def main():
args = get_config()
# seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
# path
model_dir = Path('./results') / args.env_name / args.algorithm_name
if not model_dir.exists():
curr_run = 'run1'
else:
exst_run_nums = [
int(str(folder.name).split('run')[1])
for folder in model_dir.iterdir()
if str(folder.name).startswith('run')
]
if len(exst_run_nums) == 0:
curr_run = 'run1'
else:
curr_run = 'run%i' % (max(exst_run_nums) + 1)
run_dir = model_dir / curr_run
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
#Policy network
actor_critic = []
if args.share_policy:
ac = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=args.num_agents,
base_kwargs={
'lstm': args.lstm,
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size
})
ac.to(device)
for agent_id in range(args.num_agents):
actor_critic.append(ac)
else:
for agent_id in range(args.num_agents):
ac = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=args.num_agents,
base_kwargs={
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size
})
ac.to(device)
actor_critic.append(ac)
agents = []
rollouts = []
for agent_id in range(args.num_agents):
# algorithm
agent = PPO(actor_critic[agent_id],
agent_id,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss)
#replay buffer
ro = RolloutStorage(args.num_agents, agent_id, args.episode_length,
args.n_rollout_threads,
envs.observation_space[agent_id],
envs.action_space[agent_id],
actor_critic[agent_id].recurrent_hidden_state_size)
agents.append(agent)
rollouts.append(ro)
# reset env
obs = envs.reset()
# rollout
for i in range(args.num_agents):
rollouts[i].share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollouts[i].obs[0].copy_(torch.tensor(obs[:, i, :]))
rollouts[i].recurrent_hidden_states.zero_()
rollouts[i].recurrent_hidden_states_critic.zero_()
rollouts[i].recurrent_c_states.zero_()
rollouts[i].recurrent_c_states_critic.zero_()
rollouts[i].to(device)
# run
coop_num = []
defect_num = []
coopdefect_num = []
defectcoop_num = []
gore1_num = []
gore2_num = []
gore3_num = []
hare1_num = []
hare2_num = []
hare3_num = []
collective_return = []
apple_consumption = []
waste_cleared = []
sustainability = []
fire = []
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
all_episode = 0
for episode in range(episodes):
if args.use_linear_lr_decay:
# decrease learning rate linearly
for i in range(args.num_agents):
update_linear_schedule(agents[i].optimizer, episode, episodes,
args.lr)
for step in range(args.episode_length):
# Sample actions
values = []
actions = []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
recurrent_c_statess = []
recurrent_c_statess_critic = []
with torch.no_grad():
for i in range(args.num_agents):
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic,\
recurrent_c_states, recurrent_c_states_critic =\
actor_critic[i].act(rollouts[i].share_obs[step],
rollouts[i].obs[step],
rollouts[i].recurrent_hidden_states[step],
rollouts[i].recurrent_hidden_states_critic[step],
rollouts[i].recurrent_c_states[step],
rollouts[i].recurrent_c_states_critic[step],
rollouts[i].masks[step])
values.append(value)
actions.append(action)
action_log_probs.append(action_log_prob)
recurrent_hidden_statess.append(recurrent_hidden_states)
recurrent_hidden_statess_critic.append(
recurrent_hidden_states_critic)
recurrent_c_statess.append(recurrent_c_states)
recurrent_c_statess_critic.append(
recurrent_c_states_critic)
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action_env = []
for k in range(args.num_agents):
one_hot_action = np.zeros(envs.action_space[0].n)
one_hot_action[actions[k][i]] = 1
one_hot_action_env.append(one_hot_action)
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, reward, done, infos = envs.step(actions_env)
# If done then clean the history of observations.
# insert data in buffer
masks = []
bad_masks = []
for i in range(args.num_agents):
mask = []
bad_mask = []
for done_ in done:
if done_[i]:
mask.append([0.0])
bad_mask.append([1.0])
else:
mask.append([1.0])
bad_mask.append([1.0])
masks.append(torch.FloatTensor(mask))
bad_masks.append(torch.FloatTensor(bad_mask))
for i in range(args.num_agents):
rollouts[i].insert(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)),
torch.tensor(obs[:, i, :]), recurrent_hidden_statess[i],
recurrent_hidden_statess_critic[i], recurrent_c_statess[i],
recurrent_c_statess_critic[i], actions[i],
action_log_probs[i], values[i],
torch.tensor(reward[:,
i].reshape(-1,
1)), masks[i], bad_masks[i])
with torch.no_grad():
next_values = []
for i in range(args.num_agents):
next_value = actor_critic[i].get_value(
rollouts[i].share_obs[-1], rollouts[i].obs[-1],
rollouts[i].recurrent_hidden_states[-1],
rollouts[i].recurrent_hidden_states_critic[-1],
rollouts[i].recurrent_c_states[-1],
rollouts[i].recurrent_c_states_critic[-1],
rollouts[i].masks[-1]).detach()
next_values.append(next_value)
for i in range(args.num_agents):
rollouts[i].compute_returns(next_values[i], args.use_gae,
args.gamma, args.gae_lambda,
args.use_proper_time_limits)
# update the network
value_losses = []
action_losses = []
dist_entropies = []
for i in range(args.num_agents):
value_loss, action_loss, dist_entropy = agents[i].update(
rollouts[i])
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
if args.env_name == "StagHunt":
for info in infos:
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'defect&defect_num' in info.keys():
defect_num.append(info['defect&defect_num'])
if 'coop&defect_num' in info.keys():
coopdefect_num.append(info['coop&defect_num'])
if 'defect&coop_num' in info.keys():
defectcoop_num.append(info['defect&coop_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
logger.add_scalars(
'defect&defect_num_per_episode',
{'defect&defect_num_per_episode': defect_num[all_episode]},
all_episode)
logger.add_scalars('coop&defect_num_per_episode', {
'coop&defect_num_per_episode':
coopdefect_num[all_episode]
}, all_episode)
logger.add_scalars('defect&coop_num_per_episode', {
'defect&coop_num_per_episode':
defectcoop_num[all_episode]
}, all_episode)
all_episode += 1
elif args.env_name == "StagHuntGW":
for info in infos:
if 'collective_return' in info.keys():
collective_return.append(info['collective_return'])
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'gore1_num' in info.keys():
gore1_num.append(info['gore1_num'])
if 'gore2_num' in info.keys():
gore2_num.append(info['gore2_num'])
if 'hare1_num' in info.keys():
hare1_num.append(info['hare1_num'])
if 'hare2_num' in info.keys():
hare2_num.append(info['hare2_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[all_episode]},
all_episode)
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
logger.add_scalars(
'gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[all_episode]},
all_episode)
logger.add_scalars(
'gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[all_episode]},
all_episode)
logger.add_scalars(
'hare1_num_per_episode',
{'hare1_num_per_episode': hare1_num[all_episode]},
all_episode)
logger.add_scalars(
'hare2_num_per_episode',
{'hare2_num_per_episode': hare2_num[all_episode]},
all_episode)
all_episode += 1
elif args.env_name == "EscalationGW":
for info in infos:
if 'collective_return' in info.keys():
collective_return.append(info['collective_return'])
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[all_episode]},
all_episode)
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
all_episode += 1
elif args.env_name == "multi_StagHuntGW":
for info in infos:
if 'collective_return' in info.keys():
collective_return.append(info['collective_return'])
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'gore0_num' in info.keys():
gore1_num.append(info['gore0_num'])
if 'gore1_num' in info.keys():
gore2_num.append(info['gore1_num'])
if 'gore2_num' in info.keys():
gore3_num.append(info['gore2_num'])
if 'hare0_num' in info.keys():
hare1_num.append(info['hare0_num'])
if 'hare1_num' in info.keys():
hare2_num.append(info['hare1_num'])
if 'hare2_num' in info.keys():
hare3_num.append(info['hare2_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[all_episode]},
all_episode)
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
logger.add_scalars(
'gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[all_episode]},
all_episode)
logger.add_scalars(
'gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[all_episode]},
all_episode)
logger.add_scalars(
'gore3_num_per_episode',
{'gore3_num_per_episode': gore3_num[all_episode]},
all_episode)
logger.add_scalars(
'hare1_num_per_episode',
{'hare1_num_per_episode': hare1_num[all_episode]},
all_episode)
logger.add_scalars(
'hare2_num_per_episode',
{'hare2_num_per_episode': hare2_num[all_episode]},
all_episode)
logger.add_scalars(
'hare3_num_per_episode',
{'hare3_num_per_episode': hare3_num[all_episode]},
all_episode)
all_episode += 1
# clean the buffer and reset
obs = envs.reset()
for i in range(args.num_agents):
rollouts[i].share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollouts[i].obs[0].copy_(torch.tensor(obs[:, i, :]))
rollouts[i].recurrent_hidden_states.zero_()
rollouts[i].recurrent_hidden_states_critic.zero_()
rollouts[i].recurrent_c_states.zero_()
rollouts[i].recurrent_c_states_critic.zero_()
rollouts[i].masks[0].copy_(torch.ones(args.n_rollout_threads, 1))
rollouts[i].bad_masks[0].copy_(
torch.ones(args.n_rollout_threads, 1))
rollouts[i].to(device)
for i in range(args.num_agents):
# save for every interval-th episode or for the last epoch
if (episode % args.save_interval == 0 or episode == episodes - 1):
torch.save({'model': actor_critic[i]},
str(save_dir) + "/agent%i_model" % i + ".pt")
# log information
if episode % args.log_interval == 0:
total_num_steps = (
episode + 1) * args.episode_length * args.n_rollout_threads
end = time.time()
print(
"\n Updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(episode, episodes, total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
for i in range(args.num_agents):
print("value loss of agent%i: " % i + str(value_losses[i]))
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
###----------------------------------------------------------###
###----------------------------------------------------------###
###----------------------------------------------------------###
if args.eval:
eval_dir = run_dir / 'eval'
log_dir = eval_dir / 'logs'
os.makedirs(str(log_dir))
logger = SummaryWriter(str(log_dir))
# eval best policy
eval_rewards = []
# env
if args.env_name == "StagHunt":
assert args.num_agents == 2, (
"only 2 agents is supported, check the config.py.")
env = MGEnv(args)
elif args.env_name == "StagHuntGW" or args.env_name == "EscalationGW":
assert args.num_agents == 2, (
"only 2 agent is supported in single navigation, check the config.py."
)
env = GridWorldEnv(args)
elif args.env_name == "multi_StagHuntGW":
env = multi_GridWorldEnv(args)
else:
print("Can not support the " + args.env_name + "environment.")
