def slave():
global env, test_env
env = make_env(args=config_args, dream_env=config_args.dream_env)
# doom env doesn't support mpi testing so don't bother loading
if 'DoomTakeCover-v0' != config_args.env_name:
test_env = make_env(args=config_args, dream_env=False, render_mode=False)
packet = np.empty(SOLUTION_PACKET_SIZE, dtype=np.int32)
while 1:
comm.Recv(packet, source=0)
assert(len(packet) == SOLUTION_PACKET_SIZE)
solutions = decode_solution_packet(packet)
results = []
for solution in solutions:
worker_id, jobidx, seed, train_mode, max_len, weights = solution
assert (train_mode == 1 or train_mode == 0 or train_mode == -1), str(train_mode)
worker_id = int(worker_id)
possible_error = "work_id = " + str(worker_id) + " rank = " + str(rank)
assert worker_id == rank, possible_error
jobidx = int(jobidx)
seed = int(seed)
fitness, timesteps = worker(weights, seed, train_mode, max_len)
results.append([worker_id, jobidx, fitness, timesteps])
result_packet = encode_result_packet(results)
assert len(result_packet) == RESULT_PACKET_SIZE
comm.Send(result_packet, dest=0)
def slave():
global env
if env_name == 'CarRacing-v0':
env = make_env(args=config_args, dream_env=False,
with_obs=True) # training in dreams not supported yet
else:
env = make_env(args=config_args, dream_env=True, render_mode=False)
packet = np.empty(SOLUTION_PACKET_SIZE, dtype=np.int32)
while 1:
comm.Recv(packet, source=0)
assert (len(packet) == SOLUTION_PACKET_SIZE)
solutions = decode_solution_packet(packet)
results = []
for solution in solutions:
worker_id, jobidx, seed, train_mode, max_len, weights = solution
assert (train_mode == 1 or train_mode == 0), str(train_mode)
worker_id = int(worker_id)
possible_error = "work_id = " + str(worker_id) + " rank = " + str(
rank)
assert worker_id == rank, possible_error
jobidx = int(jobidx)
seed = int(seed)
fitness, timesteps = worker(weights, seed, train_mode, max_len)
results.append([worker_id, jobidx, fitness, timesteps])
result_packet = encode_result_packet(results)
assert len(result_packet) == RESULT_PACKET_SIZE
comm.Send(result_packet, dest=0)
def main():
args = get_argparser().parse_args()
init_logging('logs')
env = make_env(args.env, args.seed, num_envs=args.num_envs, num_processes=args.num_processes)
agent = ActorCritic(env.observation_space, env.action_space, args)
train(agent, env, args, max_reward=args.max_reward)
test_env = make_env(args.env, args.seed, num_envs=1, num_processes=1)
make_fun(agent, test_env, render=True)
def initialize_settings(sigma_init=0.1, sigma_decay=0.9999):
global population, filebase, game, controller, env, num_params, es, PRECISION, SOLUTION_PACKET_SIZE, RESULT_PACKET_SIZE
population = num_worker * num_worker_trial
filebase = 'log/' + gamename + '.' + optimizer + '.' + str(
num_episode) + '.' + str(population)
controller = make_model()
env = make_env()
num_params = controller.param_count
print("size of model", num_params)
if optimizer == 'ses':
ses = PEPG(num_params,
sigma_init=sigma_init,
sigma_decay=sigma_decay,
sigma_alpha=0.2,
sigma_limit=0.02,
elite_ratio=0.1,
weight_decay=0.005,
popsize=population)
es = ses
elif optimizer == 'ga':
ga = SimpleGA(num_params,
sigma_init=sigma_init,
sigma_decay=sigma_decay,
sigma_limit=0.02,
elite_ratio=0.1,
weight_decay=0.005,
popsize=population)
es = ga
elif optimizer == 'cma':
cma = CMAES(num_params, sigma_init=sigma_init, popsize=population)
es = cma
elif optimizer == 'pepg':
pepg = PEPG(num_params,
sigma_init=sigma_init,
sigma_decay=sigma_decay,
sigma_alpha=0.20,
sigma_limit=0.02,
learning_rate=0.01,
learning_rate_decay=1.0,
learning_rate_limit=0.01,
weight_decay=0.005,
popsize=population)
es = pepg
else:
oes = OpenES(num_params,
sigma_init=sigma_init,
sigma_decay=sigma_decay,
sigma_limit=0.02,
learning_rate=0.01,
learning_rate_decay=1.0,
learning_rate_limit=0.01,
antithetic=antithetic,
weight_decay=0.005,
popsize=population)
es = oes
PRECISION = 10000
SOLUTION_PACKET_SIZE = (5 + num_params) * num_worker_trial
RESULT_PACKET_SIZE = 4 * num_worker_trial
def make_env(self, seed=-1, render_mode=False, load_model=True, lock=None):
self.render_mode = render_mode
self.env = make_env(self.env_name,
seed=seed,
render_mode=render_mode,
load_model=load_model,
lock=lock)
def make_env(self, seed=-1, render_mode=False):
self.render_mode = render_mode
self.env = make_env(self.env_name,
self.encoder,
self.max_features,
seed=seed,
render_mode=render_mode)
def __init__(self, **kwarg):
"""
Args:
kwarg: configurations for the environment.
