def __init__(self, environment, device, replay_buffer: ReplayBuffer, discount_factor: typing.Optional[float]=None,
normalize_returns: bool=False, forward_steps: int=1, action_noise: typing.Optional[nn.Module]=None):
self._environment = environment
self.device = device
self.replay_buffer = replay_buffer
self.normalize_returns = normalize_returns
self.forward_steps = forward_steps
self.discount_factor = discount_factor
self.action_noise = action_noise.to(self.device) if action_noise is not None else None
if self.normalize_returns:
assert self.discount_factor is not None, \
"TransitionReplayEnvRoller must have a discount factor defined if normalize_returns is turned on"
if self.forward_steps > 1:
assert self.discount_factor is not None, \
"TransitionReplayEnvRoller must have a discount factor defined if forward_steps is larger than one"
self.ret_rms = RunningMeanStd(shape=()) if normalize_returns else None
# Initial observation
self.last_observation_cpu = torch.from_numpy(self.environment.reset()).clone()
self.last_observation = self.last_observation_cpu.to(self.device)
# Return normalization
self.clip_obs = 5.0
self.accumulated_returns = np.zeros(environment.num_envs, dtype=np.float32)
def __init__(self, environment, device, batch_size, buffer_capacity, buffer_initial_size, noise_std_dev,
normalize_observations=False):
self.device = device
self.batch_size = batch_size
self.buffer_capacity = buffer_capacity
self.buffer_initial_size = buffer_initial_size
self.normalize_observations = normalize_observations
self.device = device
self._environment = environment
self.backend = DequeBufferBackend(
buffer_capacity=self.buffer_capacity,
observation_space=environment.observation_space,
action_space=environment.action_space
)
self.last_observation = self.environment.reset()
len_action_space = self.environment.action_space.shape[-1]
self.noise_process = OrnsteinUhlenbeckNoiseProcess(
np.zeros(len_action_space), float(noise_std_dev) * np.ones(len_action_space)
)
self.ob_rms = RunningMeanStd(shape=self.environment.observation_space.shape) if normalize_observations else None
self.clip_obs = 10.0
class VecNormalize(VecEnvWrapper):
"""
Vectorized environment base class
"""
def __init__(self,
venv,
ob=True,
ret=True,
clipob=10.,
cliprew=10.,
gamma=0.99,
epsilon=1e-8):
VecEnvWrapper.__init__(self, venv)
self.ob_rms = RunningMeanStd(
shape=self.observation_space.shape) if ob else None
self.ret_rms = RunningMeanStd(shape=()) if ret else None
self.clipob = clipob
self.cliprew = cliprew
self.ret = np.zeros(self.num_envs)
self.gamma = gamma
self.epsilon = epsilon
def step_wait(self):
"""
Apply sequence of actions to sequence of environments
actions -> (observations, rewards, news)
where 'news' is a boolean vector indicating whether each element is new.
"""
obs, rews, news, infos = self.venv.step_wait()
self.ret = self.ret * self.gamma + rews
obs = self._obfilt(obs)
if self.ret_rms:
self.ret_rms.update(self.ret)
rews = np.clip(rews / np.sqrt(self.ret_rms.var + self.epsilon),
-self.cliprew, self.cliprew)
return obs, rews, news, infos
def _obfilt(self, obs):
if self.ob_rms:
self.ob_rms.update(obs)
obs = np.clip((obs - self.ob_rms.mean) /
np.sqrt(self.ob_rms.var + self.epsilon),
-self.clipob, self.clipob)
return obs.astype(np.float32)
else:
return obs
def reset(self):
"""
Reset all environments
"""
obs = self.venv.reset()
return self._obfilt(obs)
def __init__(self, env, normalize_observations=True, normalize_returns=True,
clip_observations=10., clip_rewards=10., gamma=0.99, epsilon=1e-8):
super().__init__(env)
self.ob_rms = RunningMeanStd(shape=self.observation_space.shape) if normalize_observations else None
self.ret_rms = RunningMeanStd(shape=()) if normalize_returns else None
self.clipob = clip_observations
self.cliprew = clip_rewards
self.ret = 0.0
self.gamma = gamma
self.epsilon = epsilon
class EnvNormalize(gym.Wrapper):
"""
Single environment normalization based on VecNormalize from OpenAI baselines
"""
def __init__(self, env, normalize_observations=True, normalize_returns=True,
clip_observations=10., clip_rewards=10., gamma=0.99, epsilon=1e-8):
super().__init__(env)
self.ob_rms = RunningMeanStd(shape=self.observation_space.shape) if normalize_observations else None
self.ret_rms = RunningMeanStd(shape=()) if normalize_returns else None
self.clipob = clip_observations
self.cliprew = clip_rewards
self.ret = 0.0
self.gamma = gamma
self.epsilon = epsilon
def step(self, action):
"""
Apply sequence of actions to sequence of environments
actions -> (observations, rewards, news)
where 'news' is a boolean vector indicating whether each element is new.
