Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +1 -0
- parrot/lib/python3.10/site-packages/numpy/_core/tests/__pycache__/test_ufunc.cpython-310.pyc +3 -0
- parrot/lib/python3.10/site-packages/scipy/io/__pycache__/_idl.cpython-310.pyc +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/__pycache__/idl.cpython-310.pyc +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/__pycache__/mmio.cpython-310.pyc +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/__pycache__/netcdf.cpython-310.pyc +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/tests/data/fortran-sf8-15x10x22.dat +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/tests/data/fortran-si4-1x1x5.dat +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/tests/data/scalar_uint64.sav +0 -0
- parrot/lib/python3.10/site-packages/scipy/io/tests/data/struct_pointers_replicated.sav +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/base_env.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/remote_base_env.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/single_agent_episode.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__init__.py +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__pycache__/__init__.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__pycache__/task_settable_env.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/external_multi_agent_env.py +161 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/remote_base_env.py +462 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/utils/__init__.py +124 -0
- vllm/lib/python3.10/site-packages/ray/rllib/env/utils/infinite_lookback_buffer.py +707 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/__init__.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/action_dist.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/catalog.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/distributions.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/modelv2.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/preprocessors.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/repeated_values.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/utils.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/preprocessors.py +447 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/repeated_values.py +204 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/attention_net.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/complex_input_net.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/fcnet.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/misc.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/noop.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/recurrent_net.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/tf_distributions.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/visionnet.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/fcnet.py +148 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/__init__.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/gru_gate.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/multi_head_attention.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/noisy_layer.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/relative_multi_head_attention.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/skip_connection.cpython-310.pyc +0 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/gru_gate.py +58 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/multi_head_attention.py +61 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/noisy_layer.py +118 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/relative_multi_head_attention.py +147 -0
- vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/skip_connection.py +46 -0
.gitattributes
CHANGED
|
@@ -1452,3 +1452,4 @@ parrot/lib/python3.10/site-packages/numpy/linalg/__pycache__/_linalg.cpython-310
|
|
| 1452 |
vllm/lib/python3.10/site-packages/ray/rllib/algorithms/__pycache__/algorithm_config.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
|
| 1453 |
parrot/lib/python3.10/site-packages/numpy/_core/lib/libnpymath.a filter=lfs diff=lfs merge=lfs -text
|
| 1454 |
parrot/lib/python3.10/site-packages/numpy/_core/_multiarray_tests.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 1452 |
vllm/lib/python3.10/site-packages/ray/rllib/algorithms/__pycache__/algorithm_config.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
|
| 1453 |
parrot/lib/python3.10/site-packages/numpy/_core/lib/libnpymath.a filter=lfs diff=lfs merge=lfs -text
|
| 1454 |
parrot/lib/python3.10/site-packages/numpy/_core/_multiarray_tests.cpython-310-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
|
| 1455 |
+
parrot/lib/python3.10/site-packages/numpy/_core/tests/__pycache__/test_ufunc.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
|
parrot/lib/python3.10/site-packages/numpy/_core/tests/__pycache__/test_ufunc.cpython-310.pyc
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6a4d941a6095db86ad6e706837ad9fb6bb6ee240ed2aada924c37002d0a75470
|
| 3 |
+
size 105385
|
parrot/lib/python3.10/site-packages/scipy/io/__pycache__/_idl.cpython-310.pyc
ADDED
|
Binary file (19.2 kB). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/__pycache__/idl.cpython-310.pyc
ADDED
|
Binary file (564 Bytes). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/__pycache__/mmio.cpython-310.pyc
ADDED
|
Binary file (587 Bytes). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/__pycache__/netcdf.cpython-310.pyc
ADDED
|
Binary file (596 Bytes). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/tests/data/fortran-sf8-15x10x22.dat
ADDED
|
Binary file (26.4 kB). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/tests/data/fortran-si4-1x1x5.dat
ADDED
|
Binary file (28 Bytes). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/tests/data/scalar_uint64.sav
ADDED
|
Binary file (2.08 kB). View file
|
|
|
parrot/lib/python3.10/site-packages/scipy/io/tests/data/struct_pointers_replicated.sav
ADDED
|
Binary file (2.3 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/base_env.cpython-310.pyc
ADDED
|
Binary file (15 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/remote_base_env.cpython-310.pyc
ADDED
|
Binary file (12.4 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/__pycache__/single_agent_episode.cpython-310.pyc
ADDED
|
Binary file (74.2 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__init__.py
ADDED
|
File without changes
|
vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__pycache__/__init__.cpython-310.pyc
ADDED
|
Binary file (168 Bytes). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/apis/__pycache__/task_settable_env.cpython-310.pyc
ADDED
|
Binary file (2.09 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/env/external_multi_agent_env.py
ADDED
|
@@ -0,0 +1,161 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import uuid
|
| 2 |
+
import gymnasium as gym
|
| 3 |
+
from typing import Optional
|
| 4 |
+
|
| 5 |
+
from ray.rllib.utils.annotations import override, OldAPIStack
|
| 6 |
+
from ray.rllib.env.external_env import ExternalEnv, _ExternalEnvEpisode
|
| 7 |
+
from ray.rllib.utils.typing import MultiAgentDict
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
@OldAPIStack
|
| 11 |
+
class ExternalMultiAgentEnv(ExternalEnv):
|
| 12 |
+
"""This is the multi-agent version of ExternalEnv."""
|
| 13 |
+
|
| 14 |
+
def __init__(
|
| 15 |
+
self,
|
| 16 |
+
action_space: gym.Space,
|
| 17 |
+
observation_space: gym.Space,
|
| 18 |
+
):
|
| 19 |
+
"""Initializes an ExternalMultiAgentEnv instance.
|
| 20 |
+
|
| 21 |
+
Args:
|
| 22 |
+
action_space: Action space of the env.
|
| 23 |
+
observation_space: Observation space of the env.
|
| 24 |
+
"""
|
| 25 |
+
ExternalEnv.__init__(self, action_space, observation_space)
|
| 26 |
+
|
| 27 |
+
# We require to know all agents' spaces.
|
| 28 |
+
if isinstance(self.action_space, dict) or isinstance(
|
| 29 |
+
self.observation_space, dict
|
| 30 |
+
):
|
| 31 |
+
if not (self.action_space.keys() == self.observation_space.keys()):
|
| 32 |
+
raise ValueError(
|
| 33 |
+
"Agent ids disagree for action space and obs "
|
| 34 |
+
"space dict: {} {}".format(
|
| 35 |
+
self.action_space.keys(), self.observation_space.keys()
|
| 36 |
+
)
|
| 37 |
+
)
|
| 38 |
+
|
| 39 |
+
def run(self):
|
| 40 |
+
"""Override this to implement the multi-agent run loop.
|
| 41 |
+
|
| 42 |
+
Your loop should continuously:
|
| 43 |
+
1. Call self.start_episode(episode_id)
|
| 44 |
+
2. Call self.get_action(episode_id, obs_dict)
|
| 45 |
+
-or-
|
| 46 |
+
self.log_action(episode_id, obs_dict, action_dict)
|
| 47 |
+
3. Call self.log_returns(episode_id, reward_dict)
|
| 48 |
+
4. Call self.end_episode(episode_id, obs_dict)
|
| 49 |
+
5. Wait if nothing to do.
|
| 50 |
+
|
| 51 |
+
Multiple episodes may be started at the same time.
|
| 52 |
+
"""
|
| 53 |
+
raise NotImplementedError
|
| 54 |
+
|
| 55 |
+
@override(ExternalEnv)
|
| 56 |
+
def start_episode(
|
| 57 |
+
self, episode_id: Optional[str] = None, training_enabled: bool = True
|
| 58 |
+
) -> str:
|
| 59 |
+
if episode_id is None:
|
| 60 |
+
episode_id = uuid.uuid4().hex
|
| 61 |
+
|
| 62 |
+
if episode_id in self._finished:
|
| 63 |
+
raise ValueError("Episode {} has already completed.".format(episode_id))
|
| 64 |
+
|
| 65 |
+
if episode_id in self._episodes:
|
| 66 |
+
raise ValueError("Episode {} is already started".format(episode_id))
|
| 67 |
+
|
| 68 |
+
self._episodes[episode_id] = _ExternalEnvEpisode(
|
| 69 |
+
episode_id, self._results_avail_condition, training_enabled, multiagent=True
|
| 70 |
+
)
|
| 71 |
+
|
| 72 |
+
return episode_id
|
| 73 |
+
|
| 74 |
+
@override(ExternalEnv)
|
| 75 |
+
def get_action(
|
| 76 |
+
self, episode_id: str, observation_dict: MultiAgentDict
|
| 77 |
+
) -> MultiAgentDict:
|
| 78 |
+
"""Record an observation and get the on-policy action.
|
| 79 |
+
|
| 80 |
+
Thereby, observation_dict is expected to contain the observation
|
| 81 |
+
of all agents acting in this episode step.
|
| 82 |
+
|
| 83 |
+
Args:
|
| 84 |
+
episode_id: Episode id returned from start_episode().
|
| 85 |
+
observation_dict: Current environment observation.
|
| 86 |
+
|
| 87 |
+
Returns:
|
| 88 |
+
action: Action from the env action space.
|
| 89 |
+
"""
|
| 90 |
+
|
| 91 |
+
episode = self._get(episode_id)
|
| 92 |
+
return episode.wait_for_action(observation_dict)
|
| 93 |
+
|
| 94 |
+
@override(ExternalEnv)
|
| 95 |
+
def log_action(
|
| 96 |
+
self,
|
| 97 |
+
episode_id: str,
|
| 98 |
+
observation_dict: MultiAgentDict,
|
| 99 |
+
action_dict: MultiAgentDict,
|
| 100 |
+
) -> None:
|
| 101 |
+
"""Record an observation and (off-policy) action taken.
|
| 102 |
+
|
| 103 |
+
Args:
|
| 104 |
+
episode_id: Episode id returned from start_episode().
|
| 105 |
+
observation_dict: Current environment observation.
|
| 106 |
+
action_dict: Action for the observation.
|
| 107 |
+
"""
|
| 108 |
+
|
| 109 |
+
episode = self._get(episode_id)
|
| 110 |
+
episode.log_action(observation_dict, action_dict)
|
| 111 |
+
|
| 112 |
+
@override(ExternalEnv)
|
| 113 |
+
def log_returns(
|
| 114 |
+
self,
|
| 115 |
+
episode_id: str,
|
| 116 |
+
reward_dict: MultiAgentDict,
|
| 117 |
+
info_dict: MultiAgentDict = None,
|
| 118 |
+
multiagent_done_dict: MultiAgentDict = None,
|
| 119 |
+
) -> None:
|
| 120 |
+
"""Record returns from the environment.
|
| 121 |
+
|
| 122 |
+
The reward will be attributed to the previous action taken by the
|
| 123 |
+
episode. Rewards accumulate until the next action. If no reward is
|
| 124 |
+
logged before the next action, a reward of 0.0 is assumed.
|
| 125 |
+
|
| 126 |
+
Args:
|
| 127 |
+
episode_id: Episode id returned from start_episode().
|
| 128 |
+
reward_dict: Reward from the environment agents.
|
| 129 |
+
info_dict: Optional info dict.
|
| 130 |
+
multiagent_done_dict: Optional done dict for agents.
|
| 131 |
+
"""
|
| 132 |
+
|
| 133 |
+
episode = self._get(episode_id)
|
| 134 |
+
|
| 135 |
+
# Accumulate reward by agent.
|
| 136 |
+
# For existing agents, we want to add the reward up.
|
| 137 |
+
for agent, rew in reward_dict.items():
|
| 138 |
+
if agent in episode.cur_reward_dict:
|
| 139 |
+
episode.cur_reward_dict[agent] += rew
|
| 140 |
+
else:
|
| 141 |
+
episode.cur_reward_dict[agent] = rew
|
| 142 |
+
|
| 143 |
+
if multiagent_done_dict:
|
| 144 |
+
for agent, done in multiagent_done_dict.items():
|
| 145 |
+
episode.cur_done_dict[agent] = done
|
| 146 |
+
|
| 147 |
+
if info_dict:
|
| 148 |
+
episode.cur_info_dict = info_dict or {}
|
| 149 |
+
|
| 150 |
+
@override(ExternalEnv)
|
| 151 |
+
def end_episode(self, episode_id: str, observation_dict: MultiAgentDict) -> None:
|
| 152 |
+
"""Record the end of an episode.
|
| 153 |
+
|
| 154 |
+
Args:
|
| 155 |
+
episode_id: Episode id returned from start_episode().
|
| 156 |
+
observation_dict: Current environment observation.
|
| 157 |
+
"""
|
| 158 |
+
|
| 159 |
+
episode = self._get(episode_id)
|
| 160 |
+
self._finished.add(episode.episode_id)
|
| 161 |
+
episode.done(observation_dict)
|
vllm/lib/python3.10/site-packages/ray/rllib/env/remote_base_env.py
ADDED
|
@@ -0,0 +1,462 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gymnasium as gym
|
| 2 |
+
import logging
|
| 3 |
+
from typing import Callable, Dict, List, Optional, Set, Tuple, TYPE_CHECKING
|
| 4 |
+
|
| 5 |
+
import ray
|
| 6 |
+
from ray.util import log_once
|
| 7 |
+
from ray.rllib.env.base_env import BaseEnv, _DUMMY_AGENT_ID, ASYNC_RESET_RETURN
|
| 8 |
+
from ray.rllib.utils.annotations import override, OldAPIStack
|
| 9 |
+
from ray.rllib.utils.typing import AgentID, EnvID, EnvType, MultiEnvDict
|
| 10 |
+
|
| 11 |
+
if TYPE_CHECKING:
|
| 12 |
+
from ray.rllib.evaluation.rollout_worker import RolloutWorker
|
| 13 |
+
|
| 14 |
+
logger = logging.getLogger(__name__)
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
@OldAPIStack
|
| 18 |
+
class RemoteBaseEnv(BaseEnv):
|
| 19 |
+
"""BaseEnv that executes its sub environments as @ray.remote actors.
|
| 20 |
+
|
| 21 |
+
This provides dynamic batching of inference as observations are returned
|
| 22 |
+
from the remote simulator actors. Both single and multi-agent child envs
|
| 23 |
+
are supported, and envs can be stepped synchronously or asynchronously.
|
| 24 |
+
|
| 25 |
+
NOTE: This class implicitly assumes that the remote envs are gym.Env's
|
| 26 |
+
|
| 27 |
+
You shouldn't need to instantiate this class directly. It's automatically
|
| 28 |
+
inserted when you use the `remote_worker_envs=True` option in your
|
| 29 |
+
Algorithm's config.
|
| 30 |
+
"""
|
| 31 |
+
|
| 32 |
+
def __init__(
|
| 33 |
+
self,
|
| 34 |
+
make_env: Callable[[int], EnvType],
|
| 35 |
+
num_envs: int,
|
| 36 |
+
multiagent: bool,
|
| 37 |
+
remote_env_batch_wait_ms: int,
|
| 38 |
+
existing_envs: Optional[List[ray.actor.ActorHandle]] = None,
|
| 39 |
+
worker: Optional["RolloutWorker"] = None,
|
| 40 |
+
restart_failed_sub_environments: bool = False,
|
| 41 |
+
):
|
| 42 |
+
"""Initializes a RemoteVectorEnv instance.
|
| 43 |
+
|
| 44 |
+
Args:
|
| 45 |
+
make_env: Callable that produces a single (non-vectorized) env,
|
| 46 |
+
given the vector env index as only arg.
|
| 47 |
+
num_envs: The number of sub-environments to create for the
|
| 48 |
+
vectorization.
|
| 49 |
+
multiagent: Whether this is a multiagent env or not.
|
| 50 |
+
remote_env_batch_wait_ms: Time to wait for (ray.remote)
|
| 51 |
+
sub-environments to have new observations available when
|
| 52 |
+
polled. Only when none of the sub-environments is ready,
|
| 53 |
+
repeat the `ray.wait()` call until at least one sub-env
|
| 54 |
+
is ready. Then return only the observations of the ready
|
| 55 |
+
sub-environment(s).
|
| 56 |
+
existing_envs: Optional list of already created sub-environments.
|
| 57 |
+
These will be used as-is and only as many new sub-envs as
|
| 58 |
+
necessary (`num_envs - len(existing_envs)`) will be created.
|
| 59 |
+
worker: An optional RolloutWorker that owns the env. This is only
|
| 60 |
+
used if `remote_worker_envs` is True in your config and the
|
| 61 |
+
`on_sub_environment_created` custom callback needs to be
|
| 62 |
+
called on each created actor.
|
| 63 |
+
restart_failed_sub_environments: If True and any sub-environment (within
|
| 64 |
+
a vectorized env) throws any error during env stepping, the
|
| 65 |
+
Sampler will try to restart the faulty sub-environment. This is done
|
| 66 |
+
without disturbing the other (still intact) sub-environment and without
|
| 67 |
+
the RolloutWorker crashing.