"""
config = get_default_config()
name = kwarg['name']
for key, value in kwarg.items():
if hasattr(config, key):
setattr(config, key, value)
# create an environment
self.env = make_env(name, config)
# covert observation space
obs_space = self.env.observation_space
obs_size = sum([np.prod(v) for v in obs_space.values()])
low = -1 * np.ones(obs_size)
high = np.ones(obs_size)
self.observation_space = gym.spaces.Box(low=low, high=high)
# covert action space
dof = self.env.dof
low = -1 * np.ones(dof)
high = np.ones(dof)
self.action_space = gym.spaces.Box(low=low, high=high)
def run(args):
# env = make_env(args.env_id)
env = gym.make(env_id)
env_test = make_env(args.env_id)
buffer_exp = SerializedBuffer(
path=args.buffer, device=torch.device("cuda" if args.cuda else "cpu"))
algo = AIRL(buffer_exp=buffer_exp,
state_shape=env.observation_space.shape,
action_shape=env.action_space.shape,
device=torch.device("cuda" if args.cuda else "cpu"),
seed=args.seed,
rollout_length=args.rollout_length)
time = datetime.now().strftime("%Y%m%d-%H%M")
log_dir = os.path.join('logs', args.env_id, args.algo,
f'seed{args.seed}-{time}')
trainer = Trainer(env=env,
env_test=env_test,
algo=algo,
log_dir=log_dir,
num_steps=args.num_steps,
eval_interval=args.eval_interval,
seed=args.seed)
trainer.train()
def make_env(self, env_name, seed=-1, render_mode=False, model=None):
self.render_mode = render_mode
self.env_name = env_name
self.env = make_env(env_name,
seed=seed,
render_mode=render_mode,
model=model)
def __init__(self, env_id, lr, nstep, batch_size, n_epochs, gamma, gae_lam,
clip_range, ent_coef, vf_coef, max_grad_norm):
self.env_id = env_id
self.env = make_env(env_id, n_envs=4)
self.num_envs = self.env.num_envs if isinstance(self.env,
VecEnv) else 1
self.state_dim = self.env.observation_space.shape[0]
self.action_converter = ActionConverter(self.env.action_space)
self.lr = lr
self.nstep = nstep
self.batch_size = batch_size
self.n_epochs = n_epochs
self.gamma = gamma
self.gae_lam = gae_lam
self.clip_range = clip_range
self.ent_coef = ent_coef
self.vf_coef = vf_coef
self.max_grad_norm = max_grad_norm
self.ep_info_buffer = deque(maxlen=50)
self._n_updates = 0
self.num_timesteps = 0
self.num_episodes = 0
self.obs_rms = RunningMeanStd()
def make_env(self,
seed=-1,
render_mode=False,
full_episode=False,
worker_id=0):
self.render_mode = render_mode
self.env = make_env(self.env_name,
seed=seed,
render_mode=render_mode,
full_episode=full_episode,
worker_id=worker_id)
def main(args):
env_name = args.env_name
total_episodes = args.total_episodes
start_batch = args.start_batch
time_steps = args.time_steps
obs_data = []
action_data = []
env = make_env(env_name)
s = 0
batch = start_batch
while s < total_episodes:
for i_episode in range(200):
print('-----')
observation = env.reset()
env.render()
done = False
action = env.action_space.sample()
t = 0
obs_sequence = []
action_sequence = []
while t < time_steps:
t = t + 1
action = config.generate_data_action(t, action)
observation = config.adjust_obs(observation)
obs_sequence.append(observation)
action_sequence.append(action)
observation, reward, done, info = env.step(action)
obs_data.append(obs_sequence)
action_data.append(action_sequence)
print("Batch {} Episode {} finished after {} timesteps".format(batch, i_episode, t+1))
print("Current dataset contains {} observations".format(sum(map(len, obs_data))))
s = s + 1
print("Saving dataset for batch {}".format(batch))
np.save('./data/obs_data_' + str(batch), obs_data)
np.save('./data/action_data_' + str(batch), action_data)
batch = batch + 1
obs_data = []
action_data = []
def test_render():
env = make_env(10000, visualize=True)
obs = env.reset()
#while True:
# env.render()
zero_action = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]])
#start_time = time.time()
new_obs, rew, done, _ = env.step(zero_action)
action = np.array([[0.0, 0.0, 4.0], [0.0, 0.0, 0.0]])
while True:
new_obs, rew, done, _ = env.step(action)
env.render()
if done:
break
def init_gym(env_name):
"""
Initialize gym environment, return dimension of observation
and action spaces.
Args:
env_name: str environment name (e.g. "Humanoid-v1")
Returns: 3-tuple
gym environment (object)
number of observation dimensions (int)
number of action dimensions (int)
"""
#env = gym.make(env_name)
env = make_env(env_name)
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
return env, obs_dim, act_dim
def __init__(self, type="CarRacing", history_pick=4, seed=None, detect_edges=False, detect_grass=False, flip=False):
self.name = type + str(time.time())
random.seed(30)
self.env = make_env('CarRacing-v0', random.randint(1,10000000), render_mode = False, full_episode = True)
self.image_dimension = [64,64]
self.history_pick = history_pick
self.state_space_size = history_pick * np.prod(self.image_dimension)
self.action_space_size = 5
self.state_shape = [None, self.history_pick] + list(self.image_dimension)
self.history = []
self.action_dict = {0: [-1, 0, 0], 1: [1, 0, 0], 2: [0, 1, 0], 3: [0, 0, 0.8], 4: [0, 0, 0]}
self.seed = seed
self.detect_edges = detect_edges
self.detect_grass = detect_grass
self.flip = flip
self.flip_episode = False
self.vae = ConvVAE(batch_size=1, gpu_mode=False, is_training=False, reuse=True)
self.rnn = MDNRNN(hps_sample, gpu_mode=False, reuse=True)
self.vae.load_json('vae/vae.json')
self.rnn.load_json('rnn/rnn.json')
def test():
env = make_env()
obs = env.reset()
# Test zero action
zero_action = np.array([[0.0, 0.0, 0.0], [0.0, 0.0, 0.0]])
print(f"Before zero action = {obs}")
new_obs, rew, done, _ = env.step(zero_action)
print(f"After zero action = {new_obs}, is equal = {obs == new_obs}")
obs = new_obs
# Test dynamics
dt = 0.002
action = np.array([[1.0, 1.0, 1.0], [0.0, 0.0, 0.0]])
new_pos = deepcopy(obs)
new_pos[0] = obs[0] + (1000 * action[0, 0] * np.cos(obs[2]) +
1000 * action[0, 1] * np.sin(obs[2])) * dt
new_pos[1] = obs[1] + (1000 * action[0, 1] * np.cos(obs[2]) -
1000 * action[0, 0] * np.sin(obs[2])) * dt
new_pos[2] = obs[2] + action[0, 2] * dt
new_obs, rew, done, _ = env.step(action)
print(f"Is done = {done}, Is dynamics correct = {new_pos == new_obs}")
def main(_):
display = Display(visible=0, size=(1400, 900))
#display.start()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session() as sess:
global_step = tf.Variable(0, name='global_step', trainable=False)
env = make_env(ENV_NAME, 876, render_mode=False, full_episode=True)
np.random.seed(RANDOM_SEED)
tf.set_random_seed(RANDOM_SEED)
env.seed(RANDOM_SEED)
state_dim = [32]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
print('state_dim: ', state_dim)
print('action_dim: ', action_dim)
print('action_bound: ', action_bound)
# Ensure action bound is symmetric
# assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound, TAU)
critic = CriticNetwork(sess, state_dim, action_dim, TAU,
actor.get_num_trainable_vars())
if GYM_MONITOR_EN:
if not RENDER_ENV:
env.monitor.start(MONITOR_DIR,
video_callable=False,
force=True)
else:
env.monitor.start(MONITOR_DIR, force=True)
train(sess, env, actor, critic, global_step)
if GYM_MONITOR_EN:
env.monitor.close()
def slave():
env = make_env()
packet = np.empty(SOLUTION_PACKET_SIZE, dtype=np.int32)
while 1:
comm.Recv(packet, source=0)
assert (len(packet) == SOLUTION_PACKET_SIZE)
solutions = decode_solution_packet(packet)
results = []
for solution in solutions:
worker_id, jobidx, seed, train_mode, max_len, weights = solution
assert (train_mode == 1 or train_mode == 0), str(train_mode)
worker_id = int(worker_id)
possible_error = "work_id = " + str(worker_id) + " rank = " + str(
rank)
assert worker_id == rank, possible_error
jobidx = int(jobidx)
seed = int(seed)
fitness, timesteps = worker(weights, seed, train_mode, max_len)
results.append([worker_id, jobidx, fitness, timesteps])
result_packet = encode_result_packet(results)
assert len(result_packet) == RESULT_PACKET_SIZE
comm.Send(result_packet, dest=0)
def main(args):
exp_path = mkdir_exp(f'{args.env_id}_PPO')
export_args(args, os.path.join(exp_path, 'config.json'))
np.random.seed(args.seed)
pt.random.manual_seed(args.seed)
print("== Creating a training environment...")
