File size: 13,567 Bytes
079c32c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 |
from typing import Union, Optional, List, Any, Tuple
import torch
import os
from functools import partial
from tensorboardX import SummaryWriter
from copy import deepcopy
from ding.worker import BaseLearner, InteractionSerialEvaluator, BaseSerialCommander, create_buffer, \
get_buffer_cls, create_serial_collector
from ding.world_model import WorldModel
from ding.worker import IBuffer
from ding.envs import get_vec_env_setting, create_env_manager
from ding.config import read_config, compile_config
from ding.utils import set_pkg_seed, deep_merge_dicts
from ding.policy import create_policy
from ding.world_model import create_world_model
from ding.entry.utils import random_collect
def mbrl_entry_setup(
input_cfg: Union[str, Tuple[dict, dict]],
seed: int = 0,
env_setting: Optional[List[Any]] = None,
model: Optional[torch.nn.Module] = None,
) -> Tuple:
if isinstance(input_cfg, str):
cfg, create_cfg = read_config(input_cfg)
else:
cfg, create_cfg = deepcopy(input_cfg)
create_cfg.policy.type = create_cfg.policy.type + '_command'
env_fn = None if env_setting is None else env_setting[0]
cfg = compile_config(cfg, seed=seed, env=env_fn, auto=True, create_cfg=create_cfg, save_cfg=True)
if env_setting is None:
env_fn, collector_env_cfg, evaluator_env_cfg = get_vec_env_setting(cfg.env)
else:
env_fn, collector_env_cfg, evaluator_env_cfg = env_setting
collector_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in collector_env_cfg])
evaluator_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in evaluator_env_cfg])
collector_env.seed(cfg.seed)
evaluator_env.seed(cfg.seed, dynamic_seed=False)
set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda)
# create logger
tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial'))
# create world model
world_model = create_world_model(cfg.world_model, env_fn(cfg.env), tb_logger)
# create policy
policy = create_policy(cfg.policy, model=model, enable_field=['learn', 'collect', 'eval', 'command'])
# create worker
learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name)
collector = create_serial_collector(
cfg.policy.collect.collector,
env=collector_env,
policy=policy.collect_mode,
tb_logger=tb_logger,
exp_name=cfg.exp_name
)
evaluator = InteractionSerialEvaluator(
cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name
)
env_buffer = create_buffer(cfg.policy.other.replay_buffer, tb_logger=tb_logger, exp_name=cfg.exp_name)
commander = BaseSerialCommander(
cfg.policy.other.commander, learner, collector, evaluator, env_buffer, policy.command_mode
)
return (
cfg,
policy,
world_model,
env_buffer,
learner,
collector,
collector_env,
evaluator,
commander,
tb_logger,
)
def create_img_buffer(
cfg: dict, input_cfg: Union[str, Tuple[dict, dict]], world_model: WorldModel, tb_logger: 'SummaryWriter'
) -> IBuffer: # noqa
if isinstance(input_cfg, str):
_, create_cfg = read_config(input_cfg)
else:
_, create_cfg = input_cfg
img_buffer_cfg = cfg.world_model.other.imagination_buffer
img_buffer_cfg.update(create_cfg.imagination_buffer)
buffer_cls = get_buffer_cls(img_buffer_cfg)
cfg.world_model.other.imagination_buffer.update(deep_merge_dicts(buffer_cls.default_config(), img_buffer_cfg))
if img_buffer_cfg.type == 'elastic':
img_buffer_cfg.set_buffer_size = world_model.buffer_size_scheduler
img_buffer = create_buffer(cfg.world_model.other.imagination_buffer, tb_logger=tb_logger, exp_name=cfg.exp_name)
return img_buffer
def serial_pipeline_dyna(
input_cfg: Union[str, Tuple[dict, dict]],
seed: int = 0,
env_setting: Optional[List[Any]] = None,
model: Optional[torch.nn.Module] = None,
max_train_iter: Optional[int] = int(1e10),
max_env_step: Optional[int] = int(1e10),
) -> 'Policy': # noqa
"""
Overview:
Serial pipeline entry for dyna-style model-based RL.