|
| 68 |
+
"""
|
| 69 |
+
|
| 70 |
+
# Could be creating local or remote envs.
|
| 71 |
+
self.make_env = make_env
|
| 72 |
+
self.num_envs = num_envs
|
| 73 |
+
self.multiagent = multiagent
|
| 74 |
+
self.poll_timeout = remote_env_batch_wait_ms / 1000
|
| 75 |
+
self.worker = worker
|
| 76 |
+
self.restart_failed_sub_environments = restart_failed_sub_environments
|
| 77 |
+
|
| 78 |
+
# Already existing env objects (generated by the RolloutWorker).
|
| 79 |
+
existing_envs = existing_envs or []
|
| 80 |
+
|
| 81 |
+
# Whether the given `make_env` callable already returns ActorHandles
|
| 82 |
+
# (@ray.remote class instances) or not.
|
| 83 |
+
self.make_env_creates_actors = False
|
| 84 |
+
|
| 85 |
+
self._observation_space = None
|
| 86 |
+
self._action_space = None
|
| 87 |
+
|
| 88 |
+
# List of ray actor handles (each handle points to one @ray.remote
|
| 89 |
+
# sub-environment).
|
| 90 |
+
self.actors: Optional[List[ray.actor.ActorHandle]] = None
|
| 91 |
+
|
| 92 |
+
# `self.make_env` already produces Actors: Use it directly.
|
| 93 |
+
if len(existing_envs) > 0 and isinstance(
|
| 94 |
+
existing_envs[0], ray.actor.ActorHandle
|
| 95 |
+
):
|
| 96 |
+
self.make_env_creates_actors = True
|
| 97 |
+
self.actors = existing_envs
|
| 98 |
+
while len(self.actors) < self.num_envs:
|
| 99 |
+
self.actors.append(self._make_sub_env(len(self.actors)))
|
| 100 |
+
|
| 101 |
+
# `self.make_env` produces gym.Envs (or children thereof, such
|
| 102 |
+
# as MultiAgentEnv): Need to auto-wrap it here. The problem with
|
| 103 |
+
# this is that custom methods wil get lost. If you would like to
|
| 104 |
+
# keep your custom methods in your envs, you should provide the
|
| 105 |
+
# env class directly in your config (w/o tune.register_env()),
|
| 106 |
+
# such that your class can directly be made a @ray.remote
|
| 107 |
+
# (w/o the wrapping via `_Remote[Multi|Single]AgentEnv`).
|
| 108 |
+
# Also, if `len(existing_envs) > 0`, we have to throw those away
|
| 109 |
+
# as we need to create ray actors here.
|
| 110 |
+
else:
|
| 111 |
+
self.actors = [self._make_sub_env(i) for i in range(self.num_envs)]
|
| 112 |
+
# Utilize existing envs for inferring observation/action spaces.
|
| 113 |
+
if len(existing_envs) > 0:
|
| 114 |
+
self._observation_space = existing_envs[0].observation_space
|
| 115 |
+
self._action_space = existing_envs[0].action_space
|
| 116 |
+
# Have to call actors' remote methods to get observation/action spaces.
|
| 117 |
+
else:
|
| 118 |
+
self._observation_space, self._action_space = ray.get(
|
| 119 |
+
[
|
| 120 |
+
self.actors[0].observation_space.remote(),
|
| 121 |
+
self.actors[0].action_space.remote(),
|
| 122 |
+
]
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
# Dict mapping object refs (return values of @ray.remote calls),
|
| 126 |
+
# whose actual values we are waiting for (via ray.wait in
|
| 127 |
+
# `self.poll()`) to their corresponding actor handles (the actors
|
| 128 |
+
# that created these return values).
|
| 129 |
+
# Call `reset()` on all @ray.remote sub-environment actors.
|
| 130 |
+
self.pending: Dict[ray.actor.ActorHandle] = {
|
| 131 |
+
a.reset.remote(): a for a in self.actors
|
| 132 |
+
}
|
| 133 |
+
|
| 134 |
+
@override(BaseEnv)
|
| 135 |
+
def poll(
|
| 136 |
+
self,
|
| 137 |
+
) -> Tuple[
|
| 138 |
+
MultiEnvDict,
|
| 139 |
+
MultiEnvDict,
|
| 140 |
+
MultiEnvDict,
|
| 141 |
+
MultiEnvDict,
|
| 142 |
+
MultiEnvDict,
|
| 143 |
+
MultiEnvDict,
|
| 144 |
+
]:
|
| 145 |
+
|
| 146 |
+
# each keyed by env_id in [0, num_remote_envs)
|
| 147 |
+
obs, rewards, terminateds, truncateds, infos = {}, {}, {}, {}, {}
|
| 148 |
+
ready = []
|
| 149 |
+
|
| 150 |
+
# Wait for at least 1 env to be ready here.
|
| 151 |
+
while not ready:
|
| 152 |
+
ready, _ = ray.wait(
|
| 153 |
+
list(self.pending),
|
| 154 |
+
num_returns=len(self.pending),
|
| 155 |
+
timeout=self.poll_timeout,
|
| 156 |
+
)
|
| 157 |
+
|
| 158 |
+
# Get and return observations for each of the ready envs
|
| 159 |
+
env_ids = set()
|
| 160 |
+
for obj_ref in ready:
|
| 161 |
+
# Get the corresponding actor handle from our dict and remove the
|
| 162 |
+
# object ref (we will call `ray.get()` on it and it will no longer
|
| 163 |
+
# be "pending").
|
| 164 |
+
actor = self.pending.pop(obj_ref)
|
| 165 |
+
env_id = self.actors.index(actor)
|
| 166 |
+
env_ids.add(env_id)
|
| 167 |
+
# Get the ready object ref (this may be return value(s) of
|
| 168 |
+
# `reset()` or `step()`).
|
| 169 |
+
try:
|
| 170 |
+
ret = ray.get(obj_ref)
|
| 171 |
+
except Exception as e:
|
| 172 |
+
# Something happened on the actor during stepping/resetting.
|
| 173 |
+
# Restart sub-environment (create new actor; close old one).
|
| 174 |
+
if self.restart_failed_sub_environments:
|
| 175 |
+
logger.exception(e.args[0])
|
| 176 |
+
self.try_restart(env_id)
|
| 177 |
+
# Always return multi-agent data.
|
| 178 |
+
# Set the observation to the exception, no rewards,
|
| 179 |
+
# terminated[__all__]=True (episode will be discarded anyways),
|
| 180 |
+
# no infos.
|
| 181 |
+
ret = (
|
| 182 |
+
e,
|
| 183 |
+
{},
|
| 184 |
+
{"__all__": True},
|
| 185 |
+
{"__all__": False},
|
| 186 |
+
{},
|
| 187 |
+
)
|
| 188 |
+
# Do not try to restart. Just raise the error.
|
| 189 |
+
else:
|
| 190 |
+
raise e
|
| 191 |
+
|
| 192 |
+
# Our sub-envs are simple Actor-turned gym.Envs or MultiAgentEnvs.
|
| 193 |
+
if self.make_env_creates_actors:
|
| 194 |
+
rew, terminated, truncated, info = None, None, None, None
|
| 195 |
+
if self.multiagent:
|
| 196 |
+
if isinstance(ret, tuple):
|
| 197 |
+
# Gym >= 0.26: `step()` result: Obs, reward, terminated,
|
| 198 |
+
# truncated, info.
|
| 199 |
+
if len(ret) == 5:
|
| 200 |
+
ob, rew, terminated, truncated, info = ret
|
| 201 |
+
# Gym >= 0.26: `reset()` result: Obs and infos.
|
| 202 |
+
elif len(ret) == 2:
|
| 203 |
+
ob = ret[0]
|
| 204 |
+
info = ret[1]
|
| 205 |
+
# Gym < 0.26? Something went wrong.
|
| 206 |
+
else:
|
| 207 |
+
raise AssertionError(
|
| 208 |
+
"Your gymnasium.Env seems to NOT return the correct "
|
| 209 |
+
"number of return values for `step()` (needs to return"
|
| 210 |
+
" 5 values: obs, reward, terminated, truncated and "
|
| 211 |
+
"info) or `reset()` (needs to return 2 values: obs and "
|
| 212 |
+
"info)!"
|
| 213 |
+
)
|
| 214 |
+
# Gym < 0.26: `reset()` result: Only obs.
|
| 215 |
+
else:
|
| 216 |
+
raise AssertionError(
|
| 217 |
+
"Your gymnasium.Env seems to only return a single value "
|
| 218 |
+
"upon `reset()`! Must return 2 (obs AND infos)."
|
| 219 |
+
)
|
| 220 |
+
else:
|
| 221 |
+
if isinstance(ret, tuple):
|
| 222 |
+
# `step()` result: Obs, reward, terminated, truncated, info.
|
| 223 |
+
if len(ret) == 5:
|
| 224 |
+
ob = {_DUMMY_AGENT_ID: ret[0]}
|
| 225 |
+
rew = {_DUMMY_AGENT_ID: ret[1]}
|
| 226 |
+
terminated = {_DUMMY_AGENT_ID: ret[2], "__all__": ret[2]}
|
| 227 |
+
truncated = {_DUMMY_AGENT_ID: ret[3], "__all__": ret[3]}
|
| 228 |
+
info = {_DUMMY_AGENT_ID: ret[4]}
|
| 229 |
+
# `reset()` result: Obs and infos.
|
| 230 |
+
elif len(ret) == 2:
|
| 231 |
+
ob = {_DUMMY_AGENT_ID: ret[0]}
|
| 232 |
+
info = {_DUMMY_AGENT_ID: ret[1]}
|
| 233 |
+
# Gym < 0.26? Something went wrong.
|
| 234 |
+
else:
|
| 235 |
+
raise AssertionError(
|
| 236 |
+
"Your gymnasium.Env seems to NOT return the correct "
|
| 237 |
+
"number of return values for `step()` (needs to return"
|
| 238 |
+
" 5 values: obs, reward, terminated, truncated and "
|
| 239 |
+
"info) or `reset()` (needs to return 2 values: obs and "
|
| 240 |
+
"info)!"
|
| 241 |
+
)
|
| 242 |
+
# Gym < 0.26?
|
| 243 |
+
else:
|
| 244 |
+
raise AssertionError(
|
| 245 |
+
"Your gymnasium.Env seems to only return a single value "
|
| 246 |
+
"upon `reset()`! Must return 2 (obs and infos)."
|
| 247 |
+
)
|
| 248 |
+
|
| 249 |
+
# If this is a `reset()` return value, we only have the initial
|
| 250 |
+
# observations and infos: Set rewards, terminateds, and truncateds to
|
| 251 |
+
# dummy values.
|
| 252 |
+
if rew is None:
|
| 253 |
+
rew = {agent_id: 0 for agent_id in ob.keys()}
|
| 254 |
+
terminated = {"__all__": False}
|
| 255 |
+
truncated = {"__all__": False}
|
| 256 |
+
|
| 257 |
+
# Our sub-envs are auto-wrapped (by `_RemoteSingleAgentEnv` or
|
| 258 |
+
# `_RemoteMultiAgentEnv`) and already behave like multi-agent
|
| 259 |
+
# envs.
|
| 260 |
+
else:
|
| 261 |
+
ob, rew, terminated, truncated, info = ret
|
| 262 |
+
obs[env_id] = ob
|
| 263 |
+
rewards[env_id] = rew
|
| 264 |
+
terminateds[env_id] = terminated
|
| 265 |
+
truncateds[env_id] = truncated
|
| 266 |
+
infos[env_id] = info
|
| 267 |
+
|
| 268 |
+
logger.debug(f"Got obs batch for actors {env_ids}")
|
| 269 |
+
return obs, rewards, terminateds, truncateds, infos, {}
|
| 270 |
+
|
| 271 |
+
@override(BaseEnv)
|
| 272 |
+
def send_actions(self, action_dict: MultiEnvDict) -> None:
|
| 273 |
+
for env_id, actions in action_dict.items():
|
| 274 |
+
actor = self.actors[env_id]
|
| 275 |
+
# `actor` is a simple single-agent (remote) env, e.g. a gym.Env
|
| 276 |
+
# that was made a @ray.remote.
|
| 277 |
+
if not self.multiagent and self.make_env_creates_actors:
|
| 278 |
+
obj_ref = actor.step.remote(actions[_DUMMY_AGENT_ID])
|
| 279 |
+
# `actor` is already a _RemoteSingleAgentEnv or
|
| 280 |
+
# _RemoteMultiAgentEnv wrapper
|
| 281 |
+
# (handles the multi-agent action_dict automatically).
|
| 282 |
+
else:
|
| 283 |
+
obj_ref = actor.step.remote(actions)
|
| 284 |
+
self.pending[obj_ref] = actor
|
| 285 |
+
|
| 286 |
+
@override(BaseEnv)
|
| 287 |
+
def try_reset(
|
| 288 |
+
self,
|
| 289 |
+
env_id: Optional[EnvID] = None,
|
| 290 |
+
*,
|
| 291 |
+
seed: Optional[int] = None,
|
| 292 |
+
options: Optional[dict] = None,
|
| 293 |
+
) -> Tuple[MultiEnvDict, MultiEnvDict]:
|
| 294 |
+
actor = self.actors[env_id]
|
| 295 |
+
obj_ref = actor.reset.remote(seed=seed, options=options)
|
| 296 |
+
|
| 297 |
+
self.pending[obj_ref] = actor
|
| 298 |
+
# Because this env type does not support synchronous reset requests (with
|
| 299 |
+
# immediate return value), we return ASYNC_RESET_RETURN here to indicate
|
| 300 |
+
# that the reset results will be available via the next `poll()` call.
|
| 301 |
+
return ASYNC_RESET_RETURN, ASYNC_RESET_RETURN
|
| 302 |
+
|
| 303 |
+
@override(BaseEnv)
|
| 304 |
+
def try_restart(self, env_id: Optional[EnvID] = None) -> None:
|
| 305 |
+
# Try closing down the old (possibly faulty) sub-env, but ignore errors.
|
| 306 |
+
try:
|
| 307 |
+
# Close the env on the remote side.
|
| 308 |
+
self.actors[env_id].close.remote()
|
| 309 |
+
except Exception as e:
|
| 310 |
+
if log_once("close_sub_env"):
|
| 311 |
+
logger.warning(
|
| 312 |
+
"Trying to close old and replaced sub-environment (at vector "
|
| 313 |
+
f"index={env_id}), but closing resulted in error:\n{e}"
|
| 314 |
+
)
|
| 315 |
+
|
| 316 |
+
# Terminate the actor itself to free up its resources.
|
| 317 |
+
self.actors[env_id].__ray_terminate__.remote()
|
| 318 |
+
|
| 319 |
+
# Re-create a new sub-environment.
|
| 320 |
+
self.actors[env_id] = self._make_sub_env(env_id)
|
| 321 |
+
|
| 322 |
+
@override(BaseEnv)
|
| 323 |
+
def stop(self) -> None:
|
| 324 |
+
if self.actors is not None:
|
| 325 |
+
for actor in self.actors:
|
| 326 |
+
actor.__ray_terminate__.remote()
|
| 327 |
+
|
| 328 |
+
@override(BaseEnv)
|
| 329 |
+
def get_sub_environments(self, as_dict: bool = False) -> List[EnvType]:
|
| 330 |
+
if as_dict:
|
| 331 |
+
return {env_id: actor for env_id, actor in enumerate(self.actors)}
|
| 332 |
+
return self.actors
|
| 333 |
+
|
| 334 |
+
@property
|
| 335 |
+
@override(BaseEnv)
|
| 336 |
+
def observation_space(self) -> gym.spaces.Dict:
|
| 337 |
+
return self._observation_space
|
| 338 |
+
|
| 339 |
+
@property
|
| 340 |
+
@override(BaseEnv)
|
| 341 |
+
def action_space(self) -> gym.Space:
|
| 342 |
+
return self._action_space
|
| 343 |
+
|
| 344 |
+
def _make_sub_env(self, idx: Optional[int] = None):
|
| 345 |
+
"""Re-creates a sub-environment at the new index."""
|
| 346 |
+
|
| 347 |
+
# Our `make_env` creates ray actors directly.
|
| 348 |
+
if self.make_env_creates_actors:
|
| 349 |
+
sub_env = self.make_env(idx)
|
| 350 |
+
if self.worker is not None:
|
| 351 |
+
self.worker.callbacks.on_sub_environment_created(
|
| 352 |
+
worker=self.worker,
|
| 353 |
+
sub_environment=self.actors[idx],
|
| 354 |
+
env_context=self.worker.env_context.copy_with_overrides(
|
| 355 |
+
vector_index=idx
|
| 356 |
+
),
|
| 357 |
+
)
|
| 358 |
+
|
| 359 |
+
# Our `make_env` returns actual envs -> Have to convert them into actors
|
| 360 |
+
# using our utility wrapper classes.
|
| 361 |
+
else:
|
| 362 |
+
|
| 363 |
+
def make_remote_env(i):
|
| 364 |
+
logger.info("Launching env {} in remote actor".format(i))
|
| 365 |
+
if self.multiagent:
|
| 366 |
+
sub_env = _RemoteMultiAgentEnv.remote(self.make_env, i)
|
| 367 |
+
else:
|
| 368 |
+
sub_env = _RemoteSingleAgentEnv.remote(self.make_env, i)
|
| 369 |
+
|
| 370 |
+
if self.worker is not None:
|
| 371 |
+
self.worker.callbacks.on_sub_environment_created(
|
| 372 |
+
worker=self.worker,
|
| 373 |
+
sub_environment=sub_env,
|
| 374 |
+
env_context=self.worker.env_context.copy_with_overrides(
|
| 375 |
+
vector_index=i
|
| 376 |
+
),
|
| 377 |
+
)
|
| 378 |
+
|
| 379 |
+
return sub_env
|
| 380 |
+
|
| 381 |
+
sub_env = make_remote_env(idx)
|
| 382 |
+
|
| 383 |
+
return sub_env
|
| 384 |
+
|
| 385 |
+
@override(BaseEnv)
|
| 386 |
+
def get_agent_ids(self) -> Set[AgentID]:
|
| 387 |
+
if self.multiagent:
|
| 388 |
+
return ray.get(self.actors[0].get_agent_ids.remote())
|
| 389 |
+
else:
|
| 390 |
+
return {_DUMMY_AGENT_ID}
|
| 391 |
+
|
| 392 |
+
|
| 393 |
+
@ray.remote(num_cpus=0)
|
| 394 |
+
class _RemoteMultiAgentEnv:
|
| 395 |
+
"""Wrapper class for making a multi-agent env a remote actor."""
|
| 396 |
+
|
| 397 |
+
def __init__(self, make_env, i):
|
| 398 |
+
self.env = make_env(i)
|
| 399 |
+
self.agent_ids = set()
|
| 400 |
+
|
| 401 |
+
def reset(self, *, seed: Optional[int] = None, options: Optional[dict] = None):
|
| 402 |
+
obs, info = self.env.reset(seed=seed, options=options)
|
| 403 |
+
|
| 404 |
+
# each keyed by agent_id in the env
|
| 405 |
+
rew = {}
|
| 406 |
+
for agent_id in obs.keys():
|
| 407 |
+
self.agent_ids.add(agent_id)
|
| 408 |
+
rew[agent_id] = 0.0
|
| 409 |
+
terminated = {"__all__": False}
|
| 410 |
+
truncated = {"__all__": False}
|
| 411 |
+
return obs, rew, terminated, truncated, info
|
| 412 |
+
|
| 413 |
+
def step(self, action_dict):
|
| 414 |
+
return self.env.step(action_dict)
|
| 415 |
+
|
| 416 |
+
# Defining these 2 functions that way this information can be queried
|
| 417 |
+
# with a call to ray.get().
|
| 418 |
+
def observation_space(self):
|
| 419 |
+
return self.env.observation_space
|
| 420 |
+
|
| 421 |
+
def action_space(self):
|
| 422 |
+
return self.env.action_space
|
| 423 |
+
|
| 424 |
+
def get_agent_ids(self) -> Set[AgentID]:
|
| 425 |
+
return self.agent_ids
|
| 426 |
+
|
| 427 |
+
|
| 428 |
+
@ray.remote(num_cpus=0)
|
| 429 |
+
class _RemoteSingleAgentEnv:
|
| 430 |
+
"""Wrapper class for making a gym env a remote actor."""