raise NotImplementedError
#Policy network
coop_num = []
defect_num = []
coopdefect_num = []
defectcoop_num = []
gore1_num = []
gore2_num = []
gore3_num = []
hare1_num = []
hare2_num = []
hare3_num = []
collective_return = []
apple_consumption = []
waste_cleared = []
sustainability = []
fire = []
for episode in range(args.eval_episodes):
print("Episode %i of %i" % (episode, args.eval_episodes))
state = env.reset()
state = np.array([state])
share_obs = []
obs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
recurrent_c_statess = []
recurrent_c_statess_critic = []
masks = []
policy_reward = 0
# rollout
for i in range(args.num_agents):
share_obs.append(
(torch.tensor(state.reshape(1, -1),
dtype=torch.float32)).to(device))
obs.append((torch.tensor(state[:, i, :],
dtype=torch.float32)).to(device))
recurrent_hidden_statess.append(
torch.zeros(
1, actor_critic[i].recurrent_hidden_state_size).to(
device))
recurrent_hidden_statess_critic.append(
torch.zeros(
1, actor_critic[i].recurrent_hidden_state_size).to(
device))
recurrent_c_statess.append(
torch.zeros(
1, actor_critic[i].recurrent_hidden_state_size).to(
device))
recurrent_c_statess_critic.append(
torch.zeros(
1, actor_critic[i].recurrent_hidden_state_size).to(
device))
masks.append(torch.ones(1, 1).to(device))
for step in range(args.episode_length):
print("step %i of %i" % (step, args.episode_length))
# Sample actions
one_hot_actions = []
for i in range(args.num_agents):
one_hot_action = np.zeros(env.action_space[0].n)
with torch.no_grad():
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic, recurrent_c_states, recurrent_c_states_critic = actor_critic[
i].act(share_obs[i], obs[i],
recurrent_hidden_statess[i],
recurrent_hidden_statess_critic[i],
recurrent_c_statess[i],
recurrent_c_statess_critic[i], masks[i])
recurrent_hidden_statess[i].copy_(recurrent_hidden_states)
recurrent_hidden_statess_critic[i].copy_(
recurrent_hidden_states_critic)
recurrent_c_statess[i].copy_(recurrent_c_states)
recurrent_c_statess_critic[i].copy_(
recurrent_c_states_critic)
one_hot_action[action] = 1
one_hot_actions.append(one_hot_action)
# Obser reward and next obs
state, reward, done, infos = env.step(one_hot_actions)
for i in range(args.num_agents):
print("Reward of agent%i: " % i + str(reward[i]))
policy_reward += reward[i]
if all(done):
break
state = np.array([state])
for i in range(args.num_agents):
if len(env.observation_space[0]) == 1:
share_obs[i].copy_(
torch.tensor(state.reshape(1, -1),
dtype=torch.float32))
obs[i].copy_(
torch.tensor(state[:, i, :], dtype=torch.float32))
elif len(env.observation_space[0]) == 3:
share_obs[i].copy_(
torch.tensor(state.reshape(
1, -1, env.observation_space[0][1],
env.observation_space[0][2]),
dtype=torch.float32))
obs[i].copy_(
torch.tensor(state[:, i, :, :, :],
dtype=torch.float32))
eval_rewards.append(policy_reward)
if args.env_name == "StagHunt":
if 'coop&coop_num' in infos.keys():
coop_num.append(infos['coop&coop_num'])
if 'defect&defect_num' in infos.keys():
defect_num.append(infos['defect&defect_num'])
if 'coop&defect_num' in infos.keys():
coopdefect_num.append(infos['coop&defect_num'])
if 'defect&coop_num' in infos.keys():
defectcoop_num.append(infos['defect&coop_num'])
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[episode]}, episode)
logger.add_scalars(
'defect&defect_num_per_episode',
{'defect&defect_num_per_episode': defect_num[episode]},
episode)
logger.add_scalars(
'coop&defect_num_per_episode',
{'coop&defect_num_per_episode': coopdefect_num[episode]},
episode)
logger.add_scalars(
'defect&coop_num_per_episode',
{'defect&coop_num_per_episode': defectcoop_num[episode]},
episode)
elif args.env_name == "StagHuntGW":
if 'collective_return' in infos.keys():
collective_return.append(infos['collective_return'])
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[episode]},
episode)
if 'coop&coop_num' in infos.keys():
coop_num.append(infos['coop&coop_num'])
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[episode]},
episode)
if 'gore1_num' in infos.keys():
gore1_num.append(infos['gore1_num'])
logger.add_scalars(
'gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[episode]}, episode)
if 'gore2_num' in infos.keys():
gore2_num.append(infos['gore2_num'])
logger.add_scalars(
'gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[episode]}, episode)
if 'hare1_num' in infos.keys():
hare1_num.append(infos['hare1_num'])
logger.add_scalars(
'hare1_num_per_episode',
{'hare1_num_per_episode': hare1_num[episode]}, episode)
if 'hare2_num' in infos.keys():
hare2_num.append(infos['hare2_num'])
logger.add_scalars(
'hare2_num_per_episode',
{'hare2_num_per_episode': hare2_num[episode]}, episode)
elif args.env_name == "EscalationGW":
if 'collective_return' in infos.keys():
collective_return.append(infos['collective_return'])
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[episode]},
episode)
if 'coop&coop_num' in infos.keys():
coop_num.append(infos['coop&coop_num'])
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[episode]},
episode)
elif args.env_name == "multi_StagHuntGW":
if 'collective_return' in infos.keys():
collective_return.append(infos['collective_return'])
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[episode]},
episode)
if 'coop&coop_num' in infos.keys():
coop_num.append(infos['coop&coop_num'])
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[episode]},
episode)
if 'gore0_num' in infos.keys():
gore1_num.append(infos['gore0_num'])
logger.add_scalars(
'gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[episode]}, episode)
if 'gore1_num' in infos.keys():
gore2_num.append(infos['gore1_num'])
logger.add_scalars(
'gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[episode]}, episode)
if 'gore2_num' in infos.keys():
gore3_num.append(infos['gore2_num'])
logger.add_scalars(
'gore3_num_per_episode',
{'gore3_num_per_episode': gore3_num[episode]}, episode)
if 'hare0_num' in infos.keys():
hare1_num.append(infos['hare0_num'])
logger.add_scalars(
'hare1_num_per_episode',
{'hare1_num_per_episode': hare1_num[episode]}, episode)
if 'hare1_num' in infos.keys():
hare2_num.append(infos['hare1_num'])
logger.add_scalars(
'hare2_num_per_episode',
{'hare2_num_per_episode': hare2_num[episode]}, episode)
if 'hare2_num' in infos.keys():
hare3_num.append(infos['hare2_num'])
logger.add_scalars(
'hare3_num_per_episode',
{'hare3_num_per_episode': hare3_num[episode]}, episode)
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
def main():
args = get_config()
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# path
model_dir = Path('./results') / args.env_name / args.algorithm_name / (
"run" + str(args.seed))
if args.critic_full_obs:
run_dir = model_dir / 'adaptive'
else:
run_dir = model_dir / 'adaptive_only'
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
print(
"\n Now we have %i fixed policy! Train Single Adaptive Policy... \n" %
args.num_policy_candidates)
args.env_name = args.env_name + "Adaptive"
policy_candidates = []
for i in range(args.num_policy_candidates):
ac = torch.load(
str(model_dir) + ("/models/Policy%i" % (i + 1)) +
"-agent0_model.pt")['model'].cpu()
policy_candidates.append(ac)
# env
envs = make_parallel_env(args, policy_candidates)
#Policy network
# agent 0
actor_critic = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=args.num_agents,
base_kwargs={
'lstm': args.lstm,
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size
})
actor_critic.to(device)
agent0 = PPO(actor_critic,
0,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss)
#replay buffer
rollout = RolloutStorage(args.num_agents, 0, args.episode_length,
args.n_rollout_threads, envs.observation_space[0],
envs.action_space[0],
actor_critic.recurrent_hidden_state_size)
# reset
if args.critic_full_obs:
obs, obs_critic, select_opponent = envs.reset()
else:
obs, select_opponent = envs.reset()
# rollout
if len(envs.observation_space[0]) == 1:
if args.critic_full_obs:
rollout.share_obs[0].copy_(
torch.tensor(obs_critic.reshape(args.n_rollout_threads, -1)))
else:
rollout.share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollout.obs[0].copy_(torch.tensor(obs[:, 0, :]))
rollout.recurrent_hidden_states.zero_()
rollout.recurrent_hidden_states_critic.zero_()
rollout.recurrent_c_states.zero_()
rollout.recurrent_c_states_critic.zero_()
else:
raise NotImplementedError
rollout.to(device)
# run
collective_return = []
apple_consumption = []
waste_cleared = []
sustainability = []
fire = []
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
all_episode = 0
all_episode_adaptive = np.zeros(args.num_policy_candidates)
for episode in range(episodes):
if args.use_linear_lr_decay:
update_linear_schedule(agent0.optimizer, episode, episodes,
args.lr)
for step in range(args.episode_length):
with torch.no_grad():
value, action0, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic, recurrent_c_states, recurrent_c_states_critic = actor_critic.act(
rollout.share_obs[step], rollout.obs[step],
rollout.recurrent_hidden_states[step],
rollout.recurrent_hidden_states_critic[step],
rollout.recurrent_c_states[step],
rollout.recurrent_c_states_critic[step],
rollout.masks[step])
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action = np.zeros((1, envs.action_space[0].n))
one_hot_action[0][action0[i]] = 1
actions_env.append(one_hot_action)
# Obser reward and next obs
if args.critic_full_obs:
obs, obs_critic, select_opponent, reward, done, infos = envs.step(
actions_env)
else:
obs, select_opponent, reward, done, infos = envs.step(
actions_env)
# If done then clean the history of observations.
# insert data in buffer
masks = []
bad_masks = []
for i in range(args.num_agents):
mask = []
bad_mask = []
for done_ in done:
if done_[i]:
mask.append([0.0])
bad_mask.append([1.0])
else:
mask.append([1.0])
bad_mask.append([1.0])
masks.append(torch.FloatTensor(mask))
bad_masks.append(torch.FloatTensor(bad_mask))
if len(envs.observation_space[0]) == 1:
if args.critic_full_obs:
rollout.insert(
torch.tensor(
obs_critic.reshape(args.n_rollout_threads, -1)),
torch.tensor(obs[:, 0, :]), recurrent_hidden_states,
recurrent_hidden_states_critic, recurrent_c_states,
recurrent_c_states_critic, action0,
action_log_prob, value,
torch.tensor(reward[:, 0].reshape(-1, 1)), masks[0],
bad_masks[0])
else:
rollout.insert(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)),
torch.tensor(obs[:, 0, :]), recurrent_hidden_states,
recurrent_hidden_states_critic, recurrent_c_states,
recurrent_c_states_critic, action0,
action_log_prob, value,
torch.tensor(reward[:, 0].reshape(-1, 1)), masks[0],
bad_masks[0])
else:
raise NotImplementedError
with torch.no_grad():
next_value = actor_critic.get_value(
rollout.share_obs[-1], rollout.obs[-1],
rollout.recurrent_hidden_states[-1],
rollout.recurrent_hidden_states_critic[-1],
rollout.recurrent_c_states[-1],
rollout.recurrent_c_states_critic[-1],
rollout.masks[-1]).detach()
rollout.compute_returns(next_value, args.use_gae, args.gamma,
args.gae_lambda, args.use_proper_time_limits)
# update the network
value_loss, action_loss, dist_entropy = agent0.update(rollout)
if args.env_name == "StagHuntAdaptive":
coop_num = []
defect_num = []
coopdefect_num = []
defectcoop_num = []
rew = []
for info in infos:
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'defect&defect_num' in info.keys():
defect_num.append(info['defect&defect_num'])
if 'coop&defect_num' in info.keys():
coopdefect_num.append(info['coop&defect_num'])
if 'defect&coop_num' in info.keys():
defectcoop_num.append(info['defect&coop_num'])
for i in range(args.n_rollout_threads):
rew.append(rollout.rewards[:, i, :].sum().cpu().numpy())
for i in range(args.n_rollout_threads):
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/defect&defect_num_per_episode',
{'defect&defect_num_per_episode': defect_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/coop&defect_num_per_episode',
{'coop&defect_num_per_episode': coopdefect_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/defect&coop_num_per_episode',
{'defect&coop_num_per_episode': defectcoop_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) + '/reward',
{'reward': np.mean(np.array(rew[i]))},
all_episode_adaptive[select_opponent[i]])
all_episode_adaptive[select_opponent[i]] += 1
elif args.env_name == "StagHuntGWAdaptive":
collective_return = []
coop_num = []
gore1_num = []
gore2_num = []
hare1_num = []
hare2_num = []
for info in infos:
if 'collective_return' in info.keys():
collective_return.append(info['collective_return'])
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'gore1_num' in info.keys():
gore1_num.append(info['gore1_num'])
if 'gore2_num' in info.keys():
gore2_num.append(info['gore2_num'])
if 'hare1_num' in info.keys():
hare1_num.append(info['hare1_num'])
if 'hare2_num' in info.keys():
hare2_num.append(info['hare2_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/collective_return',
{'collective_return': collective_return[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/hare1_num_per_episode',
{'hare1_num_per_episode': hare1_num[i]},
all_episode_adaptive[select_opponent[i]])
logger.add_scalars(
'Policy-' + str(select_opponent[i] + 1) +
'/hare2_num_per_episode',
{'hare2_num_per_episode': hare2_num[i]},
all_episode_adaptive[select_opponent[i]])
all_episode_adaptive[select_opponent[i]] += 1
if args.critic_full_obs:
obs, obs_critic, select_opponent = envs.reset()
else:
obs, select_opponent = envs.reset()
if len(envs.observation_space[0]) == 1:
if args.critic_full_obs:
rollout.share_obs[0].copy_(
torch.tensor(obs_critic.reshape(args.n_rollout_threads,
-1)))
else:
rollout.share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollout.obs[0].copy_(torch.tensor(obs[:, 0, :]))
rollout.recurrent_hidden_states.zero_()
rollout.recurrent_hidden_states_critic.zero_()
rollout.recurrent_c_states.zero_()
rollout.recurrent_c_states_critic.zero_()
rollout.masks[0].copy_(torch.ones(args.n_rollout_threads, 1))
rollout.bad_masks[0].copy_(torch.ones(args.n_rollout_threads, 1))
else:
raise NotImplementedError
rollout.to(device)
if (episode % args.save_interval == 0 or episode == episodes - 1):
torch.save({'model': actor_critic},
str(save_dir) + "/agent0_model.pt")
# log information
if episode % args.log_interval == 0:
total_num_steps = (
episode + 1) * args.episode_length * args.n_rollout_threads
end = time.time()
print(
"\n Updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(episode, episodes, total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
print("value loss: agent0--" + str(value_loss))
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
def main():
args = get_config()
# seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
# path
model_dir = Path(
'./results') / args.env_name / args.scenario_name / args.algorithm_name
if not model_dir.exists():
curr_run = 'run1'
else:
exst_run_nums = [
int(str(folder.name).split('run')[1])
for folder in model_dir.iterdir()
if str(folder.name).startswith('run')
]
if len(exst_run_nums) == 0:
curr_run = 'run1'
else:
curr_run = 'run%i' % (max(exst_run_nums) + 1)
run_dir = model_dir / curr_run
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
if args.eval:
eval_env = make_eval_env(args)
num_agents = args.num_agents
all_action_space = []
all_obs_space = []
action_movement_dim = []
if args.env_name == "BlueprintConstruction":
order_obs = [
'agent_qpos_qvel', 'box_obs', 'ramp_obs', 'construction_site_obs',
'observation_self'
]
mask_order_obs = [None, None, None, None, None]
elif args.env_name == "BoxLocking":
order_obs = [
'agent_qpos_qvel', 'box_obs', 'ramp_obs', 'observation_self'
]
mask_order_obs = ['mask_aa_obs', 'mask_ab_obs', 'mask_ar_obs', None]
else:
print("Can not support the " + args.env_name + "environment.")
