def train(args, env, policy_net, value_net, running_state):
for i_episode in count(1):
memory = Memory()
num_steps = 0
reward_batch = 0
num_episodes = 0
while num_steps < args.batch_size:
state = env.reset()
state = running_state(state)
reward_sum = 0
for t in range(10000): # Don't infinite loop while learning
action = select_action(state, policy_net)
action = action.data[0].numpy()
next_state, reward, done, _ = env.step(action)
reward_sum += reward
next_state = running_state(next_state)
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state,
reward)
if args.render:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
#########################
# TRPO update parameters
#########################
update_trpo = trpo.update_params(batch, value_net, policy_net, args,
trpo_functions)
update_trpo.execute()
if i_episode % args.log_interval == 0:
print('Episode {}\tLast reward: {}\tAverage reward {:.2f}'.format(
i_episode, reward_sum, reward_batch))
def collect_samples(policy_net, min_batch_size):
memory = Memory()
num_steps = 0
reward_batch = 0
num_episodes = 0
while (num_steps < min_batch_size):
state = env.reset()
reward_sum = 0
for t in range(10000):
action = select_action(policy_net, state)
action = action.data[0].numpy()
next_state, reward, done, _ = env.step(action)
reward_sum += reward
mask = 0 if done else 1
memory.push(state, np.array([action]), mask, next_state, reward)
if render:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
print(num_episodes)
reward_batch = reward_batch / num_episodes
batch = memory.sample()
return batch, reward_batch
mem_mask.append(mask)
mem_next.append(next_state)
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch.append(reward_sum)
tb_writer.add_scalar('Mean Episode Reward',
np.sum(reward_batch) / len(reward_batch), episode)
evaluate(episode)
rewards = expert_reward(states, actions)
for idx in range(len(states)):
memory.push(states[idx][0], actions[idx], mem_mask[idx],
mem_next[idx], rewards[idx][0])
batch = memory.sample()
update_params(batch)
# ce_train_discriminator_one_step(gen_state_action_batch, expert_state_action_batch, expert_pvalue)
actions = torch.from_numpy(np.concatenate(actions))
states = torch.from_numpy(np.concatenate(states))
idx = np.random.randint(0, expert_traj.shape[0], num_steps)
expert_state_action = expert_traj[idx, :]
expert_pvalue = expert_conf[idx, :]
expert_state_action = torch.Tensor(expert_state_action).to(device)
expert_pvalue = torch.Tensor(expert_pvalue / Z).to(device)
state_action = torch.cat((states, actions), 1).to(device)
ce_loss = ce_train_discriminator_one_step(state_action,
expert_state_action,
expert_pvalue)
tb_writer.add_scalar("Disc CEloss", ce_loss, episode)
def train(rank, args, traffic_light, counter, shared_model,
shared_grad_buffers, shared_obs_stats, opt_ac):
best_result = -1000
torch.manual_seed(args.seed + rank)
torch.set_default_tensor_type('torch.DoubleTensor')
num_inputs = args.feature
num_actions = 9
last_state = [0] * 41
last_v = [0] * 10
#last_state = numpy.zeros(48)
env = RunEnv(visualize=False)
#running_state = ZFilter((num_inputs,), clip=5)
#running_reward = ZFilter((1,), demean=False, clip=10)
episode_lengths = []
PATH_TO_MODEL = '../models/' + str(args.bh)
ac_net = ActorCritic(num_inputs, num_actions)
#running_state = ZFilter((num_inputs,), clip=5)
start_time = time.time()
for i_episode in range(args.start_epoch + 1, 999999):
#print(shared_obs_stats.n[0])
#print('hei')
#if rank == 0:
# print(running_state.rs._n)
signal_init = traffic_light.get()
memory = Memory()
ac_net.load_state_dict(shared_model.state_dict())
num_steps = 0
reward_batch = 0
num_episodes = 0
#Tot_loss = 0
#Tot_num =
while num_steps < args.batch_size:
#state = env.reset()
#print(num_steps)
state = env.reset(difficulty=0)
#state = numpy.array(state)
last_state, last_v, state = process_observation(
last_state, last_v, state)
state = numpy.array(state)
#state = running_state(state)
state = Variable(torch.Tensor(state).unsqueeze(0))
shared_obs_stats.observes(state)
state = shared_obs_stats.normalize(state)
state = state.data[0].numpy()
#print(state)
#return
#print(AA)
#print(type(AA))
#print(type(state))
#print(AA.shape)
#print(state.shape)
reward_sum = 0
#timer = time.time()
for t in range(10000): # Don't infinite loop while learning
#print(t)
if args.use_sep_pol_val:
action = select_action(state)
else:
action = select_action_actor_critic(state, ac_net)
#print(action)
action = action.data[0].numpy()
if numpy.any(numpy.isnan(action)):
print(state)
print(action)
print(ac_net.affine1.weight)
print(ac_net.affine1.weight.data)
print('ERROR')
#action = select_action_actor_critic(state,ac_net)
#action = action.data[0].numpy()
#state = state + numpy.random.rand(args.feature)*0.001
raise RuntimeError('action NaN problem')
#print(action)
#print("------------------------")
#timer = time.time()
reward = 0
if args.skip:
#env.step(action)
_, A, _, _ = env.step(action)
reward += A
_, A, _, _ = env.step(action)
reward += A
BB = numpy.append(action, action)
next_state, A, done, _ = env.step(BB)
reward += A
#print(next_state)
#last_state = process_observation(state)
last_state, last_v, next_state = process_observation(
last_state, last_v, next_state)
next_state = numpy.array(next_state)
#print(next_state)
#print(next_state.shape)
#return
reward_sum += reward
#print('env:')
#print(time.time()-timer)
#last_state ,next_state = update_observation(last_state,next_state)
#next_state = running_state(next_state)
next_state = Variable(torch.Tensor(next_state).unsqueeze(0))