|
| 431 |
+
|
| 432 |
+
def __init__(self, make_env, i):
|
| 433 |
+
self.env = make_env(i)
|
| 434 |
+
|
| 435 |
+
def reset(self, *, seed: Optional[int] = None, options: Optional[dict] = None):
|
| 436 |
+
obs_and_info = self.env.reset(seed=seed, options=options)
|
| 437 |
+
|
| 438 |
+
obs = {_DUMMY_AGENT_ID: obs_and_info[0]}
|
| 439 |
+
info = {_DUMMY_AGENT_ID: obs_and_info[1]}
|
| 440 |
+
|
| 441 |
+
rew = {_DUMMY_AGENT_ID: 0.0}
|
| 442 |
+
terminated = {"__all__": False}
|
| 443 |
+
truncated = {"__all__": False}
|
| 444 |
+
return obs, rew, terminated, truncated, info
|
| 445 |
+
|
| 446 |
+
def step(self, action):
|
| 447 |
+
results = self.env.step(action[_DUMMY_AGENT_ID])
|
| 448 |
+
|
| 449 |
+
obs, rew, terminated, truncated, info = [{_DUMMY_AGENT_ID: x} for x in results]
|
| 450 |
+
|
| 451 |
+
terminated["__all__"] = terminated[_DUMMY_AGENT_ID]
|
| 452 |
+
truncated["__all__"] = truncated[_DUMMY_AGENT_ID]
|
| 453 |
+
|
| 454 |
+
return obs, rew, terminated, truncated, info
|
| 455 |
+
|
| 456 |
+
# Defining these 2 functions that way this information can be queried
|
| 457 |
+
# with a call to ray.get().
|
| 458 |
+
def observation_space(self):
|
| 459 |
+
return self.env.observation_space
|
| 460 |
+
|
| 461 |
+
def action_space(self):
|
| 462 |
+
return self.env.action_space
|
vllm/lib/python3.10/site-packages/ray/rllib/env/utils/__init__.py
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import logging
|
| 2 |
+
from typing import Type, Union
|
| 3 |
+
|
| 4 |
+
import gymnasium as gym
|
| 5 |
+
|
| 6 |
+
from ray.rllib.env.env_context import EnvContext
|
| 7 |
+
from ray.rllib.utils.error import (
|
| 8 |
+
ERR_MSG_INVALID_ENV_DESCRIPTOR,
|
| 9 |
+
EnvError,
|
| 10 |
+
)
|
| 11 |
+
from ray.util.annotations import PublicAPI
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
logger = logging.getLogger(__name__)
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
@PublicAPI
|
| 18 |
+
def try_import_pyspiel(error: bool = False):
|
| 19 |
+
"""Tries importing pyspiel and returns the module (or None).
|
| 20 |
+
|
| 21 |
+
Args:
|
| 22 |
+
error: Whether to raise an error if pyspiel cannot be imported.
|
| 23 |
+
|
| 24 |
+
Returns:
|
| 25 |
+
The pyspiel module.
|
| 26 |
+
|
| 27 |
+
Raises:
|
| 28 |
+
ImportError: If error=True and pyspiel is not installed.
|
| 29 |
+
"""
|
| 30 |
+
try:
|
| 31 |
+
import pyspiel
|
| 32 |
+
|
| 33 |
+
return pyspiel
|
| 34 |
+
except ImportError:
|
| 35 |
+
if error:
|
| 36 |
+
raise ImportError(
|
| 37 |
+
"Could not import pyspiel! Pygame is not a dependency of RLlib "
|
| 38 |
+
"and RLlib requires you to install pygame separately: "
|
| 39 |
+
"`pip install pygame`."
|
| 40 |
+
)
|
| 41 |
+
return None
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
@PublicAPI
|
| 45 |
+
def try_import_open_spiel(error: bool = False):
|
| 46 |
+
"""Tries importing open_spiel and returns the module (or None).
|
| 47 |
+
|
| 48 |
+
Args:
|
| 49 |
+
error: Whether to raise an error if open_spiel cannot be imported.
|
| 50 |
+
|
| 51 |
+
Returns:
|
| 52 |
+
The open_spiel module.
|
| 53 |
+
|
| 54 |
+
Raises:
|
| 55 |
+
ImportError: If error=True and open_spiel is not installed.
|
| 56 |
+
"""
|
| 57 |
+
try:
|
| 58 |
+
import open_spiel
|
| 59 |
+
|
| 60 |
+
return open_spiel
|
| 61 |
+
except ImportError:
|
| 62 |
+
if error:
|
| 63 |
+
raise ImportError(
|
| 64 |
+
"Could not import open_spiel! open_spiel is not a dependency of RLlib "
|
| 65 |
+
"and RLlib requires you to install open_spiel separately: "
|
| 66 |
+
"`pip install open_spiel`."
|
| 67 |
+
)
|
| 68 |
+
return None
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
def _gym_env_creator(
|
| 72 |
+
env_context: EnvContext,
|
| 73 |
+
env_descriptor: Union[str, Type[gym.Env]],
|
| 74 |
+
) -> gym.Env:
|
| 75 |
+
"""Tries to create a gym env given an EnvContext object and descriptor.
|
| 76 |
+
|
| 77 |
+
Note: This function tries to construct the env from a string descriptor
|
| 78 |
+
only using possibly installed RL env packages (such as gym, pybullet_envs,
|
| 79 |
+
etc). These packages are no installation requirements for RLlib. In case
|
| 80 |
+
you would like to support more such env packages, add the necessary imports
|
| 81 |
+
and construction logic below.
|
| 82 |
+
|
| 83 |
+
Args:
|
| 84 |
+
env_context: The env context object to configure the env.
|
| 85 |
+
Note that this is a config dict, plus the properties:
|
| 86 |
+
`worker_index`, `vector_index`, and `remote`.
|
| 87 |
+
env_descriptor: The env descriptor as a gym-registered string, e.g. CartPole-v1,
|
| 88 |
+
ALE/MsPacman-v5, or CartPoleContinuousBulletEnv-v0.
|
| 89 |
+
Alternatively, the gym.Env subclass to use.
|
| 90 |
+
|
| 91 |
+
Returns:
|
| 92 |
+
The actual gym environment object.
|
| 93 |
+
|
| 94 |
+
Raises:
|
| 95 |
+
gym.error.Error: If the env cannot be constructed.
|
| 96 |
+
"""
|
| 97 |
+
# Allow for PyBullet or envs to be used as well (via string). This allows
|
| 98 |
+
# for doing things like `env=CartPoleContinuousBulletEnv-v0`.
|
| 99 |
+
try:
|
| 100 |
+
import pybullet_envs
|
| 101 |
+
|
| 102 |
+
pybullet_envs.getList()
|
| 103 |
+
except (AttributeError, ModuleNotFoundError, ImportError):
|
| 104 |
+
pass
|
| 105 |
+
|
| 106 |
+
# If env descriptor is a str, starting with "ale_py:ALE/", for now, register all ALE
|
| 107 |
+
# envs from ale_py.
|
| 108 |
+
if isinstance(env_descriptor, str) and env_descriptor.startswith("ale_py:ALE/"):
|
| 109 |
+
import ale_py
|
| 110 |
+
|
| 111 |
+
gym.register_envs(ale_py)
|
| 112 |
+
|
| 113 |
+
# Try creating a gym env. If this fails we can output a
|
| 114 |
+
# decent error message.
|
| 115 |
+
try:
|
| 116 |
+
# If class provided, call constructor directly.
|
| 117 |
+
if isinstance(env_descriptor, type):
|
| 118 |
+
env = env_descriptor(env_context)
|
| 119 |
+
else:
|
| 120 |
+
env = gym.make(env_descriptor, **env_context)
|
| 121 |
+
except gym.error.Error:
|
| 122 |
+
raise EnvError(ERR_MSG_INVALID_ENV_DESCRIPTOR.format(env_descriptor))
|
| 123 |
+
|
| 124 |
+
return env
|
vllm/lib/python3.10/site-packages/ray/rllib/env/utils/infinite_lookback_buffer.py
ADDED
|
@@ -0,0 +1,707 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Any, Dict, List, Optional, Union
|
| 2 |
+
|
| 3 |
+
import gymnasium as gym
|
| 4 |
+
import numpy as np
|
| 5 |
+
import tree # pip install dm_tree
|
| 6 |
+
|
| 7 |
+
from ray.rllib.utils.numpy import LARGE_INTEGER, one_hot, one_hot_multidiscrete
|
| 8 |
+
from ray.rllib.utils.serialization import gym_space_from_dict, gym_space_to_dict
|
| 9 |
+
from ray.rllib.utils.spaces.space_utils import (
|
| 10 |
+
batch,
|
| 11 |
+
from_jsonable_if_needed,
|
| 12 |
+
get_dummy_batch_for_space,
|
| 13 |
+
get_base_struct_from_space,
|
| 14 |
+
to_jsonable_if_needed,
|
| 15 |
+
)
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
class InfiniteLookbackBuffer:
|
| 19 |
+
@property
|
| 20 |
+
def space(self):
|
| 21 |
+
return self._space
|
| 22 |
+
|
| 23 |
+
@space.setter
|
| 24 |
+
def space(self, value):
|
| 25 |
+
self._space = value
|
| 26 |
+
self.space_struct = get_base_struct_from_space(value)
|
| 27 |
+
|
| 28 |
+
def __init__(
|
| 29 |
+
self,
|
| 30 |
+
data: Optional[Union[List, np.ndarray]] = None,
|
| 31 |
+
lookback: int = 0,
|
| 32 |
+
space: Optional[gym.Space] = None,
|
| 33 |
+
):
|
| 34 |
+
self.data = data if data is not None else []
|
| 35 |
+
self.lookback = min(lookback, len(self.data))
|
| 36 |
+
self.finalized = not isinstance(self.data, list)
|
| 37 |
+
self.space_struct = None
|
| 38 |
+
self.space = space
|
| 39 |
+
|
| 40 |
+
def __eq__(
|
| 41 |
+
self,
|
| 42 |
+
other: "InfiniteLookbackBuffer",
|
| 43 |
+
) -> bool:
|
| 44 |
+
"""Compares two `InfiniteLookbackBuffers.
|
| 45 |
+
|
| 46 |
+
Args:
|
| 47 |
+
other: Another object. If another `LookbackBuffer` instance all
|
| 48 |
+
their attributes are compared.
|
| 49 |
+
|
| 50 |
+
Returns:
|
| 51 |
+
`True`, if `other` is an `InfiniteLookbackBuffer` instance and all
|
| 52 |
+
attributes are identical. Otherwise, returns `False`.
|
| 53 |
+
"""
|
| 54 |
+
if isinstance(other, InfiniteLookbackBuffer):
|
| 55 |
+
if (
|
| 56 |
+
self.data == other.data
|
| 57 |
+
and self.lookback == other.lookback
|
| 58 |
+
and self.finalized == other.finalized
|
| 59 |
+
and self.space_struct == other.space_struct
|
| 60 |
+
and self.space == other.space
|
| 61 |
+
):
|
| 62 |
+
return True
|
| 63 |
+
return False
|
| 64 |
+
|
| 65 |
+
def get_state(self) -> Dict[str, Any]:
|
| 66 |
+
"""Returns the pickable state of a buffer.
|
| 67 |
+
|
| 68 |
+
The data in the buffer is stored into a dictionary. Note that
|
| 69 |
+
buffers can also be generated from pickable states (see
|
| 70 |
+
`InfiniteLookbackBuffer.from_state`)
|
| 71 |
+
|
| 72 |
+
Returns:
|
| 73 |
+
A dict containing all the data and metadata from the buffer.
|
| 74 |
+
"""
|
| 75 |
+
return {
|
| 76 |
+
"data": to_jsonable_if_needed(self.data, self.space)
|
| 77 |
+
if self.space
|
| 78 |
+
else self.data,
|
| 79 |
+
"lookback": self.lookback,
|
| 80 |
+
"finalized": self.finalized,
|
| 81 |
+
"space": gym_space_to_dict(self.space) if self.space else self.space,
|
| 82 |
+
}
|
| 83 |
+
|
| 84 |
+
@staticmethod
|
| 85 |
+
def from_state(state: Dict[str, Any]) -> None:
|
| 86 |
+
"""Creates a new `InfiniteLookbackBuffer` from a state dict.
|
| 87 |
+
|
| 88 |
+
Args:
|
| 89 |
+
state: The state dict, as returned by `self.get_state`.
|
| 90 |
+
|
| 91 |
+
Returns:
|
| 92 |
+
A new `InfiniteLookbackBuffer` instance with the data and metadata
|
| 93 |
+
from the state dict.
|
| 94 |
+
"""
|
| 95 |
+
buffer = InfiniteLookbackBuffer()
|
| 96 |
+
buffer.lookback = state["lookback"]
|
| 97 |
+
buffer.finalized = state["finalized"]
|
| 98 |
+
buffer.space = gym_space_from_dict(state["space"]) if state["space"] else None
|
| 99 |
+
buffer.space_struct = (
|
| 100 |
+
get_base_struct_from_space(buffer.space) if buffer.space else None
|
| 101 |
+
)
|
| 102 |
+
buffer.data = (
|
| 103 |
+
from_jsonable_if_needed(state["data"], buffer.space)
|
| 104 |
+
if buffer.space
|
| 105 |
+
else state["data"]
|
| 106 |
+
)
|
| 107 |
+
|
| 108 |
+
return buffer
|
| 109 |
+
|
| 110 |
+
def append(self, item) -> None:
|
| 111 |
+
"""Appends the given item to the end of this buffer."""
|
| 112 |
+
if self.finalized:
|
| 113 |
+
self.data = tree.map_structure(
|
| 114 |
+
lambda d, i: np.concatenate([d, [i]], axis=0), self.data, item
|
| 115 |
+
)
|
| 116 |
+
else:
|
| 117 |
+
self.data.append(item)
|
| 118 |
+
|
| 119 |
+
def extend(self, items) -> None:
|
| 120 |
+
"""Appends all items in `items` to the end of this buffer."""
|
| 121 |
+
if self.finalized:
|
| 122 |
+
# TODO (sven): When extending with a list of structs, we should
|
| 123 |
+
# probably rather do: `tree.map_structure(..., self.data,
|
| 124 |
+
# tree.map_structure(lambda *s: np.array(*s), *items)`)??
|
| 125 |
+
self.data = tree.map_structure(
|
| 126 |
+
lambda d, i: np.concatenate([d, i], axis=0), self.data, np.array(items)
|
| 127 |
+
)
|
| 128 |
+
else:
|
| 129 |
+
for item in items:
|
| 130 |
+
self.append(item)
|
| 131 |
+
|
| 132 |
+
def concat(self, other: "InfiniteLookbackBuffer") -> None:
|
| 133 |
+
"""Concatenates the data of `other` (w/o its lookback) to `self`.
|
| 134 |
+
|
| 135 |
+
Args:
|
| 136 |
+
other: The other InfiniteLookbackBuffer to be concatenated to self.
|
| 137 |
+
"""
|
| 138 |
+
self.data.extend(other.get())
|
| 139 |
+
|
| 140 |
+
def pop(self, index: int = -1) -> None:
|
| 141 |
+
"""Removes the item at `index` from this buffer, but does NOT return it.
|
| 142 |
+
|
| 143 |
+
Args:
|
| 144 |
+
index: The index to pop out of this buffer (w/o returning it from this
|
| 145 |
+
method).
|
| 146 |
+
"""
|
| 147 |
+
if self.finalized:
|
| 148 |
+
self.data = tree.map_structure(
|
| 149 |
+
lambda s: np.delete(s, index, axis=0), self.data
|
| 150 |
+
)
|
| 151 |
+
else:
|
| 152 |
+
self.data.pop(index)
|
| 153 |
+
|
| 154 |
+
def finalize(self) -> None:
|
| 155 |
+
"""Finalizes this buffer by converting internal data lists into numpy arrays.
|
| 156 |
+
|
| 157 |
+
Thereby, if the individual items in the list are nested structures, the
|
| 158 |
+
resulting buffer content will be a nested struct of np.ndarrays (leafs).
|
| 159 |
+
"""
|
| 160 |
+
if not self.finalized:
|
| 161 |
+
self.data = batch(self.data)
|
| 162 |
+
self.finalized = True
|
| 163 |
+
|
| 164 |
+
def get(
|
| 165 |
+
self,
|
| 166 |
+
indices: Optional[Union[int, slice, List[int]]] = None,
|
| 167 |
+
*,
|
| 168 |
+
neg_index_as_lookback: bool = False,
|
| 169 |
+
fill: Optional[Any] = None,
|
| 170 |
+
one_hot_discrete: bool = False,
|
| 171 |
+
_ignore_last_ts: bool = False,
|
| 172 |
+
_add_last_ts_value: Optional[Any] = None,
|
| 173 |
+
) -> Any:
|
| 174 |
+
"""Returns data, based on the given args, from this buffer.
|
| 175 |
+
|
| 176 |
+
Args:
|
| 177 |
+
indices: A single int is interpreted as an index, from which to return the
|
| 178 |
+
individual data stored at this index.
|
| 179 |
+
A list of ints is interpreted as a list of indices from which to gather
|
| 180 |
+
individual data in a batch of size len(indices).
|
| 181 |
+
A slice object is interpreted as a range of data to be returned.
|
| 182 |
+
Thereby, negative indices by default are interpreted as "before the end"
|
| 183 |
+
unless the `neg_index_as_lookback=True` option is used, in which case
|
| 184 |
+
negative indices are interpreted as "before ts=0", meaning going back
|
| 185 |
+
into the lookback buffer.
|
| 186 |
+
neg_index_as_lookback: If True, negative values in `indices` are
|
| 187 |
+
interpreted as "before ts=0", meaning going back into the lookback
|
| 188 |
+
buffer. For example, a buffer with data [4, 5, 6, 7, 8, 9],
|
| 189 |
+
where [4, 5, 6] is the lookback buffer range (ts=0 item is 7), will
|
| 190 |
+
respond to `get(-1, neg_index_as_lookback=True)` with `6` and to
|
| 191 |
+
`get(slice(-2, 1), neg_index_as_lookback=True)` with `[5, 6, 7]`.
|
| 192 |
+
fill: An optional float value to use for filling up the returned results at
|
| 193 |
+
the boundaries. This filling only happens if the requested index range's
|
| 194 |
+
start/stop boundaries exceed the buffer's boundaries (including the
|
| 195 |
+
lookback buffer on the left side). This comes in very handy, if users
|
| 196 |
+
don't want to worry about reaching such boundaries and want to zero-pad.
|
| 197 |
+
For example, a buffer with data [10, 11, 12, 13, 14] and lookback
|
| 198 |
+
buffer size of 2 (meaning `10` and `11` are part of the lookback buffer)
|
| 199 |
+
will respond to `get(slice(-7, -2), fill=0.0)`
|
| 200 |
+
with `[0.0, 0.0, 10, 11, 12]`.
|
| 201 |
+
one_hot_discrete: If True, will return one-hot vectors (instead of
|
| 202 |
+
int-values) for those sub-components of a (possibly complex) space
|
| 203 |
+
that are Discrete or MultiDiscrete. Note that if `fill=0` and the
|
| 204 |
+
requested `indices` are out of the range of our data, the returned
|
| 205 |
+
one-hot vectors will actually be zero-hot (all slots zero).
|
| 206 |
+
_ignore_last_ts: Whether to ignore the last record in our internal
|
| 207 |
+
`self.data` when getting the provided indices.
|
| 208 |
+
_add_last_ts_value: Whether to add the value of this arg to the end of
|
| 209 |
+
the internal `self.data` buffer (just for the duration of this get
|
| 210 |
+
operation, not permanently).
