observations.append(observation)
actions.append(action)
rewards.append(reward)
observation = next_observation
if terminal:
# Finish rollout if terminal state reached
break
# We need to compute the empirical return for each time step along the
# trajectory
path = dict(
observations=np.array(observations),
actions=np.array(actions),
rewards=np.array(rewards),
)
path_baseline = baseline.predict(path)
advantages = []
returns = []
return_so_far = 0
for t in range(len(rewards) - 1, -1, -1):
return_so_far = rewards[t] + discount * return_so_far
returns.append(return_so_far)
advantage = return_so_far - path_baseline[t]
advantages.append(advantage)
# The advantages are stored backwards in time, so we need to revert it
advantages = np.array(advantages[::-1])
# And we need to do the same thing for the list of returns
returns = np.array(returns[::-1])
advantages = (advantages - np.mean(advantages)) / (np.std(advantages) + 1e-8)
def run_task(*_):
# normalize() makes sure that the actions for the environment lies
# within the range [-1, 1] (only works for environments with continuous actions)
# normalize() makes sure that the actions for the environment lies
# within the range [-1, 1] (only works for environments with continuous actions)
env = normalize(
GymEnv(env_name="LunarLanderContinuous-v2", force_reset=True))
# Initialize a neural network policy with a single hidden layer of 8 hidden units
policy = GaussianMLPPolicy(env.spec, hidden_sizes=(64, 64))
# Initialize a linear baseline estimator using default hand-crafted features
baseline = LinearFeatureBaseline(env.spec)
# We will collect 100 trajectories per iteration
N = 3
# Each trajectory will have at most 100 time steps
T = 400
# Number of iterations
n_itr = 1000
# Set the discount factor for the problem
discount = 0.99
# Learning rate for the gradient update
learning_rate = 0.001
# Construct the computation graph
# Create a Theano variable for storing the observations
# We could have simply written `observations_var = TT.matrix('observations')` instead for this example. However,
# doing it in a slightly more abstract way allows us to delegate to the environment for handling the correct data
# type for the variable. For instance, for an environment with discrete observations, we might want to use integer
# types if the observations are represented as one-hot vectors.
observations_var = env.observation_space.new_tensor_variable(
'observations',
# It should have 1 extra dimension since we want to represent a list of observations
extra_dims=1)
actions_var = env.action_space.new_tensor_variable('actions', extra_dims=1)
advantages_var = TT.vector('advantages')
# policy.dist_info_sym returns a dictionary, whose values are symbolic expressions for quantities related to the
# distribution of the actions. For a Gaussian policy, it contains the mean and (log) standard deviation.
dist_info_vars = policy.dist_info_sym(observations_var)
# policy.distribution returns a distribution object under rllab.distributions. It contains many utilities for computing
# distribution-related quantities, given the computed dist_info_vars. Below we use dist.log_likelihood_sym to compute
# the symbolic log-likelihood. For this example, the corresponding distribution is an instance of the class
# rllab.distributions.DiagonalGaussian
dist = policy.distribution
# Note that we negate the objective, since most optimizers assume a
# minimization problem
surr = -TT.mean(
dist.log_likelihood_sym(actions_var, dist_info_vars) * advantages_var)
# Get the list of trainable parameters.
params = policy.get_params(trainable=True)
grads = theano.grad(surr, params)
f_train = theano.function(
inputs=[observations_var, actions_var, advantages_var],
outputs=None,
updates=adam(grads, params, learning_rate=learning_rate),
allow_input_downcast=True)
for epoch in range(n_itr):
logger.push_prefix('epoch #%d | ' % epoch)
logger.log("Training started")
paths = []
for _ in range(N):
observations = []
actions = []
rewards = []
observation = env.reset()
for _ in range(T):