shared_obs_stats.observes(next_state)
next_state = shared_obs_stats.normalize(next_state)
next_state = next_state.data[0].numpy()
#print(next_state[41:82])
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state,
reward)
#if args.render:
# env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
#print('env:')
#print(time.time()-timer)
#timer = time.time()
update_params_actor_critic(batch, args, ac_net, opt_ac)
shared_grad_buffers.add_gradient(ac_net)
counter.increment()
epoch = i_episode
if (i_episode % args.log_interval == 0) and (rank == 0):
print(
'TrainEpisode {}\tTime{}\tLast reward: {}\tAverage reward {:.2f}'
.format(
i_episode,
time.strftime("%Hh %Mm %Ss",
time.gmtime(time.time() - start_time)),
reward_sum, reward_batch))
epoch = i_episode
if reward_batch > best_result:
best_result = reward_batch
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': shared_model.state_dict(),
'optimizer': opt_ac.state_dict(),
'obs': shared_obs_stats,
}, PATH_TO_MODEL, 'best')
if epoch % 30 == 1:
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': shared_model.state_dict(),
'optimizer': opt_ac.state_dict(),
'obs': shared_obs_stats,
}, PATH_TO_MODEL, epoch)
# wait for a new signal to continue
while traffic_light.get() == signal_init:
pass
class DDPG():
def __init__(self, parameters):
self.parameters = parameters
self.env = gym.make(self.parameters['env'])
self.nA = self.env.action_space.sample().shape[0]
self.state_size = self.env.reset().shape[0]
# Build our replay memory
self.memory = Memory(replay_size=self.parameters['replay_size'],
action_size=self.nA,
state_size=self.state_size,
batch_size=self.parameters['batch_size'])
# Create actor and critic
self.actor_critic = ActorCritic(
actor_lr=parameters['actor_learning_rate'],
critic_lr=parameters['critic_learning_rate'],
gamma=parameters['gamma'],
state_size=self.state_size,
action_size=self.nA,
tau=parameters['tau'])
def train(self):
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
config.gpu_options.allow_growth = True
# Create global step and increment operation
global_step_tensor = tf.Variable(0,
trainable=False,
name='global_step')
increment_global_step = tf.assign_add(global_step_tensor, 1)
# Create model saver
saver = tf.train.Saver()
sess = tf.Session(config=config)
if not self.parameters['restore']:
sess.run(tf.global_variables_initializer())
else:
saver.restore(sess, tf.train.latest_checkpoint('./saves'))
self.actor_critic.set_moving_to_target(sess)
run_id = np.random.randint(10000)
trainwriter = tf.summary.FileWriter(logdir='./logs/' + str(run_id),
graph=sess.graph)
# Get action noise
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(self.nA),
sigma=float(self.parameters['sigma']) * np.ones(self.nA))
# Fill Replay Memory
state = self.env.reset()
fill_amount = 0
while fill_amount < self.parameters['replay_init_size']:
action = self.env.action_space.sample()
next_state, reward, done, _ = self.env.step(action)
if done:
state = self.env.reset()
else:
fill_amount += 1
self.memory.add(state, action, reward, done, next_state)
state = next_state
# Main Loop
steps = 0
for i in range(self.parameters['num_epochs']):
avg_epoch_rewards = 0
num_epochs = 1
for e in range(self.parameters['num_episodes']):
state = self.env.reset()
ep_reward = 0
# Perform rollout
while True:
noise = action_noise()
action = self.actor_critic.pi(sess, state[None, ...])
action += noise
action = np.clip(action, self.env.action_space.low[0],
self.env.action_space.high[0])
assert action.shape == self.env.action_space.shape
"""
# UNCOMMENT TO PRINT ACTIONS
a0 = tf.Summary(value=[tf.Summary.Value(tag="action_0", simple_value=action[0,0])])
trainwriter.add_summary(a0,steps)
a1 = tf.Summary(value=[tf.Summary.Value(tag="action_1", simple_value=action[0,1])])
trainwriter.add_summary(a1,steps)
a2 = tf.Summary(value=[tf.Summary.Value(tag="action_2", simple_value=action[0,2])])
trainwriter.add_summary(a2,steps)
steps += 1
"""
next_state, reward, done, _ = self.env.step(action)
self.memory.add(state, action, reward, done, next_state)
if self.parameters['render_train']:
self.env.render()
ep_reward += reward
if done:
reward_summary = tf.Summary(value=[
tf.Summary.Value(tag="ep_rewards",
simple_value=ep_reward)
])
trainwriter.add_summary(
reward_summary,
i * self.parameters['num_episodes'] + e)
action_noise.reset()
break
state = next_state
avg_epoch_rewards = avg_epoch_rewards + (
ep_reward - avg_epoch_rewards) / num_epochs
num_epochs += 1
# Perform train
for t in range(self.parameters['num_train_steps']):
s_state, s_action, s_reward, s_next_state, s_terminal = self.memory.sample(
)
# Train actor critic model
self.actor_critic.update(sess=sess,
filewriter=trainwriter,
state_batch=s_state,
next_state_batch=s_next_state,
action_batch=s_action,
reward_batch=s_reward,
done_batch=s_terminal)
sess.run(increment_global_step)
# Print out epoch stats here
table_data = [['Epoch', 'Average Reward'],
[
str(i) + "/" +
str(self.parameters['num_epochs']),
str(avg_epoch_rewards)
]]
table = AsciiTable(table_data, "Training Run: " + str(run_id))
save_path = saver.save(sess, "./saves/model.ckpt")
os.system('clear')
print("Model saved in path: %s" % save_path + "\n" + table.table)
def test(self):
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
config.gpu_options.allow_growth = True
saver = tf.train.Saver()
sess = tf.Session(config=config)
saver.restore(sess, tf.train.latest_checkpoint('./saves'))
while True:
state = self.env.reset()
# Perform rollout
while True:
action = self.actor_critic.pi(sess, state[None, ...])
