def test_ddpg_normalization():
"""
Test that observations and returns normalizations are properly saved and loaded.
"""
param_noise = AdaptiveParamNoiseSpec(initial_stddev=0.05,
desired_action_stddev=0.05)
model = DDPG('MlpPolicy',
'Pendulum-v0',
memory_limit=50000,
normalize_observations=True,
normalize_returns=True,
nb_rollout_steps=128,
nb_train_steps=1,
batch_size=64,
param_noise=param_noise)
model.learn(1000)
obs_rms_params = model.sess.run(model.obs_rms_params)
ret_rms_params = model.sess.run(model.ret_rms_params)
model.save('./test_ddpg')
loaded_model = DDPG.load("test_ddpg")
obs_rms_params_2 = loaded_model.sess.run(loaded_model.obs_rms_params)
ret_rms_params_2 = loaded_model.sess.run(loaded_model.ret_rms_params)
for param, param_loaded in zip(obs_rms_params + ret_rms_params,
obs_rms_params_2 + ret_rms_params_2):
assert np.allclose(param, param_loaded)
del model, loaded_model
if os.path.exists("./test_ddpg"):
os.remove("./test_ddpg")
def test_ddpg_popart():
"""
Test DDPG with pop-art normalization
"""
n_actions = 1
action_noise = NormalActionNoise(mean=np.zeros(n_actions), sigma=0.1 * np.ones(n_actions))
model = DDPG('MlpPolicy', 'Pendulum-v0', memory_limit=50000, normalize_observations=True,
normalize_returns=True, nb_rollout_steps=128, nb_train_steps=1,
batch_size=64, action_noise=action_noise, enable_popart=True)
model.learn(1000)
def test_ddpg_eval_env():
"""
Additional test to check that everything is working when passing
an eval env.
"""
eval_env = gym.make("Pendulum-v0")
model = DDPG("MlpPolicy", "Pendulum-v0", nb_rollout_steps=5,
nb_train_steps=2, nb_eval_steps=10,
eval_env=eval_env, verbose=0)
model.learn(1000)
import gym
from stable_baselines.ddpg import LnMlpPolicy
from stable_baselines.ddpg import DDPG
env = gym.make('HalfCheetah-v3')
model = DDPG(LnMlpPolicy, env,gamma=.95,buffer_size=1000000,param_noise_adaption_interval=0.22,batch_size=256,
normalize_observations=True,normalize_returns=False, policy_kwargs= dict(layers=[400, 300]) ,verbose=1)
model.learn(total_timesteps=1000000)
model.save('Cheetah_model_DDPG')
obs = env.reset()
for i in range(1000):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
env.render()
env.close()
def run(env_id, seed, noise_type, layer_norm, evaluation, **kwargs):
"""
run the training of DDPG
:param env_id: (str) the environment ID
:param seed: (int) the initial random seed
:param noise_type: (str) the wanted noises ('adaptive-param', 'normal' or 'ou'), can use multiple noise type by
seperating them with commas
:param layer_norm: (bool) use layer normalization
:param evaluation: (bool) enable evaluation of DDPG training
:param kwargs: (dict) extra keywords for the training.train function
"""
# Configure things.
rank = MPI.COMM_WORLD.Get_rank()
if rank != 0:
logger.set_level(logger.DISABLED)
# Create envs.
env = gym.make(env_id)
env = bench.Monitor(
env,
logger.get_dir() and os.path.join(logger.get_dir(), str(rank)))
if evaluation and rank == 0:
eval_env = gym.make(env_id)
eval_env = bench.Monitor(eval_env,
os.path.join(logger.get_dir(), 'gym_eval'))
env = bench.Monitor(env, None)
else:
eval_env = None
# Parse noise_type
action_noise = None
param_noise = None
nb_actions = env.action_space.shape[-1]
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(
initial_stddev=float(stddev),
desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mean=np.zeros(nb_actions),
sigma=float(stddev) *
np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(
mean=np.zeros(nb_actions),
sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError(
'unknown noise type "{}"'.format(current_noise_type))