# policy.get_action() returns a pair of values. The second one returns a dictionary, whose values contains
# sufficient statistics for the action distribution. It should at least contain entries that would be
# returned by calling policy.dist_info(), which is the non-symbolic analog of policy.dist_info_sym().
# Storing these statistics is useful, e.g., when forming importance sampling ratios. In our case it is
# not needed.
action, _ = policy.get_action(observation)
# Recall that the last entry of the tuple stores diagnostic information about the environment. In our
# case it is not needed.
next_observation, reward, terminal, _ = env.step(action)
observations.append(observation)
actions.append(action)
rewards.append(reward)
observation = next_observation
if terminal:
# Finish rollout if terminal state reached
break
# We need to compute the empirical return for each time step along the
# trajectory
path = dict(
observations=np.array(observations),
actions=np.array(actions),
rewards=np.array(rewards),
)
path_baseline = baseline.predict(path)
advantages = []
returns = []
return_so_far = 0
for t in range(len(rewards) - 1, -1, -1):
return_so_far = rewards[t] + discount * return_so_far
returns.append(return_so_far)
advantage = return_so_far - path_baseline[t]
advantages.append(advantage)
# The advantages are stored backwards in time, so we need to revert it
advantages = np.array(advantages[::-1])
# And we need to do the same thing for the list of returns
returns = np.array(returns[::-1])
advantages = (advantages -
np.mean(advantages)) / (np.std(advantages) + 1e-8)
path["advantages"] = advantages
path["returns"] = returns
paths.append(path)
baseline.fit(paths)
observations = np.concatenate([p["observations"] for p in paths])
actions = np.concatenate([p["actions"] for p in paths])
advantages = np.concatenate([p["advantages"] for p in paths])
f_train(observations, actions, advantages)
returns_to_check = [sum(p["rewards"]) for p in paths]
print('Average Return:', np.mean(returns_to_check))
############################################################################
logger.log("Training finished")
logger.save_itr_params(epoch, params)
logger.dump_tabular(with_prefix=False)
logger.pop_prefix()
logger.record_tabular('Epoch', epoch)
logger.record_tabular('Steps', epoch * N * T)
logger.record_tabular('AverageReturn', np.mean(returns_to_check))
logger.record_tabular('StdReturn', np.std(returns_to_check))
logger.record_tabular('MaxReturn', np.max(returns_to_check))
logger.record_tabular('MinReturn', np.min(returns_to_check))
def doit(mode):
from rllab.envs.box2d.cartpole_env import CartpoleEnv
from rllab.baselines.linear_feature_baseline import LinearFeatureBaseline
from rllab.baselines.zero_baseline import ZeroBaseline
from rllab.policies.gaussian_mlp_policy import GaussianMLPPolicy
from rllab.envs.normalized_env import normalize
import numpy as np
import theano
import theano.tensor as TT
from lasagne.updates import adam
# normalize() makes sure that the actions for the environment lies
# within the range [-1, 1] (only works for environments with continuous actions)
env = normalize(CartpoleEnv())
# Initialize a neural network policy with a single hidden layer of 8 hidden units
policy = GaussianMLPPolicy(env.spec, hidden_sizes=(8,))
# Initialize a linear baseline estimator using default hand-crafted features
if "linbaseline" in mode:
print('linear baseline')
baseline = LinearFeatureBaseline(env.spec)
elif "vanilla" in mode:
print("zero baseline")
baseline = ZeroBaseline(env.spec)
elif mode == "batchavg":
print('batch average baseline')
# use a zero baseline but subtract the mean of the discounted returns (see below)
baseline = ZeroBaseline(env.spec)
if "_ztrans" in mode:
print('z transform advantages')
else:
print('no z transform')