"""
obs, rews, news, infos = self.env.step(action)
self.ret = self.ret * self.gamma + rews
obs = self._filter_observation(obs)
if self.ret_rms:
self.ret_rms.update(self.ret)
rews = np.clip(rews / np.sqrt(self.ret_rms.var + self.epsilon), -self.cliprew, self.cliprew)
return obs, rews, news, infos
def _filter_observation(self, obs):
if self.ob_rms:
self.ob_rms.update(obs)
obs = np.clip((obs - self.ob_rms.mean) / np.sqrt(self.ob_rms.var + self.epsilon), -self.clipob, self.clipob)
return obs.astype(np.float32)
else:
return obs
def reset(self):
"""
Reset all environments
"""
obs = self.env.reset()
return self._filter_observation(obs)
class VecNormalize(VecEnvWrapper):
"""
A vectorized wrapper that normalizes the observations
and returns from an environment.
"""
def __init__(self,
venv,
ob=True,
ret=True,
clipob=10.,
cliprew=10.,
gamma=0.99,
epsilon=1e-8):
VecEnvWrapper.__init__(self, venv)
self.ob_rms = RunningMeanStd(
shape=self.observation_space.shape) if ob else None
self.ret_rms = RunningMeanStd(shape=()) if ret else None
self.clipob = clipob
self.cliprew = cliprew
self.ret = np.zeros(self.num_envs)
self.gamma = gamma
self.epsilon = epsilon
def step_wait(self):
obs, rews, news, infos = self.venv.step_wait()
self.ret = self.ret * self.gamma + rews
obs = self._obfilt(obs)
if self.ret_rms:
self.ret_rms.update(self.ret)
rews = np.clip(rews / np.sqrt(self.ret_rms.var + self.epsilon),
-self.cliprew, self.cliprew)
self.ret[news] = 0.
return obs, rews, news, infos
def _obfilt(self, obs):
if self.ob_rms:
self.ob_rms.update(obs)
obs = np.clip((obs - self.ob_rms.mean) /
np.sqrt(self.ob_rms.var + self.epsilon),
-self.clipob, self.clipob)
return obs
else:
return obs
def reset(self):
self.ret = np.zeros(self.num_envs)
obs = self.venv.reset()
return self._obfilt(obs)
def __init__(self,
venv,
ob=True,
ret=True,
clipob=10.,
cliprew=10.,
gamma=0.99,
epsilon=1e-8):
VecEnvWrapper.__init__(self, venv)
self.ob_rms = RunningMeanStd(
shape=self.observation_space.shape) if ob else None
self.ret_rms = RunningMeanStd(shape=()) if ret else None
self.clipob = clipob
self.cliprew = cliprew
self.ret = np.zeros(self.num_envs)
self.gamma = gamma
self.epsilon = epsilon
class DequeReplayRollerOuNoise(ReplayEnvRollerBase):
"""
Enrionment roller with experience replay buffer rolling out a **single** environment
with Ornstein–Uhlenbeck noise process
"""
def __init__(self, environment, device, batch_size, buffer_capacity, buffer_initial_size, noise_std_dev,
normalize_observations=False):
self.device = device
self.batch_size = batch_size
self.buffer_capacity = buffer_capacity
self.buffer_initial_size = buffer_initial_size
self.normalize_observations = normalize_observations
self.device = device
self._environment = environment
self.backend = DequeBufferBackend(
buffer_capacity=self.buffer_capacity,
observation_space=environment.observation_space,
action_space=environment.action_space