raise NotImplementedError
for agent_id in range(num_agents):
# deal with dict action space
action_movement = envs.action_space['action_movement'][agent_id].nvec
action_movement_dim.append(len(action_movement))
action_glueall = envs.action_space['action_glueall'][agent_id].n
action_vec = np.append(action_movement, action_glueall)
if 'action_pull' in envs.action_space.spaces.keys():
action_pull = envs.action_space['action_pull'][agent_id].n
action_vec = np.append(action_vec, action_pull)
action_space = MultiDiscrete([[0, vec - 1] for vec in action_vec])
all_action_space.append(action_space)
# deal with dict obs space
obs_space = []
obs_dim = 0
for key in order_obs:
if key in envs.observation_space.spaces.keys():
space = list(envs.observation_space[key].shape)
if len(space) < 2:
space.insert(0, 1)
obs_space.append(space)
obs_dim += reduce(lambda x, y: x * y, space)
obs_space.insert(0, obs_dim)
all_obs_space.append(obs_space)
if args.share_policy:
actor_critic = Policy(all_obs_space[0],
all_action_space[0],
num_agents=num_agents,
gain=args.gain,
base_kwargs={
'naive_recurrent':
args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization':
args.use_feature_normlization,
'use_feature_popart':
args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': True
},
device=device)
actor_critic.to(device)
# algorithm
agents = PPO(actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
#replay buffer
rollouts = RolloutStorage(num_agents,
args.episode_length,
args.n_rollout_threads,
all_obs_space[0],
all_action_space[0],
args.hidden_size,
use_same_dim=True)
else:
actor_critic = []
agents = []
for agent_id in range(num_agents):
ac = Policy(all_obs_space[0],
all_action_space[0],
num_agents=num_agents,
gain=args.gain,
base_kwargs={
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization':
args.use_feature_normlization,
'use_feature_popart': args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': True
},
device=device)
ac.to(device)
# algorithm
agent = PPO(ac,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
actor_critic.append(ac)
agents.append(agent)
#replay buffer
rollouts = RolloutStorage(num_agents,
args.episode_length,
args.n_rollout_threads,
all_obs_space[0],
all_action_space[0],
args.hidden_size,
use_same_dim=True)
# reset env
dict_obs = envs.reset()
obs = []
share_obs = []
for d_o in dict_obs:
for i, key in enumerate(order_obs):
if key in envs.observation_space.spaces.keys():
if mask_order_obs[i] == None:
temp_share_obs = d_o[key].reshape(num_agents, -1).copy()
temp_obs = temp_share_obs.copy()
else:
temp_share_obs = d_o[key].reshape(num_agents, -1).copy()
temp_mask = d_o[mask_order_obs[i]].copy()
temp_obs = d_o[key].copy()
mins_temp_mask = ~temp_mask
temp_obs[mins_temp_mask] = np.zeros(
(mins_temp_mask.sum(), temp_obs.shape[2]))
temp_obs = temp_obs.reshape(num_agents, -1)
if i == 0:
reshape_obs = temp_obs.copy()
reshape_share_obs = temp_share_obs.copy()
else:
reshape_obs = np.concatenate((reshape_obs, temp_obs),
axis=1)
reshape_share_obs = np.concatenate(
(reshape_share_obs, temp_share_obs), axis=1)
obs.append(reshape_obs)
share_obs.append(reshape_share_obs)
obs = np.array(obs)
share_obs = np.array(share_obs)
# replay buffer
rollouts.share_obs[0] = share_obs.copy()
rollouts.obs[0] = obs.copy()
rollouts.recurrent_hidden_states = np.zeros(
rollouts.recurrent_hidden_states.shape).astype(np.float32)
rollouts.recurrent_hidden_states_critic = np.zeros(
rollouts.recurrent_hidden_states_critic.shape).astype(np.float32)
# run
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
timesteps = 0
for episode in range(episodes):
if args.use_linear_lr_decay: # decrease learning rate linearly
if args.share_policy:
update_linear_schedule(agents.optimizer, episode, episodes,
args.lr)
else:
for agent_id in range(num_agents):
update_linear_schedule(agents[agent_id].optimizer, episode,
episodes, args.lr)
# info list
discard_episode = 0
success = 0
trials = 0
for step in range(args.episode_length):
# Sample actions
values = []
actions = []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(
agent_id,
torch.tensor(rollouts.share_obs[step, :,
agent_id]),
torch.tensor(rollouts.obs[step, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[step, :,
agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[
step, :, agent_id]),
torch.tensor(rollouts.masks[step, :, agent_id]))
else:
actor_critic[agent_id].eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[
agent_id].act(
agent_id,
torch.tensor(rollouts.share_obs[step, :,
agent_id]),
torch.tensor(rollouts.obs[step, :, agent_id]),
torch.tensor(rollouts.recurrent_hidden_states[
step, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[
step, :, agent_id]),
torch.tensor(rollouts.masks[step, :,
agent_id]))
values.append(value.detach().cpu().numpy())
actions.append(action.detach().cpu().numpy())
action_log_probs.append(
action_log_prob.detach().cpu().numpy())
recurrent_hidden_statess.append(
recurrent_hidden_states.detach().cpu().numpy())
recurrent_hidden_statess_critic.append(
recurrent_hidden_states_critic.detach().cpu().numpy())
# rearrange action
actions_env = []
for n_rollout_thread in range(args.n_rollout_threads):
action_movement = []
action_pull = []
action_glueall = []
for agent_id in range(num_agents):
action_movement.append(actions[agent_id][n_rollout_thread]
[:action_movement_dim[agent_id]])
action_glueall.append(
int(actions[agent_id][n_rollout_thread][
action_movement_dim[agent_id]]))
if 'action_pull' in envs.action_space.spaces.keys():
action_pull.append(
int(actions[agent_id][n_rollout_thread][-1]))
action_movement = np.stack(action_movement, axis=0)
action_glueall = np.stack(action_glueall, axis=0)
if 'action_pull' in envs.action_space.spaces.keys():
action_pull = np.stack(action_pull, axis=0)
one_env_action = {
'action_movement': action_movement,
'action_pull': action_pull,
'action_glueall': action_glueall
}
actions_env.append(one_env_action)
# Obser reward and next obs
dict_obs, rewards, dones, infos = envs.step(actions_env)
if len(rewards.shape) < 3:
rewards = rewards[:, :, np.newaxis]
# If done then clean the history of observations.
# insert data in buffer
masks = []
for i, done in enumerate(dones):
if done:
if "discard_episode" in infos[i].keys():
if infos[i]['discard_episode']:
discard_episode += 1
else:
trials += 1
else:
trials += 1
if "success" in infos[i].keys():
if infos[i]['success']:
success += 1
mask = []
for agent_id in range(num_agents):
if done:
recurrent_hidden_statess[agent_id][i] = np.zeros(
args.hidden_size).astype(np.float32)
recurrent_hidden_statess_critic[agent_id][
i] = np.zeros(args.hidden_size).astype(np.float32)
mask.append([0.0])
else:
mask.append([1.0])
masks.append(mask)
obs = []
share_obs = []
for d_o in dict_obs:
for i, key in enumerate(order_obs):
if key in envs.observation_space.spaces.keys():
if mask_order_obs[i] == None:
temp_share_obs = d_o[key].reshape(num_agents,
-1).copy()
temp_obs = temp_share_obs.copy()
else:
temp_share_obs = d_o[key].reshape(num_agents,
-1).copy()
temp_mask = d_o[mask_order_obs[i]].copy()
temp_obs = d_o[key].copy()
mins_temp_mask = ~temp_mask
temp_obs[mins_temp_mask] = np.zeros(
(mins_temp_mask.sum(), temp_obs.shape[2]))
temp_obs = temp_obs.reshape(num_agents, -1)
if i == 0:
reshape_obs = temp_obs.copy()
reshape_share_obs = temp_share_obs.copy()
else:
reshape_obs = np.concatenate(
(reshape_obs, temp_obs), axis=1)
reshape_share_obs = np.concatenate(
(reshape_share_obs, temp_share_obs), axis=1)
obs.append(reshape_obs)
share_obs.append(reshape_share_obs)
obs = np.array(obs)
share_obs = np.array(share_obs)
rollouts.insert(
share_obs, obs,
np.array(recurrent_hidden_statess).transpose(1, 0, 2),
np.array(recurrent_hidden_statess_critic).transpose(1, 0, 2),
np.array(actions).transpose(1, 0, 2),
np.array(action_log_probs).transpose(1, 0, 2),
np.array(values).transpose(1, 0, 2), rewards, masks)
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
next_value, _, _ = actor_critic.get_value(
agent_id,
torch.tensor(rollouts.share_obs[-1, :, agent_id]),
torch.tensor(rollouts.obs[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[-1, :,
agent_id]),
torch.tensor(rollouts.masks[-1, :, agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id, next_value,
args.use_gae, args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents.value_normalizer)
else:
actor_critic[agent_id].eval()
next_value, _, _ = actor_critic[agent_id].get_value(
agent_id,
torch.tensor(rollouts.share_obs[-1, :, agent_id]),
torch.tensor(rollouts.obs[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[-1, :,
agent_id]),
torch.tensor(rollouts.masks[-1, :, agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id, next_value,
args.use_gae, args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents[agent_id].value_normalizer)
# update the network
if args.share_policy:
actor_critic.train()
value_loss, action_loss, dist_entropy = agents.update_share(
num_agents, rollouts)
logger.add_scalars('reward', {'reward': np.mean(rollouts.rewards)},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
else:
value_losses = []
action_losses = []
dist_entropies = []
for agent_id in range(num_agents):
actor_critic[agent_id].train()
value_loss, action_loss, dist_entropy = agents[
agent_id].update(agent_id, rollouts)
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
logger.add_scalars(
'agent%i/reward' % agent_id,
{'reward': np.mean(rollouts.rewards[:, :, agent_id])},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
# clean the buffer and reset
rollouts.after_update()
total_num_steps = (episode +
1) * args.episode_length * args.n_rollout_threads
if (episode % args.save_interval == 0 or episode == episodes -
1): # save for every interval-th episode or for the last epoch
if args.share_policy:
torch.save({'model': actor_critic},
str(save_dir) + "/agent_model.pt")
else:
for agent_id in range(num_agents):
torch.save({'model': actor_critic[agent_id]},
str(save_dir) + "/agent%i_model" % agent_id +
".pt")
# log information
if episode % args.log_interval == 0:
end = time.time()
print(
"\n Scenario {} Algo {} updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(args.scenario_name, args.algorithm_name, episode,
episodes, total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
if args.share_policy:
print("value loss of agent: " + str(value_loss))
else:
for agent_id in range(num_agents):
print("value loss of agent%i: " % agent_id +
str(value_losses[agent_id]))
logger.add_scalars('discard_episode',
{'discard_episode': discard_episode},
total_num_steps)
if trials > 0:
logger.add_scalars('success_rate',
{'success_rate': success / trials},
total_num_steps)
else:
logger.add_scalars('success_rate', {'success_rate': 0.0},
total_num_steps)
# eval
if episode % args.eval_interval == 0 and args.eval:
eval_episode = 0
eval_success = 0
eval_dict_obs = eval_env.reset()
eval_obs = []
eval_share_obs = []
for eval_d_o in eval_dict_obs:
for i, key in enumerate(order_obs):
if key in eval_env.observation_space.spaces.keys():
if mask_order_obs[i] == None:
temp_share_obs = eval_d_o[key].reshape(
num_agents, -1).copy()
temp_obs = temp_share_obs.copy()
else:
temp_share_obs = eval_d_o[key].reshape(
num_agents, -1).copy()
temp_mask = eval_d_o[mask_order_obs[i]].copy()
temp_obs = eval_d_o[key].copy()
mins_temp_mask = ~temp_mask
temp_obs[mins_temp_mask] = np.zeros(
(mins_temp_mask.sum(), temp_obs.shape[2]))
temp_obs = temp_obs.reshape(num_agents, -1)
if i == 0:
reshape_obs = temp_obs.copy()
reshape_share_obs = temp_share_obs.copy()
else:
reshape_obs = np.concatenate(
(reshape_obs, temp_obs), axis=1)
reshape_share_obs = np.concatenate(
(reshape_share_obs, temp_share_obs), axis=1)
eval_obs.append(reshape_obs)
eval_share_obs.append(reshape_share_obs)
eval_obs = np.array(eval_obs)
eval_share_obs = np.array(eval_share_obs)
eval_recurrent_hidden_states = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_recurrent_hidden_states_critic = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_masks = np.ones((1, num_agents, 1)).astype(np.float32)
while True:
eval_actions = []
actor_critic.eval()
for agent_id in range(num_agents):
_, action, _, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(
agent_id,
torch.FloatTensor(eval_share_obs[:, agent_id]),
torch.FloatTensor(eval_obs[:, agent_id]),
torch.FloatTensor(
eval_recurrent_hidden_states[:, agent_id]),
torch.FloatTensor(
eval_recurrent_hidden_states_critic[:, agent_id]),
torch.FloatTensor(eval_masks[:, agent_id]),
None,
deterministic=True)
eval_actions.append(action.detach().cpu().numpy())
eval_recurrent_hidden_states[:,
agent_id] = recurrent_hidden_states.detach(
).cpu().numpy()
eval_recurrent_hidden_states_critic[:,
agent_id] = recurrent_hidden_states_critic.detach(
).cpu().numpy()
# rearrange action
eval_actions_env = []
for n_rollout_thread in range(1):
action_movement = []
action_pull = []
action_glueall = []
for agent_id in range(num_agents):
action_movement.append(
eval_actions[agent_id][n_rollout_thread]
[:action_movement_dim[agent_id]])
action_glueall.append(
int(eval_actions[agent_id][n_rollout_thread][
action_movement_dim[agent_id]]))
if 'action_pull' in envs.action_space.spaces.keys():
action_pull.append(
int(eval_actions[agent_id][n_rollout_thread]
[-1]))
action_movement = np.stack(action_movement, axis=0)
action_glueall = np.stack(action_glueall, axis=0)
if 'action_pull' in envs.action_space.spaces.keys():
action_pull = np.stack(action_pull, axis=0)
one_env_action = {
'action_movement': action_movement,
'action_pull': action_pull,
'action_glueall': action_glueall
}
eval_actions_env.append(one_env_action)
# Obser reward and next obs
eval_dict_obs, eval_rewards, eval_dones, eval_infos = eval_env.step(
eval_actions_env)
eval_obs = []
eval_share_obs = []
for eval_d_o in eval_dict_obs:
for i, key in enumerate(order_obs):
if key in eval_env.observation_space.spaces.keys():
if mask_order_obs[i] == None:
temp_share_obs = eval_d_o[key].reshape(
num_agents, -1).copy()
temp_obs = temp_share_obs.copy()
else:
temp_share_obs = eval_d_o[key].reshape(
num_agents, -1).copy()
temp_mask = eval_d_o[mask_order_obs[i]].copy()
temp_obs = eval_d_o[key].copy()
mins_temp_mask = ~temp_mask
temp_obs[mins_temp_mask] = np.zeros(
(mins_temp_mask.sum(), temp_obs.shape[2]))
temp_obs = temp_obs.reshape(num_agents, -1)
if i == 0:
reshape_obs = temp_obs.copy()
reshape_share_obs = temp_share_obs.copy()
else:
reshape_obs = np.concatenate(
(reshape_obs, temp_obs), axis=1)
reshape_share_obs = np.concatenate(
(reshape_share_obs, temp_share_obs),
axis=1)
eval_obs.append(reshape_obs)
eval_share_obs.append(reshape_share_obs)
eval_obs = np.array(eval_obs)
eval_share_obs = np.array(eval_share_obs)
eval_recurrent_hidden_states = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_recurrent_hidden_states_critic = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_masks = np.ones((1, num_agents, 1)).astype(np.float32)
if eval_dones[0]:
eval_episode += 1
if "success" in eval_infos[0].keys():
if eval_infos[0]['success']:
eval_success += 1
for agent_id in range(num_agents):
eval_recurrent_hidden_states[0][agent_id] = np.zeros(
args.hidden_size).astype(np.float32)
eval_recurrent_hidden_states_critic[0][
agent_id] = np.zeros(args.hidden_size).astype(
np.float32)
eval_masks[0][agent_id] = 0.0
else:
for agent_id in range(num_agents):
eval_masks[0][agent_id] = 1.0
if eval_episode >= args.eval_episodes:
logger.add_scalars('eval_success_rate', {
'eval_success_rate':
eval_success / args.eval_episodes
}, total_num_steps)
print("eval_success_rate is " +
str(eval_success / args.eval_episodes))
break
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
envs.close()
if args.eval:
eval_env.close()
def main():
args = get_config()
assert (
args.share_policy == True
and args.scenario_name == 'simple_speaker_listener'
) == False, (
"The simple_speaker_listener scenario can not use shared policy. Please check the config.py."