# Seed everything to make things reproducible.
seed = seed + 1000000 * rank
logger.info('rank {}: seed={}, logdir={}'.format(rank, seed,
logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
start_time = 0
if rank == 0:
start_time = time.time()
model = DDPG(policy=MlpPolicy,
env=env,
memory_policy=Memory,
eval_env=eval_env,
param_noise=param_noise,
action_noise=action_noise,
memory_limit=int(1e6),
layer_norm=layer_norm,
verbose=2,
**kwargs)
model.learn(total_timesteps=10000)
env.close()
if eval_env is not None:
eval_env.close()
if rank == 0:
logger.info('total runtime: {}s'.format(time.time() - start_time))
def main(env,
load,
save_path,
load_path=None,
train_timesteps=1.25e6,
eval_timesteps=5e3):
# arguments
print(
"env %s; load %s; save_path %s; load_path %s; train_timesteps %s; eval_timesteps %s;"
% (env, load, save_path, load_path, train_timesteps, eval_timesteps))
train_timesteps = int(float(train_timesteps))
eval_timesteps = int(float(eval_timesteps))
# models path
model_dir = os.getcwd() + "/models/"
os.makedirs(model_dir, exist_ok=True)
# logging path
log_dir = os.getcwd() + "/log/" + save_path
os.makedirs(log_dir, exist_ok=True)
# absolute save path and models path
save_path = model_dir + save_path
if load and not load_path:
print("no load path given, exiting...")
sys.exit()
elif load:
load_path = model_dir + load_path
# make environment, flattened environment, monitor, vectorized environment
env = gym.make(env)
env = gym.wrappers.FlattenDictWrapper(
env, ['observation', 'achieved_goal', 'desired_goal'])
env = Monitor(env, log_dir, allow_early_resets=True)
env = DummyVecEnv([lambda: env])
# load model, or start from scratch
if load:
print("loading model from: " + load_path)
model = DDPG.load(load_path, env=env)
else:
print("training model from scratch")
model = DDPG(MlpPolicy, env, verbose=1)
# evaluate current model
mean_reward_before_train = evaluate(model, env, num_steps=eval_timesteps)
# train model
global best_mean_reward, n_steps
best_mean_reward, n_steps = -np.inf, 0
model.learn(total_timesteps=train_timesteps, callback=None)
# save model
print("saving model to:" + save_path)
model.save(save_path)
# evaluate post training model
mean_reward_after_train = evaluate(model, env, num_steps=eval_timesteps)
# results
print("reward before training:" + str(mean_reward_before_train))
print("reward after training:" + str(mean_reward_after_train))
print("done")
def train(env_id, num_timesteps, seed, model_path=None, images=False):
"""
Train PPO2 model for Mujoco environment, for testing purposes
:param env_id: (str) the environment id string
:param num_timesteps: (int) the number of timesteps to run
:param seed: (int) Used to seed the random generator.
"""
def make_env():
if images:
env_out = GymWrapper(
suite.make(
"SawyerLift",
use_object_obs=False,
use_camera_obs=True, # do not use pixel observations
has_offscreen_renderer=
True, # not needed since not using pixel obs
has_renderer=False, # make sure we can render to the screen
camera_depth=True,
reward_shaping=True, # use dense rewards
control_freq=
10, # control should happen fast enough so that simulation looks smooth
render_visual_mesh=False,
),
keys=["image", "depth"],
images=True,
)
else:
env_out = GymWrapper(
suite.make(
"SawyerLift",
use_object_obs=True,
use_camera_obs=False, # do not use pixel observations
has_offscreen_renderer=
False, # not needed since not using pixel obs
has_renderer=False, # make sure we can render to the screen
camera_depth=False,
reward_shaping=True, # use dense rewards
control_freq=
10, # control should happen fast enough so that simulation looks smooth
render_visual_mesh=False,
) #, keys=["image", "depth"], images=True,
)
env_out.reward_range = None
env_out.metadata = None
env_out.spec = None
env_out = bench.Monitor(env_out,
logger.get_dir(),
allow_early_resets=True)
return env_out
#env = make_env()
if images:
env = DummyVecEnv([make_env])
env = VecNormalize(env)
set_global_seeds(seed)
policy = CnnPolicy
tblog = "/cvgl2/u/surajn/workspace/tb_logs/sawyerlift_all/"
else:
env = DummyVecEnv([make_env])
env = VecNormalize(env)
set_global_seeds(seed)
policy = MlpPolicy
tblog = "/cvgl2/u/surajn/workspace/tb_logs/sawyerlift_all/"
nb_actions = env.action_space.shape[-1]
#model = PPO2(policy=policy, env=env, n_steps=2048, nminibatches=32, lam=0.95, gamma=0.99, noptepochs=10,
# ent_coef=0.0, learning_rate=3e-4, cliprange=0.2, verbose=1, tensorboard_log=tblog)
#model = TRPO(policy=policy, env, timesteps_per_batch=1024, max_kl=0.01, cg_iters=10, cg_damping=0.1, entcoeff=0.0,
# gamma=0.99, lam=0.98, vf_iters=5, vf_stepsize=1e-3, tensorboard_log=tblog, verbose=1)
model = DDPG(policy=ddpgMlpPolicy,
env=env,
memory_policy=Memory,
eval_env=None,
param_noise=AdaptiveParamNoiseSpec(initial_stddev=0.2,
desired_action_stddev=0.2),
action_noise=OrnsteinUhlenbeckActionNoise(
mean=np.zeros(nb_actions),
sigma=float(0.2) * np.ones(nb_actions)),
memory_limit=int(1e6),
verbose=2,
tensorboard_log=tblog)
model.learn(total_timesteps=num_timesteps)
env.close()
if model_path:
model.save(model_path)
#tf_util.save_state(model_path)
return model, env
def run(env_id, seed, layer_norm, evaluation, agent, delay_step, gamma=0.99, **kwargs):
