward + (1 - 0.05) * running_reward
writer.add_scalar("Running Reward", running_reward,
episode_ite)
writer.add_scalar("Episode Reward", ep_reward, episode_ite)
episode_ite += 1
# avg_ep_reward = sum(avg_episode_reward) / len(avg_episode_reward)
# Here we have collected N trajectories and prepare dataset
history.build_dataset()
data_loader = DataLoader(history,
batch_size=batch_size,
shuffle=True,
drop_last=True)
policy_loss, value_loss, train_ite = train_network(
data_loader, policy_model, value_model, policy_optimizer,
value_optimizer, n_epoch, clip, train_ite, writer,
entropy_coefficient)
for p_l, v_l in zip(policy_loss, value_loss):
epoch_ite += 1
writer.add_scalar("Policy Loss", p_l, epoch_ite)
writer.add_scalar("Value Loss", v_l, epoch_ite)
history.free_memory()
# print("\n", running_reward)
if (running_reward > env_threshold):
print("\nSolved!")
break
class SAC_Agent:
def __init__(self, load_from=None, will_train=True):
self.env = TorcsEnv(
path='/usr/local/share/games/torcs/config/raceman/quickrace.xml')
self.args = SAC_args()
self.buffer = ReplayBuffer(self.args.buffer_size)
action_dim = self.env.action_space.shape[0]
state_dim = self.env.observation_space.shape[0]
hidden_dim = 256
self.action_size = action_dim
self.state_size = state_dim
self.value_net = ValueNetwork(state_dim,
hidden_dim).to(self.args.device)
self.target_value_net = ValueNetwork(state_dim,
hidden_dim).to(self.args.device)
self.soft_q_net1 = SoftQNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.soft_q_net2 = SoftQNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.policy_net = PolicyNetwork(state_dim, action_dim,
hidden_dim).to(self.args.device)
self.target_value_net.load_state_dict(self.value_net.state_dict())
self.value_criterion = nn.MSELoss()
self.soft_q_loss1 = nn.MSELoss()
self.soft_q_loss2 = nn.MSELoss()
self.value_opt = optim.Adam(self.value_net.parameters(),
lr=self.args.lr)
self.soft_q_opt1 = optim.Adam(self.soft_q_net1.parameters(),
lr=self.args.lr)
self.soft_q_opt2 = optim.Adam(self.soft_q_net2.parameters(),
lr=self.args.lr)
self.policy_opt = optim.Adam(self.policy_net.parameters(),
lr=self.args.lr)
if will_train:
current_time = time.strftime('%d-%b-%y-%H.%M.%S', time.localtime())
self.plot_folder = f'plots/{current_time}'
self.model_save_folder = f'model/{current_time}'
make_sure_dir_exists(self.plot_folder)
make_sure_dir_exists(self.model_save_folder)
self.cp = Checkpoint(self.model_save_folder)
if load_from is not None:
try:
self.load_checkpoint(load_from)
except FileNotFoundError:
print(f'{load_from} not found. Running default.')
else:
print('Starting from scratch.')