env = make_env(args.env_id, NormalizeObservation, num_envs=args.num_envs)
print("== Creating a evaluation environment...")
eval_env = make_env(args.env_id, NormalizeObservation, num_envs=1)
obs_dim = eval_env.observation_space.shape[0]
act_dim = eval_env.action_space.shape[0]
print("== Creating an agent....")
device = pt.device('cuda' if pt.cuda.is_available() else 'cpu')
agent = ContinuousPolicyAgent(obs_dim, act_dim, args.hid_dim).to(device)
print("== Creating a data storage...")
data = TensorBook(args.env_id, args.rollout_steps)
print("== Creating a PPO optimizer...")
optimizer = ProximalPolicyOptimization(
agent,
device,
num_epochs=args.num_epochs,
batch_size=args.batch_size,
lr_max=args.lr_max,
lr_min=args.lr_min,
eps=args.eps,
gamma=args.gamma,
lam=args.lam,
alpha=args.alpha,
value_coef=args.value_coef,
entropy_coef=args.entropy_coef,
max_grad_norm=args.max_grad_norm,
target_kldiv=args.target_kldiv
)
print("== Creating a TensorBoard summary writer...")
writer = SummaryWriter(log_dir=exp_path)
print("IT'S DANGEROUS TO GO ALONE! TAKE THIS.")
obs = env.reset().to(device)
best_perf = -np.inf
num_updates = args.num_steps // args.rollout_steps // args.num_envs
for i in tqdm(range(num_updates)):
obs = agent.rollout(obs, env, data)
info = optimizer.update(data)
lr = optimizer.update_lr(i, num_updates)
# Compute mean total reward during the rollout.
reward = data.reward.sum(dim=0).mean(dim=0).item()
# Evaluate the agent.
perf = play(eval_env, agent, device, repeat=args.num_eval)
if perf > best_perf:
model_path = os.path.join(exp_path, f'{agent.__class__.__name__}.pt')
pt.save(agent.state_dict(), model_path)
best_perf = perf
# Log training progress.
step = i * args.rollout_steps * args.num_envs
writer.add_scalar('Train/lr', lr, step)
writer.add_scalar('Train/epochs', info['num_epochs'], step)
writer.add_scalar('Train/loss/policy', info['policy_loss'], step)
writer.add_scalar('Train/loss/value', info['value_loss'], step)
writer.add_scalar('Train/loss/entropy', info['entropy'], step)
writer.add_scalar('Train/loss/total', info['total_loss'], step)
writer.add_scalar('Train/reward/mean', reward, step)
writer.add_scalar('Eval/reward/mean', perf, step)
writer.add_scalar('Eval/reward/best', best_perf, step)
env.close()
eval_env.close()
writer.close()
def master():
global test_env
if env_name == 'CarRacing-v0':
test_env = make_env(args=config_args, dream_env=False, with_obs=True)
else:
test_env = make_env(args=config_args,
dream_env=False,
render_mode=False)
start_time = int(time.time())
sprint("training", env_name)
sprint("population", es.popsize)
sprint("num_worker", num_worker)
sprint("num_worker_trial", num_worker_trial)
sys.stdout.flush()
seeder = Seeder(seed_start)
filename = filebase + '.json'
filename_log = filebase + '.log.json'
filename_hist = filebase + '.hist.json'
filename_eval_hist = filebase + '.eval_hist.json'
filename_hist_best = filebase + '.hist_best.json'
filename_best = filebase + '.best.json'
t = 0
history = []
history_best = [] # stores evaluation averages every 25 steps or so
eval_log = []
eval_hist = []
best_reward_eval = 0
best_model_params_eval = None
max_len = -1 # max time steps (-1 means ignore)
while True:
solutions = es.ask()
if antithetic:
seeds = seeder.next_batch(int(es.popsize / 2))
seeds = seeds + seeds
else:
seeds = seeder.next_batch(es.popsize)
packet_list = encode_solution_packets(seeds,
solutions,
max_len=max_len)
send_packets_to_slaves(packet_list)
reward_list_total = receive_packets_from_slaves()
reward_list = reward_list_total[:, 0] # get rewards
mean_time_step = int(np.mean(reward_list_total[:, 1]) *
100) / 100. # get average time step
max_time_step = int(np.max(reward_list_total[:, 1]) *
100) / 100. # get average time step
avg_reward = int(
np.mean(reward_list) * 100) / 100. # get average time step
std_reward = int(
np.std(reward_list) * 100) / 100. # get average time step
es.tell(reward_list)
es_solution = es.result()
model_params = es_solution[0] # best historical solution
reward = es_solution[1] # best reward
curr_reward = es_solution[2] # best of the current batch
controller.set_model_params(np.array(model_params).round(4))
r_max = int(np.max(reward_list) * 100) / 100.
r_min = int(np.min(reward_list) * 100) / 100.
