-
Notifications
You must be signed in to change notification settings - Fork 7
/
maml.py
356 lines (314 loc) · 12.8 KB
/
maml.py
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
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
# -*- coding: utf-8 -*-
# @Author : Lin Lan (ryan.linlan@gmail.com)
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import sys
import os.path as osp
import logging
from collections import defaultdict
import pickle
import numpy as np
import ray
from ray.rllib.agents import Agent
from ray.rllib.agents.ppo.ppo import DEFAULT_CONFIG as ppo_default_config
from ray.rllib.utils import merge_dicts
from ray.rllib.evaluation.sample_batch import DEFAULT_POLICY_ID
from ray.rllib.env.env_context import EnvContext
from ray import tune
from ray.tune.trial import Resources
from maml.maml_policy_graph import MAMLPolicyGraph
from maml.maml_optimizer import MAMLOptimizer
from maml.maml_policy_evaluator import MAMLPolicyEvaluator
from envs.reset_wrapper import ResetArgsHolder
from utils import summarize_episodes
logger = logging.getLogger("ray.rllib.agents.maml.maml")
def on_episode_start(info):
episode = info["episode"]
episode.custom_metrics["episode_dist_reward"] = 0
def on_episode_step(info):
episode = info["episode"]
episode.custom_metrics["episode_dist_reward"] += \
episode.batch_builder.agent_builders[
"single_agent"].buffers["infos"][-1]["dist_reward"]
DEFAULT_CONFIG = merge_dicts(
ppo_default_config,
{
"random_seed": 1,
"inner_lr": 0.05,
"outer_lr": 1e-3,
"adaptive_inner_lr": False,
# "inner_lr_lower_bound": None,
# "inner_lr_upper_bound": None,
"inner_lr_bound": None,
"num_inner_updates": 3,
"inner_grad_clip": 40.0,
"num_tasks": 100,
"clip_param": 0.2,
# "linear_baseline": False,
"use_gae": True,
# "gamma": 0.99,
# "lambda": 0.97,
"gamma": 0.0,
"lambda": 0.0,
"horizon": 200,
"kl_coeff": 0.0,
"kl_target": 0.01,
"entropy_coeff": 0.0,
"vf_loss_coeff": 0.05,
"vf_clip_param": 15.0,
"model_loss_coeff": 0.0,
"num_sgd_iter": 10,
"validation": True,
"validation_freq": 5,
"sample_batch_size": 200,
"batch_mode": "complete_episodes",
"observation_filter": "NoFilter",
"num_workers": 20,
"num_envs_per_worker": 25,
"tf_session_args": {
"intra_op_parallelism_threads": 4,
"inter_op_parallelism_threads": 4
},
"callbacks": {
"on_episode_start": tune.function(on_episode_start),
"on_episode_step": tune.function(on_episode_step)
}
}
)
class MAMLAgent(Agent):
_agent_name = "MAML"
_default_config = DEFAULT_CONFIG
_policy_graph = MAMLPolicyGraph
_policy_evaluator = MAMLPolicyEvaluator
@classmethod
def default_resource_request(cls, config):
cf = merge_dicts(cls._default_config, config)
return Resources(
cpu=1,
gpu=0,
extra_cpu=cf["num_cpus_per_worker"] * cf["num_workers"] + 1,
extra_gpu=cf["num_gpus_per_worker"] * cf["num_workers"])
def make_local_evaluator(self, env_creator, policy_dict):
return self._make_evaluator(
self._policy_evaluator,
env_creator,
policy_dict,
0,
merge_dicts(
self.config, {
"tf_session_args": {
"intra_op_parallelism_threads": 8,
"inter_op_parallelism_threads": 8
}
}
))
def make_remote_evaluators(self, env_creator, policy_dict, count,
remote_args):
cls = self._policy_evaluator.as_remote(**remote_args).remote
return [
self._make_evaluator(cls, env_creator, policy_dict, i + 1,
self.config) for i in range(count)
]
def _init(self):
self._validate_config()
env = self.env_creator(env_config={})
reset_args_shape = (env.reset_args_config["shape"][0], )
self.reset_args_holder = ResetArgsHolder.remote(
(self.config["num_workers"], ) + reset_args_shape)
self.config["env_config"] = merge_dicts(
self.config["env_config"],
{"reset_args_holder": self.reset_args_holder})
self.rng = np.random.RandomState(self.config["random_seed"])