)
# seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
# path
model_dir = Path(
'./results') / args.env_name / args.scenario_name / args.algorithm_name
if not model_dir.exists():
curr_run = 'run1'
else:
exst_run_nums = [
int(str(folder.name).split('run')[1])
for folder in model_dir.iterdir()
if str(folder.name).startswith('run')
]
if len(exst_run_nums) == 0:
curr_run = 'run1'
else:
curr_run = 'run%i' % (max(exst_run_nums) + 1)
run_dir = model_dir / curr_run
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
num_agents = args.num_agents
#Policy network
if args.share_policy:
actor_critic = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=num_agents,
gain=args.gain,
base_kwargs={
'naive_recurrent':
args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization':
args.use_feature_normlization,
'use_feature_popart':
args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': args.use_same_dim
},
device=device)
actor_critic.to(device)
# algorithm
agents = PPO(actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
#replay buffer
rollouts = RolloutStorage(num_agents, args.episode_length,
args.n_rollout_threads,
envs.observation_space[0],
envs.action_space[0], args.hidden_size)
else:
actor_critic = []
agents = []
rollouts = []
for agent_id in range(num_agents):
ac = Policy(
envs.observation_space,
envs.action_space[agent_id],
num_agents=agent_id, # here is special
gain=args.gain,
base_kwargs={
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization': args.use_feature_normlization,
'use_feature_popart': args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': args.use_same_dim
},
device=device)
ac.to(device)
# algorithm
agent = PPO(ac,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
actor_critic.append(ac)
agents.append(agent)
#replay buffer
ro = SingleRolloutStorage(agent_id, args.episode_length,
args.n_rollout_threads,
envs.observation_space,
envs.action_space, args.hidden_size)
rollouts.append(ro)
# reset env
obs, _ = envs.reset()
# replay buffer
if args.share_policy:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs, 1).repeat(num_agents, axis=1)
rollouts.share_obs[0] = share_obs.copy()
rollouts.obs[0] = obs.copy()
rollouts.recurrent_hidden_states = np.zeros(
rollouts.recurrent_hidden_states.shape).astype(np.float32)
rollouts.recurrent_hidden_states_critic = np.zeros(
rollouts.recurrent_hidden_states_critic.shape).astype(np.float32)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].share_obs[0] = share_obs.copy()
rollouts[agent_id].obs[0] = np.array(list(obs[:, agent_id])).copy()
rollouts[agent_id].recurrent_hidden_states = np.zeros(
rollouts[agent_id].recurrent_hidden_states.shape).astype(
np.float32)
rollouts[agent_id].recurrent_hidden_states_critic = np.zeros(
rollouts[agent_id].recurrent_hidden_states_critic.shape
).astype(np.float32)
# run
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
timesteps = 0
for episode in range(episodes):
if args.use_linear_lr_decay: # decrease learning rate linearly
if args.share_policy:
update_linear_schedule(agents.optimizer, episode, episodes,
args.lr)
else:
for agent_id in range(num_agents):
update_linear_schedule(agents[agent_id].optimizer, episode,
episodes, args.lr)
for step in range(args.episode_length):
# Sample actions
values = []
actions = []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(
agent_id,
torch.FloatTensor(rollouts.share_obs[step, :,
agent_id]),
torch.FloatTensor(rollouts.obs[step, :, agent_id]),
torch.FloatTensor(
rollouts.recurrent_hidden_states[step, :,
agent_id]),
torch.FloatTensor(
rollouts.recurrent_hidden_states_critic[
step, :, agent_id]),
torch.FloatTensor(rollouts.masks[step, :,
agent_id]))
else:
actor_critic[agent_id].eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[
agent_id].act(
agent_id,
torch.FloatTensor(
rollouts[agent_id].share_obs[step, :]),
torch.FloatTensor(
rollouts[agent_id].obs[step, :]),
torch.FloatTensor(
rollouts[agent_id].recurrent_hidden_states[
step, :]),
torch.FloatTensor(
rollouts[agent_id].
recurrent_hidden_states_critic[step, :]),
torch.FloatTensor(
rollouts[agent_id].masks[step, :]))
values.append(value.detach().cpu().numpy())
actions.append(action.detach().cpu().numpy())
action_log_probs.append(
action_log_prob.detach().cpu().numpy())
recurrent_hidden_statess.append(
recurrent_hidden_states.detach().cpu().numpy())
recurrent_hidden_statess_critic.append(
recurrent_hidden_states_critic.detach().cpu().numpy())
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action_env = []
for agent_id in range(num_agents):
if envs.action_space[
agent_id].__class__.__name__ == 'MultiDiscrete':
uc_action = []
for j in range(envs.action_space[agent_id].shape):
uc_one_hot_action = np.zeros(
envs.action_space[agent_id].high[j] + 1)
uc_one_hot_action[actions[agent_id][i][j]] = 1
uc_action.append(uc_one_hot_action)
uc_action = np.concatenate(uc_action)
one_hot_action_env.append(uc_action)
elif envs.action_space[
agent_id].__class__.__name__ == 'Discrete':
one_hot_action = np.zeros(
envs.action_space[agent_id].n)
one_hot_action[actions[agent_id][i]] = 1
one_hot_action_env.append(one_hot_action)
else:
raise NotImplementedError
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, rewards, dones, infos, _ = envs.step(actions_env)
# If done then clean the history of observations.
# insert data in buffer
masks = []
for i, done in enumerate(dones):
mask = []
for agent_id in range(num_agents):
if done[agent_id]:
recurrent_hidden_statess[agent_id][i] = np.zeros(
args.hidden_size).astype(np.float32)
recurrent_hidden_statess_critic[agent_id][
i] = np.zeros(args.hidden_size).astype(np.float32)
mask.append([0.0])
else:
mask.append([1.0])
masks.append(mask)
if args.share_policy:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs, 1).repeat(num_agents,
axis=1)
rollouts.insert(
share_obs, obs,
np.array(recurrent_hidden_statess).transpose(1, 0, 2),
np.array(recurrent_hidden_statess_critic).transpose(
1, 0, 2),
np.array(actions).transpose(1, 0, 2),
np.array(action_log_probs).transpose(1, 0, 2),
np.array(values).transpose(1, 0, 2), rewards, masks)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].insert(
share_obs, np.array(list(obs[:, agent_id])),
np.array(recurrent_hidden_statess[agent_id]),
np.array(recurrent_hidden_statess_critic[agent_id]),
np.array(actions[agent_id]),
np.array(action_log_probs[agent_id]),
np.array(values[agent_id]), rewards[:, agent_id],
np.array(masks)[:, agent_id])
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
next_value, _, _ = actor_critic.get_value(
agent_id,
torch.FloatTensor(rollouts.share_obs[-1, :, agent_id]),
torch.FloatTensor(rollouts.obs[-1, :, agent_id]),
torch.FloatTensor(
rollouts.recurrent_hidden_states[-1, :, agent_id]),
torch.FloatTensor(
rollouts.recurrent_hidden_states_critic[-1, :,
agent_id]),
torch.FloatTensor(rollouts.masks[-1, :, agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id, next_value,
args.use_gae, args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents.value_normalizer)
else:
actor_critic[agent_id].eval()
next_value, _, _ = actor_critic[agent_id].get_value(
agent_id,
torch.FloatTensor(rollouts[agent_id].share_obs[-1, :]),
torch.FloatTensor(rollouts[agent_id].obs[-1, :]),
torch.FloatTensor(
rollouts[agent_id].recurrent_hidden_states[-1, :]),
torch.FloatTensor(
rollouts[agent_id].recurrent_hidden_states_critic[
-1, :]),
torch.FloatTensor(rollouts[agent_id].masks[-1, :]))
next_value = next_value.detach().cpu().numpy()
rollouts[agent_id].compute_returns(
next_value, args.use_gae, args.gamma, args.gae_lambda,
args.use_proper_time_limits, args.use_popart,
agents[agent_id].value_normalizer)
# update the network
if args.share_policy:
actor_critic.train()
value_loss, action_loss, dist_entropy = agents.update_share(
num_agents, rollouts)
for agent_id in range(num_agents):
rew = []
for i in range(rollouts.rewards.shape[1]):
rew.append(np.sum(rollouts.rewards[:, i, agent_id]))
logger.add_scalars('agent%i/average_episode_reward' % agent_id,
{'average_episode_reward': np.mean(rew)},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
# clean the buffer and reset
rollouts.after_update()
else:
value_losses = []
action_losses = []
dist_entropies = []
for agent_id in range(num_agents):
actor_critic[agent_id].train()
value_loss, action_loss, dist_entropy = agents[
agent_id].update_single(agent_id, rollouts[agent_id])
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
rew = []
for i in range(rollouts[agent_id].rewards.shape[1]):
rew.append(np.sum(rollouts[agent_id].rewards[:, i]))
logger.add_scalars('agent%i/average_episode_reward' % agent_id,
{'average_episode_reward': np.mean(rew)},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
rollouts[agent_id].after_update()
total_num_steps = (episode +
1) * args.episode_length * args.n_rollout_threads
if (episode % args.save_interval == 0 or episode == episodes -
1): # save for every interval-th episode or for the last epoch
if args.share_policy:
torch.save({'model': actor_critic},
str(save_dir) + "/agent_model.pt")
else:
for agent_id in range(num_agents):
torch.save({'model': actor_critic[agent_id]},
str(save_dir) + "/agent%i_model" % agent_id +
".pt")
# log information
if episode % args.log_interval == 0:
end = time.time()
print(
"\n Scenario {} Algo {} updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(args.scenario_name, args.algorithm_name, episode,
episodes, total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
if args.share_policy:
print("value loss of agent: " + str(value_loss))
else:
for agent_id in range(num_agents):
print("value loss of agent%i: " % agent_id +
str(value_losses[agent_id]))
if args.env_name == "MPE":
for agent_id in range(num_agents):
show_rewards = []
for info in infos:
if 'individual_reward' in info[agent_id].keys():
show_rewards.append(
info[agent_id]['individual_reward'])
logger.add_scalars(
'agent%i/individual_reward' % agent_id,
{'individual_reward': np.mean(show_rewards)},
total_num_steps)
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
envs.close()
def main():
args = get_args()
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
if args.gpu and torch.cuda.is_available() and args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# arguments
LAMBDA = [1.0, 0.0, 1.0,
10e-5] # for [loss_dq, loss_n_dq, loss_jeq, loss_l2]
CUDA_VISIBLE_DEVICES = 0
seed = args.seed
train = args.train
demo = args.demo
task = args.task
iteration = 3
convs = [(32, 7, 3), (64, 4, 2), (64, 3, 1)]
non_pixel_layer = [64]
in_feature = 7 * 7 * 64
hidden_actions = [128]
hidden_value = [128]
aggregator = "reduceLocalMean"
dtype = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
#if not train:
# args.num_env_steps = 50000
base_kwargs = {
'non_pixel_layer': non_pixel_layer,
'convs': convs,
'frame_history_len': args.frame_history_len,
'in_feature': in_feature,
'hidden_actions': hidden_actions,
'hidden_value': hidden_value,
'aggregator': aggregator
}
# logger
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)
# threads and device
torch.set_num_threads(1)
device = torch.device("cuda:0" if args.gpu else "cpu")
print("device:", device)
gpu = args.gpu
if (gpu == True):
print("current available gpu numbers: %d" % torch.cuda.device_count())
if torch.cuda.is_available():
torch.cuda.set_device(CUDA_VISIBLE_DEVICES)
print("CUDA Device: %d" % torch.cuda.current_device())
# envs
#envs = gym.make(task)
#obs_space = env.observation_space
#act_space = env.action_space
#action_template = env.action_space.noop()
env = gym.make(args.task)
obs_space = env.observation_space
act_space = env.action_space
action_template = env.action_space.noop()
# policy
actor_critic = Policy(obs_space, act_space, base_kwargs=base_kwargs)
actor_critic.to(device)
# algorithm
if args.algo == 'ppo':
agent = PPO(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=7e-4,
eps=1e-5,
max_grad_norm=args.max_grad_norm,
)
else:
raise NotImplementedError
# storage
replay_buffer = None
if args.frame_history_len > 1:
_, _, non_pixel_shape = parse_obs_space(obs_space)
add_non_pixel = True if non_pixel_shape > 0 else False
replay_buffer = ReplayBuffer(100000, args.frame_history_len,
non_pixel_shape, add_non_pixel)
rollouts = RolloutStorage(replay_buffer, args.frame_history_len,
args.num_steps, args.num_processes, obs_space,
act_space)
obs = env.reset()
#print("reset obs pov size: ",obs['pov'].shape)