curr_time = int(time.time()) - start_time
h = (t, curr_time, avg_reward, r_min, r_max, std_reward,
int(es.rms_stdev() * 100000) / 100000., mean_time_step + 1.,
int(max_time_step) + 1)
if cap_time_mode:
max_len = 2 * int(mean_time_step + 1.0)
else:
max_len = -1
history.append(h)
with open(filename, 'wt') as out:
res = json.dump([np.array(es.current_param()).round(4).tolist()],
out,
sort_keys=True,
indent=2,
separators=(',', ': '))
with open(filename_hist, 'wt') as out:
res = json.dump(history,
out,
sort_keys=False,
indent=0,
separators=(',', ':'))
sprint(env_name, h)
if (t == 1):
best_reward_eval = avg_reward
if (t % eval_steps == 0): # evaluate on actual task at hand
prev_best_reward_eval = best_reward_eval
model_params_quantized = np.array(es.current_param()).round(4)
reward_eval_list = evaluate_batch(model_params_quantized,
max_len=-1,
test_seed=t)
reward_eval = np.mean(reward_eval_list)
r_eval_std = np.std(reward_eval_list)
r_eval_min = np.min(reward_eval_list)
r_eval_max = np.max(reward_eval_list)
model_params_quantized = model_params_quantized.tolist()
improvement = reward_eval - best_reward_eval
eval_log.append([t, reward_eval, model_params_quantized])
e_h = (t, reward_eval, r_eval_std, r_eval_min, r_eval_max)
eval_hist.append(e_h)
with open(filename_eval_hist, 'wt') as out:
res = json.dump(eval_hist,
out,
sort_keys=False,
indent=0,
separators=(',', ':'))
with open(filename_log, 'wt') as out:
res = json.dump(eval_log, out)
if (len(eval_log) == 1 or reward_eval > best_reward_eval):
best_reward_eval = reward_eval
best_model_params_eval = model_params_quantized
else:
if retrain_mode:
sprint(
"reset to previous best params, where best_reward_eval =",
best_reward_eval)
es.set_mu(best_model_params_eval)
with open(filename_best, 'wt') as out:
res = json.dump([best_model_params_eval, best_reward_eval],
out,
sort_keys=True,
indent=0,
separators=(',', ': '))
# dump history of best
curr_time = int(time.time()) - start_time
best_record = [
t, curr_time, "improvement", improvement, "curr", reward_eval,
"prev", prev_best_reward_eval, "best", best_reward_eval
]
history_best.append(best_record)
with open(filename_hist_best, 'wt') as out:
res = json.dump(history_best,
out,
sort_keys=False,
indent=0,
separators=(',', ':'))
sprint("Eval", t, curr_time, "improvement", improvement, "curr",
reward_eval, "prev", prev_best_reward_eval, "best",
best_reward_eval)
# increment generation
t += 1
global_env = make_env(
[
("[]", (TYPE_NIL,)),
("true", (TYPE_BOOL, True)),
("false", (TYPE_BOOL, False)),
("cons", (TYPE_BUILTIN_FUNCTION, tagged_cons)),
("head", (TYPE_BUILTIN_FUNCTION, tagged_head)),
("tail", (TYPE_BUILTIN_FUNCTION, tagged_tail)),
("+", num_op(lambda x, y: x + y)),
("*", num_op(lambda x, y: x * y)),
("-", num_op(lambda x, y: x - y)),
("/", num_op(lambda x, y: x // y)),
("or", bool_op(lambda x, y: x or y)),
("and", bool_op(lambda x, y: x and y)),
("xor", bool_op(lambda x, y: x ^ y)),
("not", (TYPE_BUILTIN_FUNCTION, neg)),
("<", comp_op(lambda x, y: x < y)),
(">", comp_op(lambda x, y: x > y)),
("<=", comp_op(lambda x, y: x <= y)),
(">=", comp_op(lambda x, y: x >= y)),
("=", (TYPE_BUILTIN_FUNCTION, equal)),
("!=", (TYPE_BUILTIN_FUNCTION, unequal)),
("int_of_string", (TYPE_BUILTIN_FUNCTION, int_of_string)),
("string_of_int", (TYPE_BUILTIN_FUNCTION, string_of_int)),
("int_of_char", (TYPE_BUILTIN_FUNCTION, int_of_char)),
("char_of_int", (TYPE_BUILTIN_FUNCTION, char_of_int)),
("error", (TYPE_BUILTIN_FUNCTION, error)),
("concat", (TYPE_BUILTIN_FUNCTION, concat)),
]
)
def make_env(self, seed=-1, render_mode=False):
self.render_mode = render_mode
self.env = make_env(self.env_name, seed=seed, render_mode=render_mode)
def main(args):
"""
Inputs type of agent, observation types and simulates the environment.
"""
print("The observation tutorial will show you the various observation configurations available.")
background_name = background_names[1]
# load demo file for playback
demo = args.load_demo = input('Input path to demo file, such as demos/Sawyer_7.pkl: ')
if demo == '':
demo = args.load_demo = 'demos/Sawyer_7.pkl'
agent_name, furniture_id = demo.split('/')[-1].split('.')[0].split('_')
agent_name = agent_name[0].upper() + agent_name[1:]
furniture_id = int(furniture_id)
furniture_name = furniture_names[furniture_id]
# choose robot observation
print()
print("Include robot observation?\n")
try:
s = input("Put 1 for True or 0 for False: ")
k = int(s) == 1
except:
print("Input is not valid. Use 0 by default.")
k = False
args.robot_ob = k
# choose furniture observation
print()
print("Include furniture observation?\n")
try:
s = input("Put 1 for True or 0 for False: ")
k = int(s) == 1
except:
print("Input is not valid. Use 0 by default.")
k = False
args.object_ob = k
# choose segmentation
print()
print("Use segmentation?\n")
try:
s = input("Put 1 for True or 0 for False: ")
k = int(s) == 1
except:
print("Input is not valid. Use 0 by default.")
k = False
use_seg = k
# choose depth
print()
print("Use depth map?\n")
try:
s = input("Put 1 for True or 0 for False: ")
k = int(s) == 1
except:
print("Input is not valid. Use 0 by default.")