|
| 211 |
+
"""
|
| 212 |
+
if indices is None:
|
| 213 |
+
data = self._get_all_data(
|
| 214 |
+
one_hot_discrete=one_hot_discrete,
|
| 215 |
+
_ignore_last_ts=_ignore_last_ts,
|
| 216 |
+
)
|
| 217 |
+
elif isinstance(indices, slice):
|
| 218 |
+
data = self._get_slice(
|
| 219 |
+
indices,
|
| 220 |
+
fill=fill,
|
| 221 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 222 |
+
one_hot_discrete=one_hot_discrete,
|
| 223 |
+
_ignore_last_ts=_ignore_last_ts,
|
| 224 |
+
_add_last_ts_value=_add_last_ts_value,
|
| 225 |
+
)
|
| 226 |
+
elif isinstance(indices, list):
|
| 227 |
+
data = [
|
| 228 |
+
self._get_int_index(
|
| 229 |
+
idx,
|
| 230 |
+
fill=fill,
|
| 231 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 232 |
+
one_hot_discrete=one_hot_discrete,
|
| 233 |
+
_ignore_last_ts=_ignore_last_ts,
|
| 234 |
+
_add_last_ts_value=_add_last_ts_value,
|
| 235 |
+
)
|
| 236 |
+
for idx in indices
|
| 237 |
+
]
|
| 238 |
+
if self.finalized:
|
| 239 |
+
data = batch(data)
|
| 240 |
+
else:
|
| 241 |
+
assert isinstance(indices, int)
|
| 242 |
+
data = self._get_int_index(
|
| 243 |
+
indices,
|
| 244 |
+
fill=fill,
|
| 245 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 246 |
+
one_hot_discrete=one_hot_discrete,
|
| 247 |
+
_ignore_last_ts=_ignore_last_ts,
|
| 248 |
+
_add_last_ts_value=_add_last_ts_value,
|
| 249 |
+
)
|
| 250 |
+
|
| 251 |
+
return data
|
| 252 |
+
|
| 253 |
+
def __add__(
|
| 254 |
+
self, other: Union[List, "InfiniteLookbackBuffer", int, float, complex]
|
| 255 |
+
) -> "InfiniteLookbackBuffer":
|
| 256 |
+
"""Adds another InfiniteLookbackBuffer object or list to the end of this one.
|
| 257 |
+
|
| 258 |
+
Args:
|
| 259 |
+
other: Another `InfiniteLookbackBuffer` or a `list` or a number.
|
| 260 |
+
If a `InfiniteLookbackBuffer` its data (w/o its lookback buffer) gets
|
| 261 |
+
concatenated to self's data. If a `list`, we concat it to self's data.
|
| 262 |
+
If a number, we add this number to each element of self (if possible).
|
| 263 |
+
|
| 264 |
+
Returns:
|
| 265 |
+
A new `InfiniteLookbackBuffer` instance `self.data` containing
|
| 266 |
+
concatenated data from `self` and `other` (or adding `other` to each element
|
| 267 |
+
in self's data).
|
| 268 |
+
"""
|
| 269 |
+
|
| 270 |
+
if self.finalized:
|
| 271 |
+
raise RuntimeError(f"Cannot `add` to a finalized {type(self).__name__}.")
|
| 272 |
+
else:
|
| 273 |
+
# If `other` is an int, simply add it to all our values (if possible) and
|
| 274 |
+
# use the result as the underlying data for the returned buffer.
|
| 275 |
+
if isinstance(other, (int, float, complex)):
|
| 276 |
+
data = [
|
| 277 |
+
(d + other) if isinstance(d, (int, float, complex)) else d
|
| 278 |
+
for d in self.data
|
| 279 |
+
]
|
| 280 |
+
# If `other` is a InfiniteLookbackBuffer itself, do NOT include its
|
| 281 |
+
# lookback buffer anymore. We assume that `other`'s lookback buffer i
|
| 282 |
+
# already at the end of `self`.
|
| 283 |
+
elif isinstance(other, InfiniteLookbackBuffer):
|
| 284 |
+
data = self.data + other.data[other.lookback :]
|
| 285 |
+
# `other` is a list, simply concat the two lists and use the result as
|
| 286 |
+
# the underlying data for the returned buffer.
|
| 287 |
+
else:
|
| 288 |
+
data = self.data + other
|
| 289 |
+
|
| 290 |
+
return InfiniteLookbackBuffer(
|
| 291 |
+
data=data,
|
| 292 |
+
lookback=self.lookback,
|
| 293 |
+
space=self.space,
|
| 294 |
+
)
|
| 295 |
+
|
| 296 |
+
def __getitem__(self, item):
|
| 297 |
+
"""Support squared bracket syntax, e.g. buffer[:5]."""
|
| 298 |
+
return self.get(item)
|
| 299 |
+
|
| 300 |
+
def __setitem__(self, key, value):
|
| 301 |
+
self.set(new_data=value, at_indices=key)
|
| 302 |
+
|
| 303 |
+
def set(
|
| 304 |
+
self,
|
| 305 |
+
new_data,
|
| 306 |
+
*,
|
| 307 |
+
at_indices: Optional[Union[int, slice, List[int]]] = None,
|
| 308 |
+
neg_index_as_lookback: bool = False,
|
| 309 |
+
) -> None:
|
| 310 |
+
"""Overwrites all or some of the data in this buffer with the provided data.
|
| 311 |
+
|
| 312 |
+
Args:
|
| 313 |
+
new_data: The new data to overwrite existing records with.
|
| 314 |
+
at_indices: A single int is interpreted as an index, at which to overwrite
|
| 315 |
+
the individual record stored at this index with `new_data`.
|
| 316 |
+
A list of ints is interpreted as a list of indices, which to overwrite
|
| 317 |
+
with `new_data`, which must be a batch of size `len(at_indices)`.
|
| 318 |
+
A slice object is interpreted as a range, which to overwrite with
|
| 319 |
+
`new_data`. Thereby, negative indices by default are interpreted as
|
| 320 |
+
"before the end" unless the `neg_index_as_lookback=True` option is
|
| 321 |
+
used, in which case negative indices are interpreted as
|
| 322 |
+
"before ts=0", meaning going back into the lookback buffer.
|
| 323 |
+
neg_index_as_lookback: If True, negative values in `at_indices` are
|
| 324 |
+
interpreted as "before ts=0", meaning going back into the lookback
|
| 325 |
+
buffer. For example, a buffer with data [4, 5, 6, 7, 8, 9],
|
| 326 |
+
where [4, 5, 6] is the lookback buffer range (ts=0 item is 7), will
|
| 327 |
+
handle a call `set(99, at_indices=-1, neg_index_as_lookback=True)`
|
| 328 |
+
with `6` being replaced by 99 and to `set([98, 99, 100],
|
| 329 |
+
at_indices=slice(-2, 1), neg_index_as_lookback=True)` with
|
| 330 |
+
`[5, 6, 7]` being replaced by `[98, 99, 100]`.
|
| 331 |
+
"""
|
| 332 |
+
# `at_indices` is None -> Override all our data (excluding the lookback buffer).
|
| 333 |
+
if at_indices is None:
|
| 334 |
+
self._set_all_data(new_data)
|
| 335 |
+
|
| 336 |
+
elif isinstance(at_indices, slice):
|
| 337 |
+
self._set_slice(
|
| 338 |
+
new_data,
|
| 339 |
+
slice_=at_indices,
|
| 340 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 341 |
+
)
|
| 342 |
+
elif isinstance(at_indices, list):
|
| 343 |
+
for i, idx in enumerate(at_indices):
|
| 344 |
+
self._set_int_index(
|
| 345 |
+
new_data[i],
|
| 346 |
+
idx=idx,
|
| 347 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 348 |
+
)
|
| 349 |
+
else:
|
| 350 |
+
assert isinstance(at_indices, int)
|
| 351 |
+
self._set_int_index(
|
| 352 |
+
new_data,
|
| 353 |
+
idx=at_indices,
|
| 354 |
+
neg_index_as_lookback=neg_index_as_lookback,
|
| 355 |
+
)
|
| 356 |
+
|
| 357 |
+
def __len__(self):
|
| 358 |
+
"""Return the length of our data, excluding the lookback buffer."""
|
| 359 |
+
len_ = self.len_incl_lookback()
|
| 360 |
+
# Only count the data after the lookback.
|
| 361 |
+
return max(len_ - self.lookback, 0)
|
| 362 |
+
|
| 363 |
+
def len_incl_lookback(self):
|
| 364 |
+
if self.finalized:
|
| 365 |
+
return len(tree.flatten(self.data)[0])
|
| 366 |
+
else:
|
| 367 |
+
return len(self.data)
|
| 368 |
+
|
| 369 |
+
def __repr__(self):
|
| 370 |
+
return (
|
| 371 |
+
f"{type(self).__name__}({self.data[:self.lookback]} <- "
|
| 372 |
+
f"lookback({self.lookback}) | {self.data[self.lookback:]})"
|
| 373 |
+
)
|
| 374 |
+
|
| 375 |
+
def _get_all_data(self, one_hot_discrete=False, _ignore_last_ts=False):
|
| 376 |
+
data = self[: (None if not _ignore_last_ts else -1)]
|
| 377 |
+
if one_hot_discrete:
|
| 378 |
+
data = self._one_hot(data, space_struct=self.space_struct)
|
| 379 |
+
return data
|
| 380 |
+
|
| 381 |
+
def _set_all_data(self, new_data):
|
| 382 |
+
self._set_slice(new_data, slice(0, None))
|
| 383 |
+
|
| 384 |
+
def _get_slice(
|
| 385 |
+
self,
|
| 386 |
+
slice_,
|
| 387 |
+
fill=None,
|
| 388 |
+
neg_index_as_lookback=False,
|
| 389 |
+
one_hot_discrete=False,
|
| 390 |
+
_ignore_last_ts=False,
|
| 391 |
+
_add_last_ts_value=None,
|
| 392 |
+
):
|
| 393 |
+
data_to_use = self.data
|
| 394 |
+
if _ignore_last_ts:
|
| 395 |
+
if self.finalized:
|
| 396 |
+
data_to_use = tree.map_structure(lambda s: s[:-1], self.data)
|
| 397 |
+
else:
|
| 398 |
+
data_to_use = self.data[:-1]
|
| 399 |
+
if _add_last_ts_value is not None:
|
| 400 |
+
if self.finalized:
|
| 401 |
+
data_to_use = tree.map_structure(
|
| 402 |
+
lambda s, t: np.append(s, t),
|
| 403 |
+
data_to_use.copy(),
|
| 404 |
+
_add_last_ts_value,
|
| 405 |
+
)
|
| 406 |
+
else:
|
| 407 |
+
data_to_use = np.append(data_to_use.copy(), _add_last_ts_value)
|
| 408 |
+
|
| 409 |
+
slice_, slice_len, fill_left_count, fill_right_count = self._interpret_slice(
|
| 410 |
+
slice_,
|
| 411 |
+
neg_index_as_lookback,
|
| 412 |
+
len_self_plus_lookback=(
|
| 413 |
+
self.len_incl_lookback()
|
| 414 |
+
+ int(_add_last_ts_value is not None)
|
| 415 |
+
- int(_ignore_last_ts)
|
| 416 |
+
),
|
| 417 |
+
)
|
| 418 |
+
|
| 419 |
+
# Perform the actual slice.
|
| 420 |
+
data_slice = None
|
| 421 |
+
if slice_len > 0:
|
| 422 |
+
if self.finalized:
|
| 423 |
+
data_slice = tree.map_structure(lambda s: s[slice_], data_to_use)
|
| 424 |
+
else:
|
| 425 |
+
data_slice = data_to_use[slice_]
|
| 426 |
+
|
| 427 |
+
if one_hot_discrete:
|
| 428 |
+
data_slice = self._one_hot(data_slice, space_struct=self.space_struct)
|
| 429 |
+
|
| 430 |
+
# Data is shorter than the range requested -> Fill the rest with `fill` data.
|
| 431 |
+
if fill is not None and (fill_right_count > 0 or fill_left_count > 0):
|
| 432 |
+
if self.finalized:
|
| 433 |
+
if fill_left_count:
|
| 434 |
+
if self.space is None:
|
| 435 |
+
fill_batch = np.array([fill] * fill_left_count)
|
| 436 |
+
else:
|
| 437 |
+
fill_batch = get_dummy_batch_for_space(
|
| 438 |
+
self.space,
|
| 439 |
+
fill_value=fill,
|
| 440 |
+
batch_size=fill_left_count,
|
| 441 |
+
one_hot_discrete=one_hot_discrete,
|
| 442 |
+
)
|
| 443 |
+
if data_slice is not None:
|
| 444 |
+
data_slice = tree.map_structure(
|
| 445 |
+
lambda s0, s: np.concatenate([s0, s]),
|
| 446 |
+
fill_batch,
|
| 447 |
+
data_slice,
|
| 448 |
+
)
|
| 449 |
+
else:
|
| 450 |
+
data_slice = fill_batch
|
| 451 |
+
if fill_right_count:
|
| 452 |
+
if self.space is None:
|
| 453 |
+
fill_batch = np.array([fill] * fill_right_count)
|
| 454 |
+
else:
|
| 455 |
+
fill_batch = get_dummy_batch_for_space(
|
| 456 |
+
self.space,
|
| 457 |
+
fill_value=fill,
|
| 458 |
+
batch_size=fill_right_count,
|
| 459 |
+
one_hot_discrete=one_hot_discrete,
|
| 460 |
+
)
|
| 461 |
+
if data_slice is not None:
|
| 462 |
+
data_slice = tree.map_structure(
|
| 463 |
+
lambda s0, s: np.concatenate([s, s0]),
|
| 464 |
+
fill_batch,
|
| 465 |
+
data_slice,
|
| 466 |
+
)
|
| 467 |
+
else:
|
| 468 |
+
data_slice = fill_batch
|
| 469 |
+
|
| 470 |
+
else:
|
| 471 |
+
if self.space is None:
|
| 472 |
+
fill_batch = [fill]
|
| 473 |
+
else:
|
| 474 |
+
fill_batch = [
|
| 475 |
+
get_dummy_batch_for_space(
|
| 476 |
+
self.space,
|
| 477 |
+
fill_value=fill,
|
| 478 |
+
batch_size=0,
|
| 479 |
+
one_hot_discrete=one_hot_discrete,
|
| 480 |
+
)
|
| 481 |
+
]
|
| 482 |
+
data_slice = (
|
| 483 |
+
fill_batch * fill_left_count
|
| 484 |
+
+ (data_slice if data_slice is not None else [])
|
| 485 |
+
+ fill_batch * fill_right_count
|
| 486 |
+
)
|
| 487 |
+
|
| 488 |
+
if data_slice is None:
|
| 489 |
+
if self.finalized:
|
| 490 |
+
return tree.map_structure(lambda s: s[slice_], data_to_use)
|
| 491 |
+
else:
|
| 492 |
+
return data_to_use[slice_]
|
| 493 |
+
return data_slice
|
| 494 |
+
|
| 495 |
+
def _set_slice(
|
| 496 |
+
self,
|
| 497 |
+
new_data,
|
| 498 |
+
slice_,
|
| 499 |
+
neg_index_as_lookback=False,
|
| 500 |
+
):
|
| 501 |
+
slice_, _, _, _ = self._interpret_slice(slice_, neg_index_as_lookback)
|
| 502 |
+
|
| 503 |
+
# Check, whether the setting to new_data changes the length of self
|
| 504 |
+
# (it shouldn't). If it does, raise an error.
|
| 505 |
+
try:
|
| 506 |
+
if self.finalized:
|
| 507 |
+
|
| 508 |
+
def __set(s, n):
|
| 509 |
+
if self.space:
|
| 510 |
+
assert self.space.contains(n[0])
|
| 511 |
+
assert len(s[slice_]) == len(n)
|
| 512 |
+
s[slice_] = n
|
| 513 |
+
|
| 514 |
+
tree.map_structure(__set, self.data, new_data)
|
| 515 |
+
else:
|
| 516 |
+
assert len(self.data[slice_]) == len(new_data)
|
| 517 |
+
self.data[slice_] = new_data
|
| 518 |
+
except AssertionError:
|
| 519 |
+
raise IndexError(
|
| 520 |
+
f"Cannot `set()` value via at_indices={slice_} (option "
|
| 521 |
+
f"neg_index_as_lookback={neg_index_as_lookback})! Slice of data "
|
| 522 |
+
"does NOT have the same size as `new_data`."
|
| 523 |
+
)
|
| 524 |
+
|
| 525 |
+
def _get_int_index(
|
| 526 |
+
self,
|
| 527 |
+
idx: int,
|
| 528 |
+
fill=None,
|
| 529 |
+
neg_index_as_lookback=False,
|
| 530 |
+
one_hot_discrete=False,
|
| 531 |
+
_ignore_last_ts=False,
|
| 532 |
+
_add_last_ts_value=None,
|
| 533 |
+
):
|
| 534 |
+
data_to_use = self.data
|
| 535 |
+
if _ignore_last_ts:
|
| 536 |
+
data_to_use = self.data[:-1]
|
| 537 |
+
if _add_last_ts_value is not None:
|
| 538 |
+
data_to_use = np.append(data_to_use.copy(), _add_last_ts_value)
|
| 539 |
+
|
| 540 |
+
# If index >= 0 -> Ignore lookback buffer.
|
| 541 |
+
# Otherwise, include lookback buffer.
|
| 542 |
+
if idx >= 0 or neg_index_as_lookback:
|
| 543 |
+
idx = self.lookback + idx
|
| 544 |
+
# Negative indices mean: Go to left into lookback buffer starting from idx=0.
|
| 545 |
+
# But if we pass the lookback buffer, the index should be invalid and we will
|
| 546 |
+
# have to fill, if required. Invalidate the index by setting it to one larger
|
| 547 |
+
# than max.
|
| 548 |
+
if neg_index_as_lookback and idx < 0:
|
| 549 |
+
idx = len(self) + self.lookback - (_ignore_last_ts is True)
|
| 550 |
+
|
| 551 |
+
try:
|
| 552 |
+
if self.finalized:
|
| 553 |
+
data = tree.map_structure(lambda s: s[idx], data_to_use)
|
| 554 |
+
else:
|
| 555 |
+
data = data_to_use[idx]
|
| 556 |
+
# Out of range index -> If `fill`, use a fill dummy (B=0), if not, error out.
|
| 557 |
+
except IndexError as e:
|
| 558 |
+
if fill is not None:
|
| 559 |
+
if self.space is None:
|
| 560 |
+
return fill
|
| 561 |
+
return get_dummy_batch_for_space(
|
| 562 |
+
self.space,
|
| 563 |
+
fill_value=fill,
|
| 564 |
+
batch_size=0,
|
| 565 |
+
one_hot_discrete=one_hot_discrete,
|
| 566 |
+
)
|
| 567 |
+
else:
|
| 568 |
+
raise e
|
| 569 |
+
|
| 570 |
+
# Convert discrete/multi-discrete components to one-hot vectors, if required.
|
| 571 |
+
if one_hot_discrete:
|
| 572 |
+
data = self._one_hot(data, self.space_struct)
|
| 573 |
+
return data
|
| 574 |
+
|
| 575 |
+
def _set_int_index(self, new_data, idx, neg_index_as_lookback):
|
| 576 |
+
actual_idx = idx
|
| 577 |
+
# If index >= 0 -> Ignore lookback buffer.
|
| 578 |
+
# Otherwise, include lookback buffer.
|
| 579 |
+
if actual_idx >= 0 or neg_index_as_lookback:
|
| 580 |
+
actual_idx = self.lookback + actual_idx
|
| 581 |
+
# Negative indices mean: Go to left into lookback buffer starting from idx=0.
|
| 582 |
+
# But if we pass the lookback buffer, the index should be invalid and we will
|
| 583 |
+
# have to fill, if required. Invalidate the index by setting it to one larger
|
| 584 |
+
# than max.
|
| 585 |
+
if neg_index_as_lookback and actual_idx < 0:
|
| 586 |
+
actual_idx = len(self) + self.lookback
|
| 587 |
+
|
| 588 |
+
try:
|
| 589 |
+
if self.finalized:
|
| 590 |
+
|
| 591 |
+
def __set(s, n):
|
| 592 |
+
if self.space:
|
| 593 |
+
assert self.space.contains(n), n
|
| 594 |
+
s[actual_idx] = n
|
| 595 |
+
|
| 596 |
+
tree.map_structure(__set, self.data, new_data)
|
| 597 |
+
else:
|
| 598 |
+
self.data[actual_idx] = new_data
|
| 599 |
+
except IndexError:
|
| 600 |
+
raise IndexError(
|
| 601 |
+
f"Cannot `set()` value at index {idx} (option "
|
| 602 |
+
f"neg_index_as_lookback={neg_index_as_lookback})! Out of range "
|
| 603 |
+
f"of buffer data."