# We will collect 100 trajectories per iteration
N = 50
# Each trajectory will have at most 100 time steps
T = 50
# Number of iterations
n_itr = 50
# Set the discount factor for the problem
discount = 0.99
# Learning rate for the gradient update
learning_rate = 0.1
# Construct the computation graph
# Create a Theano variable for storing the observations
# We could have simply written `observations_var = TT.matrix('observations')` instead for this example. However,
# doing it in a slightly more abstract way allows us to delegate to the environment for handling the correct data
# type for the variable. For instance, for an environment with discrete observations, we might want to use integer
# types if the observations are represented as one-hot vectors.
observations_var = env.observation_space.new_tensor_variable(
'observations',
# It should have 1 extra dimension since we want to represent a list of observations
extra_dims=1
)
actions_var = env.action_space.new_tensor_variable(
'actions',
extra_dims=1
)
advantages_var = TT.vector('advantages')
# policy.dist_info_sym returns a dictionary, whose values are symbolic expressions for quantities related to the
# distribution of the actions. For a Gaussian policy, it contains the mean and (log) standard deviation.
dist_info_vars = policy.dist_info_sym(observations_var)
# policy.distribution returns a distribution object under rllab.distributions. It contains many utilities for computing
# distribution-related quantities, given the computed dist_info_vars. Below we use dist.log_likelihood_sym to compute
# the symbolic log-likelihood. For this example, the corresponding distribution is an instance of the class
# rllab.distributions.DiagonalGaussian
dist = policy.distribution
# Note that we negate the objective, since most optimizers assume a
# minimization problem
surr = - TT.mean(dist.log_likelihood_sym(actions_var, dist_info_vars) * advantages_var)
# Get the list of trainable parameters.
params = policy.get_params(trainable=True)
grads = theano.grad(surr, params)
f_train = theano.function(
inputs=[observations_var, actions_var, advantages_var],
outputs=None,
updates=adam(grads, params, learning_rate=learning_rate),
allow_input_downcast=True
)
results = []
for _ in range(n_itr):
paths = []
for _ in range(N):
observations = []
actions = []
rewards = []
observation = env.reset()
for _ in range(T):
# policy.get_action() returns a pair of values. The second one returns a dictionary, whose values contains
# sufficient statistics for the action distribution. It should at least contain entries that would be
# returned by calling policy.dist_info(), which is the non-symbolic analog of policy.dist_info_sym().
# Storing these statistics is useful, e.g., when forming importance sampling ratios. In our case it is
# not needed.
action, _ = policy.get_action(observation)
# Recall that the last entry of the tuple stores diagnostic information about the environment. In our
# case it is not needed.
next_observation, reward, terminal, _ = env.step(action)
observations.append(observation)
actions.append(action)
rewards.append(reward)
observation = next_observation
if terminal:
# Finish rollout if terminal state reached
break
# We need to compute the empirical return for each time step along the
# trajectory
path = dict(
observations=np.array(observations),
actions=np.array(actions),
rewards=np.array(rewards),
)
path_baseline = baseline.predict(path)
advantages = []
returns = []
return_so_far = 0
for t in range(len(rewards) - 1, -1, -1):
return_so_far = rewards[t] + discount * return_so_far
returns.append(return_so_far)
advantage = return_so_far - path_baseline[t]
advantages.append(advantage)
# The advantages are stored backwards in time, so we need to revert it
advantages = np.array(advantages[::-1])
# And we need to do the same thing for the list of returns
returns = np.array(returns[::-1])
if "_ztrans" in mode:
advantages = (advantages - np.mean(advantages)) / (np.std(advantages) + 1e-8)
path["advantages"] = advantages
path["returns"] = returns
paths.append(path)
baseline.fit(paths)
observations = np.concatenate([p["observations"] for p in paths])
actions = np.concatenate([p["actions"] for p in paths])
advantages = np.concatenate([p["advantages"] for p in paths])
if mode == 'batchavg':