k = False
use_depth = k
# set parameters for the environment (env, furniture_id, background)
env_name = 'Furniture{}Env'.format(agent_name)
args.env = env_name
args.furniture_id = furniture_id
args.background = background_name
print()
print("Creating environment (robot: {}, furniture: {}, background: {})".format(
env_name, furniture_name, background_name))
# make environment with rgb, depth map, and segmentation
args.depth_ob = True
args.segmentation_ob = True
# make environment following arguments
env = make_env(env_name, args)
ob = env.reset(args.furniture_id, args.background)
# tell user about environment observation space
print('-' * 80)
print('Observation configuration:')
print(f"Robot ob: {args.robot_ob}, Furniture ob: {args.object_ob}")
print(f"Depth Map: {use_depth}, Segmentation Map: {use_seg}")
print()
print("Observation Space:\n")
print("The observation space is a dictionary. For furniture (object) observations, it is "+
"a multiple of 7 because each part has 3 dims for position and 4 dims for quaternion. "+
"The robot_ob is dependent on the agent, and contains position, velocity, or angles of "+
"the current robot.\n")
print(env.observation_space)
print()
input("Type anything to record an episode's visual observations")
# run the trajectory, save the video
rgb_frames = []
depth_frames = []
seg_frames = []
# load demo from pickle file
with open(env._load_demo, 'rb') as f:
demo = pickle.load(f)
all_qpos = demo['qpos']
# playback first 100 frames
for qpos in all_qpos:
# set furniture part positions
for i, body in enumerate(env._object_names):
pos = qpos[body][:3]
quat = qpos[body][3:]
env._set_qpos(body, pos, quat)
env._stop_object(body, gravity=0)
# set robot positions
if env._agent_type == 'Sawyer':
env.sim.data.qpos[env._ref_joint_pos_indexes] = qpos['sawyer_qpos']
env.sim.data.qpos[env._ref_gripper_joint_pos_indexes] = qpos['l_gripper']
elif env._agent_type == 'Baxter':
env.sim.data.qpos[env._ref_joint_pos_indexes] = qpos['baxter_qpos']
env.sim.data.qpos[env._ref_gripper_right_joint_pos_indexes] = qpos['r_gripper']
env.sim.data.qpos[env._ref_gripper_left_joint_pos_indexes] = qpos['l_gripper']
elif env._agent_type == 'Cursor':
env._set_pos('cursor0', qpos['cursor0'])
env._set_pos('cursor1', qpos['cursor1'])
env.sim.forward()
env._update_unity()
img, depth = env.render('rgbd_array')
seg = I.color_segmentation(env.render('segmentation'))
rgb_frames.append(img)
depth_frames.append(depth)
seg_frames.append(seg)
env.close()
# concatenate available observation frames together and render video
wide_frames = []
L = max(len(rgb_frames), len(rgb_frames), len(seg_frames))
for l in range(L):
rgb = rgb_frames[l]
f = [rgb * 255]
if use_depth:
depth = depth_frames[l]
f.append(depth * 255)
if use_seg:
seg = seg_frames[l]
f.append(seg)
wide = np.concatenate(f, axis=1)
wide_frames.append(wide)
vr = VideoRecorder()
vr._frames = wide_frames
vr.save_video('observations.mp4')
def main(args):
print("main")
env_name = args.env_name
total_episodes = args.total_episodes
start_batch = args.start_batch
time_steps = args.time_steps
render = args.render
batch_size = args.batch_size
run_all_envs = args.run_all_envs
store_folder = args.store_folder
if not os.path.exists(store_folder):
os.makedirs(store_folder)
if run_all_envs:
envs_to_generate = config.train_envs
else:
envs_to_generate = [env_name]
print("envs:", envs_to_generate)
for current_env_name in envs_to_generate:
print("Generating data for env {}".format(current_env_name))
env = make_env(current_env_name)
s = 0
batch = start_batch
batch_size = min(batch_size, total_episodes)
total_frames = 0
while s < total_episodes:
obs_data = []
action_data = []
for i_episode in range(batch_size):
print('-----')
observation = env._reset()
#observation = config.adjust_obs(observation)
# plt.imshow(observation)
# plt.show()
env.render()
done = False
action = np.random.rand() *2.0 -1.0
t = 0
obs_sequence = []
action_sequence = []
repeat = np.random.randint(1, 11)
while t < time_steps: # and not done:
t = t + 1
if t % repeat == 0:
action = np.random.rand() * 2.0 - 1.0
repeat = np.random.randint(1, 11)
obs_sequence.append(observation)
action_sequence.append(action)
observation, reward, done, info = env._step(action)
if render:
env.render()
if done: #If we were killed
break
total_frames += t
print("dead at", t, "total recorded frames for this worker", total_frames)
obs_data.append(obs_sequence)
action_data.append(action_sequence)
print("Batch {} Episode {} finished after {} timesteps".format(batch, i_episode, t + 1))
print("Current dataset contains {} observations".format(sum(map(len, obs_data))))
s = s + 1
print("Saving dataset for batch {}".format(batch))
np.save(store_folder+'/obs_data_' + current_env_name + '_' + str(batch), obs_data)
print("Saving actions for batch {}".format(batch))
np.save(store_folder+'/action_data_' + current_env_name + '_' + str(batch), action_data)
batch = batch + 1
env.close()
def make_env(self, env_name, seed=-1, render_mode=False): self.render_mode = render_mode self.env_name = env_name self.env = make_env(env_name, seed=seed, render_mode=render_mode)
def _make_env(self):
self.render_mode = render_mode
self.env = make_env(self.env_name)
self.num_actions = self.env.action_space.n
def ddpg(env_config, ac_type, ac_kwargs, rb_type, rb_kwargs, gamma, lr, polyak,
batch_size, epochs, start_steps, steps_per_epoch, inc_ep, max_ep_len,
test_max_ep_len, number_of_tests_per_epoch, act_noise, logger_kwargs,
seed):
logger = EpochLogger(**logger_kwargs)
configs = locals().copy()
configs.pop("logger")
logger.save_config(configs)
tf.set_random_seed(seed)
np.random.seed(seed)
env, test_env = make_env(env_config), make_env(env_config)
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
# Action limit for clamping: critically, assumes all dimensions share the same bound!
act_high = env.action_space.high
# Inputs to computation graph
x_ph, a_ph, x2_ph, r_ph, d_ph = core.placeholders(obs_dim, act_dim,
obs_dim, None, None)
actor_critic = core.get_ddpg_actor_critic(ac_type)
# Main outputs from computation graph
with tf.variable_scope('main'):
pi, q, q_pi = actor_critic(x_ph, a_ph, **ac_kwargs)
# Target networks
with tf.variable_scope('target'):
pi_targ, _, q_pi_targ = actor_critic(x2_ph, a_ph, **ac_kwargs)
# Experience buffer
RB = get_replay_buffer(rb_type)
replay_buffer = RB(obs_dim, act_dim, **rb_kwargs)
# Count variables
var_counts = tuple(
core.count_vars(scope) for scope in ['main/pi', 'main/q', 'main'])
print('\nNumber of parameters: \t pi: %d, \t q: %d, \t total: %d\n' %
var_counts)
# Bellman backup for Q function
backup = tf.stop_gradient(r_ph + gamma * (1 - d_ph) * q_pi_targ)
# DDPG losses
pi_loss = -tf.reduce_mean(q_pi)
q_loss = tf.reduce_mean((q - backup)**2)
# Separate train ops for pi, q
pi_optimizer = tf.train.AdamOptimizer(learning_rate=lr)
q_optimizer = tf.train.AdamOptimizer(learning_rate=lr)
train_pi_op = pi_optimizer.minimize(pi_loss, var_list=get_vars('main/pi'))
train_q_op = q_optimizer.minimize(q_loss, var_list=get_vars('main/q'))
# Polyak averaging for target variables
target_update = tf.group([
tf.assign(v_targ, polyak * v_targ + (1 - polyak) * v_main)
for v_main, v_targ in zip(get_vars('main'), get_vars('target'))
])
# Initializing targets to match main variables
target_init = tf.group([
tf.assign(v_targ, v_main)
for v_main, v_targ in zip(get_vars('main'), get_vars('target'))
])
sess = tf.Session()
sess.run(tf.global_variables_initializer())
sess.run(target_init)
def get_action(o, noise_scale):
pi_a = sess.run(pi, feed_dict={x_ph: o.reshape(1, -1)})[0]
pi_a += noise_scale * np.random.randn(act_dim)
pi_a = np.clip(pi_a, 0, 1)
real_a = pi_a * act_high
return pi_a, real_a
def test_agent(n=10):
test_actions = []
for j in range(n):
test_actions_ep = []
o, r, d, ep_ret, ep_len = test_env.reset(), 0, False, 0, 0
while not (d or (ep_len == test_max_ep_len)):
# Take deterministic actions at test time (noise_scale=0)
_, real_a = get_action(o, 0)
test_actions_ep.append(real_a)
o, r, d, _ = test_env.step(real_a)
ep_ret += r
ep_len += 1
logger.store(TestEpRet=ep_ret, TestEpLen=ep_len)
test_actions.append(test_actions_ep)
return test_actions
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
total_steps = steps_per_epoch * epochs
actions = []
epoch_actions = []
rewards = []
rets = []
test_rets = []
max_ret = None
# Main loop: collect experience in env and update/log each epoch
for t in range(total_steps):
"""
Until start_steps have elapsed, randomly sample actions
from a uniform distribution for better exploration. Afterwards,
use the learned policy (with some noise, via act_noise).