Arguments:
- input_cfg (:obj:`Union[str, Tuple[dict, dict]]`): Config in dict type. \
``str`` type means config file path. \
``Tuple[dict, dict]`` type means [user_config, create_cfg].
- seed (:obj:`int`): Random seed.
- env_setting (:obj:`Optional[List[Any]]`): A list with 3 elements: \
``BaseEnv`` subclass, collector env config, and evaluator env config.
- model (:obj:`Optional[torch.nn.Module]`): Instance of torch.nn.Module.
- max_train_iter (:obj:`Optional[int]`): Maximum policy update iterations in training.
- max_env_step (:obj:`Optional[int]`): Maximum collected environment interaction steps.
Returns:
- policy (:obj:`Policy`): Converged policy.
"""
cfg, policy, world_model, env_buffer, learner, collector, collector_env, evaluator, commander, tb_logger = \
mbrl_entry_setup(input_cfg, seed, env_setting, model)
img_buffer = create_img_buffer(cfg, input_cfg, world_model, tb_logger)
learner.call_hook('before_run')
if cfg.policy.get('random_collect_size', 0) > 0:
random_collect(cfg.policy, policy, collector, collector_env, commander, env_buffer)
while True:
collect_kwargs = commander.step()
# eval the policy
if evaluator.should_eval(collector.envstep):
stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
if stop:
break
# fill environment buffer
data = collector.collect(train_iter=learner.train_iter, policy_kwargs=collect_kwargs)
env_buffer.push(data, cur_collector_envstep=collector.envstep)
# eval&train world model and fill imagination buffer
if world_model.should_eval(collector.envstep):
world_model.eval(env_buffer, collector.envstep, learner.train_iter)
if world_model.should_train(collector.envstep):
world_model.train(env_buffer, collector.envstep, learner.train_iter)
world_model.fill_img_buffer(
policy.collect_mode, env_buffer, img_buffer, collector.envstep, learner.train_iter
)
for i in range(cfg.policy.learn.update_per_collect):
batch_size = learner.policy.get_attribute('batch_size')
train_data = world_model.sample(env_buffer, img_buffer, batch_size, learner.train_iter)
learner.train(train_data, collector.envstep)
if cfg.policy.on_policy:
# On-policy algorithm must clear the replay buffer.
env_buffer.clear()
img_buffer.clear()
if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter:
break
learner.call_hook('after_run')
return policy
def serial_pipeline_dream(
input_cfg: Union[str, Tuple[dict, dict]],
seed: int = 0,
env_setting: Optional[List[Any]] = None,
model: Optional[torch.nn.Module] = None,
max_train_iter: Optional[int] = int(1e10),
max_env_step: Optional[int] = int(1e10),
) -> 'Policy': # noqa
"""
Overview:
Serial pipeline entry for dreamer-style model-based RL.
Arguments:
- input_cfg (:obj:`Union[str, Tuple[dict, dict]]`): Config in dict type. \
``str`` type means config file path. \
``Tuple[dict, dict]`` type means [user_config, create_cfg].
- seed (:obj:`int`): Random seed.
- env_setting (:obj:`Optional[List[Any]]`): A list with 3 elements: \
``BaseEnv`` subclass, collector env config, and evaluator env config.
- model (:obj:`Optional[torch.nn.Module]`): Instance of torch.nn.Module.
- max_train_iter (:obj:`Optional[int]`): Maximum policy update iterations in training.
- max_env_step (:obj:`Optional[int]`): Maximum collected environment interaction steps.
Returns:
- policy (:obj:`Policy`): Converged policy.