# obs: key: inventory.dirt...
# (num_processes, size)
pov, non_pixel_feature = get_obs_features(obs_space, obs)
#pov, non_pixel_feature = multi_get_obs_features(obs)
if args.frame_history_len > 1:
last_stored_frame_idx = replay_buffer.store_frame(
pov, non_pixel_feature)
pov = replay_buffer.encode_recent_observation() / 255.0 # 12 h w
pov = torch.from_numpy(pov.copy()).reshape(args.num_processes,
*pov.shape)
elif args.frame_history_len == 1:
pov = pov.transpose(2, 0, 1) / 255.0
pov = torch.from_numpy(pov.copy()).reshape(args.num_processes,
*pov.shape)
else:
raise NotImplementedError
non_pixel_feature = (torch.tensor(non_pixel_feature) / 180.0).reshape(
args.num_processes, -1)
rollouts.obs[0].copy_(pov)
rollouts.non_pixel_obs[0].copy_(non_pixel_feature)
rollouts.to(device)
# ?
episode_rewards = deque(maxlen=10)
start = time.time()
num_updates = int(
args.num_env_steps) // args.num_steps // args.num_processes
print("Total steps: ", args.num_env_steps)
ep = 0
ep_rewards = []
#mean_episode_reward = -float('nan')
best_mean_episode_reward = -float('inf')
#total_rewards = [0 for i in range(args.num_processes)]
total_rewards = 0
for j in range(num_updates):
for step in range(args.num_steps):
# num_steps = 5
# Sample actions
with torch.no_grad():
# actor_critic.act output size
# actions: torch.Tensor, not list
value, actions, action_log_probs = actor_critic.act(
rollouts.obs[step], rollouts.non_pixel_obs[step])
# value size: batch x 1
# actions size: torch.Tensor num_processes x num_branches
# action_log_probs : torch.Tensor num_processes x num_branches
#print(actions)
actions_list = actions.squeeze().tolist()
action = get_actions_continuous(actions_list, act_space,
action_template)
# step:
#print(actions)
obs, reward, done, infos = env.step(action)
#print('.',end='')
if args.num_env_steps <= 50000:
env.render()
pov, non_pixel_feature = get_obs_features(obs_space, obs)
#pov, non_pixel_feature = multi_get_obs_features(obs)
if args.frame_history_len > 1:
last_stored_frame_idx = replay_buffer.store_frame(
pov, non_pixel_feature)
pov = replay_buffer.encode_recent_observation(
) / 255.0 # 12 h w
pov = torch.from_numpy(pov.copy()).reshape(
args.num_processes, *pov.shape)
elif args.frame_history_len == 1:
pov = pov.transpose(2, 0, 1) / 255.0
pov = torch.from_numpy(pov.copy()).reshape(
args.num_processes, *pov.shape)
else:
raise NotImplementedError
non_pixel_feature = (torch.tensor(non_pixel_feature) /
180.0).reshape(args.num_processes, -1)
total_rewards += reward
#for i in range(len(reward)):
# total_rewards[i] += reward[i]
reward = torch.tensor([reward]).reshape(args.num_processes,
-1).type(dtype)
# TODO: may not need bas_masks
masks = torch.FloatTensor([[0.0] if done else [1.0]])
bad_masks = torch.FloatTensor([[1.0]])
if done:
ep += 1
ep_rewards.append(total_rewards)
best_mean_episode_reward = log(j, args.task, ep,
np.array(ep_rewards),
best_mean_episode_reward)
obs = env.reset()
pov, non_pixel_feature = get_obs_features(obs_space, obs)
#pov, non_pixel_feature = multi_get_obs_features(obs)
if args.frame_history_len > 1:
last_stored_frame_idx = replay_buffer.store_frame(
pov, non_pixel_feature)
pov = replay_buffer.encode_recent_observation(
) / 255.0 # 12 h w
pov = torch.from_numpy(pov.copy()).reshape(
args.num_processes, *pov.shape)
elif args.frame_history_len == 1:
pov = pov.transpose(2, 0, 1) / 255.0
pov = torch.from_numpy(pov.copy()).reshape(
args.num_processes, *pov.shape)
else:
raise NotImplementedError
non_pixel_feature = (torch.tensor(non_pixel_feature) /
180.0).reshape(args.num_processes, -1)
total_rewards = 0
# ？
rollouts.insert(pov, non_pixel_feature, actions, action_log_probs,
value, reward, masks, bad_masks)
with torch.no_grad():
next_value = actor_critic.get_value(rollouts.obs[-1],
rollouts.non_pixel_obs[-1])
rollouts.compute_returns(next_value, args.use_gae, args.gamma,
args.gae_lambda, args.use_proper_time_limits)
# TODO: minibathc = 32, 1 processes x 10 step should larger than 32
value_loss, action_loss, dist_entropy = agent.update(rollouts)
rollouts.after_update()
# save for every interval-th episode or for the last epoch
if (j % args.save_interval == 0
or j == num_updates - 1) and args.save_model_dir != '':
save_path = os.path.join(args.save_model_dir, args.algo)
try:
os.makedirs(save_path)
except OSError:
pass
torch.save(actor_critic, os.path.join(save_path,
args.task + ".pt"))
if j % args.log_interval == 0 and len(ep_rewards) >= 0:
total_num_steps = (j + 1) * args.num_processes * args.num_steps
end = time.time()
print("----------- Logs -------------")
if len(ep_rewards) == 0:
print(
"Updates {}, num timesteps {}, FPS {} \nThe {}th training episodes,"
.format(j, total_num_steps,
int(total_num_steps / (end - start)),
len(ep_rewards)))
else:
print(
"Updates {}, num timesteps {}, FPS {} \nThe {}th training episodes,\nmean/median reward {:.1f}/{:.1f}, min/max reward {:.1f}/{:.1f}\n"
.format(j, total_num_steps,
int(total_num_steps / (end - start)),
len(ep_rewards), np.mean(ep_rewards),
np.median(ep_rewards), np.min(ep_rewards),
np.max(ep_rewards)))
print("-----------------------Training ends-----------------------")
env.close()
def main():
args = get_config()
run = wandb.init(project='curriculum',name=str(args.algorithm_name) + "_seed" + str(args.seed))
# run = wandb.init(project='check',name='separate_reward')
assert (args.share_policy == True and args.scenario_name == 'simple_speaker_listener') == False, ("The simple_speaker_listener scenario can not use shared policy. Please check the config.py.")
# seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(1)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
# path
model_dir = Path('./results') / args.env_name / args.scenario_name / args.algorithm_name
node_dir = Path('./node') / args.env_name / args.scenario_name / args.algorithm_name
if not model_dir.exists():
curr_run = 'run1'
else:
exst_run_nums = [int(str(folder.name).split('run')[1]) for folder in model_dir.iterdir() if str(folder.name).startswith('run')]
if len(exst_run_nums) == 0:
curr_run = 'run1'
else:
curr_run = 'run%i' % (max(exst_run_nums) + 1)
if not node_dir.exists():
node_curr_run = 'run1'
else:
exst_run_nums = [int(str(folder.name).split('run')[1]) for folder in node_dir.iterdir() if str(folder.name).startswith('run')]
if len(exst_run_nums) == 0:
node_curr_run = 'run1'
else:
node_curr_run = 'run%i' % (max(exst_run_nums) + 1)
run_dir = model_dir / curr_run
save_node_dir = node_dir / node_curr_run
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
num_agents = args.num_agents
#Policy network
if args.share_policy:
actor_base = ATTBase_actor_dist_add(envs.observation_space[0].shape[0], envs.action_space[0], num_agents)
critic_base = ATTBase_critic_add(envs.observation_space[0].shape[0], num_agents)
actor_critic = Policy3(envs.observation_space[0],
envs.action_space[0],
num_agents = num_agents,
base=None,
actor_base=actor_base,
critic_base=critic_base,
base_kwargs={'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'attn': args.attn,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer':args.use_common_layer,
'use_feature_normlization':args.use_feature_normlization,
'use_feature_popart':args.use_feature_popart,
'use_orthogonal':args.use_orthogonal,
'layer_N':args.layer_N,
'use_ReLU':args.use_ReLU
},
device = device)
actor_critic.to(device)
# algorithm
agents = PPO3(actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss= args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
device=device)
# #replay buffer
# rollouts = RolloutStorage(num_agents,
# args.episode_length,
# args.n_rollout_threads,
# envs.observation_space[0],
# envs.action_space[0],
# args.hidden_size)
else:
actor_critic = []
agents = []
rollouts = []
for agent_id in range(num_agents):
ac = Policy(envs.observation_space,
envs.action_space[agent_id],
num_agents = agent_id, # here is special
base_kwargs={'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'attn': args.attn,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer':args.use_common_layer,
'use_feature_normlization':args.use_feature_normlization,
'use_feature_popart':args.use_feature_popart,
'use_orthogonal':args.use_orthogonal,
'layer_N':args.layer_N,
'use_ReLU':args.use_ReLU
},
device = device)
ac.to(device)
# algorithm
agent = PPO(ac,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss= args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
device=device)
actor_critic.append(ac)
agents.append(agent)
#replay buffer
ro = SingleRolloutStorage(agent_id,
args.episode_length,
args.n_rollout_threads,
envs.observation_space,
envs.action_space,
args.hidden_size)
rollouts.append(ro)
boundary = 3
start_boundary = [-0.3,0.3,-0.3,0.3] # 分别代表x的范围和y的范围
# start_boundary = [2.4,3.0,2.4,3.0]
max_step = 0.6
N_easy = 0
test_flag = 0
reproduce_flag = 0
target_num = 4
last_agent_num = 4
now_agent_num = num_agents
mean_cover_rate = 0
eval_frequency = 2 #需要fix几个回合
check_frequency = 1
save_node_frequency = 5
save_node_flag = True
save_90_flag = True
historical_length = 5
random.seed(args.seed)
np.random.seed(args.seed)
# init the Gan
gan_configs['goal_range'] = boundary
gan_configs['goal_center'] = np.zeros((num_agents + num_agents)* 2, dtype=float)
gan_configs['goal_size'] = (num_agents + num_agents)*2
gan = StateGAN(gan_configs = gan_configs, state_range=gan_configs['goal_range'])
feasible_goals = generate_initial_goals(num_case = 10000, start_boundary = start_boundary, agent_num = args.num_agents)
dis_loss, gen_loss = gan.pretrain(states=feasible_goals, outer_iters=gan_configs['gan_outer_iters'])
print('discriminator_loss:',str(dis_loss.cpu()), 'generator_loss:',str(gen_loss.cpu()))
# init the StateCollection
all_goals = StateCollection(distance_threshold=goal_configs['coll_eps'])
# run
begin = time.time()
episodes = int(args.num_env_steps) // args.episode_length // args.n_rollout_threads // eval_frequency
curriculum_episode = 0
current_timestep = 0
one_length = args.n_rollout_threads
starts_length = args.n_rollout_threads
num_envs = args.n_rollout_threads
for episode in range(episodes):
if args.use_linear_lr_decay:# decrease learning rate linearly
if args.share_policy:
update_linear_schedule(agents.optimizer, episode, episodes, args.lr)
else:
for agent_id in range(num_agents):
update_linear_schedule(agents[agent_id].optimizer, episode, episodes, args.lr)
raw_goals, _ = gan.sample_states_with_noise(goal_configs['num_new_goals'])
# replay buffer
if all_goals.size > 0:
old_goals = all_goals.sample(goal_configs['num_old_goals'])
goals = np.vstack([raw_goals, old_goals])
else:
goals = raw_goals
if goals.shape[0] < num_envs:
add_num = num_envs - goals.shape[0]
goals = np.vstack([goals, goals[:add_num]]) #补齐到num_new_goals+num_old_goals
# generate the starts
starts = numpy_to_list(goals, list_length=num_envs, shape=(num_agents*2,2))
for times in range(eval_frequency):
obs = envs.new_starts_obs(starts, now_agent_num, starts_length)
#replay buffer
rollouts = RolloutStorage(num_agents,
args.episode_length,
starts_length,
envs.observation_space[0],