"""
if t > start_steps:
pi_a, real_a = get_action(o, act_noise)
else:
pi_a, real_a = env.action_space.sample()
# Step the env
o2, r, d, _ = env.step(real_a)
ep_ret += r
ep_len += 1
epoch_actions.append(pi_a)
# Ignore the "done" signal if it comes from hitting the time
# horizon (that is, when it's an artificial terminal signal
# that isn't based on the agent's state)
d = False if ep_len == max_ep_len else d
# Store experience to replay buffer
replay_buffer.store(o, pi_a, r, o2, d)
# Super critical, easy to overlook step: make sure to update
# most recent observation!
o = o2
if d or (ep_len == max_ep_len):
"""
Perform all DDPG updates at the end of the trajectory,
in accordance with tuning done by TD3 paper authors.
"""
for _ in range(ep_len):
batch = replay_buffer.sample_batch(batch_size)
feed_dict = {
x_ph: batch['obs1'],
x2_ph: batch['obs2'],
a_ph: batch['acts'],
r_ph: batch['rews'],
d_ph: batch['done']
}
# Q-learning update
outs = sess.run([q_loss, q, train_q_op], feed_dict)
logger.store(LossQ=outs[0], QVals=outs[1])
# Policy update
outs = sess.run([pi_loss, train_pi_op, target_update],
feed_dict)
logger.store(LossPi=outs[0])
logger.store(EpRet=ep_ret, EpLen=ep_len)
actions.append(np.mean(epoch_actions))
epoch_actions = []
rewards.append(ep_ret)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
# End of epoch wrap-up
if (t + 1) % steps_per_epoch == 0:
epoch = (t + 1) // steps_per_epoch
# Test the performance of the deterministic version of the agent.
test_actions = test_agent(number_of_tests_per_epoch)
# Log info about epoch
logger.log_tabular('Epoch', epoch)
ret = logger.log_tabular('EpRet', average_only=True)
test_ret = logger.log_tabular('TestEpRet', average_only=True)[0]
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('TestEpLen', average_only=True)
logger.log_tabular('QVals', average_only=True)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossQ', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
rets.append(ret)
test_rets.append(test_ret)
if max_ret is None or test_ret > max_ret:
max_ret = test_ret
best_test_actions = test_actions
max_ep_len += inc_ep
util.plot_actions(test_actions, act_high,
logger.output_dir + '/actions%s.png' % epoch)
logger.save_state(
{
"actions": actions,
"rewards": rewards,
"best_test_actions": best_test_actions,
"rets": rets,
"test_rets": test_rets,
"max_ret": max_ret
}, None)
util.plot_actions(best_test_actions, act_high,
logger.output_dir + '/best_test_actions.png')
logger.log("max ret: %f" % max_ret)
exp_name = "control"
seed = 1007
env_config = '1x1_mix_ms_una.json'
iterations = 1
max_ep_len = 600
wp = 0
wi = 0.1
wd = 0.1
initial_bound = 1.1
final_bound = 1.01
bound_decay = 0.98
env = make_env(util.ENV_CONFIG_DIR + env_config)
obs = []
actions = []
action_sign = np.array([-1, -1])
for i in range(iterations):
current_bound = initial_bound
o = env.reset()
real_action = env.action_space.default() * 0.5
for t in range(max_ep_len):
o, r, d, _ = env.step(real_action)
obs.append(o)
actions.append(real_action)
vp = o
vi = np.mean(obs[-5:])
model.make_env(render_mode=render_mode)
model.load_model(filename)
else:
model = make_model(load_model=False)
print('model size', model.param_count)
model.make_env(render_mode=render_mode)
model.init_random_model_params(stdev=np.random.rand()*0.01)
N_episode = 100
if render_mode:
N_episode = 1
reward_list = []
for i in range(N_episode):
reward, steps_taken = simulate(model,
train_mode=False, render_mode=render_mode, num_episode=1)
if render_mode:
print("terminal reward", reward, "average steps taken", np.mean(steps_taken)+1)
else:
print(reward[0])
reward_list.append(reward[0])
if not render_mode:
print("seed", the_seed, "average_reward", np.mean(reward_list), "stdev", np.std(reward_list))
if __name__ == "__main__":
import env
e = env.make_env()
c = Controller()
import pdb; pdb.set_trace()
r, t = simulate(c, e, render_mode=False)
#main()
def learn(sess, n_tasks, z_size, data_dir, num_steps, max_seq_len,
batch_size_per_task=16, rnn_size=256,
grad_clip=1.0, v_lr=0.0001, vr_lr=0.0001,
min_v_lr=0.00001, v_decay=0.999, kl_tolerance=0.5,
lr=0.001, min_lr=0.00001, decay=0.999,
view="transposed",
model_dir="tf_rnn", layer_norm=False,
rnn_mmd=False, no_cor=False,
w_mmd=1.0,
alpha=1.0, beta=0.1,
recurrent_dp=1.0,
input_dp=1.0,
output_dp=1.0):
batch_size = batch_size_per_task * n_tasks
wrapper = WrapperFactory.get_wrapper(view)
if wrapper is None:
raise Exception("Such view is not available")
print("Batch size for each taks is", batch_size_per_task)
print("The total batch size is", batch_size)
check_dir(model_dir)
lf = open(model_dir + '/log_%s' % datetime.now().isoformat(), "w")
# define env
na = make_env(config.env_name).action_space.n
input_size = z_size + na
output_size = z_size
print("the environment", config.env_name, "has %i actions" % na)
seq_len = max_seq_len
fns = os.listdir(data_dir)
fns = [fn for fn in fns if '.npz' in fn]
random.shuffle(fns)
dm = get_dm(wrapper, seq_len, na, data_dir, fns, not no_cor)
tf_vrct_lr = tf.placeholder(tf.float32,
shape=[]) # learn from reconstruction.