# print("sampling goals...")
self.reset_args_train, self.reset_args_test_1, self.reset_args_test_2 \
= env.sample_reset_args(self.rng, self.config["num_tasks"])
# print("sampling finished")
self.reset_args_test = {
1: self.reset_args_test_1,
2: self.reset_args_test_2}
observation_space = env.observation_space
action_space = env.action_space
policy_dict_local = {
DEFAULT_POLICY_ID: (
self._policy_graph,
observation_space,
action_space,
{"mode": "local"})}
policy_dict_remote = {
DEFAULT_POLICY_ID: (
self._policy_graph,
observation_space,
action_space,
{"mode": "remote"})}
self.local_evaluator = self.make_local_evaluator(
self.env_creator, policy_dict_local)
self.remote_evaluators = self.make_remote_evaluators(
self.env_creator, policy_dict_remote, self.config["num_workers"], {
"num_cpus": self.config["num_cpus_per_worker"],
"num_gpus": self.config["num_gpus_per_worker"]})
self.optimizer = MAMLOptimizer(
self.local_evaluator, self.remote_evaluators, {
"num_inner_updates": self.config["num_inner_updates"],
"num_sgd_iter": self.config["num_sgd_iter"]})
def _validate_config(self):
assert not self.config["adaptive_inner_lr"]
def _train(self):
batch_reset_args_indices = \
self.rng.choice(self.reset_args_train.shape[0],
size=self.config["num_workers"],
replace=False)
batch_reset_args = self.reset_args_train[batch_reset_args_indices]
ray.get(self.reset_args_holder.set.remote(batch_reset_args))
fetches = self.optimizer.step()
# if "kl" in fetches:
# raise NotImplementedError
res = self.optimizer.collect_metrics()
res.update(
info=dict(fetches, **res.get("info", {})))
res.update({"validation": None})
if self.config["validation"]:
if self.config["validation_freq"] == "auto":
if self._iteration <= 200:
validation_freq = 100
else:
validation_freq = 25
else:
validation_freq = self.config["validation_freq"]
if (self._iteration + 1) % validation_freq == 0:
val_results = self._validation_once()
res.update({"validation": val_results})
return res
def _validation_once(self):
val_results = {}
val_results["train"] = self._test(
self.reset_args_train, self.config["num_inner_updates"])
val_results["test_1"] = self._test(
self.reset_args_test_1, self.config["num_inner_updates"])
val_results["test_2"] = self._test(
self.reset_args_test_2, self.config["num_inner_updates"])
return val_results
def train(self):
results = Agent.__base__.train(self)
return results
def _test(self, reset_args, num_inner_updates):
num_tasks = reset_args.shape[0]
free_evaluators = list(
zip(range(1, self.config["num_workers"] + 1),
self.remote_evaluators))
weights_id = ray.put(self.local_evaluator.get_weights())
running = {}
finished = {}
episodes = defaultdict(list)
i = 0
while True:
if free_evaluators and i < num_tasks:
this_index, this_evaluator = free_evaluators.pop()
this_reset_args = reset_args[i]
reset_args_holder_content = ray.get(self.reset_args_holder.get.remote())
reset_args_holder_content = np.copy(reset_args_holder_content)
reset_args_holder_content[this_index - 1] = this_reset_args
ray.get(self.reset_args_holder.set.remote(reset_args_holder_content))
ray.get(this_evaluator.set_weights.remote(weights_id))
remote = this_evaluator.inner_update.remote(num_inner_updates)
running[remote] = (
i, this_reset_args, this_index, this_evaluator)
i += 1
continue
if running:
