env.seed(args.seed)
"""define actor and critic"""
if args.model_path is None:
if is_disc_action:
policy_net = DiscretePolicy(state_dim, env.action_space.n)
else:
policy_mgr = DiscretePolicy(state_dim, 7)
policy_wrk = Policy(state_dim + subgoal_dim,
env.action_space.shape[0],
log_std=args.log_std)
value_mgr = Value(state_dim)
value_wrk = Value(state_dim + subgoal_dim)
else:
policy_net, value_net, running_state = pickle.load(
open(args.model_path, "rb"))
policy_mgr.to(device)
policy_wrk.to(device)
value_mgr.to(device)
value_wrk.to(device)
# optim_policy_m = torch.optim.Adam(policy_mgr.parameters(), lr=0.01)
# optim_policy_w = torch.optim.Adam(policy_wrk.parameters(), lr=0.01)
# optim_value_m = torch.optim.Adam(value_mgr.parameters(), lr=0.01)
# optim_value_w = torch.optim.Adam(value_wrk.parameters(), lr=0.01)
"""create agent"""
agent = Agent(env,
policy_mgr,
policy_wrk,
device,
running_state=running_state,
render=args.render,
is_disc_action = len(env_dummy.action_space.shape) == 0
running_state = ZFilter((state_dim, ), clip=5)
# running_reward = ZFilter((1,), demean=False, clip=10)
"""define actor and critic"""
if args.model_path is None:
if is_disc_action:
policy_net = DiscretePolicy(state_dim, env_dummy.action_space.n)
else:
policy_net = Policy(state_dim,
env_dummy.action_space.shape[0],
log_std=args.log_std)
value_net = Value(state_dim)
else:
policy_net, value_net, running_state = pickle.load(
open(args.model_path, "rb"))
policy_net.to(device)
value_net.to(device)
del env_dummy
optimizer_policy = torch.optim.Adam(policy_net.parameters(), lr=0.01)
optimizer_value = torch.optim.Adam(value_net.parameters(), lr=0.01)
"""create agent"""
agent = Agent(env_factory,
policy_net,
device,
running_state=running_state,
render=args.render,
num_threads=args.num_threads)
def update_params(batch):
def train(**kwargs):
print('here')
config = {
"lr": kwargs['lr'],
"gamma": kwargs['gamma']
}
dtype = torch.float64
torch.set_default_dtype(dtype)
device = torch.device('cuda', index=args.gpu_index) if torch.cuda.is_available() else torch.device('cpu')
if torch.cuda.is_available():
torch.cuda.set_device(args.gpu_index)
"""environment"""
env = gym.make(args.env_name)
state_dim = env.observation_space.shape[0]
is_disc_action = len(env.action_space.shape) == 0
running_state = ZFilter((state_dim,), clip=5)
# running_reward = ZFilter((1,), demean=False, clip=10)
"""seeding"""
np.random.seed(args.seed)
torch.manual_seed(args.seed)
env.seed(args.seed)
# """define actor and critic"""
if args.model_path is None:
if is_disc_action:
policy_net = DiscretePolicy(state_dim, env.action_space.n)
else:
policy_net = Policy(state_dim, env.action_space.shape[0], log_std=args.log_std)
value_net = Value(state_dim)
else:
policy_net, value_net, running_state = pickle.load(open(args.model_path, "rb"))
policy_net.to(device)
value_net.to(device)
# optimization epoch number and batch size for PPO
optim_epochs = 10
optim_batch_size = 64
"""create agent"""
agent = Agent(env, policy_net, device, running_state=running_state, render=args.render, num_threads=args.num_threads)
def update_params(batch, i_iter, config):
states = torch.from_numpy(np.stack(batch.state)).to(dtype).to(device)
actions = torch.from_numpy(np.stack(batch.action)).to(dtype).to(device)
rewards = torch.from_numpy(np.stack(batch.reward)).to(dtype).to(device)
masks = torch.from_numpy(np.stack(batch.mask)).to(dtype).to(device)
with torch.no_grad():
values = value_net(states)
fixed_log_probs = policy_net.get_log_prob(states, actions)
"""get advantage estimation from the trajectories"""
advantages, returns = estimate_advantages(rewards, masks, values, config['gamma'], args.tau, device)
"""perform mini-batch PPO update"""
optim_iter_num = int(math.ceil(states.shape[0] / optim_batch_size))
for _ in range(optim_epochs):
perm = np.arange(states.shape[0])
np.random.shuffle(perm)
perm = LongTensor(perm).to(device)
states, actions, returns, advantages, fixed_log_probs = \
states[perm].clone(), actions[perm].clone(), returns[perm].clone(), advantages[perm].clone(), fixed_log_probs[perm].clone()
for i in range(optim_iter_num):
ind = slice(i * optim_batch_size, min((i + 1) * optim_batch_size, states.shape[0]))
states_b, actions_b, advantages_b, returns_b, fixed_log_probs_b = \
states[ind], actions[ind], advantages[ind], returns[ind], fixed_log_probs[ind]
ppo_step(policy_net, value_net, optimizer_policy, optimizer_value, 1, states_b, actions_b, returns_b,
advantages_b, fixed_log_probs_b, args.clip_epsilon, args.l2_reg)
def main_loop(config):
optimizer_policy = torch.optim.Adam(policy_net.parameters(), lr=config['lr'])
optimizer_value = torch.optim.Adam(value_net.parameters(), lr=config['lr'])
for i_iter in range(args.max_iter_num):
"""generate multiple trajectories that reach the minimum batch_size"""
batch, log = agent.collect_samples(args.min_batch_size)
t0 = time.time()
update_params(batch, i_iter, config)
t1 = time.time()
if i_iter % args.log_interval == 0:
print('{}\tT_sample {:.4f}\tT_update {:.4f}\tR_min {:.2f}\tR_max {:.2f}\tR_avg {:.2f}'.format(
i_iter, log['sample_time'], t1-t0, log['min_reward'], log['max_reward'], log['avg_reward']))
if args.save_model_interval > 0 and (i_iter+1) % args.save_model_interval == 0:
to_device(torch.device('cpu'), policy_net, value_net)
pickle.dump((policy_net, value_net, running_state),
open(os.path.join(assets_dir(), 'learned_models/{}_ppo.p'.format(args.env_name)), 'wb'))
to_device(device, policy_net, value_net)
# """clean up gpu memory"""
torch.cuda.empty_cache()
return agent.evaluate()
print('a')
print(config)
print(args)
return main_loop(config)