|
| 604 |
+
)
|
| 605 |
+
|
| 606 |
+
def _interpret_slice(
|
| 607 |
+
self,
|
| 608 |
+
slice_,
|
| 609 |
+
neg_index_as_lookback,
|
| 610 |
+
len_self_plus_lookback=None,
|
| 611 |
+
):
|
| 612 |
+
if len_self_plus_lookback is None:
|
| 613 |
+
len_self_plus_lookback = len(self) + self.lookback
|
| 614 |
+
|
| 615 |
+
# Re-interpret slice bounds as absolute positions (>=0) within our
|
| 616 |
+
# internal data.
|
| 617 |
+
start = slice_.start
|
| 618 |
+
stop = slice_.stop
|
| 619 |
+
|
| 620 |
+
# Start is None -> Exclude lookback buffer.
|
| 621 |
+
if start is None:
|
| 622 |
+
start = self.lookback
|
| 623 |
+
# Start is negative.
|
| 624 |
+
elif start < 0:
|
| 625 |
+
# `neg_index_as_lookback=True` -> User wants to index into the lookback
|
| 626 |
+
# range.
|
| 627 |
+
if neg_index_as_lookback:
|
| 628 |
+
start = self.lookback + start
|
| 629 |
+
# Interpret index as counting "from end".
|
| 630 |
+
else:
|
| 631 |
+
start = len_self_plus_lookback + start
|
| 632 |
+
# Start is 0 or positive -> timestep right after lookback is interpreted as 0.
|
| 633 |
+
else:
|
| 634 |
+
start = self.lookback + start
|
| 635 |
+
|
| 636 |
+
# Stop is None -> Set stop to very last index + 1 of our internal data.
|
| 637 |
+
if stop is None:
|
| 638 |
+
stop = len_self_plus_lookback
|
| 639 |
+
# Stop is negative.
|
| 640 |
+
elif stop < 0:
|
| 641 |
+
# `neg_index_as_lookback=True` -> User wants to index into the lookback
|
| 642 |
+
# range. Set to 0 (beginning of lookback buffer) if result is a negative
|
| 643 |
+
# index.
|
| 644 |
+
if neg_index_as_lookback:
|
| 645 |
+
stop = self.lookback + stop
|
| 646 |
+
# Interpret index as counting "from end". Set to 0 (beginning of actual
|
| 647 |
+
# episode) if result is a negative index.
|
| 648 |
+
else:
|
| 649 |
+
stop = len_self_plus_lookback + stop
|
| 650 |
+
# Stop is positive -> Add lookback range to it.
|
| 651 |
+
else:
|
| 652 |
+
stop = self.lookback + stop
|
| 653 |
+
|
| 654 |
+
fill_left_count = fill_right_count = 0
|
| 655 |
+
# Both start and stop are on left side.
|
| 656 |
+
if start < 0 and stop < 0:
|
| 657 |
+
fill_left_count = abs(start - stop)
|
| 658 |
+
fill_right_count = 0
|
| 659 |
+
start = stop = 0
|
| 660 |
+
# Both start and stop are on right side.
|
| 661 |
+
elif start >= len_self_plus_lookback and stop >= len_self_plus_lookback:
|
| 662 |
+
fill_right_count = abs(start - stop)
|
| 663 |
+
fill_left_count = 0
|
| 664 |
+
start = stop = len_self_plus_lookback
|
| 665 |
+
# Set to 0 (beginning of actual episode) if result is a negative index.
|
| 666 |
+
elif start < 0:
|
| 667 |
+
fill_left_count = -start
|
| 668 |
+
start = 0
|
| 669 |
+
elif stop >= len_self_plus_lookback:
|
| 670 |
+
fill_right_count = stop - len_self_plus_lookback
|
| 671 |
+
stop = len_self_plus_lookback
|
| 672 |
+
# Only `stop` might be < 0, when slice has negative step and start is > 0.
|
| 673 |
+
elif stop < 0:
|
| 674 |
+
if start >= len_self_plus_lookback:
|
| 675 |
+
fill_left_count = start - len_self_plus_lookback + 1
|
| 676 |
+
start = len_self_plus_lookback - 1
|
| 677 |
+
fill_right_count = -stop - 1
|
| 678 |
+
stop = -LARGE_INTEGER
|
| 679 |
+
|
| 680 |
+
assert start >= 0 and (stop >= 0 or stop == -LARGE_INTEGER), (start, stop)
|
| 681 |
+
|
| 682 |
+
step = slice_.step if slice_.step is not None else 1
|
| 683 |
+
slice_ = slice(start, stop, step)
|
| 684 |
+
slice_len = max(0, (stop - start + (step - (1 if step > 0 else -1))) // step)
|
| 685 |
+
return slice_, slice_len, fill_left_count, fill_right_count
|
| 686 |
+
|
| 687 |
+
def _one_hot(self, data, space_struct):
|
| 688 |
+
if space_struct is None:
|
| 689 |
+
raise ValueError(
|
| 690 |
+
f"Cannot `one_hot` data in `{type(self).__name__}` if a "
|
| 691 |
+
"gym.Space was NOT provided during construction!"
|
| 692 |
+
)
|
| 693 |
+
|
| 694 |
+
def _convert(dat_, space):
|
| 695 |
+
if isinstance(space, gym.spaces.Discrete):
|
| 696 |
+
return one_hot(dat_, depth=space.n)
|
| 697 |
+
elif isinstance(space, gym.spaces.MultiDiscrete):
|
| 698 |
+
return one_hot_multidiscrete(dat_, depths=space.nvec)
|
| 699 |
+
return dat_
|
| 700 |
+
|
| 701 |
+
if isinstance(data, list):
|
| 702 |
+
data = [
|
| 703 |
+
tree.map_structure(_convert, dslice, space_struct) for dslice in data
|
| 704 |
+
]
|
| 705 |
+
else:
|
| 706 |
+
data = tree.map_structure(_convert, data, space_struct)
|
| 707 |
+
return data
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/__init__.cpython-310.pyc
ADDED
|
Binary file (492 Bytes). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/action_dist.cpython-310.pyc
ADDED
|
Binary file (4.33 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/catalog.cpython-310.pyc
ADDED
|
Binary file (21.1 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/distributions.cpython-310.pyc
ADDED
|
Binary file (8.96 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/modelv2.cpython-310.pyc
ADDED
|
Binary file (14.9 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/preprocessors.cpython-310.pyc
ADDED
|
Binary file (14.4 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/repeated_values.cpython-310.pyc
ADDED
|
Binary file (7.15 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/__pycache__/utils.cpython-310.pyc
ADDED
|
Binary file (6.51 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/preprocessors.py
ADDED
|
@@ -0,0 +1,447 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from collections import OrderedDict
|
| 2 |
+
import logging
|
| 3 |
+
import numpy as np
|
| 4 |
+
import gymnasium as gym
|
| 5 |
+
from typing import Any, List
|
| 6 |
+
|
| 7 |
+
from ray.rllib.utils.annotations import OldAPIStack, override
|
| 8 |
+
from ray.rllib.utils.spaces.repeated import Repeated
|
| 9 |
+
from ray.rllib.utils.typing import TensorType
|
| 10 |
+
from ray.rllib.utils.images import resize
|
| 11 |
+
from ray.rllib.utils.spaces.space_utils import convert_element_to_space_type
|
| 12 |
+
|
| 13 |
+
ATARI_OBS_SHAPE = (210, 160, 3)
|
| 14 |
+
ATARI_RAM_OBS_SHAPE = (128,)
|
| 15 |
+
|
| 16 |
+
# Only validate env observations vs the observation space every n times in a
|
| 17 |
+
# Preprocessor.
|
| 18 |
+
OBS_VALIDATION_INTERVAL = 100
|
| 19 |
+
|
| 20 |
+
logger = logging.getLogger(__name__)
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
@OldAPIStack
|
| 24 |
+
class Preprocessor:
|
| 25 |
+
"""Defines an abstract observation preprocessor function.
|
| 26 |
+
|
| 27 |
+
Attributes:
|
| 28 |
+
shape (List[int]): Shape of the preprocessed output.
|
| 29 |
+
"""
|
| 30 |
+
|
| 31 |
+
def __init__(self, obs_space: gym.Space, options: dict = None):
|
| 32 |
+
_legacy_patch_shapes(obs_space)
|
| 33 |
+
self._obs_space = obs_space
|
| 34 |
+
if not options:
|
| 35 |
+
from ray.rllib.models.catalog import MODEL_DEFAULTS
|
| 36 |
+
|
| 37 |
+
self._options = MODEL_DEFAULTS.copy()
|
| 38 |
+
else:
|
| 39 |
+
self._options = options
|
| 40 |
+
self.shape = self._init_shape(obs_space, self._options)
|
| 41 |
+
self._size = int(np.prod(self.shape))
|
| 42 |
+
self._i = 0
|
| 43 |
+
self._obs_for_type_matching = self._obs_space.sample()
|
| 44 |
+
|
| 45 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 46 |
+
"""Returns the shape after preprocessing."""
|
| 47 |
+
raise NotImplementedError
|
| 48 |
+
|
| 49 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 50 |
+
"""Returns the preprocessed observation."""
|
| 51 |
+
raise NotImplementedError
|
| 52 |
+
|
| 53 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 54 |
+
"""Alternative to transform for more efficient flattening."""
|
| 55 |
+
array[offset : offset + self._size] = self.transform(observation)
|
| 56 |
+
|
| 57 |
+
def check_shape(self, observation: Any) -> None:
|
| 58 |
+
"""Checks the shape of the given observation."""
|
| 59 |
+
if self._i % OBS_VALIDATION_INTERVAL == 0:
|
| 60 |
+
# Convert lists to np.ndarrays.
|
| 61 |
+
if type(observation) is list and isinstance(
|
| 62 |
+
self._obs_space, gym.spaces.Box
|
| 63 |
+
):
|
| 64 |
+
observation = np.array(observation).astype(np.float32)
|
| 65 |
+
if not self._obs_space.contains(observation):
|
| 66 |
+
observation = convert_element_to_space_type(
|
| 67 |
+
observation, self._obs_for_type_matching
|
| 68 |
+
)
|
| 69 |
+
try:
|
| 70 |
+
if not self._obs_space.contains(observation):
|
| 71 |
+
raise ValueError(
|
| 72 |
+
"Observation ({} dtype={}) outside given space ({})!".format(
|
| 73 |
+
observation,
|
| 74 |
+
observation.dtype
|
| 75 |
+
if isinstance(self._obs_space, gym.spaces.Box)
|
| 76 |
+
else None,
|
| 77 |
+
self._obs_space,
|
| 78 |
+
)
|
| 79 |
+
)
|
| 80 |
+
except AttributeError as e:
|
| 81 |
+
raise ValueError(
|
| 82 |
+
"Observation for a Box/MultiBinary/MultiDiscrete space "
|
| 83 |
+
"should be an np.array, not a Python list.",
|
| 84 |
+
observation,
|
| 85 |
+
) from e
|
| 86 |
+
self._i += 1
|
| 87 |
+
|
| 88 |
+
@property
|
| 89 |
+
def size(self) -> int:
|
| 90 |
+
return self._size
|
| 91 |
+
|
| 92 |
+
@property
|
| 93 |
+
def observation_space(self) -> gym.Space:
|
| 94 |
+
obs_space = gym.spaces.Box(-1.0, 1.0, self.shape, dtype=np.float32)
|
| 95 |
+
# Stash the unwrapped space so that we can unwrap dict and tuple spaces
|
| 96 |
+
# automatically in modelv2.py
|
| 97 |
+
classes = (
|
| 98 |
+
DictFlatteningPreprocessor,
|
| 99 |
+
OneHotPreprocessor,
|
| 100 |
+
RepeatedValuesPreprocessor,
|
| 101 |
+
TupleFlatteningPreprocessor,
|
| 102 |
+
AtariRamPreprocessor,
|
| 103 |
+
GenericPixelPreprocessor,
|
| 104 |
+
)
|
| 105 |
+
if isinstance(self, classes):
|
| 106 |
+
obs_space.original_space = self._obs_space
|
| 107 |
+
return obs_space
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
@OldAPIStack
|
| 111 |
+
class GenericPixelPreprocessor(Preprocessor):
|
| 112 |
+
"""Generic image preprocessor.
|
| 113 |
+
|
| 114 |
+
Note: for Atari games, use config {"preprocessor_pref": "deepmind"}
|
| 115 |
+
instead for deepmind-style Atari preprocessing.
|
| 116 |
+
"""
|
| 117 |
+
|
| 118 |
+
@override(Preprocessor)
|
| 119 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 120 |
+
self._grayscale = options.get("grayscale")
|
| 121 |
+
self._zero_mean = options.get("zero_mean")
|
| 122 |
+
self._dim = options.get("dim")
|
| 123 |
+
if self._grayscale:
|
| 124 |
+
shape = (self._dim, self._dim, 1)
|
| 125 |
+
else:
|
| 126 |
+
shape = (self._dim, self._dim, 3)
|
| 127 |
+
|
| 128 |
+
return shape
|
| 129 |
+
|
| 130 |
+
@override(Preprocessor)
|
| 131 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 132 |
+
"""Downsamples images from (210, 160, 3) by the configured factor."""
|
| 133 |
+
self.check_shape(observation)
|
| 134 |
+
scaled = observation[25:-25, :, :]
|
| 135 |
+
if self._dim < 84:
|
| 136 |
+
scaled = resize(scaled, height=84, width=84)
|
| 137 |
+
# OpenAI: Resize by half, then down to 42x42 (essentially mipmapping).
|
| 138 |
+
# If we resize directly we lose pixels that, when mapped to 42x42,
|
| 139 |
+
# aren't close enough to the pixel boundary.
|
| 140 |
+
scaled = resize(scaled, height=self._dim, width=self._dim)
|
| 141 |
+
if self._grayscale:
|
| 142 |
+
scaled = scaled.mean(2)
|
| 143 |
+
scaled = scaled.astype(np.float32)
|
| 144 |
+
# Rescale needed for maintaining 1 channel
|
| 145 |
+
scaled = np.reshape(scaled, [self._dim, self._dim, 1])
|
| 146 |
+
if self._zero_mean:
|
| 147 |
+
scaled = (scaled - 128) / 128
|
| 148 |
+
else:
|
| 149 |
+
scaled *= 1.0 / 255.0
|
| 150 |
+
return scaled
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
@OldAPIStack
|
| 154 |
+
class AtariRamPreprocessor(Preprocessor):
|
| 155 |
+
@override(Preprocessor)
|
| 156 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 157 |
+
return (128,)
|
| 158 |
+
|
| 159 |
+
@override(Preprocessor)
|
| 160 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 161 |
+
self.check_shape(observation)
|
| 162 |
+
return (observation.astype("float32") - 128) / 128
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
@OldAPIStack
|
| 166 |
+
class OneHotPreprocessor(Preprocessor):
|
| 167 |
+
"""One-hot preprocessor for Discrete and MultiDiscrete spaces.
|
| 168 |
+
|
| 169 |
+
.. testcode::
|
| 170 |
+
:skipif: True
|
| 171 |
+
|
| 172 |
+
self.transform(Discrete(3).sample())
|
| 173 |
+
|
| 174 |
+
.. testoutput::
|
| 175 |
+
|
| 176 |
+
np.array([0.0, 1.0, 0.0])
|
| 177 |
+
|
| 178 |
+
.. testcode::
|
| 179 |
+
:skipif: True
|
| 180 |
+
|
| 181 |
+
self.transform(MultiDiscrete([2, 3]).sample())
|
| 182 |
+
|
| 183 |
+
.. testoutput::
|
| 184 |
+
|
| 185 |
+
np.array([0.0, 1.0, 0.0, 0.0, 1.0])
|
| 186 |
+
"""
|
| 187 |
+
|
| 188 |
+
@override(Preprocessor)
|
| 189 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 190 |
+
if isinstance(obs_space, gym.spaces.Discrete):
|
| 191 |
+
return (self._obs_space.n,)
|
| 192 |
+
else:
|
| 193 |
+
return (np.sum(self._obs_space.nvec),)
|
| 194 |
+
|
| 195 |
+
@override(Preprocessor)
|
| 196 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 197 |
+
self.check_shape(observation)
|
| 198 |
+
return gym.spaces.utils.flatten(self._obs_space, observation).astype(np.float32)
|
| 199 |
+
|
| 200 |
+
@override(Preprocessor)
|
| 201 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 202 |
+
array[offset : offset + self.size] = self.transform(observation)
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
@OldAPIStack
|
| 206 |
+
class NoPreprocessor(Preprocessor):
|
| 207 |
+
@override(Preprocessor)
|
| 208 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 209 |
+
return self._obs_space.shape
|
| 210 |
+
|
| 211 |
+
@override(Preprocessor)
|
| 212 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 213 |
+
self.check_shape(observation)
|
| 214 |
+
return observation
|
| 215 |
+
|
| 216 |
+
@override(Preprocessor)
|
| 217 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 218 |
+
array[offset : offset + self._size] = np.array(observation, copy=False).ravel()
|
| 219 |
+
|
| 220 |
+
@property
|
| 221 |
+
@override(Preprocessor)
|
| 222 |
+
def observation_space(self) -> gym.Space:
|
| 223 |
+
return self._obs_space
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
@OldAPIStack
|
| 227 |
+
class MultiBinaryPreprocessor(Preprocessor):
|
| 228 |
+
"""Preprocessor that turns a MultiBinary space into a Box.