envs.action_space[0],
args.hidden_size)
# replay buffer init
if args.share_policy:
share_obs = obs.reshape(starts_length, -1)
share_obs = np.expand_dims(share_obs,1).repeat(num_agents,axis=1)
rollouts.share_obs[0] = share_obs.copy()
rollouts.obs[0] = obs.copy()
rollouts.recurrent_hidden_states = np.zeros(rollouts.recurrent_hidden_states.shape).astype(np.float32)
rollouts.recurrent_hidden_states_critic = np.zeros(rollouts.recurrent_hidden_states_critic.shape).astype(np.float32)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].share_obs[0] = share_obs.copy()
rollouts[agent_id].obs[0] = np.array(list(obs[:,agent_id])).copy()
rollouts[agent_id].recurrent_hidden_states = np.zeros(rollouts[agent_id].recurrent_hidden_states.shape).astype(np.float32)
rollouts[agent_id].recurrent_hidden_states_critic = np.zeros(rollouts[agent_id].recurrent_hidden_states_critic.shape).astype(np.float32)
step_cover_rate = np.zeros(shape=(one_length,args.episode_length))
for step in range(args.episode_length):
# Sample actions
values = []
actions= []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(agent_id,
torch.FloatTensor(rollouts.share_obs[step,:,agent_id]),
torch.FloatTensor(rollouts.obs[step,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states[step,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states_critic[step,:,agent_id]),
torch.FloatTensor(rollouts.masks[step,:,agent_id]))
else:
actor_critic[agent_id].eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[agent_id].act(agent_id,
torch.FloatTensor(rollouts[agent_id].share_obs[step,:]),
torch.FloatTensor(rollouts[agent_id].obs[step,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states[step,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states_critic[step,:]),
torch.FloatTensor(rollouts[agent_id].masks[step,:]))
values.append(value.detach().cpu().numpy())
actions.append(action.detach().cpu().numpy())
action_log_probs.append(action_log_prob.detach().cpu().numpy())
recurrent_hidden_statess.append(recurrent_hidden_states.detach().cpu().numpy())
recurrent_hidden_statess_critic.append(recurrent_hidden_states_critic.detach().cpu().numpy())
# rearrange action
actions_env = []
for i in range(starts_length):
one_hot_action_env = []
for agent_id in range(num_agents):
if envs.action_space[agent_id].__class__.__name__ == 'MultiDiscrete':
uc_action = []
for j in range(envs.action_space[agent_id].shape):
uc_one_hot_action = np.zeros(envs.action_space[agent_id].high[j]+1)
uc_one_hot_action[actions[agent_id][i][j]] = 1
uc_action.append(uc_one_hot_action)
uc_action = np.concatenate(uc_action)
one_hot_action_env.append(uc_action)
elif envs.action_space[agent_id].__class__.__name__ == 'Discrete':
one_hot_action = np.zeros(envs.action_space[agent_id].n)
one_hot_action[actions[agent_id][i]] = 1
one_hot_action_env.append(one_hot_action)
else:
raise NotImplementedError
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, rewards, dones, infos, _ = envs.step(actions_env, starts_length, num_agents)
cover_rate_list = []
for env_id in range(one_length):
cover_rate_list.append(infos[env_id][0]['cover_rate'])
step_cover_rate[:,step] = np.array(cover_rate_list)
# step_cover_rate[:,step] = np.array(infos)[0:one_length,0]
# If done then clean the history of observations.
# insert data in buffer
masks = []
for i, done in enumerate(dones):
mask = []
for agent_id in range(num_agents):
if done[agent_id]:
recurrent_hidden_statess[agent_id][i] = np.zeros(args.hidden_size).astype(np.float32)
recurrent_hidden_statess_critic[agent_id][i] = np.zeros(args.hidden_size).astype(np.float32)
mask.append([0.0])
else:
mask.append([1.0])
masks.append(mask)
if args.share_policy:
share_obs = obs.reshape(starts_length, -1)
share_obs = np.expand_dims(share_obs,1).repeat(num_agents,axis=1)
rollouts.insert(share_obs,
obs,
np.array(recurrent_hidden_statess).transpose(1,0,2),
np.array(recurrent_hidden_statess_critic).transpose(1,0,2),
np.array(actions).transpose(1,0,2),
np.array(action_log_probs).transpose(1,0,2),
np.array(values).transpose(1,0,2),
rewards,
masks)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].insert(share_obs,
np.array(list(obs[:,agent_id])),
np.array(recurrent_hidden_statess[agent_id]),
np.array(recurrent_hidden_statess_critic[agent_id]),
np.array(actions[agent_id]),
np.array(action_log_probs[agent_id]),
np.array(values[agent_id]),
rewards[:,agent_id],
np.array(masks)[:,agent_id])
# logger.add_scalars('agent/training_cover_rate',{'training_cover_rate': np.mean(np.mean(step_cover_rate[:,-historical_length:],axis=1))}, current_timestep)
wandb.log({'training_cover_rate': np.mean(np.mean(step_cover_rate[:,-historical_length:],axis=1))}, current_timestep)
print('training_cover_rate: ', np.mean(np.mean(step_cover_rate[:,-historical_length:],axis=1)))
current_timestep += args.episode_length * starts_length
curriculum_episode += 1
#region train the gan
if times == 1:
start_time = time.time()
filtered_raw_goals = []
labels = np.zeros((num_envs, 1), dtype = int)
for i in range(num_envs):
R_i = np.mean(step_cover_rate[i, -goal_configs['historical_length']:])
if R_i < goal_configs['R_max'] and R_i > goal_configs['R_min']:
labels[i] = 1
filtered_raw_goals.append(goals[i])
gan.train(goals, labels)
all_goals.append(filtered_raw_goals)
end_time = time.time()
print("Gan training time: %.2f"%(end_time-start_time))
with torch.no_grad(): # get value and com
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
next_value,_,_ = actor_critic.get_value(agent_id,
torch.FloatTensor(rollouts.share_obs[-1,:,agent_id]),
torch.FloatTensor(rollouts.obs[-1,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states[-1,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states_critic[-1,:,agent_id]),
torch.FloatTensor(rollouts.masks[-1,:,agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id,
next_value,
args.use_gae,
args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents.value_normalizer)
else:
actor_critic[agent_id].eval()
next_value,_,_ = actor_critic[agent_id].get_value(agent_id,
torch.FloatTensor(rollouts[agent_id].share_obs[-1,:]),
torch.FloatTensor(rollouts[agent_id].obs[-1,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states[-1,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states_critic[-1,:]),
torch.FloatTensor(rollouts[agent_id].masks[-1,:]))
next_value = next_value.detach().cpu().numpy()
rollouts[agent_id].compute_returns(next_value,
args.use_gae,
args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents[agent_id].value_normalizer)
# update the network
if args.share_policy:
actor_critic.train()
value_loss, action_loss, dist_entropy = agents.update_share_asynchronous(last_agent_num, rollouts, False, initial_optimizer=False)
print('value_loss: ', value_loss)
wandb.log(
{'value_loss': value_loss},
current_timestep)
rew = []
for i in range(rollouts.rewards.shape[1]):
rew.append(np.sum(rollouts.rewards[:,i]))
wandb.log(
{'average_episode_reward': np.mean(rew)},
current_timestep)
# clean the buffer and reset
rollouts.after_update()
else:
value_losses = []
action_losses = []
dist_entropies = []
for agent_id in range(num_agents):
actor_critic[agent_id].train()
value_loss, action_loss, dist_entropy = agents[agent_id].update_single(agent_id, rollouts[agent_id])
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
rew = []
for i in range(rollouts[agent_id].rewards.shape[1]):
rew.append(np.sum(rollouts[agent_id].rewards[:,i]))
logger.add_scalars('agent%i/average_episode_reward'%agent_id,
{'average_episode_reward': np.mean(rew)},
(episode+1) * args.episode_length * one_length*eval_frequency)
rollouts[agent_id].after_update()
# test
if episode % check_frequency==0:
obs, _ = envs.reset(num_agents)
episode_length = 70
#replay buffer
rollouts = RolloutStorage(num_agents,
episode_length,
args.n_rollout_threads,
envs.observation_space[0],
envs.action_space[0],
args.hidden_size)
# replay buffer init
if args.share_policy:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs,1).repeat(num_agents,axis=1)
rollouts.share_obs[0] = share_obs.copy()
rollouts.obs[0] = obs.copy()
rollouts.recurrent_hidden_states = np.zeros(rollouts.recurrent_hidden_states.shape).astype(np.float32)
rollouts.recurrent_hidden_states_critic = np.zeros(rollouts.recurrent_hidden_states_critic.shape).astype(np.float32)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].share_obs[0] = share_obs.copy()
rollouts[agent_id].obs[0] = np.array(list(obs[:,agent_id])).copy()
rollouts[agent_id].recurrent_hidden_states = np.zeros(rollouts[agent_id].recurrent_hidden_states.shape).astype(np.float32)
rollouts[agent_id].recurrent_hidden_states_critic = np.zeros(rollouts[agent_id].recurrent_hidden_states_critic.shape).astype(np.float32)
test_cover_rate = np.zeros(shape=(args.n_rollout_threads,episode_length))
for step in range(episode_length):
# Sample actions
values = []
actions= []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(agent_id,
torch.FloatTensor(rollouts.share_obs[step,:,agent_id]),
torch.FloatTensor(rollouts.obs[step,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states[step,:,agent_id]),
torch.FloatTensor(rollouts.recurrent_hidden_states_critic[step,:,agent_id]),
torch.FloatTensor(rollouts.masks[step,:,agent_id]),deterministic=True)
else:
actor_critic[agent_id].eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[agent_id].act(agent_id,
torch.FloatTensor(rollouts[agent_id].share_obs[step,:]),
torch.FloatTensor(rollouts[agent_id].obs[step,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states[step,:]),
torch.FloatTensor(rollouts[agent_id].recurrent_hidden_states_critic[step,:]),
torch.FloatTensor(rollouts[agent_id].masks[step,:]),deterministic=True)
values.append(value.detach().cpu().numpy())
actions.append(action.detach().cpu().numpy())
action_log_probs.append(action_log_prob.detach().cpu().numpy())
recurrent_hidden_statess.append(recurrent_hidden_states.detach().cpu().numpy())
recurrent_hidden_statess_critic.append(recurrent_hidden_states_critic.detach().cpu().numpy())
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action_env = []
for agent_id in range(num_agents):
if envs.action_space[agent_id].__class__.__name__ == 'MultiDiscrete':
uc_action = []
for j in range(envs.action_space[agent_id].shape):
uc_one_hot_action = np.zeros(envs.action_space[agent_id].high[j]+1)
uc_one_hot_action[actions[agent_id][i][j]] = 1
uc_action.append(uc_one_hot_action)
uc_action = np.concatenate(uc_action)
one_hot_action_env.append(uc_action)
elif envs.action_space[agent_id].__class__.__name__ == 'Discrete':
one_hot_action = np.zeros(envs.action_space[agent_id].n)
one_hot_action[actions[agent_id][i]] = 1
one_hot_action_env.append(one_hot_action)
else:
raise NotImplementedError
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, rewards, dones, infos, _ = envs.step(actions_env, args.n_rollout_threads, num_agents)
cover_rate_list = []
for env_id in range(args.n_rollout_threads):
cover_rate_list.append(infos[env_id][0]['cover_rate'])
test_cover_rate[:,step] = np.array(cover_rate_list)