def train(self):
remove_log_file()
clear_action_logs()
eps_n = 0
rewards = []
test_rewards = []
best_reward = -np.inf
info = None
for eps_n in range(1, self.args.max_eps + 1): # Train loop
self.set_mode('train')
relaunch = (eps_n - 1) % (20 / self.args.test_rate) == 0
state = self.env.reset(relaunch=relaunch,
render=False,
sampletrack=False)
eps_r = 0
sigma = (self.args.start_sigma - self.args.end_sigma) * (max(
0, 1 - (eps_n - 1) / self.args.max_eps)) + self.args.end_sigma
randomprocess = OrnsteinUhlenbeckProcess(self.args.theta, sigma,
self.action_size)
for step in range(self.args.max_eps_time): # Episode
action = self.policy_net.get_train_action(state, randomprocess)
next_state, reward, done, info = self.env.step(action)
self.buffer.push(state, action, reward, next_state, done)
state = next_state
eps_r += reward
if len(self.buffer) > self.args.batch_size:
self.update()
if done:
break
rewards.append(eps_r)
test_reward = self.test(eps_n)
test_rewards.append(test_reward)
if test_reward > best_reward:
best_reward = test_reward
self.save_checkpoint(eps_n, best_reward)
info_str = ', '.join(
[key for key in info.keys() if key != 'place'])
info_str += f", {info['place']}. place"
log(f'Episode {eps_n:<4} Reward: {eps_r:>7.2f} Test Reward: {test_reward:>7.2f} Info: {info_str}'
)
if eps_n % self.args.plot_per == 0:
self.plot(rewards, test_rewards, eps_n)
def update(self):
state, action, reward, next_state, done = self.buffer.sample(
self.args.batch_size)
state = FloatTensor(state).to(self.args.device)
next_state = FloatTensor(next_state).to(self.args.device)
action = FloatTensor(action).to(self.args.device)
reward = FloatTensor(reward).unsqueeze(1).to(self.args.device)
done = FloatTensor(np.float32(done)).unsqueeze(1).to(self.args.device)
predicted_q_value1 = self.soft_q_net1(state, action)
predicted_q_value2 = self.soft_q_net2(state, action)
predicted_value = self.value_net(state)
new_action, log_prob, epsilon, mean, log_std = self.policy_net.evaluate(
state)
# Training Q function
target_value = self.target_value_net(next_state)
target_q_value = reward + (1 - done) * self.args.gamma * target_value
q_value_loss1 = self.soft_q_loss1(predicted_q_value1,
target_q_value.detach())
q_value_loss2 = self.soft_q_loss2(predicted_q_value2,
target_q_value.detach())
self.soft_q_opt1.zero_grad()
q_value_loss1.backward()
if self.args.clipgrad:
self.clip_grad(self.soft_q_net1.parameters())
self.soft_q_opt1.step()
self.soft_q_opt2.zero_grad()
q_value_loss2.backward()
if self.args.clipgrad:
self.clip_grad(self.soft_q_net2.parameters())
self.soft_q_opt2.step()
# Training Value function
predicted_new_q_value = torch.min(self.soft_q_net1(state, new_action),
self.soft_q_net2(state, new_action))
target_value_func = predicted_new_q_value - self.args.alpha * log_prob.sum(
)
value_loss = self.value_criterion(predicted_value,
target_value_func.detach())
self.value_opt.zero_grad()
value_loss.backward()
if self.args.clipgrad:
self.clip_grad(self.value_net.parameters())
self.value_opt.step()
# Training Policy function
policy_loss = (log_prob - predicted_new_q_value).mean()
self.policy_opt.zero_grad()
policy_loss.backward()
if self.args.clipgrad:
self.clip_grad(self.policy_net.parameters())
self.policy_opt.step()
# Updating target value network
for target_param, param in zip(self.target_value_net.parameters(),
self.value_net.parameters()):
target_param.data.copy_(target_param.data *
(1.0 - self.args.soft_tau) +
param.data * self.args.soft_tau)
def test(self, eps_n):
self.set_mode('eval')
rewards = []
for step in range(self.args.test_rate):
render = (eps_n % 30 == 0) and (step == 0)
relaunch = render or ((eps_n % 30 == 0) and (step == 1))
state = self.env.reset(relaunch=relaunch,
render=render,
sampletrack=False)
running_reward = 0
for t in range(self.args.max_eps_time):
action = self.policy_net.get_test_action(state)
state, reward, done, info = self.env.step(action)
store(action, eps_n, reward, info, t == 0)
running_reward += reward
if done:
break
rewards.append(running_reward)
avg_reward = sum(rewards) / self.args.test_rate
return avg_reward
def plot(self, rewards, test_rewards, eps_n):
torch.save({
'train_rewards': rewards,
'test_rewards': test_rewards
}, f'{self.plot_folder}/{eps_n}.pth')
figure = plt.figure()
plt.plot(rewards, label='Train Rewards')
plt.plot(test_rewards, label='Test Rewards')
plt.xlabel('Episode')
plt.legend()
plt.savefig(f'{self.plot_folder}/{eps_n}.png')
try:
send_mail(f'Improved Torcs SAC | Episode {eps_n}',
f'{self.plot_folder}/{eps_n}.png')
log('Mail has been sent.')