|
| 229 |
+
|
| 230 |
+
Note: Before RLModules were introduced, RLlib's ModelCatalogV2 would produce
|
| 231 |
+
ComplexInputNetworks that treat MultiBinary spaces as Boxes. This preprocessor is
|
| 232 |
+
needed to get rid of the ComplexInputNetworks and use RLModules instead because
|
| 233 |
+
RLModules lack the logic to handle MultiBinary or other non-Box spaces.
|
| 234 |
+
"""
|
| 235 |
+
|
| 236 |
+
@override(Preprocessor)
|
| 237 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 238 |
+
return self._obs_space.shape
|
| 239 |
+
|
| 240 |
+
@override(Preprocessor)
|
| 241 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 242 |
+
# The shape stays the same, but the dtype changes.
|
| 243 |
+
self.check_shape(observation)
|
| 244 |
+
return observation.astype(np.float32)
|
| 245 |
+
|
| 246 |
+
@override(Preprocessor)
|
| 247 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 248 |
+
array[offset : offset + self._size] = np.array(observation, copy=False).ravel()
|
| 249 |
+
|
| 250 |
+
@property
|
| 251 |
+
@override(Preprocessor)
|
| 252 |
+
def observation_space(self) -> gym.Space:
|
| 253 |
+
obs_space = gym.spaces.Box(0.0, 1.0, self.shape, dtype=np.float32)
|
| 254 |
+
obs_space.original_space = self._obs_space
|
| 255 |
+
return obs_space
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
@OldAPIStack
|
| 259 |
+
class TupleFlatteningPreprocessor(Preprocessor):
|
| 260 |
+
"""Preprocesses each tuple element, then flattens it all into a vector.
|
| 261 |
+
|
| 262 |
+
RLlib models will unpack the flattened output before _build_layers_v2().
|
| 263 |
+
"""
|
| 264 |
+
|
| 265 |
+
@override(Preprocessor)
|
| 266 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 267 |
+
assert isinstance(self._obs_space, gym.spaces.Tuple)
|
| 268 |
+
size = 0
|
| 269 |
+
self.preprocessors = []
|
| 270 |
+
for i in range(len(self._obs_space.spaces)):
|
| 271 |
+
space = self._obs_space.spaces[i]
|
| 272 |
+
logger.debug("Creating sub-preprocessor for {}".format(space))
|
| 273 |
+
preprocessor_class = get_preprocessor(space)
|
| 274 |
+
if preprocessor_class is not None:
|
| 275 |
+
preprocessor = preprocessor_class(space, self._options)
|
| 276 |
+
size += preprocessor.size
|
| 277 |
+
else:
|
| 278 |
+
preprocessor = None
|
| 279 |
+
size += int(np.prod(space.shape))
|
| 280 |
+
self.preprocessors.append(preprocessor)
|
| 281 |
+
return (size,)
|
| 282 |
+
|
| 283 |
+
@override(Preprocessor)
|
| 284 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 285 |
+
self.check_shape(observation)
|
| 286 |
+
array = np.zeros(self.shape, dtype=np.float32)
|
| 287 |
+
self.write(observation, array, 0)
|
| 288 |
+
return array
|
| 289 |
+
|
| 290 |
+
@override(Preprocessor)
|
| 291 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 292 |
+
assert len(observation) == len(self.preprocessors), observation
|
| 293 |
+
for o, p in zip(observation, self.preprocessors):
|
| 294 |
+
p.write(o, array, offset)
|
| 295 |
+
offset += p.size
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
@OldAPIStack
|
| 299 |
+
class DictFlatteningPreprocessor(Preprocessor):
|
| 300 |
+
"""Preprocesses each dict value, then flattens it all into a vector.
|
| 301 |
+
|
| 302 |
+
RLlib models will unpack the flattened output before _build_layers_v2().
|
| 303 |
+
"""
|
| 304 |
+
|
| 305 |
+
@override(Preprocessor)
|
| 306 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 307 |
+
assert isinstance(self._obs_space, gym.spaces.Dict)
|
| 308 |
+
size = 0
|
| 309 |
+
self.preprocessors = []
|
| 310 |
+
for space in self._obs_space.spaces.values():
|
| 311 |
+
logger.debug("Creating sub-preprocessor for {}".format(space))
|
| 312 |
+
preprocessor_class = get_preprocessor(space)
|
| 313 |
+
if preprocessor_class is not None:
|
| 314 |
+
preprocessor = preprocessor_class(space, self._options)
|
| 315 |
+
size += preprocessor.size
|
| 316 |
+
else:
|
| 317 |
+
preprocessor = None
|
| 318 |
+
size += int(np.prod(space.shape))
|
| 319 |
+
self.preprocessors.append(preprocessor)
|
| 320 |
+
return (size,)
|
| 321 |
+
|
| 322 |
+
@override(Preprocessor)
|
| 323 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 324 |
+
self.check_shape(observation)
|
| 325 |
+
array = np.zeros(self.shape, dtype=np.float32)
|
| 326 |
+
self.write(observation, array, 0)
|
| 327 |
+
return array
|
| 328 |
+
|
| 329 |
+
@override(Preprocessor)
|
| 330 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 331 |
+
if not isinstance(observation, OrderedDict):
|
| 332 |
+
observation = OrderedDict(sorted(observation.items()))
|
| 333 |
+
assert len(observation) == len(self.preprocessors), (
|
| 334 |
+
len(observation),
|
| 335 |
+
len(self.preprocessors),
|
| 336 |
+
)
|
| 337 |
+
for o, p in zip(observation.values(), self.preprocessors):
|
| 338 |
+
p.write(o, array, offset)
|
| 339 |
+
offset += p.size
|
| 340 |
+
|
| 341 |
+
|
| 342 |
+
@OldAPIStack
|
| 343 |
+
class RepeatedValuesPreprocessor(Preprocessor):
|
| 344 |
+
"""Pads and batches the variable-length list value."""
|
| 345 |
+
|
| 346 |
+
@override(Preprocessor)
|
| 347 |
+
def _init_shape(self, obs_space: gym.Space, options: dict) -> List[int]:
|
| 348 |
+
assert isinstance(self._obs_space, Repeated)
|
| 349 |
+
child_space = obs_space.child_space
|
| 350 |
+
self.child_preprocessor = get_preprocessor(child_space)(
|
| 351 |
+
child_space, self._options
|
| 352 |
+
)
|
| 353 |
+
# The first slot encodes the list length.
|
| 354 |
+
size = 1 + self.child_preprocessor.size * obs_space.max_len
|
| 355 |
+
return (size,)
|
| 356 |
+
|
| 357 |
+
@override(Preprocessor)
|
| 358 |
+
def transform(self, observation: TensorType) -> np.ndarray:
|
| 359 |
+
array = np.zeros(self.shape)
|
| 360 |
+
if isinstance(observation, list):
|
| 361 |
+
for elem in observation:
|
| 362 |
+
self.child_preprocessor.check_shape(elem)
|
| 363 |
+
else:
|
| 364 |
+
pass # ValueError will be raised in write() below.
|
| 365 |
+
self.write(observation, array, 0)
|
| 366 |
+
return array
|
| 367 |
+
|
| 368 |
+
@override(Preprocessor)
|
| 369 |
+
def write(self, observation: TensorType, array: np.ndarray, offset: int) -> None:
|
| 370 |
+
if not isinstance(observation, (list, np.ndarray)):
|
| 371 |
+
raise ValueError(
|
| 372 |
+
"Input for {} must be list type, got {}".format(self, observation)
|
| 373 |
+
)
|
| 374 |
+
elif len(observation) > self._obs_space.max_len:
|
| 375 |
+
raise ValueError(
|
| 376 |
+
"Input {} exceeds max len of space {}".format(
|
| 377 |
+
observation, self._obs_space.max_len
|
| 378 |
+
)
|
| 379 |
+
)
|
| 380 |
+
# The first slot encodes the list length.
|
| 381 |
+
array[offset] = len(observation)
|
| 382 |
+
for i, elem in enumerate(observation):
|
| 383 |
+
offset_i = offset + 1 + i * self.child_preprocessor.size
|
| 384 |
+
self.child_preprocessor.write(elem, array, offset_i)
|
| 385 |
+
|
| 386 |
+
|
| 387 |
+
@OldAPIStack
|
| 388 |
+
def get_preprocessor(space: gym.Space, include_multi_binary=False) -> type:
|
| 389 |
+
"""Returns an appropriate preprocessor class for the given space."""
|
| 390 |
+
|
| 391 |
+
_legacy_patch_shapes(space)
|
| 392 |
+
obs_shape = space.shape
|
| 393 |
+
|
| 394 |
+
if isinstance(space, (gym.spaces.Discrete, gym.spaces.MultiDiscrete)):
|
| 395 |
+
preprocessor = OneHotPreprocessor
|
| 396 |
+
elif obs_shape == ATARI_OBS_SHAPE:
|
| 397 |
+
logger.debug(
|
| 398 |
+
"Defaulting to RLlib's GenericPixelPreprocessor because input "
|
| 399 |
+
"space has the atari-typical shape {}. Turn this behaviour off by setting "
|
| 400 |
+
"`preprocessor_pref=None` or "
|
| 401 |
+
"`preprocessor_pref='deepmind'` or disabling the preprocessing API "
|
| 402 |
+
"altogether with `_disable_preprocessor_api=True`.".format(ATARI_OBS_SHAPE)
|
| 403 |
+
)
|
| 404 |
+
preprocessor = GenericPixelPreprocessor
|
| 405 |
+
elif obs_shape == ATARI_RAM_OBS_SHAPE:
|
| 406 |
+
logger.debug(
|
| 407 |
+
"Defaulting to RLlib's AtariRamPreprocessor because input "
|
| 408 |
+
"space has the atari-typical shape {}. Turn this behaviour off by setting "
|
| 409 |
+
"`preprocessor_pref=None` or "
|
| 410 |
+
"`preprocessor_pref='deepmind' or disabling the preprocessing API "
|
| 411 |
+
"altogether with `_disable_preprocessor_api=True`."
|
| 412 |
+
"`.".format(ATARI_OBS_SHAPE)
|
| 413 |
+
)
|
| 414 |
+
preprocessor = AtariRamPreprocessor
|
| 415 |
+
elif isinstance(space, gym.spaces.Tuple):
|
| 416 |
+
preprocessor = TupleFlatteningPreprocessor
|
| 417 |
+
elif isinstance(space, gym.spaces.Dict):
|
| 418 |
+
preprocessor = DictFlatteningPreprocessor
|
| 419 |
+
elif isinstance(space, Repeated):
|
| 420 |
+
preprocessor = RepeatedValuesPreprocessor
|
| 421 |
+
# We usually only want to include this when using RLModules
|
| 422 |
+
elif isinstance(space, gym.spaces.MultiBinary) and include_multi_binary:
|
| 423 |
+
preprocessor = MultiBinaryPreprocessor
|
| 424 |
+
else:
|
| 425 |
+
preprocessor = NoPreprocessor
|
| 426 |
+
|
| 427 |
+
return preprocessor
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
def _legacy_patch_shapes(space: gym.Space) -> List[int]:
|
| 431 |
+
"""Assigns shapes to spaces that don't have shapes.
|
| 432 |
+
|
| 433 |
+
This is only needed for older gym versions that don't set shapes properly
|
| 434 |
+
for Tuple and Discrete spaces.
|
| 435 |
+
"""
|
| 436 |
+
|
| 437 |
+
if not hasattr(space, "shape"):
|
| 438 |
+
if isinstance(space, gym.spaces.Discrete):
|
| 439 |
+
space.shape = ()
|
| 440 |
+
elif isinstance(space, gym.spaces.Tuple):
|
| 441 |
+
shapes = []
|
| 442 |
+
for s in space.spaces:
|
| 443 |
+
shape = _legacy_patch_shapes(s)
|
| 444 |
+
shapes.append(shape)
|
| 445 |
+
space.shape = tuple(shapes)
|
| 446 |
+
|
| 447 |
+
return space.shape
|
vllm/lib/python3.10/site-packages/ray/rllib/models/repeated_values.py
ADDED
|
@@ -0,0 +1,204 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import List
|
| 2 |
+
|
| 3 |
+
from ray.rllib.utils.annotations import OldAPIStack
|
| 4 |
+
from ray.rllib.utils.typing import TensorType, TensorStructType
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
@OldAPIStack
|
| 8 |
+
class RepeatedValues:
|
| 9 |
+
"""Represents a variable-length list of items from spaces.Repeated.
|
| 10 |
+
|
| 11 |
+
RepeatedValues are created when you use spaces.Repeated, and are
|
| 12 |
+
accessible as part of input_dict["obs"] in ModelV2 forward functions.
|
| 13 |
+
|
| 14 |
+
Example:
|
| 15 |
+
Suppose the gym space definition was:
|
| 16 |
+
Repeated(Repeated(Box(K), N), M)
|
| 17 |
+
|
| 18 |
+
Then in the model forward function, input_dict["obs"] is of type:
|
| 19 |
+
RepeatedValues(RepeatedValues(<Tensor shape=(B, M, N, K)>))
|
| 20 |
+
|
| 21 |
+
The tensor is accessible via:
|
| 22 |
+
input_dict["obs"].values.values
|
| 23 |
+
|
| 24 |
+
And the actual data lengths via:
|
| 25 |
+
# outer repetition, shape [B], range [0, M]
|
| 26 |
+
input_dict["obs"].lengths
|
| 27 |
+
-and-
|
| 28 |
+
# inner repetition, shape [B, M], range [0, N]
|
| 29 |
+
input_dict["obs"].values.lengths
|
| 30 |
+
|
| 31 |
+
Attributes:
|
| 32 |
+
values: The padded data tensor of shape [B, max_len, ..., sz],
|
| 33 |
+
where B is the batch dimension, max_len is the max length of this
|
| 34 |
+
list, followed by any number of sub list max lens, followed by the
|
| 35 |
+
actual data size.
|
| 36 |
+
lengths (List[int]): Tensor of shape [B, ...] that represents the
|
| 37 |
+
number of valid items in each list. When the list is nested within
|
| 38 |
+
other lists, there will be extra dimensions for the parent list
|
| 39 |
+
max lens.
|
| 40 |
+
max_len: The max number of items allowed in each list.
|
| 41 |
+
|
| 42 |
+
TODO(ekl): support conversion to tf.RaggedTensor.
|
| 43 |
+
"""
|
| 44 |
+
|
| 45 |
+
def __init__(self, values: TensorType, lengths: List[int], max_len: int):
|
| 46 |
+
self.values = values
|
| 47 |
+
self.lengths = lengths
|
| 48 |
+
self.max_len = max_len
|
| 49 |
+
self._unbatched_repr = None
|
| 50 |
+
|
| 51 |
+
def unbatch_all(self) -> List[List[TensorType]]:
|
| 52 |
+
"""Unbatch both the repeat and batch dimensions into Python lists.
|
| 53 |
+
|
| 54 |
+
This is only supported in PyTorch / TF eager mode.
|
| 55 |
+
|
| 56 |
+
This lets you view the data unbatched in its original form, but is
|
| 57 |
+
not efficient for processing.
|
| 58 |
+
|
| 59 |
+
.. testcode::
|
| 60 |
+
:skipif: True
|
| 61 |
+
|
| 62 |
+
batch = RepeatedValues(<Tensor shape=(B, N, K)>)
|
| 63 |
+
items = batch.unbatch_all()
|
| 64 |
+
print(len(items) == B)
|
| 65 |
+
|
| 66 |
+
.. testoutput::
|
| 67 |
+
|
| 68 |
+
True
|
| 69 |
+
|
| 70 |
+
.. testcode::
|
| 71 |
+
:skipif: True
|
| 72 |
+
|
| 73 |
+
print(max(len(x) for x in items) <= N)
|
| 74 |
+
|
| 75 |
+
.. testoutput::
|
| 76 |
+
|
| 77 |
+
True
|
| 78 |
+
|
| 79 |
+
.. testcode::
|
| 80 |
+
:skipif: True
|
| 81 |
+
|
| 82 |
+
print(items)
|
| 83 |
+
|
| 84 |
+
.. testoutput::
|
| 85 |
+
|
| 86 |
+
[[<Tensor_1 shape=(K)>, ..., <Tensor_N, shape=(K)>],
|
| 87 |
+
...
|
| 88 |
+
[<Tensor_1 shape=(K)>, <Tensor_2 shape=(K)>],
|
| 89 |
+
...
|
| 90 |
+
[<Tensor_1 shape=(K)>],
|
| 91 |
+
...
|
| 92 |
+
[<Tensor_1 shape=(K)>, ..., <Tensor_N shape=(K)>]]
|
| 93 |
+
"""
|
| 94 |
+
|
| 95 |
+
if self._unbatched_repr is None:
|
| 96 |
+
B = _get_batch_dim_helper(self.values)
|
| 97 |
+
if B is None:
|
| 98 |
+
raise ValueError(
|
| 99 |
+
"Cannot call unbatch_all() when batch_dim is unknown. "
|
| 100 |
+
"This is probably because you are using TF graph mode."
|
| 101 |
+
)
|
| 102 |
+
else:
|
| 103 |
+
B = int(B)
|
| 104 |
+
slices = self.unbatch_repeat_dim()
|
| 105 |
+
result = []
|
| 106 |
+
for i in range(B):
|
| 107 |
+
if hasattr(self.lengths[i], "item"):
|
| 108 |
+
dynamic_len = int(self.lengths[i].item())
|
| 109 |
+
else:
|
| 110 |
+
dynamic_len = int(self.lengths[i].numpy())
|
| 111 |
+
dynamic_slice = []
|
| 112 |
+
for j in range(dynamic_len):
|
| 113 |
+
dynamic_slice.append(_batch_index_helper(slices, i, j))
|
| 114 |
+
result.append(dynamic_slice)
|
| 115 |
+
self._unbatched_repr = result
|
| 116 |
+
|
| 117 |
+
return self._unbatched_repr
|
| 118 |
+
|
| 119 |
+
def unbatch_repeat_dim(self) -> List[TensorType]:
|
| 120 |
+
"""Unbatches the repeat dimension (the one `max_len` in size).
|
| 121 |
+
|
| 122 |
+
This removes the repeat dimension. The result will be a Python list of
|
| 123 |
+
with length `self.max_len`. Note that the data is still padded.