# Create envs.
env = create_env(env_id, delay_step, str(0))
print(env.observation_space, env.action_space)
if evaluation:
eval_env = create_env(env_id, delay_step, "eval_env")
else:
eval_env = None
# Seed everything to make things reproducible.
logger.info('seed={}, logdir={}'.format(seed, logger.get_dir()))
tf.reset_default_graph()
set_global_seeds(seed)
env.seed(seed)
if eval_env is not None:
eval_env.seed(seed)
# Disable logging for rank != 0 to avoid noise.
start_time = time.time()
policy = 'MlpPolicy'
td3_variants = {
"TD3": TD3,
"TD3SIL": TD3SIL,
"TD3NSTEP": TD3NSTEP,
"TD3REDQ": TD3REDQ,
"TD3DoubleTwin": TD3DoubleTwin,
}
if td3_variants.get(agent, None):
model_func = td3_variants[agent]
model = model_func(policy=policy, env=env, eval_env=eval_env, gamma=gamma, batch_size=128,
tau=0.005, policy_delay=2, learning_starts=25000,
action_noise=create_action_noise(env, "normal_0.1"), buffer_size=100000, verbose=2,
n_cpu_tf_sess=10,
policy_kwargs={"layers": [400, 300]})
elif agent == "DDPG":
model = DDPG(policy=policy, env=env, eval_env=eval_env, gamma=gamma, nb_eval_steps=5, batch_size=100,
nb_train_steps=100, nb_rollout_steps=100, learning_starts=10000,
actor_lr=1e-3, critic_lr=1e-3, critic_l2_reg=0,
tau=0.005, normalize_observations=False,
action_noise=create_action_noise(env, "normal_0.1"), buffer_size=int(1e6),
verbose=2, n_cpu_tf_sess=10,
policy_kwargs={"layers": [400, 300]})
elif agent == "SAC":
model = SAC(policy=policy, env=env, eval_env=eval_env, gamma=gamma, batch_size=256,
action_noise=create_action_noise(env, "normal_0.1"), buffer_size=int(1e6), verbose=2,
n_cpu_tf_sess=10, learning_starts=10000,
policy_kwargs={"layers": [256, 256]})
elif agent == "GEM":
policy = 'TD3LnMlpPolicy'
model = TD3MemGEM(policy=policy, env=env, eval_env=eval_env, gamma=gamma, batch_size=128,
tau=0.005, policy_delay=2, learning_starts=25000,
action_noise=create_action_noise(env, "normal_0.1"), buffer_size=100000, verbose=2,
n_cpu_tf_sess=10,
alpha=0.5, beta=-1, iterative_q=-1,
num_q=4, gradient_steps=200, max_step=kwargs['max_steps'], reward_scale=1., nb_eval_steps=10,
policy_kwargs={"layers": [400, 300]})
elif agent == "BP":
policy = 'TD3LnMlpPolicy'
model = TD3MemBackProp(policy=policy, env=env, eval_env=eval_env, gamma=gamma, batch_size=128,
tau=0.005, policy_delay=2, learning_starts=25000,
action_noise=create_action_noise(env, "normal_0.1"), buffer_size=100000, verbose=2,
n_cpu_tf_sess=10,
alpha=0.5, beta=-1, gradient_steps=200, max_step=kwargs['max_steps'], reward_scale=1., nb_eval_steps=10,
policy_kwargs={"layers": [400, 300]})
else:
raise NotImplementedError
print("model building finished")
model.learn(total_timesteps=kwargs['num_timesteps'])
env.close()
if eval_env is not None:
eval_env.close()
logger.info('total runtime: {}s'.format(time.time() - start_time))