action = np.clip(action, self.env.action_space.low[0],
self.env.action_space.high[0])
assert action.shape == self.env.action_space.shape
next_state, reward, done, _ = self.env.step(action)
self.env.render()
if done:
break
state = next_state
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state,
reward)
if args.render:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample() # 1200 transition [5*1200*...]
loss = update_params(batch)
if i_episode % args.log_interval == 0:
episode_list.append(i_episode)
reward_tmp.append(reward_batch)
reward_list.append(reward_tmp)
args.batch_size += 1000
print("batchsize", batchsize_list)
print("episode", episode_list)
print("reward", reward_list)
parazhexian(episode_list, batchsize_list, reward_list, 0.995)
class TD3:
def __init__(self,
env_id,
render=False,
num_process=1,
memory_size=1000000,
lr_p=1e-3,
lr_v=1e-3,
gamma=0.99,
polyak=0.995,
action_noise=0.1,
target_action_noise_std=0.2,
target_action_noise_clip=0.5,
explore_size=10000,
step_per_iter=3000,
batch_size=100,
min_update_step=1000,
update_step=50,
policy_update_delay=2,
seed=1,
model_path=None):
self.env_id = env_id
self.gamma = gamma
self.polyak = polyak
self.action_noise = action_noise
self.target_action_noise_std = target_action_noise_std
self.target_action_noise_clip = target_action_noise_clip
self.memory = Memory(memory_size)
self.explore_size = explore_size
self.step_per_iter = step_per_iter
self.render = render
self.num_process = num_process
self.lr_p = lr_p
self.lr_v = lr_v
self.batch_size = batch_size
self.min_update_step = min_update_step
self.update_step = update_step
self.policy_update_delay = policy_update_delay
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, self.num_actions = get_env_info(
self.env_id)
assert env_continuous, "TD3 is only applicable to continuous environment !!!!"
self.action_low, self.action_high = self.env.action_space.low[
0], self.env.action_space.high[0]
# seeding
np.random.seed(self.seed)
torch.manual_seed(self.seed)
self.env.seed(self.seed)
self.policy_net = Policy(num_states, self.num_actions,
self.action_high).to(device)
self.policy_net_target = Policy(num_states, self.num_actions,
self.action_high).to(device)
self.value_net_1 = Value(num_states, self.num_actions).to(device)
self.value_net_target_1 = Value(num_states,
self.num_actions).to(device)
self.value_net_2 = Value(num_states, self.num_actions).to(device)
self.value_net_target_2 = Value(num_states,
self.num_actions).to(device)
self.running_state = ZFilter((num_states, ), clip=5)
self.num_states = num_states
if self.model_path:
print("Loading Saved Model {}_td3.p".format(self.env_id))
self.policy_net, self.value_net_1, self.value_net_2, self.running_state = pickle.load(
open('{}/{}_td3.p'.format(self.model_path, self.env_id), "rb"))
self.policy_net_target.load_state_dict(self.policy_net.state_dict())
self.value_net_target_1.load_state_dict(self.value_net_1.state_dict())
self.value_net_target_2.load_state_dict(self.value_net_2.state_dict())
self.optimizer_p = optim.Adam(self.policy_net.parameters(),
lr=self.lr_p)
self.optimizer_v_1 = optim.Adam(self.value_net_1.parameters(),
lr=self.lr_v)
self.optimizer_v_2 = optim.Adam(self.value_net_2.parameters(),
lr=self.lr_v)
def choose_action(self, state, noise_scale):
"""select action"""
state = FLOAT(state).unsqueeze(0).to(device)
with torch.no_grad():
action, log_prob = self.policy_net.get_action_log_prob(state)
action = action.cpu().numpy()[0]
# add noise
noise = noise_scale * np.random.randn(self.num_actions)
action += noise
action = np.clip(action, -self.action_high, self.action_high)
return action
def eval(self, i_iter, render=False):
"""evaluate model"""
state = self.env.reset()
test_reward = 0
while True:
if render:
self.env.render()
action = self.choose_action(state, 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):
"""interact"""
global_steps = (i_iter - 1) * self.step_per_iter
log = dict()
num_steps = 0
num_episodes = 0
total_reward = 0
min_episode_reward = float('inf')
max_episode_reward = float('-inf')
while num_steps < self.step_per_iter:
state = self.env.reset()
episode_reward = 0
for t in range(10000):
if self.render:
self.env.render()
if global_steps < self.explore_size: # explore
action = self.env.action_space.sample()
else: # action with noise
action = self.choose_action(state, self.action_noise)
next_state, reward, done, _ = self.env.step(action)
# next_state = self.running_state(next_state)
mask = 0 if done else 1
# ('state', 'action', 'reward', 'next_state', 'mask', 'log_prob')
self.memory.push(state, action, reward, next_state, mask, None)
episode_reward += reward
global_steps += 1
num_steps += 1
if global_steps >= self.min_update_step and global_steps % self.update_step == 0:
for k in range(self.update_step):
batch, permuted_batch = self.memory.sample(
self.batch_size) # random sample batch
self.update(batch, permuted_batch, k)
if done or num_steps >= self.step_per_iter:
break
state = next_state
num_episodes += 1
total_reward += episode_reward
min_episode_reward = min(episode_reward, min_episode_reward)
max_episode_reward = max(episode_reward, max_episode_reward)
self.env.close()
log['num_steps'] = num_steps
log['num_episodes'] = num_episodes
log['total_reward'] = total_reward
log['avg_reward'] = total_reward / num_episodes
log['max_episode_reward'] = max_episode_reward
log['min_episode_reward'] = min_episode_reward
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}")
# record reward information
writer.add_scalar("rewards/total_reward", log['total_reward'], i_iter)
writer.add_scalar("rewards/average_reward", log['avg_reward'], i_iter)
writer.add_scalar("rewards/min_reward", log['min_episode_reward'],
i_iter)
writer.add_scalar("rewards/max_reward", log['max_episode_reward'],
i_iter)
writer.add_scalar("rewards/num_steps", log['num_steps'], i_iter)