|
| 124 |
+
|
| 125 |
+
.. testcode::
|
| 126 |
+
:skipif: True
|
| 127 |
+
|
| 128 |
+
batch = RepeatedValues(<Tensor shape=(B, N, K)>)
|
| 129 |
+
items = batch.unbatch()
|
| 130 |
+
len(items) == batch.max_len
|
| 131 |
+
|
| 132 |
+
.. testoutput::
|
| 133 |
+
|
| 134 |
+
True
|
| 135 |
+
|
| 136 |
+
.. testcode::
|
| 137 |
+
:skipif: True
|
| 138 |
+
|
| 139 |
+
print(items)
|
| 140 |
+
|
| 141 |
+
.. testoutput::
|
| 142 |
+
|
| 143 |
+
[<Tensor_1 shape=(B, K)>, ..., <Tensor_N shape=(B, K)>]
|
| 144 |
+
"""
|
| 145 |
+
return _unbatch_helper(self.values, self.max_len)
|
| 146 |
+
|
| 147 |
+
def __repr__(self):
|
| 148 |
+
return "RepeatedValues(value={}, lengths={}, max_len={})".format(
|
| 149 |
+
repr(self.values), repr(self.lengths), self.max_len
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
def __str__(self):
|
| 153 |
+
return repr(self)
|
| 154 |
+
|
| 155 |
+
|
| 156 |
+
def _get_batch_dim_helper(v: TensorStructType) -> int:
|
| 157 |
+
"""Tries to find the batch dimension size of v, or None."""
|
| 158 |
+
if isinstance(v, dict):
|
| 159 |
+
for u in v.values():
|
| 160 |
+
return _get_batch_dim_helper(u)
|
| 161 |
+
elif isinstance(v, tuple):
|
| 162 |
+
return _get_batch_dim_helper(v[0])
|
| 163 |
+
elif isinstance(v, RepeatedValues):
|
| 164 |
+
return _get_batch_dim_helper(v.values)
|
| 165 |
+
else:
|
| 166 |
+
B = v.shape[0]
|
| 167 |
+
if hasattr(B, "value"):
|
| 168 |
+
B = B.value # TensorFlow
|
| 169 |
+
return B
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def _unbatch_helper(v: TensorStructType, max_len: int) -> TensorStructType:
|
| 173 |
+
"""Recursively unpacks the repeat dimension (max_len)."""
|
| 174 |
+
if isinstance(v, dict):
|
| 175 |
+
return {k: _unbatch_helper(u, max_len) for (k, u) in v.items()}
|
| 176 |
+
elif isinstance(v, tuple):
|
| 177 |
+
return tuple(_unbatch_helper(u, max_len) for u in v)
|
| 178 |
+
elif isinstance(v, RepeatedValues):
|
| 179 |
+
unbatched = _unbatch_helper(v.values, max_len)
|
| 180 |
+
return [
|
| 181 |
+
RepeatedValues(u, v.lengths[:, i, ...], v.max_len)
|
| 182 |
+
for i, u in enumerate(unbatched)
|
| 183 |
+
]
|
| 184 |
+
else:
|
| 185 |
+
return [v[:, i, ...] for i in range(max_len)]
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
def _batch_index_helper(v: TensorStructType, i: int, j: int) -> TensorStructType:
|
| 189 |
+
"""Selects the item at the ith batch index and jth repetition."""
|
| 190 |
+
if isinstance(v, dict):
|
| 191 |
+
return {k: _batch_index_helper(u, i, j) for (k, u) in v.items()}
|
| 192 |
+
elif isinstance(v, tuple):
|
| 193 |
+
return tuple(_batch_index_helper(u, i, j) for u in v)
|
| 194 |
+
elif isinstance(v, list):
|
| 195 |
+
# This is the output of unbatch_repeat_dim(). Unfortunately we have to
|
| 196 |
+
# process it here instead of in unbatch_all(), since it may be buried
|
| 197 |
+
# under a dict / tuple.
|
| 198 |
+
return _batch_index_helper(v[j], i, j)
|
| 199 |
+
elif isinstance(v, RepeatedValues):
|
| 200 |
+
unbatched = v.unbatch_all()
|
| 201 |
+
# Don't need to select j here; that's already done in unbatch_all.
|
| 202 |
+
return unbatched[i]
|
| 203 |
+
else:
|
| 204 |
+
return v[i, ...]
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/attention_net.cpython-310.pyc
ADDED
|
Binary file (15.8 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/complex_input_net.cpython-310.pyc
ADDED
|
Binary file (4.7 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/fcnet.cpython-310.pyc
ADDED
|
Binary file (3.49 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/misc.cpython-310.pyc
ADDED
|
Binary file (2.54 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/noop.cpython-310.pyc
ADDED
|
Binary file (956 Bytes). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/recurrent_net.cpython-310.pyc
ADDED
|
Binary file (8.72 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/tf_distributions.cpython-310.pyc
ADDED
|
Binary file (20.9 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/__pycache__/visionnet.cpython-310.pyc
ADDED
|
Binary file (5.02 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/fcnet.py
ADDED
|
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
import gymnasium as gym
|
| 3 |
+
from typing import Dict
|
| 4 |
+
|
| 5 |
+
from ray.rllib.models.tf.misc import normc_initializer
|
| 6 |
+
from ray.rllib.models.tf.tf_modelv2 import TFModelV2
|
| 7 |
+
from ray.rllib.models.utils import get_activation_fn
|
| 8 |
+
from ray.rllib.utils.annotations import OldAPIStack
|
| 9 |
+
from ray.rllib.utils.framework import try_import_tf
|
| 10 |
+
from ray.rllib.utils.typing import TensorType, List, ModelConfigDict
|
| 11 |
+
|
| 12 |
+
tf1, tf, tfv = try_import_tf()
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
@OldAPIStack
|
| 16 |
+
class FullyConnectedNetwork(TFModelV2):
|
| 17 |
+
"""Generic fully connected network implemented in ModelV2 API."""
|
| 18 |
+
|
| 19 |
+
def __init__(
|
| 20 |
+
self,
|
| 21 |
+
obs_space: gym.spaces.Space,
|
| 22 |
+
action_space: gym.spaces.Space,
|
| 23 |
+
num_outputs: int,
|
| 24 |
+
model_config: ModelConfigDict,
|
| 25 |
+
name: str,
|
| 26 |
+
):
|
| 27 |
+
super(FullyConnectedNetwork, self).__init__(
|
| 28 |
+
obs_space, action_space, num_outputs, model_config, name
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
hiddens = list(model_config.get("fcnet_hiddens", [])) + list(
|
| 32 |
+
model_config.get("post_fcnet_hiddens", [])
|
| 33 |
+
)
|
| 34 |
+
activation = model_config.get("fcnet_activation")
|
| 35 |
+
if not model_config.get("fcnet_hiddens", []):
|
| 36 |
+
activation = model_config.get("post_fcnet_activation")
|
| 37 |
+
activation = get_activation_fn(activation)
|
| 38 |
+
no_final_linear = model_config.get("no_final_linear")
|
| 39 |
+
vf_share_layers = model_config.get("vf_share_layers")
|
| 40 |
+
free_log_std = model_config.get("free_log_std")
|
| 41 |
+
|
| 42 |
+
# Generate free-floating bias variables for the second half of
|
| 43 |
+
# the outputs.
|
| 44 |
+
if free_log_std:
|
| 45 |
+
assert num_outputs % 2 == 0, (
|
| 46 |
+
"num_outputs must be divisible by two",
|
| 47 |
+
num_outputs,
|
| 48 |
+
)
|
| 49 |
+
num_outputs = num_outputs // 2
|
| 50 |
+
self.log_std_var = tf.Variable(
|
| 51 |
+
[0.0] * num_outputs, dtype=tf.float32, name="log_std"
|
| 52 |
+
)
|
| 53 |
+
|
| 54 |
+
# We are using obs_flat, so take the flattened shape as input.
|
| 55 |
+
inputs = tf.keras.layers.Input(
|
| 56 |
+
shape=(int(np.prod(obs_space.shape)),), name="observations"
|
| 57 |
+
)
|
| 58 |
+
# Last hidden layer output (before logits outputs).
|
| 59 |
+
last_layer = inputs
|
| 60 |
+
# The action distribution outputs.
|
| 61 |
+
logits_out = None
|
| 62 |
+
i = 1
|
| 63 |
+
|
| 64 |
+
# Create layers 0 to second-last.
|
| 65 |
+
for size in hiddens[:-1]:
|
| 66 |
+
last_layer = tf.keras.layers.Dense(
|
| 67 |
+
size,
|
| 68 |
+
name="fc_{}".format(i),
|
| 69 |
+
activation=activation,
|
| 70 |
+
kernel_initializer=normc_initializer(1.0),
|
| 71 |
+
)(last_layer)
|
| 72 |
+
i += 1
|
| 73 |
+
|
| 74 |
+
# The last layer is adjusted to be of size num_outputs, but it's a
|
| 75 |
+
# layer with activation.
|
| 76 |
+
if no_final_linear and num_outputs:
|
| 77 |
+
logits_out = tf.keras.layers.Dense(
|
| 78 |
+
num_outputs,
|
| 79 |
+
name="fc_out",
|
| 80 |
+
activation=activation,
|
| 81 |
+
kernel_initializer=normc_initializer(1.0),
|
| 82 |
+
)(last_layer)
|
| 83 |
+
# Finish the layers with the provided sizes (`hiddens`), plus -
|
| 84 |
+
# iff num_outputs > 0 - a last linear layer of size num_outputs.
|
| 85 |
+
else:
|
| 86 |
+
if len(hiddens) > 0:
|
| 87 |
+
last_layer = tf.keras.layers.Dense(
|
| 88 |
+
hiddens[-1],
|
| 89 |
+
name="fc_{}".format(i),
|
| 90 |
+
activation=activation,
|
| 91 |
+
kernel_initializer=normc_initializer(1.0),
|
| 92 |
+
)(last_layer)
|
| 93 |
+
if num_outputs:
|
| 94 |
+
logits_out = tf.keras.layers.Dense(
|
| 95 |
+
num_outputs,
|
| 96 |
+
name="fc_out",
|
| 97 |
+
activation=None,
|
| 98 |
+
kernel_initializer=normc_initializer(0.01),
|
| 99 |
+
)(last_layer)
|
| 100 |
+
# Adjust num_outputs to be the number of nodes in the last layer.
|
| 101 |
+
else:
|
| 102 |
+
self.num_outputs = ([int(np.prod(obs_space.shape))] + hiddens[-1:])[-1]
|
| 103 |
+
|
| 104 |
+
# Concat the log std vars to the end of the state-dependent means.
|
| 105 |
+
if free_log_std and logits_out is not None:
|
| 106 |
+
|
| 107 |
+
def tiled_log_std(x):
|
| 108 |
+
return tf.tile(tf.expand_dims(self.log_std_var, 0), [tf.shape(x)[0], 1])
|
| 109 |
+
|
| 110 |
+
log_std_out = tf.keras.layers.Lambda(tiled_log_std)(inputs)
|
| 111 |
+
logits_out = tf.keras.layers.Concatenate(axis=1)([logits_out, log_std_out])
|
| 112 |
+
|
| 113 |
+
last_vf_layer = None
|
| 114 |
+
if not vf_share_layers:
|
| 115 |
+
# Build a parallel set of hidden layers for the value net.
|
| 116 |
+
last_vf_layer = inputs
|
| 117 |
+
i = 1
|
| 118 |
+
for size in hiddens:
|
| 119 |
+
last_vf_layer = tf.keras.layers.Dense(
|
| 120 |
+
size,
|
| 121 |
+
name="fc_value_{}".format(i),
|
| 122 |
+
activation=activation,
|
| 123 |
+
kernel_initializer=normc_initializer(1.0),
|
| 124 |
+
)(last_vf_layer)
|
| 125 |
+
i += 1
|
| 126 |
+
|
| 127 |
+
value_out = tf.keras.layers.Dense(
|
| 128 |
+
1,
|
| 129 |
+
name="value_out",
|
| 130 |
+
activation=None,
|
| 131 |
+
kernel_initializer=normc_initializer(0.01),
|
| 132 |
+
)(last_vf_layer if last_vf_layer is not None else last_layer)
|
| 133 |
+
|
| 134 |
+
self.base_model = tf.keras.Model(
|
| 135 |
+
inputs, [(logits_out if logits_out is not None else last_layer), value_out]
|
| 136 |
+
)
|
| 137 |
+
|
| 138 |
+
def forward(
|
| 139 |
+
self,
|
| 140 |
+
input_dict: Dict[str, TensorType],
|
| 141 |
+
state: List[TensorType],
|
| 142 |
+
seq_lens: TensorType,
|
| 143 |
+
) -> (TensorType, List[TensorType]):
|
| 144 |
+
model_out, self._value_out = self.base_model(input_dict["obs_flat"])
|
| 145 |
+
return model_out, state
|
| 146 |
+
|
| 147 |
+
def value_function(self) -> TensorType:
|
| 148 |
+
return tf.reshape(self._value_out, [-1])
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/__init__.cpython-310.pyc
ADDED
|
Binary file (675 Bytes). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/gru_gate.cpython-310.pyc
ADDED
|
Binary file (2.2 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/multi_head_attention.cpython-310.pyc
ADDED
|
Binary file (2.3 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/noisy_layer.cpython-310.pyc
ADDED
|
Binary file (3.43 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/relative_multi_head_attention.cpython-310.pyc
ADDED
|
Binary file (4.7 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/__pycache__/skip_connection.cpython-310.pyc
ADDED
|
Binary file (1.85 kB). View file
|
|
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/gru_gate.py
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from ray.rllib.utils.framework import try_import_tf
|
| 2 |
+
from ray.rllib.utils.typing import TensorType, TensorShape
|
| 3 |
+
from ray.rllib.utils.deprecation import deprecation_warning
|
| 4 |
+
from ray.util import log_once
|
| 5 |
+
|
| 6 |
+
tf1, tf, tfv = try_import_tf()
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class GRUGate(tf.keras.layers.Layer if tf else object):
|
| 10 |
+
def __init__(self, init_bias: float = 0.0, **kwargs):
|
| 11 |
+
super().__init__(**kwargs)
|
| 12 |
+
self._init_bias = init_bias
|
| 13 |
+
if log_once("gru_gate"):
|
| 14 |
+
deprecation_warning(
|
| 15 |
+
old="rllib.models.tf.layers.GRUGate",
|
| 16 |
+
)
|
| 17 |
+
|
| 18 |
+
def build(self, input_shape: TensorShape):
|
| 19 |
+
h_shape, x_shape = input_shape
|
| 20 |
+
if x_shape[-1] != h_shape[-1]:
|
| 21 |
+
raise ValueError(
|
| 22 |
+
"Both inputs to GRUGate must have equal size in last axis!"
|
| 23 |
+
)
|
| 24 |
+
|
| 25 |
+
dim = int(h_shape[-1])
|
| 26 |
+
self._w_r = self.add_weight(shape=(dim, dim))
|
| 27 |
+
self._w_z = self.add_weight(shape=(dim, dim))
|
| 28 |
+
self._w_h = self.add_weight(shape=(dim, dim))
|
| 29 |
+
|
| 30 |
+
self._u_r = self.add_weight(shape=(dim, dim))
|
| 31 |
+
self._u_z = self.add_weight(shape=(dim, dim))
|
| 32 |
+
self._u_h = self.add_weight(shape=(dim, dim))
|
| 33 |
+
|
| 34 |
+
def bias_initializer(shape, dtype):
|
| 35 |
+
return tf.fill(shape, tf.cast(self._init_bias, dtype=dtype))
|
| 36 |
+
|
| 37 |
+
self._bias_z = self.add_weight(shape=(dim,), initializer=bias_initializer)
|
| 38 |
+
|
| 39 |
+
def call(self, inputs: TensorType, **kwargs) -> TensorType:
|
| 40 |
+
# Pass in internal state first.
|
| 41 |
+
h, X = inputs
|
| 42 |
+
|
| 43 |
+
r = tf.tensordot(X, self._w_r, axes=1) + tf.tensordot(h, self._u_r, axes=1)
|
| 44 |
+
r = tf.nn.sigmoid(r)
|
| 45 |
+
|
| 46 |
+
z = (
|
| 47 |
+
tf.tensordot(X, self._w_z, axes=1)
|
| 48 |
+
+ tf.tensordot(h, self._u_z, axes=1)
|
| 49 |
+
- self._bias_z
|
| 50 |
+
)
|
| 51 |
+
z = tf.nn.sigmoid(z)
|
| 52 |
+
|
| 53 |
+
h_next = tf.tensordot(X, self._w_h, axes=1) + tf.tensordot(
|
| 54 |
+
(h * r), self._u_h, axes=1
|
| 55 |
+
)
|
| 56 |
+
h_next = tf.nn.tanh(h_next)
|
| 57 |
+
|
| 58 |
+
return (1 - z) * h + z * h_next
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/multi_head_attention.py
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
[1] - Attention Is All You Need - Vaswani, Jones, Shazeer, Parmar,
|
| 3 |
+
Uszkoreit, Gomez, Kaiser - Google Brain/Research, U Toronto - 2017.
|
| 4 |
+
https://arxiv.org/pdf/1706.03762.pdf
|
| 5 |
+
"""
|
| 6 |
+
from ray.rllib.utils.framework import try_import_tf
|
| 7 |
+
from ray.rllib.utils.typing import TensorType
|
| 8 |
+
from ray.rllib.utils.deprecation import deprecation_warning
|
| 9 |
+
from ray.util import log_once
|
| 10 |
+
|
| 11 |
+
tf1, tf, tfv = try_import_tf()
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
class MultiHeadAttention(tf.keras.layers.Layer if tf else object):
|
| 15 |
+
"""A multi-head attention layer described in [1]."""