)
self.last_observation = self.environment.reset()
len_action_space = self.environment.action_space.shape[-1]
self.noise_process = OrnsteinUhlenbeckNoiseProcess(
np.zeros(len_action_space), float(noise_std_dev) * np.ones(len_action_space)
)
self.ob_rms = RunningMeanStd(shape=self.environment.observation_space.shape) if normalize_observations else None
self.clip_obs = 10.0
@property
def environment(self):
""" Return environment of this env roller """
return self._environment
def is_ready_for_sampling(self) -> bool:
""" If buffer is ready for drawing samples from it (usually checks if there is enough data) """
return self.backend.current_size >= self.buffer_initial_size
@torch.no_grad()
def rollout(self, batch_info, model) -> Rollout:
""" Roll-out the environment and return it """
observation_tensor = torch.from_numpy(self.last_observation).to(self.device)
step = model.step(observation_tensor[None])
action = step['actions'].detach().cpu().numpy()[0]
noise = self.noise_process()
action_perturbed = np.clip(
action + noise, self.environment.action_space.low, self.environment.action_space.high
)
observation, reward, done, info = self.environment.step(action_perturbed)
if self.ob_rms is not None:
self.ob_rms.update(observation)
self.backend.store_transition(self.last_observation, action_perturbed, reward, done)
# Usual, reset on done
if done:
observation = self.environment.reset()
self.noise_process.reset()
self.last_observation = observation
return Transitions(
size=1,
environment_information=[info],
transition_tensors={
'actions': step['actions'],
'values': step['values']
},
)
def _filter_observation(self, obs):
""" Potentially normalize observation """
if self.ob_rms is not None:
obs = np.clip((obs - self.ob_rms.mean) / np.sqrt(self.ob_rms.var + 1e-8), -self.clip_obs, self.clip_obs)
return obs.astype(np.float32)
else:
return obs
def sample(self, batch_info, model) -> Transitions:
""" Sample experience from replay buffer and return a batch """
indexes = self.backend.sample_batch_uniform(self.batch_size, history_length=1)
batch = self.backend.get_batch(indexes, history_length=1)
observations = torch.from_numpy(self._filter_observation(batch['states'])).to(self.device)
observations_plus1 = torch.from_numpy(self._filter_observation(batch['states+1'])).to(self.device)
dones = torch.from_numpy(batch['dones'].astype(np.float32)).to(self.device)
rewards = torch.from_numpy(batch['rewards'].astype(np.float32)).to(self.device)
actions = torch.from_numpy(batch['actions']).to(self.device)
return Transitions(
size=self.batch_size,
environment_information=[],
transition_tensors={
'observations': observations,
'observations_next': observations_plus1,
'dones': dones,
'rewards': rewards,
'actions': actions
}
)
class TransitionReplayEnvRoller(ReplayEnvRollerBase):
"""
Calculate environment rollouts using a replay buffer for experience replay.