# test_cover_rate[:,step] = np.array(infos)[:,0]
# If done then clean the history of observations.
# insert data in buffer
masks = []
for i, done in enumerate(dones):
mask = []
for agent_id in range(num_agents):
if done[agent_id]:
recurrent_hidden_statess[agent_id][i] = np.zeros(args.hidden_size).astype(np.float32)
recurrent_hidden_statess_critic[agent_id][i] = np.zeros(args.hidden_size).astype(np.float32)
mask.append([0.0])
else:
mask.append([1.0])
masks.append(mask)
if args.share_policy:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs,1).repeat(num_agents,axis=1)
rollouts.insert(share_obs,
obs,
np.array(recurrent_hidden_statess).transpose(1,0,2),
np.array(recurrent_hidden_statess_critic).transpose(1,0,2),
np.array(actions).transpose(1,0,2),
np.array(action_log_probs).transpose(1,0,2),
np.array(values).transpose(1,0,2),
rewards,
masks)
else:
share_obs = []
for o in obs:
share_obs.append(list(itertools.chain(*o)))
share_obs = np.array(share_obs)
for agent_id in range(num_agents):
rollouts[agent_id].insert(share_obs,
np.array(list(obs[:,agent_id])),
np.array(recurrent_hidden_statess[agent_id]),
np.array(recurrent_hidden_statess_critic[agent_id]),
np.array(actions[agent_id]),
np.array(action_log_probs[agent_id]),
np.array(values[agent_id]),
rewards[:,agent_id],
np.array(masks)[:,agent_id])
# logger.add_scalars('agent/cover_rate_1step',{'cover_rate_1step': np.mean(test_cover_rate[:,-1])},current_timestep)
# logger.add_scalars('agent/cover_rate_5step',{'cover_rate_5step': np.mean(np.mean(test_cover_rate[:,-historical_length:],axis=1))}, current_timestep)
rew = []
for i in range(rollouts.rewards.shape[1]):
rew.append(np.sum(rollouts.rewards[:,i]))
wandb.log(
{'eval_episode_reward': np.mean(rew)},
current_timestep)
wandb.log({'cover_rate_1step': np.mean(test_cover_rate[:,-1])},current_timestep)
wandb.log({'cover_rate_5step': np.mean(np.mean(test_cover_rate[:,-historical_length:],axis=1))}, current_timestep)
mean_cover_rate = np.mean(np.mean(test_cover_rate[:,-historical_length:],axis=1))
if mean_cover_rate >= 0.9 and args.algorithm_name=='ours' and save_90_flag:
torch.save({'model': actor_critic}, str(save_dir) + "/cover09_agent_model.pt")
save_90_flag = False
print('test_agent_num: ', last_agent_num)
print('test_mean_cover_rate: ', mean_cover_rate)
total_num_steps = current_timestep
if (episode % args.save_interval == 0 or episode == episodes - 1):# save for every interval-th episode or for the last epoch
if args.share_policy:
torch.save({
'model': actor_critic
},
str(save_dir) + "/agent_model.pt")
else:
for agent_id in range(num_agents):
torch.save({
'model': actor_critic[agent_id]
},
str(save_dir) + "/agent%i_model" % agent_id + ".pt")
# log information
if episode % args.log_interval == 0:
end = time.time()
print("\n Scenario {} Algo {} updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(args.scenario_name,
args.algorithm_name,
episode,
episodes,
total_num_steps,
args.num_env_steps,
int(total_num_steps / (end - begin))))
if args.share_policy:
print("value loss of agent: " + str(value_loss))
else:
for agent_id in range(num_agents):
print("value loss of agent%i: " % agent_id + str(value_losses[agent_id]))
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
envs.close()
def main():
args = get_config()
# seed
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
# path
model_dir = Path(
'./results') / args.env_name / args.map_name / args.algorithm_name
if not model_dir.exists():
curr_run = 'run1'
else:
exst_run_nums = [
int(str(folder.name).split('run')[1])
for folder in model_dir.iterdir()
if str(folder.name).startswith('run')
]
if len(exst_run_nums) == 0:
curr_run = 'run1'
else:
curr_run = 'run%i' % (max(exst_run_nums) + 1)
run_dir = model_dir / curr_run
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
if args.eval:
eval_env = make_eval_env(args)
num_agents = get_map_params(args.map_name)["n_agents"]
#Policy network
if args.share_policy:
actor_critic = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=num_agents,
gain=args.gain,
base_kwargs={
'naive_recurrent':
args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization':
args.use_feature_normlization,
'use_feature_popart':
args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': args.use_same_dim
},
device=device)
actor_critic.to(device)
# algorithm
agents = PPO(actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
#replay buffer
rollouts = RolloutStorage(num_agents, args.episode_length,
args.n_rollout_threads,
envs.observation_space[0],
envs.action_space[0], args.hidden_size)
else:
actor_critic = []
agents = []
for agent_id in range(num_agents):
ac = Policy(envs.observation_space[0],
envs.action_space[0],
num_agents=num_agents,
gain=args.gain,
base_kwargs={
'naive_recurrent': args.naive_recurrent_policy,
'recurrent': args.recurrent_policy,
'hidden_size': args.hidden_size,
'recurrent_N': args.recurrent_N,
'attn': args.attn,
'attn_only_critic': args.attn_only_critic,
'attn_size': args.attn_size,
'attn_N': args.attn_N,
'attn_heads': args.attn_heads,
'dropout': args.dropout,
'use_average_pool': args.use_average_pool,
'use_common_layer': args.use_common_layer,
'use_feature_normlization':
args.use_feature_normlization,
'use_feature_popart': args.use_feature_popart,
'use_orthogonal': args.use_orthogonal,
'layer_N': args.layer_N,
'use_ReLU': args.use_ReLU,
'use_same_dim': args.use_same_dim
},
device=device)
ac.to(device)
# algorithm
agent = PPO(ac,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
weight_decay=args.weight_decay,
max_grad_norm=args.max_grad_norm,
use_max_grad_norm=args.use_max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss,
use_common_layer=args.use_common_layer,
use_huber_loss=args.use_huber_loss,
huber_delta=args.huber_delta,
use_popart=args.use_popart,
use_value_high_masks=args.use_value_high_masks,
device=device)
actor_critic.append(ac)
agents.append(agent)
#replay buffer
rollouts = RolloutStorage(num_agents, args.episode_length,
args.n_rollout_threads,
envs.observation_space[0],
envs.action_space[0], args.hidden_size)
# reset env
obs, available_actions = envs.reset()
# replay buffer
if len(envs.observation_space[0]) == 3:
share_obs = obs.reshape(args.n_rollout_threads, -1,
envs.observation_space[0][1],
envs.observation_space[0][2])
else:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs, 1).repeat(num_agents, axis=1)
rollouts.share_obs[0] = share_obs.copy()
rollouts.obs[0] = obs.copy()
rollouts.available_actions[0] = available_actions.copy()
rollouts.recurrent_hidden_states = np.zeros(
rollouts.recurrent_hidden_states.shape).astype(np.float32)
rollouts.recurrent_hidden_states_critic = np.zeros(
rollouts.recurrent_hidden_states_critic.shape).astype(np.float32)
# run
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
timesteps = 0
last_battles_game = np.zeros(args.n_rollout_threads)
last_battles_won = np.zeros(args.n_rollout_threads)
for episode in range(episodes):
if args.use_linear_lr_decay: # decrease learning rate linearly
if args.share_policy:
update_linear_schedule(agents.optimizer, episode, episodes,
args.lr)
else:
for agent_id in range(num_agents):
update_linear_schedule(agents[agent_id].optimizer, episode,
episodes, args.lr)
for step in range(args.episode_length):
# Sample actions
values = []
actions = []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(
agent_id,
torch.tensor(rollouts.share_obs[step, :,
agent_id]),
torch.tensor(rollouts.obs[step, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[step, :,
agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[
step, :, agent_id]),
torch.tensor(rollouts.masks[step, :, agent_id]),
torch.tensor(rollouts.available_actions[step, :,
agent_id]))
else:
actor_critic[agent_id].eval()
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[
agent_id].act(
agent_id,
torch.tensor(rollouts.share_obs[step, :,
agent_id]),
torch.tensor(rollouts.obs[step, :, agent_id]),
torch.tensor(rollouts.recurrent_hidden_states[
step, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[
step, :, agent_id]),
torch.tensor(rollouts.masks[step, :,
agent_id]),
torch.tensor(
rollouts.available_actions[step, :,
agent_id]))
values.append(value.detach().cpu().numpy())
actions.append(action.detach().cpu().numpy())
action_log_probs.append(
action_log_prob.detach().cpu().numpy())
recurrent_hidden_statess.append(
recurrent_hidden_states.detach().cpu().numpy())
recurrent_hidden_statess_critic.append(
recurrent_hidden_states_critic.detach().cpu().numpy())
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action_env = []
for agent_id in range(num_agents):
one_hot_action = np.zeros(envs.action_space[agent_id].n)
one_hot_action[actions[agent_id][i]] = 1
one_hot_action_env.append(one_hot_action)
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, reward, dones, infos, available_actions = envs.step(
actions_env)
# If done then clean the history of observations.
# insert data in buffer
masks = []
for i, done in enumerate(dones):
mask = []
for agent_id in range(num_agents):
if done:
recurrent_hidden_statess[agent_id][i] = np.zeros(
args.hidden_size).astype(np.float32)
recurrent_hidden_statess_critic[agent_id][
i] = np.zeros(args.hidden_size).astype(np.float32)
mask.append([0.0])
else:
mask.append([1.0])
masks.append(mask)
bad_masks = []
high_masks = []
for info in infos:
bad_mask = []
high_mask = []
for agent_id in range(num_agents):
if info[agent_id]['bad_transition']:
bad_mask.append([0.0])
else:
bad_mask.append([1.0])
if info[agent_id]['high_masks']:
high_mask.append([1.0])
else:
high_mask.append([0.0])
bad_masks.append(bad_mask)
high_masks.append(high_mask)
if len(envs.observation_space[0]) == 3:
share_obs = obs.reshape(args.n_rollout_threads, -1,
envs.observation_space[0][1],
envs.observation_space[0][2])
share_obs = np.expand_dims(share_obs, 1).repeat(num_agents,
axis=1)
rollouts.insert(
share_obs, obs,
np.array(recurrent_hidden_statess).transpose(1, 0, 2),
np.array(recurrent_hidden_statess_critic).transpose(
1, 0, 2),
np.array(actions).transpose(1, 0, 2),
np.array(action_log_probs).transpose(1, 0, 2),
np.array(values).transpose(1, 0, 2), reward, masks,
bad_masks, high_masks, available_actions)
else:
share_obs = obs.reshape(args.n_rollout_threads, -1)
share_obs = np.expand_dims(share_obs, 1).repeat(num_agents,
axis=1)
rollouts.insert(
share_obs, obs,
np.array(recurrent_hidden_statess).transpose(1, 0, 2),
np.array(recurrent_hidden_statess_critic).transpose(
1, 0, 2),
np.array(actions).transpose(1, 0, 2),
np.array(action_log_probs).transpose(1, 0, 2),
np.array(values).transpose(1, 0, 2), reward, masks,
bad_masks, high_masks, available_actions)
with torch.no_grad():
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
next_value, _, _ = actor_critic.get_value(
agent_id,
torch.tensor(rollouts.share_obs[-1, :, agent_id]),
torch.tensor(rollouts.obs[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[-1, :,
agent_id]),
torch.tensor(rollouts.masks[-1, :, agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id, next_value,
args.use_gae, args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents.value_normalizer)
else:
actor_critic[agent_id].eval()
next_value, _, _ = actor_critic[agent_id].get_value(
agent_id,
torch.tensor(rollouts.share_obs[-1, :, agent_id]),
torch.tensor(rollouts.obs[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states[-1, :, agent_id]),
torch.tensor(
rollouts.recurrent_hidden_states_critic[-1, :,
agent_id]),
torch.tensor(rollouts.masks[-1, :, agent_id]))
next_value = next_value.detach().cpu().numpy()
rollouts.compute_returns(agent_id, next_value,
args.use_gae, args.gamma,
args.gae_lambda,
args.use_proper_time_limits,
args.use_popart,
agents[agent_id].value_normalizer)
# update the network
if args.share_policy:
actor_critic.train()
value_loss, action_loss, dist_entropy = agents.update_share(
num_agents, rollouts)
logger.add_scalars('reward', {'reward': np.mean(rollouts.rewards)},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
else:
value_losses = []
action_losses = []
dist_entropies = []
for agent_id in range(num_agents):
actor_critic[agent_id].train()
value_loss, action_loss, dist_entropy = agents[
agent_id].update(agent_id, rollouts)
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
logger.add_scalars(
'agent%i/reward' % agent_id,
{'reward': np.mean(rollouts.rewards[:, :, agent_id])},
(episode + 1) * args.episode_length *
args.n_rollout_threads)
# clean the buffer and reset
rollouts.after_update()
total_num_steps = (episode +
1) * args.episode_length * args.n_rollout_threads
if (episode % args.save_interval == 0 or episode == episodes -
1): # save for every interval-th episode or for the last epoch
if args.share_policy:
torch.save({'model': actor_critic},
str(save_dir) + "/agent_model.pt")
else:
for agent_id in range(num_agents):