except (KeyboardInterrupt, SystemExit):
print('KeyboardInterrupt or SystemExit')
raise
except Exception as e:
print('Mail Exception occured:', e)
emsg = e.args[-1]
emsg = emsg[:1].lower() + emsg[1:]
log('Couldn\'t send mail because', emsg)
def clip_grad(self, parameters):
for param in parameters:
param.grad.data.clamp_(-1, 1)
def set_mode(self, mode):
if mode == 'train':
self.value_net.train()
self.target_value_net.train()
self.soft_q_net1.train()
self.soft_q_net2.train()
self.policy_net.train()
elif mode == 'eval':
self.value_net.eval()
self.target_value_net.eval()
self.soft_q_net1.eval()
self.soft_q_net2.eval()
self.policy_net.eval()
else:
raise ValueError('mode should be either train or eval')
def save_checkpoint(self, eps_n, test_reward):
self.cp.update(self.value_net, self.soft_q_net1, self.soft_q_net2,
self.policy_net)
self.cp.save(f'e{eps_n}-r{test_reward:.4f}.pth')
log(f'Saved checkpoint at episode {eps_n}.')
def load_checkpoint(self, load_from):
state_dicts = torch.load(load_from)
self.value_net.load_state_dict(state_dicts['best_value'])
self.soft_q_net1.load_state_dict(state_dicts['best_q1'])
self.soft_q_net2.load_state_dict(state_dicts['best_q2'])
self.policy_net.load_state_dict(state_dicts['best_policy'])
print(f'Loaded from {load_from}.')
def race(self, sampletrack=True):
with torch.no_grad():
state = self.env.reset(relaunch=True,
render=True,
sampletrack=sampletrack)
running_reward = 0
done = False
while not done:
action = self.policy_net.get_test_action(state)
state, reward, done, info = self.env.step(action)
running_reward += reward
print('Reward:', running_reward)
class Agent():
def __init__(self,
alpha,
beta,
input_dims,
tau,
env,
env_id,
gamma=0.99,
n_actions=2,
max_size=1000000,
layer_1_size=256,
layer_2_size=256,
batch_size=100,
reward_scale=2):
self.gamma = gamma
self.tau = tau
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.batch_size = batch_size
self.n_actions = n_actions
self.scale = reward_scale
self.actor = ActorNetwork(alpha,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_actor',
max_action=env.action_space.high)
self.critic_1 = CriticNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_critic_1')
self.critic_2 = CriticNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
n_actions=n_actions,
name=env_id + '_critic_2')
self.value = ValueNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
name=env_id + '_value')
self.target_value = ValueNetwork(beta,
input_dims,
layer_1_size,
layer_2_size,
name=env_id + '_target_value')
self.update_network_parameters(tau=1)
def choose_action(self, observation):
state = T.tensor([observation], dtype=T.float).to(self.actor.device)
actions, _ = self.actor.sample_normal(state, reparameterize=False)
return actions.cpu().detach().numpy()[0]
def remember(self, state, action, reward, state_, done):
self.memory.store_transitions(state, action, reward, state_, done)
def update_network_parameters(self, tau=None):
if tau is None:
tau = self.tau
value_params = self.value.named_parameters()
target_value_params = self.target_value.named_parameters()
value_state_dict = dict(value_params)
target_value_state_dict = dict(target_value_params)
for name in value_state_dict:
value_state_dict[name] = tau*value_state_dict[name].clone() \
+ (1-tau)*target_value_state_dict[name].clone()
self.target_value.load_state_dict(value_state_dict)
def save_models(self):
print('.... saving models ....')
