Spaces:
Sleeping
Sleeping
File size: 11,646 Bytes
e11e4fe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 |
from typing import List, NamedTuple
import numpy as np
from mlagents.trainers.buffer import (
AgentBuffer,
ObservationKeyPrefix,
AgentBufferKey,
BufferKey,
)
from mlagents_envs.base_env import ActionTuple
from mlagents.trainers.torch_entities.action_log_probs import LogProbsTuple
class AgentStatus(NamedTuple):
"""
Stores observation, action, and reward for an agent. Does not have additional
fields that are present in AgentExperience.
"""
obs: List[np.ndarray]
reward: float
action: ActionTuple
done: bool
class AgentExperience(NamedTuple):
"""
Stores the full amount of data for an agent in one timestep. Includes
the status' of group mates and the group reward, as well as the probabilities
outputted by the policy.
"""
obs: List[np.ndarray]
reward: float
done: bool
action: ActionTuple
action_probs: LogProbsTuple
action_mask: np.ndarray
prev_action: np.ndarray
interrupted: bool
memory: np.ndarray
group_status: List[AgentStatus]
group_reward: float
class ObsUtil:
@staticmethod
def get_name_at(index: int) -> AgentBufferKey:
"""
returns the name of the observation given the index of the observation
"""
return ObservationKeyPrefix.OBSERVATION, index
@staticmethod
def get_name_at_next(index: int) -> AgentBufferKey:
"""
returns the name of the next observation given the index of the observation
"""
return ObservationKeyPrefix.NEXT_OBSERVATION, index
@staticmethod
def from_buffer(batch: AgentBuffer, num_obs: int) -> List[np.array]:
"""
Creates the list of observations from an AgentBuffer
"""
result: List[np.array] = []
for i in range(num_obs):
result.append(batch[ObsUtil.get_name_at(i)])
return result
@staticmethod
def from_buffer_next(batch: AgentBuffer, num_obs: int) -> List[np.array]:
"""
Creates the list of next observations from an AgentBuffer
"""
result = []
for i in range(num_obs):
result.append(batch[ObsUtil.get_name_at_next(i)])
return result
class GroupObsUtil:
@staticmethod
def get_name_at(index: int) -> AgentBufferKey:
"""
returns the name of the observation given the index of the observation
"""
return ObservationKeyPrefix.GROUP_OBSERVATION, index
@staticmethod
def get_name_at_next(index: int) -> AgentBufferKey:
"""
returns the name of the next team observation given the index of the observation
"""
return ObservationKeyPrefix.NEXT_GROUP_OBSERVATION, index
@staticmethod
def _transpose_list_of_lists(
list_list: List[List[np.ndarray]],
) -> List[List[np.ndarray]]:
return list(map(list, zip(*list_list)))
@staticmethod
def from_buffer(batch: AgentBuffer, num_obs: int) -> List[np.array]:
"""
Creates the list of observations from an AgentBuffer
"""
separated_obs: List[np.array] = []
for i in range(num_obs):
separated_obs.append(
batch[GroupObsUtil.get_name_at(i)].padded_to_batch(pad_value=np.nan)
)
# separated_obs contains a List(num_obs) of Lists(num_agents), we want to flip
# that and get a List(num_agents) of Lists(num_obs)
result = GroupObsUtil._transpose_list_of_lists(separated_obs)
return result
@staticmethod
def from_buffer_next(batch: AgentBuffer, num_obs: int) -> List[np.array]:
"""
Creates the list of observations from an AgentBuffer
"""
separated_obs: List[np.array] = []
for i in range(num_obs):
separated_obs.append(
batch[GroupObsUtil.get_name_at_next(i)].padded_to_batch(
pad_value=np.nan
)
)
# separated_obs contains a List(num_obs) of Lists(num_agents), we want to flip
# that and get a List(num_agents) of Lists(num_obs)
result = GroupObsUtil._transpose_list_of_lists(separated_obs)
return result
class Trajectory(NamedTuple):
steps: List[AgentExperience]
next_obs: List[
np.ndarray
] # Observation following the trajectory, for bootstrapping
next_group_obs: List[List[np.ndarray]]
agent_id: str
behavior_id: str
def to_agentbuffer(self) -> AgentBuffer:
"""
Converts a Trajectory to an AgentBuffer
:param trajectory: A Trajectory
:returns: AgentBuffer. Note that the length of the AgentBuffer will be one
less than the trajectory, as the next observation need to be populated from the last
step of the trajectory.
