def train(params):
rank = MPI.COMM_WORLD.Get_rank()
if rank == 0:
logger.configure()
else:
logger.configure(format_strs=[])
# setup config
if params.get("policy") == 'mlp':
policy = MlpPolicy
env = gym.make(params.get("environment"))
env.configure(envConfig)
env.reset()
else:
policy = CnnPolicy
env = gym.make(params.get("environment"))
env.configure(CnnNet)
env.reset()
exp_name = ("{0}_{1}_{2}".format(params.get("model_name"),
params.get("policy"),
params.get("environment")))
log_dir = './logs/' + exp_name
if params.get("seed") > 0:
workerseed = params.get("seed"), +10000 * MPI.COMM_WORLD.Get_rank()
set_global_seeds(workerseed)
env.seed(workerseed)
# create model
model = TRPO(policy,
env,
verbose=1,
tensorboard_log=log_dir,
timesteps_per_batch=params.get("timesteps_per_batch"),
max_kl=params.get("max_kl"),
cg_iters=params.get("cg_iters"),
cg_damping=params.get("cg_damping"),
entcoeff=params.get("entcoeff"),
gamma=params.get("gamma"),
lam=params.get("lam"),
vf_iters=params.get("vf_iters"),
vf_stepsize=params.get("vf_stepsize")
# ,policy_kwargs=policy_kwargs
)
model.learn(total_timesteps=params.get("train_steps"))
model.save(exp_name)
env.close()
del env
def train(env_id, num_timesteps, seed, algorithm, model_save_file=None, log_dir=None):
with tf_util.single_threaded_session():
logger.configure(folder=log_dir, format_strs=['stdout', 'log', 'csv'])
workerseed = seed + MPI.COMM_WORLD.Get_rank()
env = make_mujoco_env(env_id, workerseed)
if algorithm == "TRPO":
model = TRPO(MlpPolicy, env, seed=workerseed, verbose=1)
else:
# Algorithm is PPO
model = PPO1(MlpPolicy, env, seed=workerseed, verbose=1)
model.learn(total_timesteps=num_timesteps)
if model_save_file is not None:
model.save(model_save_file)
env.close()
def main():
"""
Runs the test
"""
"""
Create an argparse.ArgumentParser for run_mujoco.py.
:return: (ArgumentParser) parser {'--env': 'Reacher-v2', '--seed': 0, '--num-timesteps': int(1e6), '--play': False}
parser = arg_parser()
parser.add_argument('--env', help='environment ID', type=str, default='Reacher-v2')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--num-timesteps', type=int, default=int(1e6))
parser.add_argument('--play', default=False, action='store_true')
return parse
"""
env_id = 'UR5Gripper-v0'
model_path = '/tmp/gym/trpo_mpi/'
# args = mujoco_arg_parser().parse_args()
# train(args.env, num_timesteps=args.num_timesteps, seed=args.seed)
# train(env_id=env_id, num_timesteps=int(1e7), seed=0, model_path=model_path)
env = gym.make(env_id)
env = Monitor(env, model_path, allow_early_resets=True)
model = TRPO(MlpPolicy, env, verbose=1, tensorboard_log=model_path)
model = model.load(model_path + "trpo.pkl")
model.learn(total_timesteps=int(1e5), callback=callback)
model.save(model_path + "trpo.pkl")
# tf_util.save_state(model_path)
# Enjoy trained agent
obs = env.reset()
for i in range(100):
obs = env.reset()
env.render()
for i in range(200):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
env.render()
class GAIL(ActorCriticRLModel):
"""
Generative Adversarial Imitation Learning (GAIL)
.. warning::
Images are not yet handled properly by the current implementation
:param policy: (ActorCriticPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, CnnLstmPolicy, ...)
:param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str)
:param expert_dataset: (ExpertDataset) the dataset manager
:param gamma: (float) the discount value
:param timesteps_per_batch: (int) the number of timesteps to run per batch (horizon)
:param max_kl: (float) the kullback leiber loss threashold
:param cg_iters: (int) the number of iterations for the conjugate gradient calculation
:param lam: (float) GAE factor
:param entcoeff: (float) the weight for the entropy loss
:param cg_damping: (float) the compute gradient dampening factor
:param vf_stepsize: (float) the value function stepsize
:param vf_iters: (int) the value function's number iterations for learning
:param hidden_size: ([int]) the hidden dimension for the MLP
:param g_step: (int) number of steps to train policy in each epoch
:param d_step: (int) number of steps to train discriminator in each epoch
:param d_stepsize: (float) the reward giver stepsize
:param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
:param full_tensorboard_log: (bool) enable additional logging when using tensorboard
WARNING: this logging can take a lot of space quickly
"""
def __init__(self, policy, env, expert_dataset=None,
hidden_size_adversary=100, adversary_entcoeff=1e-3,
g_step=3, d_step=1, d_stepsize=3e-4, verbose=0,
_init_setup_model=True, **kwargs):
super().__init__(policy=policy, env=env, verbose=verbose, requires_vec_env=False,
_init_setup_model=_init_setup_model)
self.trpo = TRPO(policy, env, verbose=verbose, _init_setup_model=False, **kwargs)
self.trpo.using_gail = True
self.trpo.expert_dataset = expert_dataset
self.trpo.g_step = g_step
self.trpo.d_step = d_step
self.trpo.d_stepsize = d_stepsize
self.trpo.hidden_size_adversary = hidden_size_adversary
self.trpo.adversary_entcoeff = adversary_entcoeff
self.env = self.trpo.env
if _init_setup_model:
self.setup_model()
def _get_pretrain_placeholders(self):
pass
def pretrain(self, dataset, n_epochs=10, learning_rate=1e-4,
adam_epsilon=1e-8, val_interval=None):
self.trpo.pretrain(dataset, n_epochs=n_epochs, learning_rate=learning_rate,
adam_epsilon=adam_epsilon, val_interval=val_interval)
return self
def set_env(self, env):
self.trpo.set_env(env)
self.env = self.trpo.env
def setup_model(self):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the GAIL model must be an " \
"instance of common.policies.ActorCriticPolicy."