self.actor.save_checkpoint()
self.value.save_checkpoint()
self.target_value.save_checkpoint()
self.critic_1.save_checkpoint()
self.critic_2.save_checkpoint()
def load_models(self):
print('.... loading models ....')
self.actor.load_checkpoint()
self.value.load_checkpoint()
self.target_value.load_checkpoint()
self.critic_1.load_checkpoint()
self.critic_2.load_checkpoint()
def learn(self):
if self.memory.mem_cntr < self.batch_size:
return
state, action, reward, state_, done =\
self.memory.sample_buffer(self.batch_size)
state = T.tensor(state, dtype=T.float).to(self.critic_1.device)
state_ = T.tensor(state_, dtype=T.float).to(self.critic_1.device)
action = T.tensor(action, dtype=T.float).to(self.critic_1.device)
reward = T.tensor(reward, dtype=T.float).to(self.critic_1.device)
done = T.tensor(done).to(self.critic_1.device)
value = self.value(state).view(-1)
value_ = self.target_value(state_).view(-1)
value_[done] = 0.0
actions, log_probs = self.actor.sample_normal(state,
reparameterize=False)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic_1(state, actions)
q2_new_policy = self.critic_2(state, actions)
critic_value = T.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
self.value.optimizer.zero_grad()
value_target = critic_value - log_probs
value_loss = 0.5 * F.mse_loss(value, value_target)
value_loss.backward(retain_graph=True)
self.value.optimizer.step()
actions, log_probs = self.actor.sample_normal(state,
reparameterize=True)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic_1(state, actions)
q2_new_policy = self.critic_2(state, actions)
critic_value = T.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
actor_loss = log_probs - critic_value
actor_loss = T.mean(actor_loss)
self.actor.optimizer.zero_grad()
actor_loss.backward(retain_graph=True)
self.actor.optimizer.step()
self.critic_1.optimizer.zero_grad()
self.critic_2.optimizer.zero_grad()
q_hat = self.scale * reward + self.gamma * value_
q1_old_policy = self.critic_1(state, action).view(-1)
q2_old_policy = self.critic_2(state, action).view(-1)
critic_1_loss = 0.5 * F.mse_loss(q1_old_policy, q_hat)
critic_2_loss = 0.5 * F.mse_loss(q2_old_policy, q_hat)
critic_loss = critic_1_loss + critic_2_loss
critic_loss.backward()
self.critic_1.optimizer.step()
self.critic_2.optimizer.step()
self.update_network_parameters()
numrun = 1
for run in range(numrun):
env = make_env()
in_size = env.observation_space.shape[0]
num_actions = env.action_space.n
network = network_factory(in_size, num_actions, env)
network.to(device)
pe = PolicyNetwork(network)
# Load policy to test
#pe.network.load_state_dict(torch.load('saved_network_50000_baseline.pkl'))
ve = ValueNetwork(in_size)
ep_returns = reinforce(env, pe, ve, episodes) #,ve , loss_policy, loss_value
#fwrite = open('runs_data/'+str(run)+'.pkl','wb')
#fwrite = open('runs_data/0.pkl','wb')
#pickle.dump(ep_returns, fwrite)
#fwrite.close()
window = 100
plt.figure(figsize=(12,8))
plt.plot(sliding_window(ep_returns, window))
plt.title("Episode Return")
plt.xlabel("Episode")
plt.ylabel("Average Return (Sliding Window 100)")
class Agent():
def __init__(self,
alpha=0.0003,
beta=.0003,
input_dims=[8],
env=None,
gamma=.99,
n_actions=2,
max_size=1000000,
layer1_size=256,
layer2_size=256,
tau=.005,
batch_size=256,
reward_scale=2):
# reward scales depends on action convention for the environment
self.gamma = gamma
self.tau = tau
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.batch_size = batch_size
self.n_actions = n_actions
# set up classes
self.actor = ActorNetwork(alpha,
input_dims,
max_action=env.action_space.high,
n_actions=n_actions,
name='actor')
self.critic1 = CriticNetwork(beta,