Replay buffer is parametrized
Samples transitions from the replay buffer (individual frame transitions)
"""
def __init__(self, environment, device, replay_buffer: ReplayBuffer, discount_factor: typing.Optional[float]=None,
normalize_returns: bool=False, forward_steps: int=1, action_noise: typing.Optional[nn.Module]=None):
self._environment = environment
self.device = device
self.replay_buffer = replay_buffer
self.normalize_returns = normalize_returns
self.forward_steps = forward_steps
self.discount_factor = discount_factor
self.action_noise = action_noise.to(self.device) if action_noise is not None else None
if self.normalize_returns:
assert self.discount_factor is not None, \
"TransitionReplayEnvRoller must have a discount factor defined if normalize_returns is turned on"
if self.forward_steps > 1:
assert self.discount_factor is not None, \
"TransitionReplayEnvRoller must have a discount factor defined if forward_steps is larger than one"
self.ret_rms = RunningMeanStd(shape=()) if normalize_returns else None
# Initial observation
self.last_observation_cpu = torch.from_numpy(self.environment.reset()).clone()
self.last_observation = self.last_observation_cpu.to(self.device)
# Return normalization
self.clip_obs = 5.0
self.accumulated_returns = np.zeros(environment.num_envs, dtype=np.float32)
@property
def environment(self):
""" Return environment of this env roller """
return self._environment
@torch.no_grad()
def rollout(self, batch_info: BatchInfo, model: RlModel, number_of_steps: int) -> Rollout:
""" Calculate env rollout """
assert not model.is_recurrent, "Replay env roller does not support recurrent models"
accumulator = TensorAccumulator()
episode_information = [] # List of dictionaries with episode information
for step_idx in range(number_of_steps):
step = model.step(self.last_observation)
if self.action_noise is not None:
step['actions'] = self.action_noise(step['actions'], batch_info=batch_info)
replay_extra_information = {}
accumulator.add('observations', self.last_observation_cpu)
# Add step to the tensor accumulator
for name, tensor in step.items():
tensor_cpu = tensor.cpu()
accumulator.add(name, tensor_cpu)
if name != 'actions':
replay_extra_information[name] = tensor_cpu.numpy()
actions_numpy = step['actions'].detach().cpu().numpy()
new_obs, new_rewards, new_dones, new_infos = self.environment.step(actions_numpy)
# Store rollout in the experience replay buffer
self.replay_buffer.store_transition(
frame=self.last_observation_cpu.numpy(),
action=actions_numpy,
reward=new_rewards,
done=new_dones,
extra_info=replay_extra_information
)
if self.ret_rms is not None:
self.accumulated_returns = new_rewards + self.discount_factor * self.accumulated_returns
self.ret_rms.update(self.accumulated_returns)
# Done is flagged true when the episode has ended AND the frame we see is already a first frame from the
# next episode
dones_tensor = torch.from_numpy(new_dones.astype(np.float32)).clone()
accumulator.add('dones', dones_tensor)
if self.action_noise is not None:
self.action_noise.reset_training_state(dones_tensor, batch_info=batch_info)
self.accumulated_returns = self.accumulated_returns * (1.0 - new_dones.astype(np.float32))
self.last_observation_cpu = torch.from_numpy(new_obs).clone()
self.last_observation = self.last_observation_cpu.to(self.device)
if self.ret_rms is not None:
new_rewards = np.clip(new_rewards / np.sqrt(self.ret_rms.var + 1e-8), -self.clip_obs, self.clip_obs)
accumulator.add('rewards', torch.from_numpy(new_rewards.astype(np.float32)).clone())
episode_information.append(new_infos)
accumulated_tensors = accumulator.result()
return Trajectories(
num_steps=accumulated_tensors['observations'].size(0),
num_envs=accumulated_tensors['observations'].size(1),
environment_information=episode_information,
transition_tensors=accumulated_tensors,
rollout_tensors={}
).to_transitions()
def sample(self, batch_info: BatchInfo, model: RlModel, number_of_steps: int) -> Rollout:
""" Sample experience from replay buffer and return a batch """
if self.forward_steps > 1:
transitions = self.replay_buffer.sample_forward_transitions(
batch_size=number_of_steps, batch_info=batch_info, forward_steps=self.forward_steps,
discount_factor=self.discount_factor
)
else:
transitions = self.replay_buffer.sample_transitions(batch_size=number_of_steps, batch_info=batch_info)
if self.ret_rms is not None:
rewards = transitions.transition_tensors['rewards']
new_rewards = torch.clamp(rewards / np.sqrt(self.ret_rms.var + 1e-8), -self.clip_obs, self.clip_obs)
transitions.transition_tensors['rewards'] = new_rewards
return transitions
def is_ready_for_sampling(self) -> bool:
""" If buffer is ready for drawing samples from it (usually checks if there is enough data) """
return self.replay_buffer.is_ready_for_sampling()
def initial_memory_size_hint(self) -> typing.Optional[int]:
""" Hint how much data is needed to begin sampling, required only for diagnostics """
return self.replay_buffer.initial_memory_size_hint()
def update(self, rollout, batch_info):
""" Perform update of the internal state of the buffer - e.g. for the prioritized replay weights """
self.replay_buffer.update(rollout, batch_info)