|
| 16 |
+
|
| 17 |
+
def __init__(self, out_dim: int, num_heads: int, head_dim: int, **kwargs):
|
| 18 |
+
super().__init__(**kwargs)
|
| 19 |
+
|
| 20 |
+
# No bias or non-linearity.
|
| 21 |
+
self._num_heads = num_heads
|
| 22 |
+
self._head_dim = head_dim
|
| 23 |
+
self._qkv_layer = tf.keras.layers.Dense(
|
| 24 |
+
3 * num_heads * head_dim, use_bias=False
|
| 25 |
+
)
|
| 26 |
+
self._linear_layer = tf.keras.layers.TimeDistributed(
|
| 27 |
+
tf.keras.layers.Dense(out_dim, use_bias=False)
|
| 28 |
+
)
|
| 29 |
+
if log_once("multi_head_attention"):
|
| 30 |
+
deprecation_warning(
|
| 31 |
+
old="rllib.models.tf.layers.MultiHeadAttention",
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
def call(self, inputs: TensorType) -> TensorType:
|
| 35 |
+
L = tf.shape(inputs)[1] # length of segment
|
| 36 |
+
H = self._num_heads # number of attention heads
|
| 37 |
+
D = self._head_dim # attention head dimension
|
| 38 |
+
|
| 39 |
+
qkv = self._qkv_layer(inputs)
|
| 40 |
+
|
| 41 |
+
queries, keys, values = tf.split(qkv, 3, -1)
|
| 42 |
+
queries = queries[:, -L:] # only query based on the segment
|
| 43 |
+
|
| 44 |
+
queries = tf.reshape(queries, [-1, L, H, D])
|
| 45 |
+
keys = tf.reshape(keys, [-1, L, H, D])
|
| 46 |
+
values = tf.reshape(values, [-1, L, H, D])
|
| 47 |
+
|
| 48 |
+
score = tf.einsum("bihd,bjhd->bijh", queries, keys)
|
| 49 |
+
score = score / D**0.5
|
| 50 |
+
|
| 51 |
+
# causal mask of the same length as the sequence
|
| 52 |
+
mask = tf.sequence_mask(tf.range(1, L + 1), dtype=score.dtype)
|
| 53 |
+
mask = mask[None, :, :, None]
|
| 54 |
+
|
| 55 |
+
masked_score = score * mask + 1e30 * (mask - 1.0)
|
| 56 |
+
wmat = tf.nn.softmax(masked_score, axis=2)
|
| 57 |
+
|
| 58 |
+
out = tf.einsum("bijh,bjhd->bihd", wmat, values)
|
| 59 |
+
shape = tf.concat([tf.shape(out)[:2], [H * D]], axis=0)
|
| 60 |
+
out = tf.reshape(out, shape)
|
| 61 |
+
return self._linear_layer(out)
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/noisy_layer.py
ADDED
|
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
|
| 3 |
+
from ray.rllib.models.utils import get_activation_fn
|
| 4 |
+
from ray.rllib.utils.framework import (
|
| 5 |
+
get_variable,
|
| 6 |
+
try_import_tf,
|
| 7 |
+
TensorType,
|
| 8 |
+
TensorShape,
|
| 9 |
+
)
|
| 10 |
+
from ray.rllib.utils.deprecation import deprecation_warning
|
| 11 |
+
from ray.util import log_once
|
| 12 |
+
|
| 13 |
+
tf1, tf, tfv = try_import_tf()
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class NoisyLayer(tf.keras.layers.Layer if tf else object):
|
| 17 |
+
r"""A Layer that adds learnable Noise to some previous layer's outputs.
|
| 18 |
+
|
| 19 |
+
Consists of:
|
| 20 |
+
- a common dense layer: y = w^{T}x + b
|
| 21 |
+
- a noisy layer: y = (w + \epsilon_w*\sigma_w)^{T}x +
|
| 22 |
+
(b+\epsilon_b*\sigma_b)
|
| 23 |
+
, where \epsilon are random variables sampled from factorized normal
|
| 24 |
+
distributions and \sigma are trainable variables which are expected to
|
| 25 |
+
vanish along the training procedure.
|
| 26 |
+
"""
|
| 27 |
+
|
| 28 |
+
def __init__(
|
| 29 |
+
self, prefix: str, out_size: int, sigma0: float, activation: str = "relu"
|
| 30 |
+
):
|
| 31 |
+
"""Initializes a NoisyLayer object.
|
| 32 |
+
|
| 33 |
+
Args:
|
| 34 |
+
prefix:
|
| 35 |
+
out_size: Output size for Noisy Layer
|
| 36 |
+
sigma0: Initialization value for sigma_b (bias noise)
|
| 37 |
+
non_linear: Non-linear activation for Noisy Layer
|
| 38 |
+
"""
|
| 39 |
+
super().__init__()
|
| 40 |
+
self.prefix = prefix
|
| 41 |
+
self.out_size = out_size
|
| 42 |
+
# TF noise generation can be unreliable on GPU
|
| 43 |
+
# If generating the noise on the CPU,
|
| 44 |
+
# lowering sigma0 to 0.1 may be helpful
|
| 45 |
+
self.sigma0 = sigma0 # 0.5~GPU, 0.1~CPU
|
| 46 |
+
self.activation = activation
|
| 47 |
+
# Variables.
|
| 48 |
+
self.w = None # Weight matrix.
|
| 49 |
+
self.b = None # Biases.
|
| 50 |
+
self.sigma_w = None # Noise for weight matrix
|
| 51 |
+
self.sigma_b = None # Noise for biases.
|
| 52 |
+
if log_once("noisy_layer"):
|
| 53 |
+
deprecation_warning(
|
| 54 |
+
old="rllib.models.tf.layers.NoisyLayer",
|
| 55 |
+
)
|
| 56 |
+
|
| 57 |
+
def build(self, input_shape: TensorShape):
|
| 58 |
+
in_size = int(input_shape[1])
|
| 59 |
+
|
| 60 |
+
self.sigma_w = get_variable(
|
| 61 |
+
value=tf.keras.initializers.RandomUniform(
|
| 62 |
+
minval=-1.0 / np.sqrt(float(in_size)),
|
| 63 |
+
maxval=1.0 / np.sqrt(float(in_size)),
|
| 64 |
+
),
|
| 65 |
+
trainable=True,
|
| 66 |
+
tf_name=self.prefix + "_sigma_w",
|
| 67 |
+
shape=[in_size, self.out_size],
|
| 68 |
+
dtype=tf.float32,
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
self.sigma_b = get_variable(
|
| 72 |
+
value=tf.keras.initializers.Constant(self.sigma0 / np.sqrt(float(in_size))),
|
| 73 |
+
trainable=True,
|
| 74 |
+
tf_name=self.prefix + "_sigma_b",
|
| 75 |
+
shape=[self.out_size],
|
| 76 |
+
dtype=tf.float32,
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
self.w = get_variable(
|
| 80 |
+
value=tf.keras.initializers.GlorotUniform(),
|
| 81 |
+
tf_name=self.prefix + "_fc_w",
|
| 82 |
+
trainable=True,
|
| 83 |
+
shape=[in_size, self.out_size],
|
| 84 |
+
dtype=tf.float32,
|
| 85 |
+
)
|
| 86 |
+
|
| 87 |
+
self.b = get_variable(
|
| 88 |
+
value=tf.keras.initializers.Zeros(),
|
| 89 |
+
tf_name=self.prefix + "_fc_b",
|
| 90 |
+
trainable=True,
|
| 91 |
+
shape=[self.out_size],
|
| 92 |
+
dtype=tf.float32,
|
| 93 |
+
)
|
| 94 |
+
|
| 95 |
+
def call(self, inputs: TensorType) -> TensorType:
|
| 96 |
+
in_size = int(inputs.shape[1])
|
| 97 |
+
epsilon_in = tf.random.normal(shape=[in_size])
|
| 98 |
+
epsilon_out = tf.random.normal(shape=[self.out_size])
|
| 99 |
+
epsilon_in = self._f_epsilon(epsilon_in)
|
| 100 |
+
epsilon_out = self._f_epsilon(epsilon_out)
|
| 101 |
+
epsilon_w = tf.matmul(
|
| 102 |
+
a=tf.expand_dims(epsilon_in, -1), b=tf.expand_dims(epsilon_out, 0)
|
| 103 |
+
)
|
| 104 |
+
epsilon_b = epsilon_out
|
| 105 |
+
|
| 106 |
+
action_activation = (
|
| 107 |
+
tf.matmul(inputs, self.w + self.sigma_w * epsilon_w)
|
| 108 |
+
+ self.b
|
| 109 |
+
+ self.sigma_b * epsilon_b
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
fn = get_activation_fn(self.activation, framework="tf")
|
| 113 |
+
if fn is not None:
|
| 114 |
+
action_activation = fn(action_activation)
|
| 115 |
+
return action_activation
|
| 116 |
+
|
| 117 |
+
def _f_epsilon(self, x: TensorType) -> TensorType:
|
| 118 |
+
return tf.math.sign(x) * tf.math.sqrt(tf.math.abs(x))
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/relative_multi_head_attention.py
ADDED
|
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Optional
|
| 2 |
+
|
| 3 |
+
from ray.rllib.utils.framework import try_import_tf
|
| 4 |
+
from ray.rllib.utils.typing import TensorType
|
| 5 |
+
from ray.rllib.utils.deprecation import deprecation_warning
|
| 6 |
+
from ray.util import log_once
|
| 7 |
+
|
| 8 |
+
tf1, tf, tfv = try_import_tf()
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class RelativeMultiHeadAttention(tf.keras.layers.Layer if tf else object):
|
| 12 |
+
"""A RelativeMultiHeadAttention layer as described in [3].
|
| 13 |
+
|
| 14 |
+
Uses segment level recurrence with state reuse.
|
| 15 |
+
"""
|
| 16 |
+
|
| 17 |
+
def __init__(
|
| 18 |
+
self,
|
| 19 |
+
out_dim: int,
|
| 20 |
+
num_heads: int,
|
| 21 |
+
head_dim: int,
|
| 22 |
+
input_layernorm: bool = False,
|
| 23 |
+
output_activation: Optional["tf.nn.activation"] = None,
|
| 24 |
+
**kwargs
|
| 25 |
+
):
|
| 26 |
+
"""Initializes a RelativeMultiHeadAttention keras Layer object.
|
| 27 |
+
|
| 28 |
+
Args:
|
| 29 |
+
out_dim: The output dimensions of the multi-head attention
|
| 30 |
+
unit.
|
| 31 |
+
num_heads: The number of attention heads to use.
|
| 32 |
+
Denoted `H` in [2].
|
| 33 |
+
head_dim: The dimension of a single(!) attention head within
|
| 34 |
+
a multi-head attention unit. Denoted as `d` in [3].
|
| 35 |
+
input_layernorm: Whether to prepend a LayerNorm before
|
| 36 |
+
everything else. Should be True for building a GTrXL.
|
| 37 |
+
output_activation (Optional[tf.nn.activation]): Optional tf.nn
|
| 38 |
+
activation function. Should be relu for GTrXL.
|
| 39 |
+
**kwargs:
|
| 40 |
+
"""
|
| 41 |
+
if log_once("relative_multi_head_attention"):
|
| 42 |
+
deprecation_warning(
|
| 43 |
+
old="rllib.models.tf.layers.RelativeMultiHeadAttention",
|
| 44 |
+
)
|
| 45 |
+
super().__init__(**kwargs)
|
| 46 |
+
|
| 47 |
+
# No bias or non-linearity.
|
| 48 |
+
self._num_heads = num_heads
|
| 49 |
+
self._head_dim = head_dim
|
| 50 |
+
# 3=Query, key, and value inputs.
|
| 51 |
+
self._qkv_layer = tf.keras.layers.Dense(
|
| 52 |
+
3 * num_heads * head_dim, use_bias=False
|
| 53 |
+
)
|
| 54 |
+
self._linear_layer = tf.keras.layers.TimeDistributed(
|
| 55 |
+
tf.keras.layers.Dense(out_dim, use_bias=False, activation=output_activation)
|
| 56 |
+
)
|
| 57 |
+
|
| 58 |
+
self._uvar = self.add_weight(shape=(num_heads, head_dim))
|
| 59 |
+
self._vvar = self.add_weight(shape=(num_heads, head_dim))
|
| 60 |
+
|
| 61 |
+
# Constant (non-trainable) sinusoid rel pos encoding matrix, which
|
| 62 |
+
# depends on this incoming time dimension.
|
| 63 |
+
# For inference, we prepend the memory to the current timestep's
|
| 64 |
+
# input: Tau + 1. For training, we prepend the memory to the input
|
| 65 |
+
# sequence: Tau + T.
|
| 66 |
+
self._pos_embedding = PositionalEmbedding(out_dim)
|
| 67 |
+
self._pos_proj = tf.keras.layers.Dense(num_heads * head_dim, use_bias=False)
|
| 68 |
+
|
| 69 |
+
self._input_layernorm = None
|
| 70 |
+
if input_layernorm:
|
| 71 |
+
self._input_layernorm = tf.keras.layers.LayerNormalization(axis=-1)
|
| 72 |
+
|
| 73 |
+
def call(
|
| 74 |
+
self, inputs: TensorType, memory: Optional[TensorType] = None
|
| 75 |
+
) -> TensorType:
|
| 76 |
+
T = tf.shape(inputs)[1] # length of segment (time)
|
| 77 |
+
H = self._num_heads # number of attention heads
|
| 78 |
+
d = self._head_dim # attention head dimension
|
| 79 |
+
|
| 80 |
+
# Add previous memory chunk (as const, w/o gradient) to input.
|
| 81 |
+
# Tau (number of (prev) time slices in each memory chunk).
|
| 82 |
+
Tau = tf.shape(memory)[1]
|
| 83 |
+
inputs = tf.concat([tf.stop_gradient(memory), inputs], axis=1)
|
| 84 |
+
|
| 85 |
+
# Apply the Layer-Norm.
|
| 86 |
+
if self._input_layernorm is not None:
|
| 87 |
+
inputs = self._input_layernorm(inputs)
|
| 88 |
+
|
| 89 |
+
qkv = self._qkv_layer(inputs)
|
| 90 |
+
|
| 91 |
+
queries, keys, values = tf.split(qkv, 3, -1)
|
| 92 |
+
# Cut out memory timesteps from query.
|
| 93 |
+
queries = queries[:, -T:]
|
| 94 |
+
|
| 95 |
+
# Splitting up queries into per-head dims (d).
|
| 96 |
+
queries = tf.reshape(queries, [-1, T, H, d])
|
| 97 |
+
keys = tf.reshape(keys, [-1, Tau + T, H, d])
|
| 98 |
+
values = tf.reshape(values, [-1, Tau + T, H, d])
|
| 99 |
+
|
| 100 |
+
R = self._pos_embedding(Tau + T)
|
| 101 |
+
R = self._pos_proj(R)
|
| 102 |
+
R = tf.reshape(R, [Tau + T, H, d])
|
| 103 |
+
|
| 104 |
+
# b=batch
|
| 105 |
+
# i and j=time indices (i=max-timesteps (inputs); j=Tau memory space)
|
| 106 |
+
# h=head
|
| 107 |
+
# d=head-dim (over which we will reduce-sum)
|
| 108 |
+
score = tf.einsum("bihd,bjhd->bijh", queries + self._uvar, keys)
|
| 109 |
+
pos_score = tf.einsum("bihd,jhd->bijh", queries + self._vvar, R)
|
| 110 |
+
score = score + self.rel_shift(pos_score)
|
| 111 |
+
score = score / d**0.5
|
| 112 |
+
|
| 113 |
+
# Causal mask of the same length as the sequence.
|
| 114 |
+
mask = tf.sequence_mask(tf.range(Tau + 1, Tau + T + 1), dtype=score.dtype)
|
| 115 |
+
mask = mask[None, :, :, None]
|
| 116 |
+
|
| 117 |
+
masked_score = score * mask + 1e30 * (mask - 1.0)
|
| 118 |
+
wmat = tf.nn.softmax(masked_score, axis=2)
|
| 119 |
+
|
| 120 |
+
out = tf.einsum("bijh,bjhd->bihd", wmat, values)
|
| 121 |
+
out = tf.reshape(out, tf.concat((tf.shape(out)[:2], [H * d]), axis=0))
|
| 122 |
+
return self._linear_layer(out)
|
| 123 |
+
|
| 124 |
+
@staticmethod
|
| 125 |
+
def rel_shift(x: TensorType) -> TensorType:
|
| 126 |
+
# Transposed version of the shift approach described in [3].
|
| 127 |
+
# https://github.com/kimiyoung/transformer-xl/blob/
|
| 128 |
+
# 44781ed21dbaec88b280f74d9ae2877f52b492a5/tf/model.py#L31
|
| 129 |
+
x_size = tf.shape(x)
|
| 130 |
+
|
| 131 |
+
x = tf.pad(x, [[0, 0], [0, 0], [1, 0], [0, 0]])
|
| 132 |
+
x = tf.reshape(x, [x_size[0], x_size[2] + 1, x_size[1], x_size[3]])
|
| 133 |
+
x = x[:, 1:, :, :]
|
| 134 |
+
x = tf.reshape(x, x_size)
|
| 135 |
+
|
| 136 |
+
return x
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
class PositionalEmbedding(tf.keras.layers.Layer if tf else object):
|
| 140 |
+
def __init__(self, out_dim, **kwargs):
|
| 141 |
+
super().__init__(**kwargs)
|
| 142 |
+
self.inverse_freq = 1 / (10000 ** (tf.range(0, out_dim, 2.0) / out_dim))
|
| 143 |
+
|
| 144 |
+
def call(self, seq_length):
|
| 145 |
+
pos_offsets = tf.cast(tf.range(seq_length - 1, -1, -1), tf.float32)
|
| 146 |
+
inputs = pos_offsets[:, None] * self.inverse_freq[None, :]
|
| 147 |
+
return tf.concat((tf.sin(inputs), tf.cos(inputs)), axis=-1)
|
vllm/lib/python3.10/site-packages/ray/rllib/models/tf/layers/skip_connection.py
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Optional, Any
|
| 2 |
+
|
| 3 |
+
from ray.rllib.utils.framework import try_import_tf
|
| 4 |
+
from ray.rllib.utils.typing import TensorType
|
| 5 |
+
from ray.rllib.utils.deprecation import deprecation_warning
|
| 6 |
+
from ray.util import log_once
|
| 7 |
+
|
| 8 |
+
tf1, tf, tfv = try_import_tf()
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class SkipConnection(tf.keras.layers.Layer if tf else object):
|
| 12 |
+
"""Skip connection layer.
|
| 13 |
+
|
| 14 |
+
Adds the original input to the output (regular residual layer) OR uses
|
| 15 |
+
input as hidden state input to a given fan_in_layer.
|
| 16 |
+
"""
|
| 17 |
+
|
| 18 |
+
def __init__(self, layer: Any, fan_in_layer: Optional[Any] = None, **kwargs):
|
| 19 |
+
"""Initializes a SkipConnection keras layer object.
|
| 20 |
+
|
| 21 |
+
Args:
|
| 22 |
+
layer (tf.keras.layers.Layer): Any layer processing inputs.
|
| 23 |
+
fan_in_layer (Optional[tf.keras.layers.Layer]): An optional
|
| 24 |
+
layer taking two inputs: The original input and the output
|
| 25 |
+
of `layer`.
|
| 26 |
+
"""
|
| 27 |
+
if log_once("skip_connection"):
|
| 28 |
+
deprecation_warning(
|
| 29 |
+
old="rllib.models.tf.layers.SkipConnection",
|
| 30 |
+
)
|
| 31 |
+
super().__init__(**kwargs)
|
| 32 |
+
self._layer = layer
|
| 33 |
+
self._fan_in_layer = fan_in_layer
|
| 34 |
+
|
| 35 |
+
def call(self, inputs: TensorType, **kwargs) -> TensorType:
|
| 36 |
+
# del kwargs
|
| 37 |
+
outputs = self._layer(inputs, **kwargs)
|
| 38 |
+
# Residual case, just add inputs to outputs.
|
| 39 |
+
if self._fan_in_layer is None:
|
| 40 |
+
outputs = outputs + inputs
|
| 41 |
+
# Fan-in e.g. RNN: Call fan-in with `inputs` and `outputs`.
|
| 42 |
+
else:
|
| 43 |
+
# NOTE: In the GRU case, `inputs` is the state input.
|
| 44 |
+
outputs = self._fan_in_layer((inputs, outputs))
|
| 45 |
+
|
| 46 |
+
return outputs
|