"""
agent_buffer_trajectory = AgentBuffer()
obs = self.steps[0].obs
for step, exp in enumerate(self.steps):
is_last_step = step == len(self.steps) - 1
if not is_last_step:
next_obs = self.steps[step + 1].obs
else:
next_obs = self.next_obs
num_obs = len(obs)
for i in range(num_obs):
agent_buffer_trajectory[ObsUtil.get_name_at(i)].append(obs[i])
agent_buffer_trajectory[ObsUtil.get_name_at_next(i)].append(next_obs[i])
# Take care of teammate obs and actions
teammate_continuous_actions, teammate_discrete_actions, teammate_rewards = (
[],
[],
[],
)
for group_status in exp.group_status:
teammate_rewards.append(group_status.reward)
teammate_continuous_actions.append(group_status.action.continuous)
teammate_discrete_actions.append(group_status.action.discrete)
# Team actions
agent_buffer_trajectory[BufferKey.GROUP_CONTINUOUS_ACTION].append(
teammate_continuous_actions
)
agent_buffer_trajectory[BufferKey.GROUP_DISCRETE_ACTION].append(
teammate_discrete_actions
)
agent_buffer_trajectory[BufferKey.GROUPMATE_REWARDS].append(
teammate_rewards
)
agent_buffer_trajectory[BufferKey.GROUP_REWARD].append(exp.group_reward)
# Next actions
teammate_cont_next_actions = []
teammate_disc_next_actions = []
if not is_last_step:
next_exp = self.steps[step + 1]
for group_status in next_exp.group_status:
teammate_cont_next_actions.append(group_status.action.continuous)
teammate_disc_next_actions.append(group_status.action.discrete)
else:
for group_status in exp.group_status:
teammate_cont_next_actions.append(group_status.action.continuous)
teammate_disc_next_actions.append(group_status.action.discrete)
agent_buffer_trajectory[BufferKey.GROUP_NEXT_CONT_ACTION].append(
teammate_cont_next_actions
)
agent_buffer_trajectory[BufferKey.GROUP_NEXT_DISC_ACTION].append(
teammate_disc_next_actions
)
for i in range(num_obs):
ith_group_obs = []
for _group_status in exp.group_status:
# Assume teammates have same obs space
ith_group_obs.append(_group_status.obs[i])
agent_buffer_trajectory[GroupObsUtil.get_name_at(i)].append(
ith_group_obs
)
ith_group_obs_next = []
if is_last_step:
for _obs in self.next_group_obs:
ith_group_obs_next.append(_obs[i])
else:
next_group_status = self.steps[step + 1].group_status
for _group_status in next_group_status:
# Assume teammates have same obs space
ith_group_obs_next.append(_group_status.obs[i])
agent_buffer_trajectory[GroupObsUtil.get_name_at_next(i)].append(
ith_group_obs_next
)
if exp.memory is not None:
agent_buffer_trajectory[BufferKey.MEMORY].append(exp.memory)
agent_buffer_trajectory[BufferKey.MASKS].append(1.0)
agent_buffer_trajectory[BufferKey.DONE].append(exp.done)
agent_buffer_trajectory[BufferKey.GROUP_DONES].append(
[_status.done for _status in exp.group_status]
)
# Adds the log prob and action of continuous/discrete separately
agent_buffer_trajectory[BufferKey.CONTINUOUS_ACTION].append(
exp.action.continuous
)
agent_buffer_trajectory[BufferKey.DISCRETE_ACTION].append(
exp.action.discrete
)
if not is_last_step:
next_action = self.steps[step + 1].action
cont_next_actions = next_action.continuous
disc_next_actions = next_action.discrete
else:
cont_next_actions = np.zeros_like(exp.action.continuous)
disc_next_actions = np.zeros_like(exp.action.discrete)
agent_buffer_trajectory[BufferKey.NEXT_CONT_ACTION].append(
cont_next_actions
)
agent_buffer_trajectory[BufferKey.NEXT_DISC_ACTION].append(
disc_next_actions
)
agent_buffer_trajectory[BufferKey.CONTINUOUS_LOG_PROBS].append(
exp.action_probs.continuous
)
agent_buffer_trajectory[BufferKey.DISCRETE_LOG_PROBS].append(
exp.action_probs.discrete
)
# Store action masks if necessary. Note that 1 means active, while
# in AgentExperience False means active.
if exp.action_mask is not None:
mask = 1 - np.concatenate(exp.action_mask)
agent_buffer_trajectory[BufferKey.ACTION_MASK].append(
mask, padding_value=1
)
else:
# This should never be needed unless the environment somehow doesn't supply the
# action mask in a discrete space.
action_shape = exp.action.discrete.shape
agent_buffer_trajectory[BufferKey.ACTION_MASK].append(
np.ones(action_shape, dtype=np.float32), padding_value=1
)
agent_buffer_trajectory[BufferKey.PREV_ACTION].append(exp.prev_action)
agent_buffer_trajectory[BufferKey.ENVIRONMENT_REWARDS].append(exp.reward)
# Store the next visual obs as the current
obs = next_obs
return agent_buffer_trajectory
@property
def done_reached(self) -> bool:
"""
Returns true if trajectory is terminated with a Done.
"""
return self.steps[-1].done
@property
def all_group_dones_reached(self) -> bool:
"""
Returns true if all other agents in this trajectory are done at the end of the trajectory.
Combine with done_reached to check if the whole team is done.
"""
return all(_status.done for _status in self.steps[-1].group_status)
@property
def interrupted(self) -> bool:
"""
Returns true if trajectory was terminated because max steps was reached.
"""
return self.steps[-1].interrupted
|