input_dims,
n_actions=n_actions,
name='critic_1')
self.critic2 = CriticNetwork(beta,
input_dims,
n_actions=n_actions,
name='critic_2')
self.value = ValueNetwork(beta, input_dims, name='value')
# target value
self.target_value = ValueNetwork(beta, input_dims, name='target_value')
self.scale = reward_scale
self.update_network_parameters(tau=1)
def choose_action(self, observation):
# here we turn into a tensor
state = T.tensor([observation]).to(self.actor.device).float()
# print(type(state))
actions, _ = self.actor.sample_normal(state, reparameterize=False)
return actions.cpu().detach().numpy()[0]
def remember(self, state, action, reward, new_state, done):
self.memory.store_transition(state, action, reward, new_state, done)
def update_network_parameters(self, tau=None):
if tau is None:
tau = self.tau
target_value_params = self.target_value.named_parameters()
value_params = self.value.named_parameters()
target_value_state_dict = dict(target_value_params)
value_state_dict = dict(value_params)
for name in value_state_dict:
value_state_dict[name] = tau * value_state_dict[name].clone() + (
1 - tau) * target_value_state_dict[name].clone()
self.target_value.load_state_dict(value_state_dict)
def save_models(self):
print("saving models:")
self.actor.save_checkpoint()
self.value.save_checkpoint()
self.target_value.save_checkpoint()
self.critic1.save_checkpoint()
self.critic2.save_checkpoint()
def load_models(self):
print("loading models:")
self.actor.load_checkpoint()
self.value.load_checkpoint()
self.target_value.load_checkpoint()
self.critic1.load_checkpoint()
self.critic2.load_checkpoint()
def learn(self):
# must fully load up memory, otherwise must keep learning
if self.memory.mem_cntr < self.batch_size:
return
state, action, reward, new_state, done = self.memory.sample_buffer(
self.batch_size)
reward = T.tensor(reward, dtype=T.float).to(self.actor.device)
done = T.tensor(done).to(self.actor.device)
state_ = T.tensor(new_state, dtype=T.float).to(self.actor.device)
state = T.tensor(state, dtype=T.float).to(self.actor.device)
action = T.tensor(action, dtype=T.float).to(self.actor.device)
value = self.value(state).view(-1)
value_ = self.target_value(state_).view(-1)
value_[done] = 0.0
actions, log_probs = self.actor.sample_normal(state,
reparameterize=False)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic1.forward(state, actions)
q2_new_policy = self.critic2.forward(state, actions)
critic_value = T.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
self.value.optimizer.zero_grad()
value_target = critic_value - log_probs
value_loss = .5 * F.mse_loss(value, value_target)
value_loss.backward(retain_graph=True)
self.value.optimizer.step()
actions, log_probs = self.actor.sample_normal(state,
reparameterize=True)
log_probs = log_probs.view(-1)
q1_new_policy = self.critic1.forward(state, actions)
q2_new_policy = self.critic2.forward(state, actions)
critic_value = T.min(q1_new_policy, q2_new_policy)
critic_value = critic_value.view(-1)
actor_loss = log_probs - critic_value
actor_loss = T.mean(actor_loss)
self.actor.optimizer.zero_grad()
actor_loss.backward(retain_graph=True)
self.actor.optimizer.step()
self.critic1.optimizer.zero_grad()
self.critic2.optimizer.zero_grad()
q_hat = self.scale * reward + self.gamma * value_
q1_old_policy = self.critic1.forward(state, action).view(-1)
q2_old_policy = self.critic2.forward(state, action).view(-1)
critic_1_loss = .5 * F.mse_loss(q1_old_policy, q_hat)
critic_2_loss = .5 * F.mse_loss(q2_old_policy, q_hat)
critic_loss = critic_1_loss + critic_2_loss
critic_loss.backward()
self.critic1.optimizer.step()
self.critic2.optimizer.step()
self.update_network_parameters()