def update(self, batch, batch2, k_iter):
"""learn model"""
batch_state = FLOAT(batch.state).to(device)
batch_action = FLOAT(batch.action).to(device)
batch_reward = FLOAT(batch.reward).to(device)
batch_next_state = FLOAT(batch.next_state).to(device)
batch_mask = FLOAT(batch.mask).to(device)
# update by TD3
alg_step_stats = td3_step(
self.policy_net, self.policy_net_target, self.value_net_1,
self.value_net_target_1, self.value_net_2, self.value_net_target_2,
self.optimizer_p, self.optimizer_v_1, self.optimizer_v_2,
batch_state, batch_action, batch_reward, batch_next_state,
batch_mask, self.gamma, self.polyak, self.target_action_noise_std,
self.target_action_noise_clip, self.action_high,
k_iter % self.policy_update_delay == 0)
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.policy_net, self.value_net_1, self.value_net_2,
self.running_state),
open('{}/{}_td3.p'.format(save_path, self.env_id), 'wb'))
class DDPG():
def __init__(self, parameters):
self.parameters = parameters
self.env = gym.make(
self.parameters['env'][:self.parameters['env'].find('_')])
self.nA = self.env.action_space.sample().shape[0]
self.state_size = self.env.reset().shape[0]
# Build our replay memory
self.memory = Memory(replay_size=self.parameters['replay_size'],
action_size=self.nA,
state_size=self.state_size,
batch_size=self.parameters['batch_size'])
# Create actor and critic
self.actor_critic = ActorCritic(
actor_lr=parameters['actor_learning_rate'],
critic_lr=parameters['critic_learning_rate'],
gamma=parameters['gamma'],
state_size=self.state_size,
action_size=self.nA,
tau=parameters['tau'])
def train(self):
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
config.gpu_options.allow_growth = True
# Create global step and increment operation
global_step_tensor = tf.Variable(0,
trainable=False,
name='global_step')
increment_global_step = tf.assign_add(global_step_tensor, 1)
# Create model saver
saver = tf.train.Saver(max_to_keep=None)
sess = tf.Session(config=config)
if not self.parameters['restore']:
sess.run(tf.global_variables_initializer())
else:
saver.restore(sess, tf.train.latest_checkpoint('./saves'))
self.actor_critic.set_moving_to_target(sess)
run_id = np.random.randint(10000)
trainwriter = tf.summary.FileWriter(logdir='./logs/' + str(run_id),
graph=sess.graph)
# Get action noise
action_noise = OrnsteinUhlenbeckActionNoise(
mu=np.zeros(self.nA),
sigma=float(self.parameters['sigma']) * np.ones(self.nA))
# Fill Replay Memory
state = self.env.reset()
fill_amount = 0
while fill_amount < self.parameters['replay_init_size']:
action = self.env.action_space.sample()
next_state, reward, done, _ = self.env.step(action)
if done:
state = self.env.reset()
else:
fill_amount += 1
self.memory.add(state, action, reward, done, next_state)
state = next_state
# Main Loop
plots = {'critic_loss': [], 'actor_loss': [], 'episode_reward': []}
plots_dir = './plots/'
weights_dir = './weights/'
graph_dir = './graph/'
if not os.path.exists(plots_dir):
os.makedirs(plots_dir)
if not os.path.exists(weights_dir):
os.makedirs(weights_dir)
if not os.path.exists(graph_dir):
os.makedirs(graph_dir)
saver.export_meta_graph(graph_dir + self.parameters['env'] +
'/graph.meta')
#cumulative step counter
cumu_step = 0
for i in range(self.parameters['num_epochs']):
avg_epoch_rewards = 0
n_epochs = 1
for e in range(self.parameters['num_episodes']):
state = self.env.reset()
ep_reward = 0
ep_n_action = 0
# Perform rollout
for _ in range(500):
noise = action_noise()
action = self.actor_critic.pi(sess, state[None, ...])
action += noise
action = np.clip(action, self.env.action_space.low[0],
self.env.action_space.high[0])
assert action.shape == self.env.action_space.shape
next_state, reward, done, _ = self.env.step(action)
# print(action)
# print(next_state)
# print(reward)
self.memory.add(state, action, reward, done, next_state)
if self.parameters['render_train']: self.env.render()
ep_reward += reward
ep_n_action += 1
cumu_step += 1
state = next_state
# Perform train
avg_critic_loss = 0.0
avg_actor_loss = 0.0
for t in range(self.parameters['num_train_steps']):
s_state, s_action, s_reward, s_next_state, s_terminal = self.memory.sample(
)
# Train actor critic model
_, _, critic_loss, actor_loss = self.actor_critic.update(
sess=sess,
filewriter=trainwriter,
state_batch=s_state,
next_state_batch=s_next_state,
action_batch=s_action,
reward_batch=s_reward,
done_batch=s_terminal)
avg_critic_loss += critic_loss
avg_actor_loss += actor_loss
sess.run(increment_global_step)
avg_critic_loss /= self.parameters['num_train_steps']
avg_actor_loss /= self.parameters['num_train_steps']
if done:
reward_summary = tf.Summary(value=[
tf.Summary.Value(tag="ep_rewards",
simple_value=ep_reward)
])
trainwriter.add_summary(
reward_summary,
i * self.parameters['num_episodes'] + e)
action_noise.reset()
break
avg_epoch_rewards = avg_epoch_rewards + (
ep_reward - avg_epoch_rewards) / n_epochs
n_epochs += 1
print('Epoch: {:d} | Reward: {:d} | Avg_Q_loss: {:.4f} | Avg_a_loss: {:.4f} | Episode: {:d} | Step: {:d} | Cumu Step: {:d}'\
.format(i+1, int(ep_reward), avg_critic_loss, avg_actor_loss, e+1, ep_n_action, cumu_step))
if e % 19 == 0:
save_path = saver.save(
sess,
weights_dir + self.parameters['env'] + '/model.ckpt',
global_step=i * e + 1)
plots['episode_reward'].append(ep_reward)
plots['critic_loss'].append(critic_loss)
plots['actor_loss'].append(critic_loss)
pickle.dump(
plots,
open(plots_dir + self.parameters['env'] + '_plot.pickle',
'wb'))
def test(self):
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
config.gpu_options.allow_growth = True
saver = tf.train.Saver()
sess = tf.Session(config=config)
saver.restore(
sess,
tf.train.latest_checkpoint(
'./weights/HalfCheetah-v2_kirkiles_train50episode_noise_norm_bufsize1Mi1k'
))
while True:
state = self.env.reset()
# Perform rollout
while True:
action = self.actor_critic.pi(sess, state[None, ...])