vaes, vcomps = build_vaes(n_tasks, na, z_size, seq_len, tf_vrct_lr,
kl_tolerance)
vae_losses = [vcomp.loss for vcomp in vcomps]
transform_loss = get_transform_loss(vcomps[0], vaes[1], wrapper)
old_vae0 = ConvVAE(name="old_vae0", z_size=z_size)
old_vcomp0 = build_vae("old_vae0", old_vae0, na, z_size, seq_len,
tf_vrct_lr, kl_tolerance)
assign_old_eq_new = tf.group([tf.assign(oldv, newv)
for (oldv, newv) in
zip(old_vcomp0.var_list, vcomps[0].var_list)])
vmmd_losses = get_vmmd_losses(n_tasks, old_vcomp0, vcomps, alpha, beta)
vrec_ops = get_vae_rec_ops(n_tasks, vcomps, vmmd_losses, w_mmd)
vrec_all_op = tf.group(vrec_ops)
# Meta RNN.
rnn = VRNN("rnn", max_seq_len, input_size, output_size, batch_size_per_task,
rnn_size, layer_norm, recurrent_dp, input_dp, output_dp)
global_step = tf.Variable(0, name='global_step', trainable=False)
tf_rpred_lr = tf.placeholder(tf.float32, shape=[])
rcomp0 = build_rnn("rnn", rnn, na, z_size, batch_size_per_task, seq_len)
print("The basic rnn has been built")
rcomps = build_rnns(n_tasks, rnn, vaes, vcomps, kl_tolerance)
rnn_losses = [rcomp.loss for rcomp in rcomps]
if rnn_mmd:
rmmd_losses = get_rmmd_losses(n_tasks, old_vcomp0, vcomps, alpha, beta)
for i in range(n_tasks):
rnn_losses[i] += 0.1 * rmmd_losses[i]
ptransform_loss = get_predicted_transform_loss(vcomps[0], rcomps[0],
vaes[1],
wrapper, batch_size_per_task,
seq_len)
print("RNN has been connected to each VAE")
rnn_total_loss = tf.reduce_mean(rnn_losses)
rpred_opt = tf.train.AdamOptimizer(tf_rpred_lr, name="rpred_opt")
gvs = rpred_opt.compute_gradients(rnn_total_loss, rcomp0.var_list)
clip_gvs = [(tf.clip_by_value(grad, -grad_clip, grad_clip), var) for
grad, var in gvs if grad is not None]
rpred_op = rpred_opt.apply_gradients(clip_gvs, global_step=global_step,
name='rpred_op')
# VAE in prediction phase
vpred_ops, tf_vpred_lrs = get_vae_pred_ops(n_tasks, vcomps, rnn_losses)
vpred_all_op = tf.group(vpred_ops)
rpred_lr = lr
vrct_lr = v_lr
vpred_lr = vr_lr
sess.run(tf.global_variables_initializer())
for i in range(num_steps):
step = sess.run(global_step)
rpred_lr = (rpred_lr - min_lr) * decay + min_lr
vrct_lr = (vrct_lr - min_v_lr) * v_decay + min_v_lr
vpred_lr = (vpred_lr - min_v_lr) * v_decay + min_v_lr
ratio = 1.0
data_buffer = []
for it in range(config.psteps_per_it):
raw_obs_list, raw_a_list = dm.random_batch(batch_size_per_task)
data_buffer.append((raw_obs_list, raw_a_list))
feed = {tf_rpred_lr: rpred_lr, tf_vrct_lr: vrct_lr,
tf_vpred_lrs[0]: vpred_lr,
tf_vpred_lrs[1]: vpred_lr * ratio}
feed[old_vcomp0.x] = raw_obs_list[0]
for j in range(n_tasks):
vcomp = vcomps[j]
feed[vcomp.x] = raw_obs_list[j]
feed[vcomp.a] = raw_a_list[j][:, :-1, :]
(rnn_cost, rnn_cost2, vae_cost, vae_cost2,
transform_cost, ptransform_cost, _, _) = sess.run(
[rnn_losses[0], rnn_losses[1],
vae_losses[0], vae_losses[1],
transform_loss, ptransform_loss,
rpred_op, vpred_all_op], feed)
ratio = rnn_cost2 / rnn_cost
if i % config.log_interval == 0:
output_log = get_output_log(step, rpred_lr, [vae_cost], [rnn_cost], [transform_cost], [ptransform_cost])
lf.write(output_log)
data_order = np.arange(len(data_buffer))
nd = len(data_order)
np.random.shuffle(data_order)
for it in range(config.rsteps_per_it):
if (it + 1) % nd == 0:
np.random.shuffle(data_order)
rid = data_order[it % nd]
raw_obs_list, raw_a_list = data_buffer[rid]
# raw_obs_list, raw_a_list = dm.random_batch(batch_size_per_task)
feed = {tf_rpred_lr: rpred_lr, tf_vrct_lr: vrct_lr}
feed[old_vcomp0.x] = raw_obs_list[0]
for j in range(n_tasks):
vcomp = vcomps[j]
feed[vcomp.x] = raw_obs_list[j]
feed[vcomp.a] = raw_a_list[j][:, :-1, :]
(rnn_cost, rnn_cost2, vae_cost, vae_cost2, transform_cost,
ptransform_cost, _) = sess.run([
rnn_losses[0], rnn_losses[1],
vae_losses[0], vae_losses[1],
transform_loss, ptransform_loss,
vrec_all_op], feed)
if i % config.log_interval == 0:
output_log = get_output_log(step, rpred_lr, [vae_cost], [rnn_cost], [transform_cost], [ptransform_cost])
lf.write(output_log)
lf.flush()
if (i + 1) % config.target_update_interval == 0:
sess.run(assign_old_eq_new)
if i % config.model_save_interval == 0:
tmp_dir = model_dir + '/it_%i' % i
check_dir(tmp_dir)
saveToFlat(rcomp0.var_list, tmp_dir + '/rnn.p')
for j in range(n_tasks):
vcomp = vcomps[j]
saveToFlat(vcomp.var_list, tmp_dir + '/vae%i.p' % j)
saveToFlat(rcomp0.var_list, model_dir + '/final_rnn.p')
for i in range(n_tasks):
vcomp = vcomps[i]
saveToFlat(vcomp.var_list, model_dir + '/final_vae%i.p' % i)
def actor():
print(f"STARTING ACTOR with rank {rank}")
sys.stdout.flush()
# GAE hyper-parameters
lam = 0.95
gamma = 0.99
# Build network architecture
nav = Navigation(1, training=False)
nav.call_build()
# Get agent type
agent_type = np.where(np.array(actors) == rank)[0][0]
# Setup environment
env = make_env()
obs = env.reset()
dones = False
while True:
weights = comm.recv(source=learners[agent_type])
nav.set_weights(weights)
mb_rewards = np.zeros([nsteps, 1], dtype=np.float32)
mb_values = np.zeros([nsteps, 1], dtype=np.float32)
mb_neglogpacs = np.zeros([nsteps, 1], dtype=np.float32)
mb_dones = np.zeros([nsteps, 1], dtype=np.float32)
mb_obs = np.zeros([nsteps, 16], dtype=np.float32)
mb_actions = {
'x1': np.zeros([nsteps, 1], dtype=np.int32),