torch.save({'model': actor_critic[agent_id]},
str(save_dir) + "/agent%i_model" % agent_id +
".pt")
# log information
if episode % args.log_interval == 0:
end = time.time()
print(
"\n Map {} Algo {} updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(args.map_name, args.algorithm_name, episode, episodes,
total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
if args.share_policy:
print("value loss of agent: " + str(value_loss))
else:
for agent_id in range(num_agents):
print("value loss of agent%i: " % agent_id +
str(value_losses[agent_id]))
if args.env_name == "StarCraft2":
battles_won = []
battles_game = []
incre_battles_won = []
incre_battles_game = []
for i, info in enumerate(infos):
if 'battles_won' in info[0].keys():
battles_won.append(info[0]['battles_won'])
incre_battles_won.append(info[0]['battles_won'] -
last_battles_won[i])
if 'battles_game' in info[0].keys():
battles_game.append(info[0]['battles_game'])
incre_battles_game.append(info[0]['battles_game'] -
last_battles_game[i])
if np.sum(incre_battles_game) > 0:
logger.add_scalars(
'incre_win_rate', {
'incre_win_rate':
np.sum(incre_battles_won) /
np.sum(incre_battles_game)
}, total_num_steps)
else:
logger.add_scalars('incre_win_rate', {'incre_win_rate': 0},
total_num_steps)
last_battles_game = battles_game
last_battles_won = battles_won
if episode % args.eval_interval == 0 and args.eval:
eval_battles_won = 0
eval_episode = 0
eval_obs, eval_available_actions = eval_env.reset()
eval_share_obs = eval_obs.reshape(1, -1)
eval_recurrent_hidden_states = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_recurrent_hidden_states_critic = np.zeros(
(1, num_agents, args.hidden_size)).astype(np.float32)
eval_masks = np.ones((1, num_agents, 1)).astype(np.float32)
while True:
eval_actions = []
for agent_id in range(num_agents):
if args.share_policy:
actor_critic.eval()
_, action, _, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic.act(
agent_id,
torch.tensor(eval_share_obs),
torch.tensor(eval_obs[:, agent_id]),
torch.tensor(
eval_recurrent_hidden_states[:, agent_id]),
torch.tensor(
eval_recurrent_hidden_states_critic[:,
agent_id]),
torch.tensor(eval_masks[:, agent_id]),
torch.tensor(eval_available_actions[:,
agent_id, :]),
deterministic=True)
else:
actor_critic[agent_id].eval()
_, action, _, recurrent_hidden_states, recurrent_hidden_states_critic = actor_critic[
agent_id].act(
agent_id,
torch.tensor(eval_share_obs),
torch.tensor(eval_obs[:, agent_id]),
torch.tensor(
eval_recurrent_hidden_states[:, agent_id]),
torch.tensor(
eval_recurrent_hidden_states_critic[:,
agent_id]
),
torch.tensor(eval_masks[:, agent_id]),
torch.tensor(
eval_available_actions[:, agent_id, :]),
deterministic=True)
eval_actions.append(action.detach().cpu().numpy())
eval_recurrent_hidden_states[:,
agent_id] = recurrent_hidden_states.detach(
).cpu().numpy()
eval_recurrent_hidden_states_critic[:,
agent_id] = recurrent_hidden_states_critic.detach(
).cpu().numpy()
# rearrange action
eval_actions_env = []
for agent_id in range(num_agents):
one_hot_action = np.zeros(
eval_env.action_space[agent_id].n)
one_hot_action[eval_actions[agent_id][0]] = 1
eval_actions_env.append(one_hot_action)
# Obser reward and next obs
eval_obs, eval_rewards, eval_dones, eval_infos, eval_available_actions = eval_env.step(
[eval_actions_env])
eval_share_obs = eval_obs.reshape(1, -1)
if eval_dones[0]:
eval_episode += 1
if eval_infos[0][0]['won']:
eval_battles_won += 1
for agent_id in range(num_agents):
eval_recurrent_hidden_states[0][agent_id] = np.zeros(
args.hidden_size).astype(np.float32)
eval_recurrent_hidden_states_critic[0][
agent_id] = np.zeros(args.hidden_size).astype(
np.float32)
eval_masks[0][agent_id] = 0.0
else:
for agent_id in range(num_agents):
eval_masks[0][agent_id] = 1.0
if eval_episode >= args.eval_episodes:
logger.add_scalars(
'eval_win_rate',
{'eval_win_rate': eval_battles_won / eval_episode},
total_num_steps)
break
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()
envs.close()
if args.eval:
eval_env.close()
def main():
args = get_config()
# cuda
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda:0")
torch.set_num_threads(args.n_training_threads)
if args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
else:
device = torch.device("cpu")
torch.set_num_threads(args.n_training_threads)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
args.reward_randomization = False
# path
model_dir = Path('./results') / args.env_name / args.algorithm_name / (
"run" + str(args.seed))
run_dir = model_dir / 'finetune'
log_dir = run_dir / 'logs'
save_dir = run_dir / 'models'
os.makedirs(str(log_dir))
os.makedirs(str(save_dir))
logger = SummaryWriter(str(log_dir))
# env
envs = make_parallel_env(args)
#Policy network
actor_critic = []
if args.share_policy:
ac = torch.load(str(model_dir / 'models') +
"/agent0_model.pt")['model'].to(device)
for i in range(args.num_agents):
actor_critic.append(ac)
else:
for i in range(args.num_agents):
ac = torch.load(
str(model_dir / 'models') + "/agent%i_model" % i +
".pt")['model'].to(device)
actor_critic.append(ac)
agents = []
rollouts = []
for agent_id in range(args.num_agents):
# algorithm
agent = PPO(actor_critic[agent_id],
agent_id,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.data_chunk_length,
args.value_loss_coef,
args.entropy_coef,
logger,
lr=args.lr,
eps=args.eps,
max_grad_norm=args.max_grad_norm,
use_clipped_value_loss=args.use_clipped_value_loss)
#replay buffer
ro = RolloutStorage(args.num_agents, agent_id, args.episode_length,
args.n_rollout_threads,
envs.observation_space[agent_id],
envs.action_space[agent_id],
actor_critic[agent_id].recurrent_hidden_state_size)
agents.append(agent)
rollouts.append(ro)
# reset env
obs = envs.reset()
# rollout
for i in range(args.num_agents):
if len(envs.observation_space[0]) == 1:
rollouts[i].share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollouts[i].obs[0].copy_(torch.tensor(obs[:, i, :]))
rollouts[i].recurrent_hidden_states.zero_()
rollouts[i].recurrent_hidden_states_critic.zero_()
rollouts[i].recurrent_c_states.zero_()
rollouts[i].recurrent_c_states_critic.zero_()
else:
raise NotImplementedError
rollouts[i].to(device)
# run
coop_num = []
defect_num = []
coopdefect_num = []
defectcoop_num = []
gore1_num = []
gore2_num = []
collective_return = []
apple_consumption = []
waste_cleared = []
sustainability = []
fire = []
start = time.time()
episodes = int(
args.num_env_steps) // args.episode_length // args.n_rollout_threads
all_episode = 0
turn_on = False
for episode in range(episodes):
if (episode > episodes / 2) and (turn_on == False):
print("\n Turn off fixed actor...")
# actor
turn_on = True
if args.use_linear_lr_decay:
# decrease learning rate linearly
for i in range(args.num_agents):
update_linear_schedule(agents[i].optimizer, episode, episodes,
args.lr)
for step in range(args.episode_length):
# Sample actions
values = []
actions = []
action_log_probs = []
recurrent_hidden_statess = []
recurrent_hidden_statess_critic = []
recurrent_c_statess = []
recurrent_c_statess_critic = []
with torch.no_grad():
for i in range(args.num_agents):
value, action, action_log_prob, recurrent_hidden_states, recurrent_hidden_states_critic, recurrent_c_states, recurrent_c_states_critic = actor_critic[
i].act(
rollouts[i].share_obs[step], rollouts[i].obs[step],
rollouts[i].recurrent_hidden_states[step],
rollouts[i].recurrent_hidden_states_critic[step],
rollouts[i].recurrent_c_states[step],
rollouts[i].recurrent_c_states_critic[step],
rollouts[i].masks[step])
values.append(value)
actions.append(action)
action_log_probs.append(action_log_prob)
recurrent_hidden_statess.append(recurrent_hidden_states)
recurrent_hidden_statess_critic.append(
recurrent_hidden_states_critic)
recurrent_c_statess.append(recurrent_c_states)
recurrent_c_statess_critic.append(
recurrent_c_states_critic)
# rearrange action
actions_env = []
for i in range(args.n_rollout_threads):
one_hot_action_env = []
for k in range(args.num_agents):
one_hot_action = np.zeros(envs.action_space[0].n)
one_hot_action[actions[k][i]] = 1
one_hot_action_env.append(one_hot_action)
actions_env.append(one_hot_action_env)
# Obser reward and next obs
obs, reward, done, infos = envs.step(actions_env)
# If done then clean the history of observations.
# insert data in buffer
masks = []
bad_masks = []
masks_critic = []
bad_masks_critic = []
for i in range(args.num_agents):
mask = []
bad_mask = []
for done_ in done:
if done_[i]:
mask.append([0.0])
bad_mask.append([1.0])
else:
mask.append([1.0])
bad_mask.append([1.0])
masks.append(torch.FloatTensor(mask))
bad_masks.append(torch.FloatTensor(bad_mask))
for i in range(args.num_agents):
if len(envs.observation_space[0]) == 1:
rollouts[i].insert(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)),
torch.tensor(obs[:,
i, :]), recurrent_hidden_statess[i],
recurrent_hidden_statess_critic[i],
recurrent_c_statess[i], recurrent_c_statess_critic[i],
actions[i], action_log_probs[i], values[i],
torch.tensor(reward[:, i].reshape(-1, 1)), masks[i],
bad_masks[i])
else:
raise NotImplementedError
with torch.no_grad():
next_values = []
for i in range(args.num_agents):
next_value = actor_critic[i].get_value(
rollouts[i].share_obs[-1], rollouts[i].obs[-1],
rollouts[i].recurrent_hidden_states[-1],
rollouts[i].recurrent_hidden_states_critic[-1],
rollouts[i].recurrent_c_states[-1],
rollouts[i].recurrent_c_states_critic[-1],
rollouts[i].masks[-1]).detach()
next_values.append(next_value)
for i in range(args.num_agents):
rollouts[i].compute_returns(next_values[i], args.use_gae,
args.gamma, args.gae_lambda,
args.use_proper_time_limits)
# update the network
value_losses = []
action_losses = []
dist_entropies = []
for i in range(args.num_agents):
value_loss, action_loss, dist_entropy = agents[i].update(
rollouts[i], turn_on)
value_losses.append(value_loss)
action_losses.append(action_loss)
dist_entropies.append(dist_entropy)
if args.env_name == "StagHunt":
for info in infos:
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'defect&defect_num' in info.keys():
defect_num.append(info['defect&defect_num'])
if 'coop&defect_num' in info.keys():
coopdefect_num.append(info['coop&defect_num'])
if 'defect&coop_num' in info.keys():
defectcoop_num.append(info['defect&coop_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
logger.add_scalars(
'defect&defect_num_per_episode',
{'defect&defect_num_per_episode': defect_num[all_episode]},
all_episode)
logger.add_scalars('coop&defect_num_per_episode', {
'coop&defect_num_per_episode':
coopdefect_num[all_episode]
}, all_episode)
logger.add_scalars('defect&coop_num_per_episode', {
'defect&coop_num_per_episode':
defectcoop_num[all_episode]
}, all_episode)
all_episode += 1
elif args.env_name == "StagHuntGW":
for info in infos:
if 'collective_return' in info.keys():
collective_return.append(info['collective_return'])
if 'coop&coop_num' in info.keys():
coop_num.append(info['coop&coop_num'])
if 'gore1_num' in info.keys():
gore1_num.append(info['gore1_num'])
if 'gore2_num' in info.keys():
gore2_num.append(info['gore2_num'])
for i in range(args.n_rollout_threads):
logger.add_scalars(
'collective_return',
{'collective_return': collective_return[all_episode]},
all_episode)
logger.add_scalars(
'coop&coop_num_per_episode',
{'coop&coop_num_per_episode': coop_num[all_episode]},
all_episode)
logger.add_scalars(
'gore1_num_per_episode',
{'gore1_num_per_episode': gore1_num[all_episode]},
all_episode)
logger.add_scalars(
'gore2_num_per_episode',
{'gore2_num_per_episode': gore2_num[all_episode]},
all_episode)
all_episode += 1
# clean the buffer and reset
obs = envs.reset()
for i in range(args.num_agents):
if len(envs.observation_space[0]) == 1:
rollouts[i].share_obs[0].copy_(
torch.tensor(obs.reshape(args.n_rollout_threads, -1)))
rollouts[i].obs[0].copy_(torch.tensor(obs[:, i, :]))
rollouts[i].recurrent_hidden_states.zero_()
rollouts[i].recurrent_hidden_states_critic.zero_()
rollouts[i].recurrent_c_states.zero_()
rollouts[i].recurrent_c_states_critic.zero_()
else:
raise NotImplementedError
rollouts[i].to(device)
for i in range(args.num_agents):
# save for every interval-th episode or for the last epoch
if (episode % args.save_interval == 0 or episode == episodes - 1):
torch.save({'model': actor_critic[i]},
str(save_dir) + "/agent%i_model" % i + ".pt")
# log information
if episode % args.log_interval == 0:
total_num_steps = (
episode + 1) * args.episode_length * args.n_rollout_threads
end = time.time()
print(
"\n updates {}/{} episodes, total num timesteps {}/{}, FPS {}.\n"
.format(episode, episodes, total_num_steps, args.num_env_steps,
int(total_num_steps / (end - start))))
for i in range(args.num_agents):
print("value loss of agent%i: " % i + str(value_losses[i]))
logger.export_scalars_to_json(str(log_dir / 'summary.json'))
logger.close()