self.trpo.setup_model()
def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="GAIL",
reset_num_timesteps=True):
assert self.trpo.expert_dataset is not None, "You must pass an expert dataset to GAIL for training"
self.trpo.learn(total_timesteps, callback, seed, log_interval, tb_log_name, reset_num_timesteps)
return self
def predict(self, observation, state=None, mask=None, deterministic=False):
return self.trpo.predict(observation, state=state, mask=mask, deterministic=deterministic)
def action_probability(self, observation, state=None, mask=None, actions=None):
return self.trpo.action_probability(observation, state=state, mask=mask, actions=actions)
def save(self, save_path):
self.trpo.save(save_path)
@classmethod
def load(cls, load_path, env=None, **kwargs):
data, params = cls._load_from_file(load_path)
model = cls(policy=data["policy"], env=None, _init_setup_model=False)
model.trpo.__dict__.update(data)
model.trpo.__dict__.update(kwargs)
model.set_env(env)
model.setup_model()
restores = []
for param, loaded_p in zip(model.trpo.params, params):
restores.append(param.assign(loaded_p))
model.trpo.sess.run(restores)
return model
class GAIL(ActorCriticRLModel):
"""
Generative Adversarial Imitation Learning (GAIL)
:param policy: (ActorCriticPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, CnnLstmPolicy, ...)
:param env: (Gym environment or str) The environment to learn from (if registered in Gym, can be str)
:param gamma: (float) the discount value
:param timesteps_per_batch: (int) the number of timesteps to run per batch (horizon)
:param max_kl: (float) the kullback leiber loss threashold
:param cg_iters: (int) the number of iterations for the conjugate gradient calculation
:param lam: (float) GAE factor
:param entcoeff: (float) the weight for the entropy loss
:param cg_damping: (float) the compute gradient dampening factor
:param vf_stepsize: (float) the value function stepsize
:param vf_iters: (int) the value function's number iterations for learning
:param pretrained_weight: (str) the save location for the pretrained weights
:param hidden_size: ([int]) the hidden dimension for the MLP
:param expert_dataset: (Dset) the dataset manager
:param save_per_iter: (int) the number of iterations before saving
:param checkpoint_dir: (str) the location for saving checkpoints
:param g_step: (int) number of steps to train policy in each epoch
:param d_step: (int) number of steps to train discriminator in each epoch
:param task_name: (str) the name of the task (can be None)
:param d_stepsize: (float) the reward giver stepsize
:param verbose: (int) the verbosity level: 0 none, 1 training information, 2 tensorflow debug
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
"""
def __init__(self,
policy,
env,
pretrained_weight=False,
hidden_size_adversary=100,
adversary_entcoeff=1e-3,
expert_dataset=None,
save_per_iter=1,
checkpoint_dir="/tmp/gail/ckpt/",
g_step=1,
d_step=1,
task_name="task_name",
d_stepsize=3e-4,
verbose=0,
_init_setup_model=True,
**kwargs):
super().__init__(policy=policy,
env=env,
verbose=verbose,
requires_vec_env=False,
_init_setup_model=_init_setup_model)
self.trpo = TRPO(policy,
env,
verbose=verbose,
_init_setup_model=False,
**kwargs)
self.trpo.using_gail = True
self.trpo.pretrained_weight = pretrained_weight
self.trpo.expert_dataset = expert_dataset
self.trpo.save_per_iter = save_per_iter
self.trpo.checkpoint_dir = checkpoint_dir
self.trpo.g_step = g_step
self.trpo.d_step = d_step
self.trpo.task_name = task_name
self.trpo.d_stepsize = d_stepsize
self.trpo.hidden_size_adversary = hidden_size_adversary
self.trpo.adversary_entcoeff = adversary_entcoeff
if _init_setup_model:
self.setup_model()
def set_env(self, env):
super().set_env(env)
self.trpo.set_env(env)
def setup_model(self):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the GAIL model must be an " \
"instance of common.policies.ActorCriticPolicy."
assert isinstance(
self.action_space,
gym.spaces.Box), "Error: GAIL requires a continuous action space."
self.trpo.setup_model()
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="GAIL"):
self.trpo.learn(total_timesteps, callback, seed, log_interval,
tb_log_name)
return self
def predict(self, observation, state=None, mask=None, deterministic=False):
return self.trpo.predict(observation,
state,
mask,
deterministic=deterministic)
def action_probability(self, observation, state=None, mask=None):
return self.trpo.action_probability(observation, state, mask)
def save(self, save_path):
self.trpo.save(save_path)
@classmethod
def load(cls, load_path, env=None, **kwargs):
data, params = cls._load_from_file(load_path)
model = cls(policy=data["policy"], env=None, _init_setup_model=False)
model.trpo.__dict__.update(data)
model.trpo.__dict__.update(kwargs)
model.set_env(env)
model.setup_model()
restores = []
for param, loaded_p in zip(model.trpo.params, params):
restores.append(param.assign(loaded_p))
model.trpo.sess.run(restores)
return model