'x2': np.zeros([nsteps, 1], dtype=np.int32),
'y1': np.zeros([nsteps, 1], dtype=np.int32),
'y2': np.zeros([nsteps, 1], dtype=np.int32),
'w1': np.zeros([nsteps, 1], dtype=np.int32),
'w2': np.zeros([nsteps, 1], dtype=np.int32)
}
mb_logits = {
'x1': np.zeros([nsteps, 21], dtype=np.float32),
'x2': np.zeros([nsteps, 21], dtype=np.float32),
'y1': np.zeros([nsteps, 21], dtype=np.float32),
'y2': np.zeros([nsteps, 21], dtype=np.float32),
'w1': np.zeros([nsteps, 21], dtype=np.float32),
'w2': np.zeros([nsteps, 21], dtype=np.float32)
}
for i in range(nsteps):
# Get actions of training agent
actions, neglogp, entropy, value, logits = nav(
np.expand_dims(obs, axis=0))
mb_values[i] = value
mb_neglogpacs[i] = neglogp
mb_obs[i] = obs
for k in actions.keys():
mb_actions[k][i] = actions[k]
mb_logits[k][i] = logits[k]
mb_dones[i] = dones
# Take actions in env and look at the results
actions = {k: (v[0] - 10) / 10 for k, v in actions.items()}
agent_actions = np.array(
[[actions['x1'], actions['y1'], actions['w1']],
[actions['x2'], actions['y2'], actions['w2']]])
obs, rewards, dones, infos = env.step(agent_actions)
# Handle rewards
mb_rewards[i] = rewards
if dones:
obs = env.reset()
# get last value for bootstrap
_, _, _, last_values, _ = nav(np.expand_dims(obs, axis=0))
# discount/bootstrap off value fn
mb_returns = np.zeros_like(mb_rewards)
mb_advs = np.zeros_like(mb_rewards)
lastgaelam = 0
# perform GAE calculation
for t in reversed(range(nsteps)):
if t == nsteps - 1:
nextnonterminal = 1.0 - dones
nextvalues = last_values
else:
nextnonterminal = 1.0 - mb_dones[t + 1]
nextvalues = mb_values[t + 1]
delta = mb_rewards[
t] + gamma * nextvalues * nextnonterminal - mb_values[t]
mb_advs[
t] = lastgaelam = delta + gamma * lam * nextnonterminal * lastgaelam
mb_returns = mb_advs + mb_values
# Send trajectory to learner
mb_values = np.squeeze(mb_values, axis=-1)
mb_rewards = np.squeeze(mb_rewards, axis=-1)
mb_neglogpacs = np.squeeze(mb_neglogpacs, axis=-1)
mb_returns = np.squeeze(mb_returns, axis=-1)
mb_dones = np.squeeze(mb_dones, axis=-1)
trajectory = {
'mb_obs': mb_obs,
'mb_actions': mb_actions,
'mb_logits': mb_logits,
'mb_returns': mb_returns,
'mb_dones': mb_dones,
'mb_values': mb_values,
'mb_neglogpacs': mb_neglogpacs,
'mb_rewards': mb_rewards
}
comm.send(trajectory, dest=learners[agent_type])
def main(args):
env_name = args.env_name
total_episodes = args.total_episodes
time_steps = args.time_steps
render = args.render
run_all_envs = args.run_all_envs
action_refresh_rate = args.action_refresh_rate
if run_all_envs:
envs_to_generate = config.train_envs
else:
envs_to_generate = [env_name]
for current_env_name in envs_to_generate:
print("Generating data for env {}".format(current_env_name))
env = make_env(current_env_name) # <1>
s = 0
while s < total_episodes:
episode_id = random.randint(0, 2**31 - 1)
filename = DIR_NAME + str(episode_id) + ".npz"
observation = env.reset()
env.render()
t = 0
obs_sequence = []
action_sequence = []
reward_sequence = []
done_sequence = []
reward = -0.1
done = False
while t < time_steps: # and not done:
if t % action_refresh_rate == 0:
action = config.generate_data_action(t, env) # <2>
observation = config.adjust_obs(observation) # <3>
obs_sequence.append(observation)
action_sequence.append(action)
reward_sequence.append(reward)
done_sequence.append(done)
observation, reward, done, info = env.step(action) # <4>
t = t + 1
if render:
env.render()
print("Episode {} finished after {} timesteps".format(s, t))
np.savez_compressed(filename,
obs=obs_sequence,
action=action_sequence,
reward=reward_sequence,
done=done_sequence) # <4>
s = s + 1
env.close()
def main(args):
env_name = args.env_name
total_episodes = args.total_episodes
start_batch = args.start_batch
time_steps = args.time_steps
render = args.render
batch_size = args.batch_size
run_all_envs = args.run_all_envs
if run_all_envs:
envs_to_generate = config.train_envs
else:
envs_to_generate = [env_name]
for current_env_name in envs_to_generate:
print("Generating data for env {}".format(current_env_name))
env = make_env(current_env_name)
s = 0
batch = start_batch
batch_size = min(batch_size, total_episodes)
while s < total_episodes:
obs_data = []
action_data = []
for i_episode in range(batch_size):
print('-----')
observation = env.reset()
observation = config.adjust_obs(observation)
# plt.imshow(observation)
# plt.show()
env.render()
done = False
action = env.action_space.sample()
t = 0
obs_sequence = []
action_sequence = []
while t < time_steps: #and not done:
t = t + 1
action = config.generate_data_action(t, action)
obs_sequence.append(observation)
action_sequence.append(action)
observation, reward, done, info = env.step(action)
observation = config.adjust_obs(observation)
if render:
env.render()
obs_data.append(obs_sequence)
action_data.append(action_sequence)
print("Batch {} Episode {} finished after {} timesteps".format(batch, i_episode, t+1))
print("Current dataset contains {} observations".format(sum(map(len, obs_data))))
s = s + 1
print("Saving dataset for batch {}".format(batch))
np.save('./data/obs_data_' + current_env_name + '_' + str(batch), obs_data)
np.save('./data/action_data_' + current_env_name + '_' + str(batch), action_data)
batch = batch + 1
env.close()