action = np.clip(action, self.env.action_space.low[0],
self.env.action_space.high[0])
#print(action)
assert action.shape == self.env.action_space.shape
next_state, reward, done, _ = self.env.step(action)
self.env.render()
if done:
break
state = next_state
if args.render and i_episode >= 20:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
# print('came')
#avergae policy reward
reward_batch /= num_episodes
batch = memory.sample()
models_path = str(args.env_name) + "_tests" + "/seed" + str(
args.seed) + "/parameters/episode_" + str(i_episode) + "/" + str(
args.plot_name)
torch.save(policy_net.state_dict(), models_path + "_policy")
torch.save(value_net.state_dict(), models_path + "_value")
if i_episode % args.parameter_save_interval == 0:
if args.use_parameter_noise:
post_evaluation_noise1.append(
post_evaluate(models_path, sigma, add_noise=True))
post_evaluation_noise2.append(
post_evaluate(models_path, sigma, add_noise=False))
def main(gamma=0.995, env_name="Walker2d-v2", tau=0.97, number_of_batches=500,\
batch_size=5000, maximum_steps=10000, render=False,\
seed=543, log_interval=1, entropy_coeff=0.0, clip_epsilon=0.2):
env = gym.make(env_name)
#Get number of inputs for A3CActor
num_inputs = env.observation_space.shape[0]
#Get number of outputs required for describing action
num_actions = env.action_space.shape[0]
env.seed(seed)
torch.manual_seed(seed)
actor_net = A3CActor(num_inputs, num_actions)
actor_optimizer = optim.Adam(actor_net.parameters(), lr=0.001)
running_state = ZFilter((num_inputs, ), clip=5)
running_reward = ZFilter((1, ), demean=False, clip=10)
episode_lengths = []
for i_episode in range(number_of_batches):
memory = Memory()
num_steps = 0
reward_batch = 0
num_episodes = 0
while num_steps < batch_size:
state = env.reset()
state = running_state(state)
reward_sum = 0
for t in range(maximum_steps):
action = select_action(state, actor_net)
action = action.data[0].numpy()
next_state, reward, done, _ = env.step(action)
reward_sum += reward
next_state = running_state(next_state)
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state,
reward)
if render:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
update_params(batch, actor_net, actor_optimizer, gamma, tau,
clip_epsilon)
if i_episode % log_interval == 0:
print('Episode {}\t Last reward: {}\tAverage reward {:.2f}'.format(
i_episode, reward_sum, reward_batch))
return
-1000.0 * reward)
if args.render:
env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
train_interval = i_episode % gan_interval
# if train_interval < 100:
batch_pro = memory_pro.sample()
print("Protogonist Training")
update_params(policy_net_pro, value_net_pro, batch_pro)
# else:
batch_adv = memory_adv.sample()
# print("Adversary Training")
update_params(policy_net_adv, value_net_adv, batch_adv)
reward_plot.append(reward_batch)
plt.title(
str(args.env_name) + r' Training Rewards Plot for $\varepsilon$ = ' +
str(epsilon))
plt.xlabel("Iterations")
plt.ylabel("Reward")
plt.plot(np.arange(i_episode),
reward_plot,
class SAC:
def __init__(self, env_name, n_states, n_actions, memory_size, batch_size,
gamma, alpha, lr, action_bounds, reward_scale):
self.env_name = env_name
self.n_states = n_states
self.n_actions = n_actions
self.memory_size = memory_size
self.batch_size = batch_size
self.gamma = gamma
self.alpha = alpha
self.lr = lr
self.action_bounds = action_bounds
self.reward_scale = reward_scale
self.memory = Memory(memory_size=self.memory_size)
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.policy_network = PolicyNetwork(
n_states=self.n_states,
n_actions=self.n_actions,
action_bounds=self.action_bounds).to(self.device)
self.q_value_network1 = QvalueNetwork(n_states=self.n_states,
n_actions=self.n_actions).to(
self.device)
self.q_value_network2 = QvalueNetwork(n_states=self.n_states,
n_actions=self.n_actions).to(
self.device)
self.value_network = ValueNetwork(n_states=self.n_states).to(
self.device)
self.value_target_network = ValueNetwork(n_states=self.n_states).to(
self.device)
self.value_target_network.load_state_dict(
self.value_network.state_dict())
self.value_target_network.eval()
self.value_loss = torch.nn.MSELoss()
self.q_value_loss = torch.nn.MSELoss()
self.value_opt = Adam(self.value_network.parameters(), lr=self.lr)
self.q_value1_opt = Adam(self.q_value_network1.parameters(),
lr=self.lr)
self.q_value2_opt = Adam(self.q_value_network2.parameters(),
lr=self.lr)
self.policy_opt = Adam(self.policy_network.parameters(), lr=self.lr)
def store(self, state, reward, done, action, next_state):
state = from_numpy(state).float().to("cpu")
reward = torch.Tensor([reward]).to("cpu")
done = torch.Tensor([done]).to("cpu")
action = torch.Tensor([action]).to("cpu")
next_state = from_numpy(next_state).float().to("cpu")
self.memory.add(state, reward, done, action, next_state)
def unpack(self, batch):
batch = Transition(*zip(*batch))
states = torch.cat(batch.state).view(self.batch_size,
self.n_states).to(self.device)
rewards = torch.cat(batch.reward).view(self.batch_size,
1).to(self.device)
dones = torch.cat(batch.done).view(self.batch_size, 1).to(self.device)
actions = torch.cat(batch.action).view(-1,
self.n_actions).to(self.device)
next_states = torch.cat(batch.next_state).view(
self.batch_size, self.n_states).to(self.device)
return states, rewards, dones, actions, next_states