ready_ids, _ = ray.wait(list(running.keys()),
num_returns=1,
timeout=1000)
if ready_ids:
assert len(ready_ids) == 1
ready_id = ready_ids[0]
task_id, this_reset_args, this_index, this_evaluator = \
running.pop(ready_id)
assert np.array_equal(ray.get(ready_id), this_reset_args)
assert np.array_equal(reset_args[task_id], this_reset_args)
this_episodes = ray.get(
this_evaluator.apply.remote(lambda e: e.episodes))
for k, v in this_episodes.items():
episodes[k].extend(v)
finished[task_id] = (this_reset_args, this_episodes)
free_evaluators.append((this_index, this_evaluator))
continue
break
return {k: summarize_episodes(v, v, 0) for k, v in episodes.items()}
def _stop(self):
self.reset_args_holder.__ray_terminate__.remote()
Agent._stop(self)
def _save(self, checkpoint_dir):
checkpoint_path = osp.join(checkpoint_dir,
f"checkpoint-{self.iteration}")
with open(checkpoint_path, "wb") as f:
pickle.dump(self.__getstate__(), f)
pickle.dump({"reset_args_train": self.reset_args_train,
"reset_args_test": self.reset_args_test}, f)
return checkpoint_path
def _restore(self, checkpoint_path):
with open(checkpoint_path, "rb") as f:
extra_data = pickle.load(f)
reset_args = pickle.load(f)
self.__setstate__(extra_data)
self.reset_args_train = reset_args["reset_args_train"]
self.reset_args_test = reset_args["reset_args_test"]
if __name__ == "__main__":
import time
import ray
import numpy as np
import tensorflow as tf
from ray.tune.registry import register_env
from ray.rllib.models.catalog import ModelCatalog
from ray.rllib.evaluation.metrics import summarize_episodes
from ray.tune.logger import pretty_print
from rllib_models.rllib_mlp import RLlibMLP
from envs.point_env import PointEnv
from envs.mujoco.reacher import ReacherEnv
from envs.reset_wrapper import ResetWrapper
logger = logging.getLogger("ray.rllib.agents.maml")
logger.setLevel(logging.DEBUG)
ray.init()
# ray.init(redis_address="localhost:32222")
env_cls = PointEnv
env_cls = ReacherEnv
register_env(env_cls.__name__,
lambda env_config: ResetWrapper(env_cls(env_config), env_config))
# register_env("PointEnv", lambda env_config: PointEnv(env_config))
ModelCatalog.register_custom_model("maml_mlp", RLlibMLP)
config = {
"env_config": {"ctrl_cost_coeff": 0.0},
# "num_workers": 20,
"model": {
"custom_model": "maml_mlp",
"fcnet_hiddens": [100, 100],
"fcnet_activation": "tanh",
"custom_options": {"vf_share_layers": True},
# "squash_to_range": True,
# "free_log_std": True
}
}
agent = MAMLAgent(config=config, env=env_cls.__name__)
evaluator = agent.local_evaluator
policy = evaluator.policy_map[DEFAULT_POLICY_ID]
optimizer = agent.optimizer
for i in range(10):
st = time.time()
logger.info(f"\n{i}")
res = agent.train()
logger.info(f'\n{pretty_print(res["inner_update_metrics"])}')
# only perform inner update in the local evaluator
# policy.clear_grad_buffer()
# def func():
# grads, infos, samples = evaluator._inner_update_once()
# policy.update_grad_buffer(grads)
# episodes = evaluator.sampler.get_metrics()
# logger.info(
# f'\n{pretty_print(summarize_episodes(episodes, episodes))}')
# logger.info(f"\n{pretty_print(infos)}")
# return grads, samples
# for i in range(1000):
# print(i)
# grads, samples = func()
writer = tf.summary.FileWriter(logdir="./summary", graph=evaluator.tf_sess.graph)
writer.flush()