"""
cfg, policy, world_model, env_buffer, learner, collector, collector_env, evaluator, commander, tb_logger = \
mbrl_entry_setup(input_cfg, seed, env_setting, model)
learner.call_hook('before_run')
if cfg.policy.get('random_collect_size', 0) > 0:
random_collect(cfg.policy, policy, collector, collector_env, commander, env_buffer)
while True:
collect_kwargs = commander.step()
# eval the policy
if evaluator.should_eval(collector.envstep):
stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
if stop:
break
# fill environment buffer
data = collector.collect(train_iter=learner.train_iter, policy_kwargs=collect_kwargs)
env_buffer.push(data, cur_collector_envstep=collector.envstep)
# eval&train world model and fill imagination buffer
if world_model.should_eval(collector.envstep):
world_model.eval(env_buffer, collector.envstep, learner.train_iter)
if world_model.should_train(collector.envstep):
world_model.train(env_buffer, collector.envstep, learner.train_iter)
update_per_collect = cfg.policy.learn.update_per_collect // world_model.rollout_length_scheduler(
collector.envstep
)
update_per_collect = max(1, update_per_collect)
for i in range(update_per_collect):
batch_size = learner.policy.get_attribute('batch_size')
train_data = env_buffer.sample(batch_size, learner.train_iter)
# dreamer-style: use pure on-policy imagined rollout to train policy,
# which depends on the current envstep to decide the rollout length
learner.train(
train_data, collector.envstep, policy_kwargs=dict(world_model=world_model, envstep=collector.envstep)
)
if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter:
break
learner.call_hook('after_run')
return policy
def serial_pipeline_dreamer(
input_cfg: Union[str, Tuple[dict, dict]],
seed: int = 0,
env_setting: Optional[List[Any]] = None,
model: Optional[torch.nn.Module] = None,
max_train_iter: Optional[int] = int(1e10),
max_env_step: Optional[int] = int(1e10),
) -> 'Policy': # noqa
"""
Overview:
Serial pipeline entry for dreamerv3.
Arguments:
- input_cfg (:obj:`Union[str, Tuple[dict, dict]]`): Config in dict type. \
``str`` type means config file path. \
``Tuple[dict, dict]`` type means [user_config, create_cfg].
- seed (:obj:`int`): Random seed.
- env_setting (:obj:`Optional[List[Any]]`): A list with 3 elements: \
``BaseEnv`` subclass, collector env config, and evaluator env config.
- model (:obj:`Optional[torch.nn.Module]`): Instance of torch.nn.Module.
- max_train_iter (:obj:`Optional[int]`): Maximum policy update iterations in training.
- max_env_step (:obj:`Optional[int]`): Maximum collected environment interaction steps.
Returns:
- policy (:obj:`Policy`): Converged policy.
"""
cfg, policy, world_model, env_buffer, learner, collector, collector_env, evaluator, commander, tb_logger = \
mbrl_entry_setup(input_cfg, seed, env_setting, model)
learner.call_hook('before_run')
# prefill environment buffer
if cfg.policy.get('random_collect_size', 0) > 0:
cfg.policy.random_collect_size = cfg.policy.random_collect_size // cfg.policy.collect.unroll_len
random_collect(cfg.policy, policy, collector, collector_env, commander, env_buffer)
while True:
collect_kwargs = commander.step()
# eval the policy
if evaluator.should_eval(collector.envstep):
stop, reward = evaluator.eval(
learner.save_checkpoint,
learner.train_iter,
collector.envstep,
policy_kwargs=dict(world_model=world_model)
)
if stop:
break
# train world model and fill imagination buffer
steps = (
cfg.world_model.pretrain
if world_model.should_pretrain() else int(world_model.should_train(collector.envstep))
)
for _ in range(steps):
batch_size = learner.policy.get_attribute('batch_size')
batch_length = cfg.policy.learn.batch_length
post, context = world_model.train(
env_buffer, collector.envstep, learner.train_iter, batch_size, batch_length
)
start = post
learner.train(
start, collector.envstep, policy_kwargs=dict(world_model=world_model, envstep=collector.envstep)
)
# fill environment buffer
data = collector.collect(
train_iter=learner.train_iter,
policy_kwargs=dict(world_model=world_model, envstep=collector.envstep, **collect_kwargs)
)
env_buffer.push(data, cur_collector_envstep=collector.envstep)
if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter:
break
learner.call_hook('after_run')
return policy
|