def train(self):
if len(self.memory) < self.batch_size:
return 0, 0, 0
else:
batch = self.memory.sample(self.batch_size)
states, rewards, dones, actions, next_states = self.unpack(batch)
# Calculating the value target
reparam_actions, log_probs = self.policy_network.sample_or_likelihood(
states)
q1 = self.q_value_network1(states, reparam_actions)
q2 = self.q_value_network2(states, reparam_actions)
q = torch.min(q1, q2)
target_value = q.detach() - self.alpha * log_probs.detach()
value = self.value_network(states)
value_loss = self.value_loss(value, target_value)
# Calculating the Q-Value target
with torch.no_grad():
target_q = self.reward_scale * rewards + \
self.gamma * self.value_target_network(next_states) * (1 - dones)
q1 = self.q_value_network1(states, actions)
q2 = self.q_value_network2(states, actions)
q1_loss = self.q_value_loss(q1, target_q)
q2_loss = self.q_value_loss(q2, target_q)
policy_loss = (self.alpha * log_probs - q).mean()
self.policy_opt.zero_grad()
policy_loss.backward()
self.policy_opt.step()
self.value_opt.zero_grad()
value_loss.backward()
self.value_opt.step()
self.q_value1_opt.zero_grad()
q1_loss.backward()
self.q_value1_opt.step()
self.q_value2_opt.zero_grad()
q2_loss.backward()
self.q_value2_opt.step()
self.soft_update_target_network(self.value_network,
self.value_target_network)
return value_loss.item(), 0.5 * (
q1_loss + q2_loss).item(), policy_loss.item()
def choose_action(self, states):
states = np.expand_dims(states, axis=0)
states = from_numpy(states).float().to(self.device)
action, _ = self.policy_network.sample_or_likelihood(states)
return action.detach().cpu().numpy()[0]
@staticmethod
def soft_update_target_network(local_network, target_network, tau=0.005):
for target_param, local_param in zip(target_network.parameters(),
local_network.parameters()):
target_param.data.copy_(tau * local_param.data +
(1 - tau) * target_param.data)
def save_weights(self):
torch.save(self.policy_network.state_dict(),
self.env_name + "_weights.pth")
def load_weights(self):
self.policy_network.load_state_dict(
torch.load(self.env_name + "_weights.pth"))
def set_to_eval_mode(self):
self.policy_network.eval()
def main(gamma=0.995, env_name="Walker2d-v2", tau=0.97, number_of_batches=500,\
batch_size=5000, maximum_steps=10000, render=False,\
seed=543, log_interval=1, entropy_coeff=0.0, clip_epsilon=0.2):
env = gym.make(env_name)
#Get number of inputs for A3CActor
num_inputs = env.observation_space.shape[0]
#Get number of outputs required for describing action
num_actions = env.action_space.shape[0]
env.seed(seed)
torch.manual_seed(seed)
actor_net = A3CActor(num_inputs, num_actions)
actor_optimizer = optim.Adam(actor_net.parameters(), lr=0.001)
running_state = ZFilter((num_inputs,), clip=5)
running_reward = ZFilter((1, ), demean=False, clip=10)
episode_lengths = []
plot_rew = []
for i_episode in range(number_of_batches):
memory = Memory()
num_steps = 0
reward_batch = 0
num_episodes = 0
while num_steps < batch_size:
state = env.reset()
state = running_state(state)
reward_sum = 0
for t in range(maximum_steps):
action = select_action(state, actor_net)
action = action.data[0].numpy()
next_state, reward, done, _ = env.step(action)
reward_sum += reward
next_state = running_state(next_state)
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state, reward)
if render:
env.render()
if done:
break
state = next_state
num_steps += (t-1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
plot_rew.append(reward_batch)
update_params(batch, actor_net, actor_optimizer, gamma, tau, clip_epsilon)
if i_episode % log_interval == 0:
print('Episode {}\t Last reward: {}\tAverage reward {:.2f}'.format(
i_episode, reward_sum, reward_batch))
plot_epi = []
for i in range (number_of_batches):
plot_epi.append(i)
trace = go.Scatter( x = plot_epi, y = plot_rew)
layout = go.Layout(title='A2C',xaxis=dict(title='Episodes', titlefont=dict(family='Courier New, monospace',size=18,color='#7f7f7f')),
yaxis=dict(title='Average Reward', titlefont=dict(family='Courier New, monospace',size=18,color='#7f7f7f')))
plotly.offline.plot({"data": [trace], "layout": layout},filename='PPO.html',image='jpeg')
return
def train(rank, args, shared_model, opt_ac, can_save, shared_obs_stats):
best_result = -1000
torch.manual_seed(args.seed + rank)
torch.set_default_tensor_type('torch.DoubleTensor')
num_inputs = args.feature
num_actions = 9
last_state = [1] * 48
if args.render and can_save:
env = RunEnv(visualize=True)
else:
env = RunEnv(visualize=False)
#running_state = ZFilter((num_inputs,), clip=5)
#running_reward = ZFilter((1,), demean=False, clip=10)
episode_lengths = []
PATH_TO_MODEL = '../models/' + str(args.bh)
ac_net = ActorCritic(num_inputs, num_actions)
start_time = time.time()
for i_episode in count(1):
memory = Memory()
ac_net.load_state_dict(shared_model.state_dict())
ac_net.zero_grad()
num_steps = 0
reward_batch = 0
num_episodes = 0
#Tot_loss = 0
#Tot_num =
while num_steps < args.batch_size:
#state = env.reset()
#print(num_steps)
state = env.reset(difficulty=0)
last_state = process_observation(state)
state = process_observation(state)
last_state, state = transform_observation(last_state, state)
state = numpy.array(state)
#global last_state
#last_state,_ = update_observation(last_state,state)
#last_state,state = update_observation(last_state,state)