# in this case `advantages` up to here are just our good old returns, without baseline or z transformation.
# now we subtract their mean across all episodes.
advantages = advantages - np.mean(advantages)
f_train(observations, actions, advantages)
avgr = np.mean([sum(p["rewards"]) for p in paths])
print(('Average Return:',avgr))
results.append(avgr)
return results
class LowSampler(Sampler):
def __init__(self):
"""
:type algo: BatchPolopt
"""
env_low = normalize(AntEnv(ego_obs=True))
# baseline_low = LinearFeatureBaseline(env_spec=env_low.spec)
# low_policy = env.low_policy
pkl_path = '/home/lsy/Desktop/rllab/data/local/Ant-snn1000/Ant-snn_10MI_5grid_6latCat_bil_0040/params.pkl'
data = joblib.load(os.path.join(config.PROJECT_PATH, pkl_path))
low_policy = data['policy']
self.baseline = LinearFeatureBaseline(env_spec=env_low.spec)
self.discount = 0.99
self.gae_lambda = 1.0
self.center_adv = True
self.positive_adv = False
self.policy = low_policy
def process_samples(self, itr, paths):
baselines = []
returns = []
if hasattr(self.baseline, "predict_n"):
all_path_baselines = self.baseline.predict_n(paths)
else:
all_path_baselines = [self.baseline.predict(path) for path in paths]
for idx, path in enumerate(paths):
path_baselines = np.append(all_path_baselines[idx], 0)
deltas = path["rewards"] + \
self.discount * path_baselines[1:] - \
path_baselines[:-1]
path["advantages"] = special.discount_cumsum(
deltas, self.discount * self.gae_lambda)
path["returns"] = special.discount_cumsum(path["rewards"], self.discount)
baselines.append(path_baselines[:-1])
returns.append(path["returns"])
ev = special.explained_variance_1d(
np.concatenate(baselines),
np.concatenate(returns)
)
# if not self.algo.policy.recurrent:
observations = tensor_utils.concat_tensor_list([path["observations"] for path in paths])
actions = tensor_utils.concat_tensor_list([path["actions"] for path in paths])
rewards = tensor_utils.concat_tensor_list([path["rewards"] for path in paths])
returns = tensor_utils.concat_tensor_list([path["returns"] for path in paths])
advantages = tensor_utils.concat_tensor_list([path["advantages"] for path in paths])
env_infos = tensor_utils.concat_tensor_dict_list([path["env_infos"] for path in paths])
agent_infos = tensor_utils.concat_tensor_dict_list([path["agent_infos"] for path in paths])
if self.center_adv:
advantages = util.center_advantages(advantages)
if self.positive_adv:
advantages = util.shift_advantages_to_positive(advantages)
average_discounted_return = \
np.mean([path["returns"][0] for path in paths])
undiscounted_returns = [sum(path["rewards"]) for path in paths]
ent = np.mean(self.policy.distribution.entropy(agent_infos))
samples_data = dict(
observations=observations,
actions=actions,
rewards=rewards,
returns=returns,
advantages=advantages,
env_infos=env_infos,
agent_infos=agent_infos,
paths=paths,
)
logger.log("fitting baseline...")
if hasattr(self.baseline, 'fit_with_samples'):
self.baseline.fit_with_samples(paths, samples_data)
else:
self.baseline.fit(paths)
logger.log("fitted")
with logger.tabular_prefix('Low_'):
logger.record_tabular('Iteration', itr)
logger.record_tabular('AverageDiscountedReturn',
average_discounted_return)
logger.record_tabular('AverageReturn', np.mean(undiscounted_returns))
logger.record_tabular('ExplainedVariance', ev)
logger.record_tabular('NumTrajs', len(paths))
logger.record_tabular('Entropy', ent)
logger.record_tabular('Perplexity', np.exp(ent))
logger.record_tabular('StdReturn', np.std(undiscounted_returns))
logger.record_tabular('MaxReturn', np.max(undiscounted_returns))
logger.record_tabular('MinReturn', np.min(undiscounted_returns))
return samples_data
d_rewards = [p["rewards"] for p in paths]
temp = list()
tempa = list()
p_4b = []
for kx in range(N):
x = d_rewards[kx]
path = dict(
observations=np.array(observations[kx]),
actions=np.array(actions[kx]),
rewards=np.array(x),
)
z = list()
za = list()
z_rew = list()
t = 1
path_baseline = baseline.predict(path)
path_baseline = np.append(path_baseline, 0)
for yk in range(len(x)):
y = x[yk]
z.append(y * t)
za.append(
t *
(y -
(path_baseline[yk] - discount * path_baseline[yk + 1])))
t *= discount
return_so_far = 0
for t in range(len(x) - 1, -1, -1):
return_so_far = x[t] + discount * return_so_far
z_rew.append(return_so_far)
z_rew = np.array(z_rew[::-1])
temp.append(np.array(z))