#print(state.shape[0])
#print(state[41])
state = Variable(torch.Tensor(state).unsqueeze(0))
shared_obs_stats.observes(state)
state = shared_obs_stats.normalize(state)
state = state.data[0].numpy()
#state = running_state(state)
reward_sum = 0
#timer = time.time()
for t in range(10000): # Don't infinite loop while learning
#print(t)
if args.use_sep_pol_val:
action = select_action(state)
else:
action = select_action_actor_critic(state, ac_net)
#print(action)
action = action.data[0].numpy()
if numpy.any(numpy.isnan(action)):
print(state)
print(action)
print('ERROR')
raise RuntimeError('action NaN problem')
#print(action)
#print("------------------------")
#timer = time.time()
BB = numpy.append(action, action)
#print(BB)
reward = 0
if args.skip:
#env.step(action)
_, A, _, _ = env.step(BB)
reward += A
_, A, _, _ = env.step(BB)
reward += A
next_state, A, done, _ = env.step(BB)
reward += A
next_state = process_observation(next_state)
last_state, next_state = transform_observation(
last_state, next_state)
next_state = numpy.array(next_state)
reward_sum += reward
#print('env:')
#print(time.time()-timer)
#last_state ,next_state = update_observation(last_state,next_state)
#next_state = running_state(next_state)
next_state = Variable(torch.Tensor(next_state).unsqueeze(0))
shared_obs_stats.observes(next_state)
next_state = shared_obs_stats.normalize(next_state)
next_state = next_state.data[0].numpy()
#print(next_state[41:82])
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state,
reward)
#if args.render:
# env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
#print('env:')
#print(time.time()-timer)
#timer = time.time()
update_params_actor_critic(batch, args, shared_model, ac_net, opt_ac)
#print('backpropagate:')
#print(time.time()-timer)
epoch = i_episode
if (i_episode % args.log_interval == 0) and (rank == 0):
print('TrainEpisode {}\tLast reward: {}\tAverage reward {:.2f}'.
format(i_episode, reward_sum, reward_batch))
if reward_batch > best_result:
best_result = reward_batch
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': ac_net.state_dict(),
'optimizer': opt_ac,
'obs': shared_obs_stats,
}, PATH_TO_MODEL, 'best')
if epoch % 30 == 1:
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': ac_net.state_dict(),
'optimizer': opt_ac,
'obs': shared_obs_stats,
}, PATH_TO_MODEL, epoch)
def main(gamma=0.995, env_name='Walker2d-v2', tau=0.97, seed=543, number_of_batches=500,\
batch_size=5000, maximum_steps=10000, render=False, log_interval=1, entropy_coeff=0.0,\
clip_epsilon=0.2, use_joint_pol_val=False):
torch.set_default_tensor_type('torch.DoubleTensor')
PI = torch.DoubleTensor([3.1415926])
env = gym.make(env_name)
num_inputs = env.observation_space.shape[0]
num_actions = env.action_space.shape[0]
env.seed(seed)
torch.manual_seed(seed)
policy_net = Policy(num_inputs, num_actions)
value_net = Value(num_inputs)
opt_policy = optim.Adam(policy_net.parameters(), lr=0.001)
opt_value = optim.Adam(value_net.parameters(), lr=0.001)
running_state = ZFilter((num_inputs,), clip=5)
running_reward = ZFilter((1,), demean=False, clip=10)
episode_lengths = []
plot_rew = []
for i_episode in range(number_of_batches):
memory = Memory()
num_steps = 0
reward_batch = 0
num_episodes = 0
while num_steps < batch_size:
state = env.reset()
state = running_state(state)
reward_sum = 0
for t in range(maximum_steps): # Don't infinite loop while learning
action = select_action(state, policy_net)
action = action.data[0].numpy()
next_state, reward, done, _ = env.step(action)
reward_sum += reward
next_state = running_state(next_state)
mask = 1
if done:
mask = 0
memory.push(state, np.array([action]), mask, next_state, reward)
if render:
env.render()
if done:
break
state = next_state
num_steps += (t-1)
num_episodes += 1
reward_batch += reward_sum
reward_batch /= num_episodes
batch = memory.sample()
plot_rew.append(reward_batch)
update_params(batch, policy_net, value_net, gamma, opt_policy, opt_value)
if i_episode % args.log_interval == 0:
print('Episode {}\tLast reward: {}\tAverage reward {:.2f}'.format(
i_episode, reward_sum, reward_batch))
plot_epi = []
for i in range (number_of_batches):
plot_epi.append(i)
trace = go.Scatter( x = plot_epi, y = plot_rew)
layout = go.Layout(title='PPO',xaxis=dict(title='Episodes', titlefont=dict(family='Courier New, monospace',size=18,color='#7f7f7f')),
yaxis=dict(title='Average Reward', titlefont=dict(family='Courier New, monospace',size=18,color='#7f7f7f')))
plotly.offline.plot({"data": [trace], "layout": layout},filename='PPO.html',image='jpeg')
def test(rank, args, shared_model, opt_ac):
best_result = -1000
torch.manual_seed(args.seed + rank)
torch.set_default_tensor_type('torch.DoubleTensor')
num_inputs = args.feature
num_actions = 9
last_state = numpy.zeros(41)
if args.render:
env = RunEnv(visualize=True)
else:
env = RunEnv(visualize=False)
running_state = ZFilter((num_inputs, ), clip=5)
running_reward = ZFilter((1, ), demean=False, clip=10)
episode_lengths = []
PATH_TO_MODEL = '../models/' + str(args.bh)
ac_net = ActorCritic(num_inputs, num_actions)
start_time = time.time()
for i_episode in count(1):
memory = Memory()
ac_net.load_state_dict(shared_model.state_dict())
num_steps = 0
reward_batch = 0
num_episodes = 0
while num_steps < args.batch_size:
#state = env.reset()
#print(num_steps)
state = env.reset(difficulty=0)
state = numpy.array(state)
#global last_state
#last_state = state
#last_state,_ = update_observation(last_state,state)
#last_state,state = update_observation(last_state,state)
#print(state.shape[0])
#print(state[41])
state = running_state(state)
reward_sum = 0
for t in range(10000): # Don't infinite loop while learning
#print(t)
#timer = time.time()
if args.use_sep_pol_val:
action = select_action(state)
else:
action = select_action_actor_critic(state, ac_net)
#print(action)
action = action.data[0].numpy()
if numpy.any(numpy.isnan(action)):
print(action)
puts('ERROR')
return
#print('NN take:')
#print(time.time()-timer)
#print(action)
#print("------------------------")
#timer = time.time()
if args.skip:
#env.step(action)
_, reward, _, _ = env.step(action)
reward_sum += reward
next_state, reward, done, _ = env.step(action)
next_state = numpy.array(next_state)
reward_sum += reward
#print('env take:')
#print(time.time()-timer)
#timer = time.time()
#last_state ,next_state = update_observation(last_state,next_state)
next_state = running_state(next_state)
#print(next_state[41:82])
mask = 1
if done:
mask = 0
#print('update take:')
#print(time.time()-timer)
#timer = time.time()
memory.push(state, np.array([action]), mask, next_state,
reward)
#print('memory take:')
#print(time.time()-timer)
#if args.render:
# env.render()
if done:
break
state = next_state
num_steps += (t - 1)
num_episodes += 1
#print(num_episodes)
reward_batch += reward_sum
#print(num_episodes)
reward_batch /= num_episodes
batch = memory.sample()
#update_params_actor_critic(batch,args,shared_model,ac_net,opt_ac)
time.sleep(60)
if i_episode % args.log_interval == 0:
File = open(PATH_TO_MODEL + '/record.txt', 'a+')
File.write("Time {}, episode reward {}, Average reward {}".format(
time.strftime("%Hh %Mm %Ss",
time.gmtime(time.time() - start_time)),
reward_sum, reward_batch))
File.close()
#print('TestEpisode {}\tLast reward: {}\tAverage reward {:.2f}'.format(
# i_episode, reward_sum, reward_batch))
print("Time {}, episode reward {}, Average reward {}".format(
time.strftime("%Hh %Mm %Ss",
time.gmtime(time.time() - start_time)),
reward_sum, reward_batch))
#print('!!!!')
epoch = i_episode
if reward_batch > best_result:
best_result = reward_batch
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': shared_model.state_dict(),
'optimizer': opt_ac.state_dict(),
}, PATH_TO_MODEL, 'best')
if epoch % 30 == 1:
save_model(
{
'epoch': epoch,
'bh': args.bh,
'state_dict': shared_model.state_dict(),
'optimizer': opt_ac.state_dict(),
}, PATH_TO_MODEL, epoch)
class Agent:
def __init__(self, n_bits, lr, memory_size, batch_size, gamma):
self.n_bits = n_bits
self.lr = lr
self.gamma = gamma
self.batch_size = batch_size
self.memory_size = memory_size
self.memory = Memory(self.memory_size)
self.device = "cpu"
self.model = DQN(n_inputs=2 * self.n_bits, n_outputs=n_bits).to(self.device)
self.target_model = DQN(n_inputs=2 * self.n_bits, n_outputs=n_bits).to(self.device)
self.target_model.load_state_dict(self.model.state_dict())
self.target_model.eval()
self.opt = Adam(self.model.parameters(), lr=self.lr)
self.loss_fn = MSELoss()
self.epsilon = 1.0
self.epsilon_decay = 0.001
def choose_action(self, states, goals):
if np.random.random() < self.epsilon:
action = np.random.randint(low=0, high=self.n_bits)
else:
states = torch.Tensor(states, device=self.device)
goals = torch.Tensor(goals, device=self.device)
action = self.model(states, goals).max(dim=-1)[1].item()
return action
def update_epsilon(self):
self.epsilon = max(self.epsilon - self.epsilon_decay, 0)
def store(self, state, action, reward, done, next_state, goal):
state = torch.Tensor(state, device=self.device)
reward = torch.Tensor([reward], device=self.device)
action = torch.Tensor([action], device=self.device)
next_state = torch.Tensor(next_state, device=self.device)
done = torch.Tensor([done], device=self.device)
goal = torch.Tensor(goal, device=self.device)
self.memory.push(state, action, reward, done, next_state, goal)
def unpack_batch(self, batch):
batch = Transition(*zip(*batch))
states = torch.cat(batch.state).to(self.device).view(self.batch_size, self.n_bits)
actions = torch.cat(batch.action).to(self.device).view((-1, 1))
rewards = torch.cat(batch.reward).to(self.device)
next_states = torch.cat(batch.next_state).to(self.device).view(self.batch_size, self.n_bits)
dones = torch.cat(batch.done).to(self.device)
goals = torch.cat(batch.goal).to(self.device).view(self.batch_size, self.n_bits)
return states, actions, rewards, dones, next_states, goals
def learn(self):
if len(self.memory) < self.batch_size:
return 0
batch = self.memory.sample(self.batch_size)
states, actions, rewards, dones, next_states, goals = self.unpack_batch(batch)
with torch.no_grad():
target_q = rewards + self.gamma * self.target_model(next_states, goals).max(-1)[0] * (1 - dones)
q = self.model(states, goals).gather(1, actions.long())
loss = self.loss_fn(q, target_q.view(self.batch_size, 1))
self.opt.zero_grad()
loss.backward()
self.opt.step()
self.soft_update_of_target_network(self.model, self.target_model)
return loss.item()
@staticmethod
def soft_update_of_target_network(local_model, target_model, tau=0.05):
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(tau * local_param.data + (1.0 - tau) * target_param.data)